diff --git a/Cargo.lock b/Cargo.lock
index b3f88a420..8f85ebe17 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -24,41 +24,6 @@ version = "2.0.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "320119579fcad9c21884f5c4861d16174d0e06250625266f50fe6898340abefa"
 
-[[package]]
-name = "aead"
-version = "0.5.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d122413f284cf2d62fb1b7db97e02edb8cda96d769b16e443a4f6195e35662b0"
-dependencies = [
- "crypto-common 0.1.7",
- "generic-array",
-]
-
-[[package]]
-name = "aes"
-version = "0.8.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b169f7a6d4742236a0a00c541b845991d0ac43e546831af1249753ab4c3aa3a0"
-dependencies = [
- "cfg-if",
- "cipher",
- "cpufeatures 0.2.17",
-]
-
-[[package]]
-name = "aes-gcm"
-version = "0.10.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "831010a0f742e1209b3bcea8fab6a8e149051ba6099432c8cb2cc117dec3ead1"
-dependencies = [
- "aead",
- "aes",
- "cipher",
- "ctr",
- "ghash",
- "subtle",
-]
-
 [[package]]
 name = "aho-corasick"
 version = "1.1.4"
@@ -128,22 +93,13 @@ checksum = "cc7bb162ec39d46ab1ca8c77bf72e890535becd1751bb45f64c597edb4c8c6b3"
 
 [[package]]
 name = "alloc-stdlib"
-version = "0.2.2"
+version = "0.2.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "94fb8275041c72129eb51b7d0322c29b8387a0386127718b096429201a5d6ece"
+checksum = "0e76a019e91224d279006ff972f1e984179a6e9feb050adba6ce8274aef23195"
 dependencies = [
  "alloc-no-stdlib",
 ]
 
-[[package]]
-name = "android_system_properties"
-version = "0.1.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "819e7219dbd41043ac279b19830f2efc897156490d7fd6ea916720117ee66311"
-dependencies = [
- "libc",
-]
-
 [[package]]
 name = "anstream"
 version = "1.0.0"
@@ -180,7 +136,7 @@ version = "1.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "40c48f72fd53cd289104fc64099abca73db4166ad86ea0b4341abe65af83dadc"
 dependencies = [
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -191,14 +147,14 @@ checksum = "291e6a250ff86cd4a820112fb8898808a366d8f9f58ce16d1f538353ad55747d"
 dependencies = [
  "anstyle",
  "once_cell_polyfill",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
 name = "anyhow"
-version = "1.0.102"
+version = "1.0.103"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
+checksum = "2a4385e2e34eb35d6b3efe798b9eb88096925d87726c0798709bf56d9ed84af3"
 
 [[package]]
 name = "approx"
@@ -228,15 +184,9 @@ dependencies = [
 
 [[package]]
 name = "arrayvec"
-version = "0.7.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7c02d123df017efcdfbd739ef81735b36c5ba83ec3c59c80a9d7ecc718f92e50"
-
-[[package]]
-name = "as-any"
-version = "0.3.2"
+version = "0.7.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b0f477b951e452a0b6b4a10b53ccd569042d1d01729b519e02074a9c0958a063"
+checksum = "f02882884d3e1bc524fb12c79f107f6ad0e1cfd498c536ffb494301740995dfe"
 
 [[package]]
 name = "as-slice"
@@ -259,28 +209,6 @@ dependencies = [
  "tokio",
 ]
 
-[[package]]
-name = "async-stream"
-version = "0.3.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0b5a71a6f37880a80d1d7f19efd781e4b5de42c88f0722cc13bcb6cc2cfe8476"
-dependencies = [
- "async-stream-impl",
- "futures-core",
- "pin-project-lite",
-]
-
-[[package]]
-name = "async-stream-impl"
-version = "0.3.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c7c24de15d275a1ecfd47a380fb4d5ec9bfe0933f309ed5e705b775596a3574d"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
-]
-
 [[package]]
 name = "async-trait"
 version = "0.1.89"
@@ -472,9 +400,15 @@ checksum = "1e4b40c7323adcfc0a41c4b88143ed58346ff65a288fc144329c5c45e05d70c6"
 
 [[package]]
 name = "bitflags"
-version = "2.12.1"
+version = "1.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bef38d45163c2f1dde094a7dfd33ccf595c92905c8f8f4fdc18d06fb1037718a"
+
+[[package]]
+name = "bitflags"
+version = "2.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "84d7ced0ae9557296835c32bf1b1e02b44c746701f898460fb000d7eaa84f00a"
+checksum = "b4388bee8683e3d04af747c73422af53102d2bd24d9eadb6cbc100baef4b43f8"
 
 [[package]]
 name = "bitstream-io"
@@ -496,27 +430,18 @@ dependencies = [
 
 [[package]]
 name = "block-buffer"
-version = "0.12.0"
+version = "0.12.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cdd35008169921d80bc60d3d0ab416eecb028c4cd653352907921d95084790be"
+checksum = "d2f6c7dbe95a6ed67ad9f18e57daf93a2f034c524b99fd2b76d18fdfeb6660aa"
 dependencies = [
  "hybrid-array",
 ]
 
-[[package]]
-name = "block-padding"
-version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a8894febbff9f758034a5b8e12d87918f56dfc64a8e1fe757d65e29041538d93"
-dependencies = [
- "generic-array",
-]
-
 [[package]]
 name = "brotli"
-version = "8.0.3"
+version = "8.0.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8119e4516436f5708bbc474a9d395bf12f1b5395e93a92a56e647ac3388c8610"
+checksum = "5cc91aac060a7a1e25823bdccbfb6af1875b88f17c6daac97894eed8207166b3"
 dependencies = [
  "alloc-no-stdlib",
  "alloc-stdlib",
@@ -525,9 +450,9 @@ dependencies = [
 
 [[package]]
 name = "brotli-decompressor"
-version = "5.0.1"
+version = "5.0.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5962523e1b92ce1b5e793d9169b9943eece10d39f62550bc04bb605d75b94924"
+checksum = "3a32acac15fe1967bc3986b2a6347dffc965602354ea6f450ad07e8bfd253583"
 dependencies = [
  "alloc-no-stdlib",
  "alloc-stdlib",
@@ -575,9 +500,9 @@ checksum = "8f1fe948ff07f4bd06c30984e69f5b4899c516a3ef74f34df92a2df2ab535495"
 
 [[package]]
 name = "bytes"
-version = "1.11.1"
+version = "1.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1e748733b7cbc798e1434b6ac524f0c1ff2ab456fe201501e6497c8417a4fc33"
+checksum = "8ae3f5d315924270530207e2a68396c3cc547f6dca3fbdca317cfb1a51edb593"
 dependencies = [
  "serde",
 ]
@@ -588,20 +513,11 @@ version = "0.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1c53ba0f290bfc610084c05582d9c5d421662128fc69f4bf236707af6fd321b9"
 
-[[package]]
-name = "cbc"
-version = "0.1.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "26b52a9543ae338f279b96b0b9fed9c8093744685043739079ce85cd58f289a6"
-dependencies = [
- "cipher",
-]
-
 [[package]]
 name = "cc"
-version = "1.2.63"
+version = "1.2.65"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "556e016178bb5662a08681bbe0f00f8e17631781a4dfc8c45e466e4b185ec27f"
+checksum = "e228eec9be7c17ccb640b59b36a5cd805ea2a564a4c5e162c2f659fea30d3b96"
 dependencies = [
  "find-msvc-tools",
  "jobserver",
@@ -619,17 +535,6 @@ dependencies = [
  "serde",
 ]
 
-[[package]]
-name = "cfb"
-version = "0.7.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d38f2da7a0a2c4ccf0065be06397cc26a81f4e528be095826eee9d4adbb8c60f"
-dependencies = [
- "byteorder",
- "fnv",
- "uuid",
-]
-
 [[package]]
 name = "cfg-if"
 version = "1.0.4"
@@ -644,38 +549,15 @@ checksum = "613afe47fcd5fac7ccf1db93babcb082c5994d996f20b8b159f2ad1658eb5724"
 
 [[package]]
 name = "chacha20"
-version = "0.10.0"
+version = "0.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6f8d983286843e49675a4b7a2d174efe136dc93a18d69130dd18198a6c167601"
+checksum = "d524456ba66e72eb8b115ff89e01e497f8e6d11d78b70b1aa13c0fbd97540a81"
 dependencies = [
  "cfg-if",
  "cpufeatures 0.3.0",
  "rand_core 0.10.1",
 ]
 
-[[package]]
-name = "chrono"
-version = "0.4.45"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1aa79e62e7697b8e29b513a68abacf485adcd1fe8284a4316c5ae868e6633327"
-dependencies = [
- "iana-time-zone",
- "js-sys",
- "num-traits",
- "wasm-bindgen",
- "windows-link",
-]
-
-[[package]]
-name = "cipher"
-version = "0.4.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "773f3b9af64447d2ce9850330c473515014aa235e6a783b02db81ff39e4a3dad"
-dependencies = [
- "crypto-common 0.1.7",
- "inout",
-]
-
 [[package]]
 name = "clap"
 version = "4.6.1"
@@ -695,7 +577,7 @@ dependencies = [
  "anstream",
  "anstyle",
  "clap_lex",
- "strsim 0.11.1",
+ "strsim",
 ]
 
 [[package]]
@@ -704,7 +586,7 @@ version = "4.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f2ce8604710f6733aa641a2b3731eaa1e8b3d9973d5e3565da11800813f997a9"
 dependencies = [
- "heck 0.5.0",
+ "heck",
  "proc-macro2",
  "quote",
  "syn",
@@ -782,26 +664,6 @@ version = "0.4.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "cc14f565cf027a105f7a44ccf9e5b424348421a1d8952a8fc9d499d313107789"
 
-[[package]]
-name = "console_error_panic_hook"
-version = "0.1.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a06aeb73f470f66dcdbf7223caeebb85984942f22f1adb2a088cf9668146bbbc"
-dependencies = [
- "cfg-if",
- "wasm-bindgen",
-]
-
-[[package]]
-name = "console_log"
-version = "1.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "be8aed40e4edbf4d3b4431ab260b63fdc40f5780a4766824329ea0f1eefe3c0f"
-dependencies = [
- "log",
- "web-sys",
-]
-
 [[package]]
 name = "const-oid"
 version = "0.10.2"
@@ -817,15 +679,6 @@ dependencies = [
  "unicode-segmentation",
 ]
 
-[[package]]
-name = "convert_case"
-version = "0.11.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "affbf0190ed2caf063e3def54ff444b449371d55c58e513a95ab98eca50adb49"
-dependencies = [
- "unicode-segmentation",
-]
-
 [[package]]
 name = "core-foundation"
 version = "0.10.1"
@@ -842,15 +695,6 @@ version = "0.8.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "773648b94d0e5d620f64f280777445740e61fe701025087ec8b57f45c791888b"
 
-[[package]]
-name = "counter"
-version = "0.7.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "66e8e052be91f1c8aae2c1d81307d9f6e67f5f37001e3ddee419e971e73f03bc"
-dependencies = [
- "num-traits",
-]
-
 [[package]]
 name = "cpufeatures"
 version = "0.2.17"
@@ -932,7 +776,6 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "78c8292055d1c1df0cce5d180393dc8cce0abec0a7102adb6c7b1eef6016d60a"
 dependencies = [
  "generic-array",
- "rand_core 0.6.4",
  "typenum",
 ]
 
@@ -989,15 +832,6 @@ dependencies = [
  "memchr",
 ]
 
-[[package]]
-name = "ctr"
-version = "0.9.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0369ee1ad671834580515889b80f2ea915f23b8be8d0daa4bbaf2ac5c7590835"
-dependencies = [
- "cipher",
-]
-
 [[package]]
 name = "darling"
 version = "0.20.11"
@@ -1028,7 +862,7 @@ dependencies = [
  "ident_case",
  "proc-macro2",
  "quote",
- "strsim 0.11.1",
+ "strsim",
  "syn",
 ]
 
@@ -1041,7 +875,7 @@ dependencies = [
  "ident_case",
  "proc-macro2",
  "quote",
- "strsim 0.11.1",
+ "strsim",
  "syn",
 ]
 
@@ -1082,59 +916,35 @@ dependencies = [
 ]
 
 [[package]]
-name = "data-encoding"
-version = "2.11.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a4ae5f15dda3c708c0ade84bfee31ccab44a3da4f88015ed22f63732abe300c8"
-
-[[package]]
-name = "deluxe"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8ed332aaf752b459088acf3dd4eca323e3ef4b83c70a84ca48fb0ec5305f1488"
-dependencies = [
- "deluxe-core",
- "deluxe-macros",
- "once_cell",
- "proc-macro2",
- "syn",
-]
-
-[[package]]
-name = "deluxe-core"
-version = "0.5.0"
+name = "defmt"
+version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "eddada51c8576df9d6a8450c351ff63042b092c9458b8ac7d20f89cbd0ffd313"
+checksum = "a6e524506490a1953d237cb87b1cfc1e46f88c18f10a22dfe0f507dc6bfc7f7f"
 dependencies = [
- "arrayvec",
- "proc-macro2",
- "quote",
- "strsim 0.10.0",
- "syn",
+ "bitflags 1.3.2",
+ "defmt-macros",
 ]
 
 [[package]]
-name = "deluxe-macros"
-version = "0.5.0"
+name = "defmt-macros"
+version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f87546d9c837f0b7557e47b8bd6eae52c3c223141b76aa233c345c9ab41d9117"
+checksum = "f0a27770e9c8f719a79d8b638281f4d828f77d8fd61e0bd94451b9b85e576a0b"
 dependencies = [
- "deluxe-core",
- "heck 0.4.1",
- "if_chain",
- "proc-macro-crate 1.3.1",
+ "defmt-parser",
+ "proc-macro-error2",
  "proc-macro2",
  "quote",
  "syn",
 ]
 
 [[package]]
-name = "deranged"
-version = "0.5.8"
+name = "defmt-parser"
+version = "1.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7cd812cc2bc1d69d4764bd80df88b4317eaef9e773c75226407d9bc0876b211c"
+checksum = "10d60334b3b2e7c9d91ef8150abfb6fa4c1c39ebbcf4a81c2e346aad939fee3e"
 dependencies = [
- "powerfmt",
+ "thiserror",
 ]
 
 [[package]]
@@ -1183,7 +993,7 @@ version = "2.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "799a97264921d8623a957f6c3b9011f3b5492f557bbb7a5a19b7fa6d06ba8dcb"
 dependencies = [
- "convert_case 0.10.0",
+ "convert_case",
  "proc-macro2",
  "quote",
  "rustc_version",
@@ -1191,12 +1001,6 @@ dependencies = [
  "unicode-xid",
 ]
 
-[[package]]
-name = "deunicode"
-version = "1.6.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "abd57806937c9cc163efc8ea3910e00a62e2aeb0b8119f1793a978088f8f6b04"
-
 [[package]]
 name = "digest"
 version = "0.10.7"
@@ -1213,7 +1017,7 @@ version = "0.11.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f1dd6dbb5841937940781866fa1281a1ff7bd3bf827091440879f9994983d5c2"
 dependencies = [
- "block-buffer 0.12.0",
+ "block-buffer 0.12.1",
  "const-oid",
  "crypto-common 0.2.2",
 ]
@@ -1256,15 +1060,6 @@ version = "1.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d0881ea181b1df73ff77ffaaf9c7544ecc11e82fba9b5f27b262a3c73a332555"
 
-[[package]]
-name = "ecb"
-version = "0.1.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1a8bfa975b1aec2145850fcaa1c6fe269a16578c44705a532ae3edc92b8881c7"
-dependencies = [
- "cipher",
-]
-
 [[package]]
 name = "ego-tree"
 version = "0.11.0"
@@ -1278,83 +1073,269 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "91622ff5e7162018101f2fea40d6ebf4a78bbe5a49736a2020649edf9693679e"
 
 [[package]]
-name = "embedded-io"
-version = "0.4.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ef1a6892d9eef45c8fa6b9e0086428a2cca8491aca8f787c534a3d6d0bcb3ced"
-
-[[package]]
-name = "embedded-io"
-version = "0.6.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "edd0f118536f44f5ccd48bcb8b111bdc3de888b58c74639dfb034a357d0f206d"
+name = "elide"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
+dependencies = [
+ "elide-codec",
+ "elide-context",
+ "elide-core",
+ "elide-detection",
+ "elide-llm",
+ "elide-ner",
+ "elide-ocr",
+ "elide-orchestration",
+ "elide-pattern",
+ "elide-redaction",
+ "elide-stt",
+]
 
 [[package]]
-name = "encoding_rs"
-version = "0.8.35"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "75030f3c4f45dafd7586dd6780965a8c7e8e285a5ecb86713e63a79c5b2766f3"
+name = "elide-bento"
+version = "0.1.0"
 dependencies = [
- "cfg-if",
+ "async-trait",
+ "base64",
+ "bentoml",
+ "elide-core",
+ "elide-ner",
+ "elide-ocr",
+ "hipstr",
+ "serde",
+ "thiserror",
+ "tokio",
 ]
 
 [[package]]
-name = "enum_dispatch"
-version = "0.3.13"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "aa18ce2bc66555b3218614519ac839ddb759a7d6720732f979ef8d13be147ecd"
+name = "elide-codec"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
 dependencies = [
- "once_cell",
- "proc-macro2",
- "quote",
- "syn",
+ "bytes",
+ "csv",
+ "ego-tree",
+ "elide-core",
+ "hex",
+ "hound",
+ "image",
+ "imageproc",
+ "mp3lame-encoder",
+ "quick-xml 0.40.1",
+ "scraper",
+ "serde_json",
+ "sha2 0.11.0",
+ "symphonia",
+ "tracing",
 ]
 
 [[package]]
-name = "equator"
-version = "0.4.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4711b213838dfee0117e3be6ac926007d7f433d7bbe33595975d4190cb07e6fc"
+name = "elide-context"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
 dependencies = [
- "equator-macro",
+ "elide-core",
+ "hipstr",
+ "unicode-segmentation",
 ]
 
 [[package]]
-name = "equator-macro"
-version = "0.4.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "44f23cf4b44bfce11a86ace86f8a73ffdec849c9fd00a386a53d278bd9e81fb3"
+name = "elide-core"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
 dependencies = [
- "proc-macro2",
- "quote",
- "syn",
+ "bytes",
+ "celes",
+ "derive_builder",
+ "hipstr",
+ "jiff",
+ "oxilangtag",
+ "serde",
+ "type-map",
+ "uuid",
 ]
 
 [[package]]
-name = "equivalent"
-version = "1.0.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "877a4ace8713b0bcf2a4e7eec82529c029f1d0619886d18145fea96c3ffe5c0f"
+name = "elide-detection"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
+dependencies = [
+ "elide-core",
+ "futures",
+ "tracing",
+]
 
 [[package]]
-name = "errno"
-version = "0.3.14"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
+name = "elide-llm"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
 dependencies = [
- "libc",
- "windows-sys 0.52.0",
+ "async-trait",
+ "derive_builder",
+ "elide-core",
+ "schemars",
+ "serde",
+ "tracing",
+ "unicode-normalization",
 ]
 
 [[package]]
-name = "eventsource-stream"
-version = "0.2.3"
+name = "elide-ner"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
+dependencies = [
+ "async-trait",
+ "derive_builder",
+ "elide-core",
+ "hipstr",
+ "serde",
+ "uuid",
+]
+
+[[package]]
+name = "elide-ocr"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
+dependencies = [
+ "async-trait",
+ "elide-core",
+ "uuid",
+]
+
+[[package]]
+name = "elide-orchestration"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
+dependencies = [
+ "bytes",
+ "elide-codec",
+ "elide-core",
+ "elide-detection",
+ "elide-redaction",
+ "erased-serde",
+ "serde",
+]
+
+[[package]]
+name = "elide-pattern"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
+dependencies = [
+ "aho-corasick",
+ "bs58",
+ "csv",
+ "derive_builder",
+ "derive_more",
+ "elide-context",
+ "elide-core",
+ "iban_validate",
+ "phonenumber",
+ "regex",
+ "serde",
+ "toml",
+]
+
+[[package]]
+name = "elide-redaction"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
+dependencies = [
+ "bytes",
+ "elide-core",
+ "hex",
+ "hipstr",
+ "sha2 0.11.0",
+ "tracing",
+ "uuid",
+]
+
+[[package]]
+name = "elide-stt"
+version = "0.1.0"
+source = "git+https://github.com/nvisycom/elide?branch=main#70e6e8bff19a64cef7653fe4a1b1b7975f95c249"
+dependencies = [
+ "async-trait",
+ "elide-core",
+ "uuid",
+]
+
+[[package]]
+name = "embedded-io"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ef1a6892d9eef45c8fa6b9e0086428a2cca8491aca8f787c534a3d6d0bcb3ced"
+
+[[package]]
+name = "embedded-io"
+version = "0.6.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "edd0f118536f44f5ccd48bcb8b111bdc3de888b58c74639dfb034a357d0f206d"
+
+[[package]]
+name = "encoding_rs"
+version = "0.8.35"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "74fef4569247a5f429d9156b9d0a2599914385dd189c539334c625d8099d90ab"
+checksum = "75030f3c4f45dafd7586dd6780965a8c7e8e285a5ecb86713e63a79c5b2766f3"
 dependencies = [
- "futures-core",
- "nom 7.1.3",
- "pin-project-lite",
+ "cfg-if",
+]
+
+[[package]]
+name = "enum_dispatch"
+version = "0.3.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "aa18ce2bc66555b3218614519ac839ddb759a7d6720732f979ef8d13be147ecd"
+dependencies = [
+ "once_cell",
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "equator"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4711b213838dfee0117e3be6ac926007d7f433d7bbe33595975d4190cb07e6fc"
+dependencies = [
+ "equator-macro",
+]
+
+[[package]]
+name = "equator-macro"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "44f23cf4b44bfce11a86ace86f8a73ffdec849c9fd00a386a53d278bd9e81fb3"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "equivalent"
+version = "1.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "877a4ace8713b0bcf2a4e7eec82529c029f1d0619886d18145fea96c3ffe5c0f"
+
+[[package]]
+name = "erased-serde"
+version = "0.4.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d2add8a07dd6a8d93ff627029c51de145e12686fbc36ecb298ac22e74cf02dec"
+dependencies = [
+ "serde",
+ "serde_core",
+ "typeid",
+]
+
+[[package]]
+name = "errno"
+version = "0.3.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
+dependencies = [
+ "libc",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -1378,25 +1359,11 @@ version = "0.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "af9673d8203fcb076b19dfd17e38b3d4ae9f44959416ea532ce72415a6020365"
 
-[[package]]
-name = "fake"
-version = "5.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ea6be833b323a56361118a747470a45a1bcd5c52a2ec9b1e40c83dafe687e453"
-dependencies = [
- "deunicode",
- "either",
- "rand 0.10.1",
-]
-
 [[package]]
 name = "fastrand"
 version = "2.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9f1f227452a390804cdb637b74a86990f2a7d7ba4b7d5693aac9b4dd6defd8d6"
-dependencies = [
- "getrandom 0.3.4",
-]
 
 [[package]]
 name = "fax"
@@ -1461,12 +1428,6 @@ version = "1.0.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1"
 
-[[package]]
-name = "foldhash"
-version = "0.1.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d9c4f5dac5e15c24eb999c26181a6ca40b39fe946cbe4c263c7209467bc83af2"
-
 [[package]]
 name = "form_urlencoded"
 version = "1.2.2"
@@ -1482,12 +1443,6 @@ version = "1.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c"
 
-[[package]]
-name = "fst"
-version = "0.4.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7ab85b9b05e3978cc9a9cf8fea7f01b494e1a09ed3037e16ba39edc7a29eb61a"
-
 [[package]]
 name = "futures"
 version = "0.3.32"
@@ -1559,12 +1514,6 @@ version = "0.3.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "037711b3d59c33004d3856fbdc83b99d4ff37a24768fa1be9ce3538a1cde4393"
 
-[[package]]
-name = "futures-timer"
-version = "3.0.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "af43fadb8a98512d547e37b4e92e0ced13e205c061b87b4623eff01d918d6968"
-
 [[package]]
 name = "futures-util"
 version = "0.3.32"
@@ -1645,30 +1594,18 @@ dependencies = [
 
 [[package]]
 name = "getrandom"
-version = "0.4.2"
+version = "0.4.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0de51e6874e94e7bf76d726fc5d13ba782deca734ff60d5bb2fb2607c7406555"
+checksum = "300e883d756b2e4ec94e02791f39b04b522276138852cfc41d9fb7e904106099"
 dependencies = [
  "cfg-if",
  "js-sys",
  "libc",
  "r-efi 6.0.0",
  "rand_core 0.10.1",
- "wasip2",
- "wasip3",
  "wasm-bindgen",
 ]
 
-[[package]]
-name = "ghash"
-version = "0.5.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f0d8a4362ccb29cb0b265253fb0a2728f592895ee6854fd9bc13f2ffda266ff1"
-dependencies = [
- "opaque-debug",
- "polyval",
-]
-
 [[package]]
 name = "gif"
 version = "0.14.2"
@@ -1703,31 +1640,6 @@ version = "0.33.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "898f5a568a84989b6c0f8caa50a93074b97dbdc58fc6d9543157bb4562758933"
 
-[[package]]
-name = "glob"
-version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0cc23270f6e1808e30a928bdc84dea0b9b4136a8bc82338574f23baf47bbd280"
-
-[[package]]
-name = "h2"
-version = "0.4.14"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "171fefbc92fe4a4de27e0698d6a5b392d6a0e333506bc49133760b3bcf948733"
-dependencies = [
- "atomic-waker",
- "bytes",
- "fnv",
- "futures-core",
- "futures-sink",
- "http",
- "indexmap",
- "slab",
- "tokio",
- "tokio-util",
- "tracing",
-]
-
 [[package]]
 name = "half"
 version = "2.7.1"
@@ -1754,15 +1666,6 @@ version = "0.14.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e5274423e17b7c9fc20b6e7e208532f9b19825d82dfd615708b70edd83df41f1"
 
-[[package]]
-name = "hashbrown"
-version = "0.15.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9229cfe53dfd69f0609a49f65461bd93001ea1ef889cd5529dd176593f5338a1"
-dependencies = [
- "foldhash",
-]
-
 [[package]]
 name = "hashbrown"
 version = "0.16.1"
@@ -1789,12 +1692,6 @@ dependencies = [
  "stable_deref_trait",
 ]
 
-[[package]]
-name = "heck"
-version = "0.4.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "95505c38b4572b2d910cecb0281560f54b440a19336cbbcb27bf6ce6adc6f5a8"
-
 [[package]]
 name = "heck"
 version = "0.5.0"
@@ -1835,9 +1732,9 @@ dependencies = [
 
 [[package]]
 name = "http"
-version = "1.4.1"
+version = "1.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8be7462df143984c4598a256ef469b251d7d7f9e271135073e78fc535414f3d0"
+checksum = "6970f50e31d6fc17d3fa27329444bfa74e196cf62e95052a3f6fee181dba6425"
 dependencies = [
  "bytes",
  "itoa",
@@ -1913,7 +1810,6 @@ dependencies = [
  "bytes",
  "futures-channel",
  "futures-core",
- "h2",
  "http",
  "http-body",
  "httparse",
@@ -1963,30 +1859,6 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "iana-time-zone"
-version = "0.1.65"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e31bc9ad994ba00e440a8aa5c9ef0ec67d5cb5e5cb0cc7f8b744a35b389cc470"
-dependencies = [
- "android_system_properties",
- "core-foundation-sys",
- "iana-time-zone-haiku",
- "js-sys",
- "log",
- "wasm-bindgen",
- "windows-core",
-]
-
-[[package]]
-name = "iana-time-zone-haiku"
-version = "0.1.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f31827a206f56af32e590ba56d5d2d085f558508192593743f16b2306495269f"
-dependencies = [
- "cc",
-]
-
 [[package]]
 name = "iban_validate"
 version = "5.0.1"
@@ -2078,12 +1950,6 @@ dependencies = [
  "zerovec",
 ]
 
-[[package]]
-name = "id-arena"
-version = "2.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3d3067d79b975e8844ca9eb072e16b31c3c1c36928edf9c6789548c524d0d954"
-
 [[package]]
 name = "ident_case"
 version = "1.0.1"
@@ -2111,12 +1977,6 @@ dependencies = [
  "icu_properties",
 ]
 
-[[package]]
-name = "if_chain"
-version = "1.0.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cd62e6b5e86ea8eeeb8db1de02880a6abc01a397b2ebb64b5d74ac255318f5cb"
-
 [[package]]
 name = "image"
 version = "0.25.10"
@@ -2159,7 +2019,7 @@ checksum = "c7b27bc0867dc40df08deb53d6e96342db6e0702e7ae33ed09a4eba33e594b05"
 dependencies = [
  "ab_glyph",
  "approx",
- "getrandom 0.4.2",
+ "getrandom 0.4.3",
  "image",
  "itertools",
  "nalgebra",
@@ -2172,28 +2032,9 @@ dependencies = [
 
 [[package]]
 name = "imgref"
-version = "1.12.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "40fac9d56ed6437b198fddba683305e8e2d651aa42647f00f5ae542e7f5c94a2"
-
-[[package]]
-name = "include_dir"
-version = "0.7.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "923d117408f1e49d914f1a379a309cffe4f18c05cf4e3d12e613a15fc81bd0dd"
-dependencies = [
- "include_dir_macros",
-]
-
-[[package]]
-name = "include_dir_macros"
-version = "0.7.4"
+version = "1.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7cab85a7ed0bd5f0e76d93846e0147172bed2e2d3f859bcc33a8d9699cad1a75"
-dependencies = [
- "proc-macro2",
- "quote",
-]
+checksum = "89194689a993ab15268672e99e7b0e19da2da3268ac682e8f02d29d4d1434cd7"
 
 [[package]]
 name = "indexmap"
@@ -2207,34 +2048,6 @@ dependencies = [
  "serde_core",
 ]
 
-[[package]]
-name = "indoc"
-version = "2.0.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "79cf5c93f93228cf8efb3ba362535fb11199ac548a09ce117c9b1adc3030d706"
-dependencies = [
- "rustversion",
-]
-
-[[package]]
-name = "infer"
-version = "0.19.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a588916bfdfd92e71cacef98a63d9b1f0d74d6599980d11894290e7ddefffcf7"
-dependencies = [
- "cfb",
-]
-
-[[package]]
-name = "inout"
-version = "0.1.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "879f10e63c20629ecabbb64a8010319738c66a5cd0c29b02d63d272b03751d01"
-dependencies = [
- "block-padding",
- "generic-array",
-]
-
 [[package]]
 name = "interpolate_name"
 version = "0.2.4"
@@ -2284,10 +2097,11 @@ checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
 
 [[package]]
 name = "jiff"
-version = "0.2.28"
+version = "0.2.29"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4603d3033e49e2b0e31229fcab20a5d40089c607d975cd9c80551dc69eed9102"
+checksum = "34f877a98676d2fb664698d74cc6a51ce6c484ce8c770f05d0108ec9090aeb46"
 dependencies = [
+ "defmt",
  "jiff-static",
  "jiff-tzdb-platform",
  "log",
@@ -2299,9 +2113,9 @@ dependencies = [
 
 [[package]]
 name = "jiff-static"
-version = "0.2.28"
+version = "0.2.29"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "782d32378dddf207193ac91cefb848ad41abb58195c95168e1291227a0832b47"
+checksum = "0666b5ab5ecaca213fc2a85b8c0083d9004e84ee2d5f9a7e0017aaf50986f25f"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -2384,13 +2198,12 @@ dependencies = [
 
 [[package]]
 name = "js-sys"
-version = "0.3.99"
+version = "0.3.103"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "142bc4740e452c1e57ade0cbc129f139c9093e354346f0872ef985f4f5cf5f11"
+checksum = "53b44bfcdb3f8d5837a46dae1ca9660a837176eee74a28b229bc626816589102"
 dependencies = [
  "cfg-if",
  "futures-util",
- "once_cell",
  "wasm-bindgen",
 ]
 
@@ -2400,12 +2213,6 @@ version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
 
-[[package]]
-name = "leb128fmt"
-version = "0.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09edd9e8b54e49e587e4f6295a7d29c3ea94d469cb40ab8ca70b288248a81db2"
-
 [[package]]
 name = "lebe"
 version = "0.5.3"
@@ -2428,54 +2235,12 @@ dependencies = [
  "cc",
 ]
 
-[[package]]
-name = "libloading"
-version = "0.9.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "754ca22de805bb5744484a5b151a9e1a8e837d5dc232c2d7d8c2e3492edc8b60"
-dependencies = [
- "cfg-if",
- "windows-link",
-]
-
 [[package]]
 name = "libm"
 version = "0.2.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6d2cec3eae94f9f509c767b45932f1ada8350c4bdb85af2fcab4a3c14807981"
 
-[[package]]
-name = "lingua"
-version = "1.8.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f40d9129bb9fe42c95d1bd420d6891607eaff17df16ee15674aed2d05b0ec8f4"
-dependencies = [
- "counter",
- "dashmap",
- "fastrand",
- "fst",
- "include_dir",
- "itertools",
- "lingua-english-language-model",
- "maplit",
- "rayon",
- "regex",
- "serde",
- "serde-wasm-bindgen",
- "strum 0.27.2",
- "strum_macros 0.27.2",
- "wasm-bindgen",
-]
-
-[[package]]
-name = "lingua-english-language-model"
-version = "1.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "97102de08b134a49f1cce05a1b6f5bf08ef21fe858074ae2b794e7892c43dd4b"
-dependencies = [
- "include_dir",
-]
-
 [[package]]
 name = "linked-hash-map"
 version = "0.5.6"
@@ -2505,9 +2270,9 @@ dependencies = [
 
 [[package]]
 name = "log"
-version = "0.4.32"
+version = "0.4.33"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "953f07c43838f8e6f9758cab68bf5bed85465e7587ebe0b823f1bcd81978ad3a"
+checksum = "0ceec5bc11778974d1bcb055b18002eba7f4b3518b6a0081b3af5f21666da9ad"
 
 [[package]]
 name = "loom"
@@ -2531,37 +2296,6 @@ dependencies = [
  "imgref",
 ]
 
-[[package]]
-name = "lopdf"
-version = "0.41.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "67513274c50a2b51e5f75d9e682fcf4ab064a8a9c9ae2c3c59309084882bb24d"
-dependencies = [
- "aes",
- "bitflags",
- "cbc",
- "chrono",
- "ecb",
- "encoding_rs",
- "flate2",
- "getrandom 0.4.2",
- "indexmap",
- "itoa",
- "jiff",
- "log",
- "md-5",
- "nom 8.0.0",
- "rand 0.10.1",
- "rangemap",
- "rayon",
- "sha2 0.10.9",
- "stringprep",
- "thiserror",
- "time",
- "ttf-parser",
- "weezl",
-]
-
 [[package]]
 name = "lru-cache"
 version = "0.1.2"
@@ -2608,12 +2342,6 @@ dependencies = [
  "twox-hash",
 ]
 
-[[package]]
-name = "maplit"
-version = "1.0.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3e2e65a1a2e43cfcb47a895c4c8b10d1f4a61097f9f254f183aee60cad9c651d"
-
 [[package]]
 name = "markup5ever"
 version = "0.39.0"
@@ -2650,12 +2378,6 @@ dependencies = [
  "rawpointer",
 ]
 
-[[package]]
-name = "maybe-owned"
-version = "0.3.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4facc753ae494aeb6e3c22f839b158aebd4f9270f55cd3c79906c45476c47ab4"
-
 [[package]]
 name = "maybe-rayon"
 version = "0.1.1"
@@ -2666,27 +2388,11 @@ dependencies = [
  "rayon",
 ]
 
-[[package]]
-name = "md-5"
-version = "0.10.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d89e7ee0cfbedfc4da3340218492196241d89eefb6dab27de5df917a6d2e78cf"
-dependencies = [
- "cfg-if",
- "digest 0.10.7",
-]
-
 [[package]]
 name = "memchr"
-version = "2.8.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6b947ae49db0d222b1dbc6b113ce7248a3fc3a6ca21b696717bfc000ba4484d8"
-
-[[package]]
-name = "memo-map"
-version = "0.3.3"
+version = "2.8.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "38d1115007560874e373613744c6fba374c17688327a71c1476d1a5954cc857b"
+checksum = "88904434abc2901f197fe8cc55f0445e7ded921dba5911dad2e2b39b48e663c4"
 
 [[package]]
 name = "mime"
@@ -2704,16 +2410,6 @@ dependencies = [
  "unicase",
 ]
 
-[[package]]
-name = "minijinja"
-version = "2.20.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2929e494b2280e1e18959bb2e121da03347ae896896fdfaceaab43c88a02803f"
-dependencies = [
- "memo-map",
- "serde",
-]
-
 [[package]]
 name = "minimal-lexical"
 version = "0.2.1"
@@ -2807,15 +2503,6 @@ dependencies = [
  "typenum",
 ]
 
-[[package]]
-name = "nanoid"
-version = "0.4.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3ffa00dec017b5b1a8b7cf5e2c008bfda1aa7e0697ac1508b491fdf2622fb4d8"
-dependencies = [
- "rand 0.8.6",
-]
-
 [[package]]
 name = "new_debug_unreachable"
 version = "1.0.6"
@@ -2862,7 +2549,7 @@ version = "0.50.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
 dependencies = [
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -2897,12 +2584,6 @@ dependencies = [
  "num-traits",
 ]
 
-[[package]]
-name = "num-conv"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "521739c6d2bac4aa25192232afe6841231376b2b26d4d9fae5ecf8ca5772e441"
-
 [[package]]
 name = "num-derive"
 version = "0.4.2"
@@ -2964,7 +2645,7 @@ dependencies = [
  "clap",
  "humantime",
  "humantime-serde",
- "nvisy-engine",
+ "nvisy-core",
  "nvisy-server",
  "serde",
  "tokio",
@@ -2973,62 +2654,23 @@ dependencies = [
  "tracing-subscriber",
 ]
 
-[[package]]
-name = "nvisy-codec"
-version = "0.1.0"
-dependencies = [
- "async-trait",
- "bytes",
- "csv",
- "derive_more",
- "ego-tree",
- "hex",
- "hound",
- "image",
- "imageproc",
- "infer",
- "lopdf",
- "mp3lame-encoder",
- "nvisy-core",
- "pdfium-render",
- "rayon",
- "schemars",
- "scraper",
- "serde",
- "serde_json",
- "sha2 0.11.0",
- "symphonia",
- "tokio",
- "tracing",
-]
-
-[[package]]
-name = "nvisy-context"
-version = "0.1.0"
-dependencies = [
- "async-trait",
- "hipstr",
- "nvisy-core",
- "unicode-segmentation",
-]
-
 [[package]]
 name = "nvisy-core"
 version = "0.1.0"
 dependencies = [
  "async-trait",
- "bytes",
- "celes",
  "derive_builder",
  "derive_more",
+ "elide",
+ "elide-core",
  "hipstr",
- "oxilangtag",
+ "jiff",
  "schemars",
+ "semver",
  "serde",
  "serde_json",
  "strum 0.28.0",
  "thiserror",
- "type-map",
  "uuid",
 ]
 
@@ -3036,129 +2678,26 @@ dependencies = [
 name = "nvisy-engine"
 version = "0.1.0"
 dependencies = [
- "aes-gcm",
- "anyhow",
- "async-trait",
- "base64",
  "bytes",
- "derive_builder",
  "derive_more",
+ "elide",
+ "elide-bento",
+ "elide-core",
+ "elide-ocr",
+ "elide-stt",
  "fjall",
  "futures",
  "hipstr",
- "humantime-serde",
  "jiff",
- "nvisy-codec",
  "nvisy-core",
- "nvisy-engine",
- "nvisy-llm",
- "nvisy-ner",
- "nvisy-ocr",
- "nvisy-pattern",
- "nvisy-stt",
- "nvisy-toolkit",
- "rand 0.10.1",
- "schemars",
  "semver",
  "serde",
  "serde_json",
- "strum 0.28.0",
- "tempfile",
- "tokio",
- "tokio-util",
- "toml",
- "tracing",
- "uuid",
- "validator",
-]
-
-[[package]]
-name = "nvisy-fake"
-version = "0.1.0"
-dependencies = [
- "async-trait",
- "fake",
- "nvisy-core",
- "nvisy-toolkit",
- "tokio",
- "uuid",
-]
-
-[[package]]
-name = "nvisy-llm"
-version = "0.1.0"
-dependencies = [
- "async-trait",
- "base64",
- "derive_builder",
- "derive_more",
- "humantime-serde",
- "minijinja",
- "nvisy-core",
- "reqwest-middleware",
- "reqwest-retry",
- "reqwest-tracing",
- "rig",
- "schemars",
- "serde",
- "serde_json",
- "thiserror",
- "tokio",
- "toml",
- "tracing",
- "unicode-normalization",
-]
-
-[[package]]
-name = "nvisy-ner"
-version = "0.1.0"
-dependencies = [
- "async-trait",
- "bentoml",
- "derive_builder",
- "lingua",
- "nvisy-core",
- "serde",
- "tokio",
- "tracing",
- "type-map",
- "uuid",
-]
-
-[[package]]
-name = "nvisy-ocr"
-version = "0.1.0"
-dependencies = [
- "async-trait",
- "bentoml",
- "bytes",
- "futures",
- "nvisy-core",
+ "sha2 0.11.0",
  "tokio",
- "tracing",
  "uuid",
 ]
 
-[[package]]
-name = "nvisy-pattern"
-version = "0.1.0"
-dependencies = [
- "aho-corasick",
- "async-trait",
- "bs58",
- "csv",
- "derive_builder",
- "derive_more",
- "iban_validate",
- "nvisy-context",
- "nvisy-core",
- "phonenumber",
- "regex",
- "serde",
- "tokio",
- "toml",
-]
-
 [[package]]
 name = "nvisy-server"
 version = "0.1.0"
@@ -3166,12 +2705,14 @@ dependencies = [
  "aide",
  "axum",
  "derive_more",
+ "elide-core",
  "futures",
  "humantime-serde",
  "jiff",
  "nvisy-core",
  "nvisy-engine",
  "schemars",
+ "semver",
  "serde",
  "tokio",
  "tower",
@@ -3180,45 +2721,6 @@ dependencies = [
  "uuid",
 ]
 
-[[package]]
-name = "nvisy-stt"
-version = "0.1.0"
-dependencies = [
- "async-trait",
- "nvisy-core",
- "tokio",
- "tracing",
- "uuid",
-]
-
-[[package]]
-name = "nvisy-toolkit"
-version = "0.1.0"
-dependencies = [
- "aes-gcm",
- "async-trait",
- "base64",
- "nvisy-codec",
- "nvisy-context",
- "nvisy-core",
- "nvisy-fake",
- "nvisy-llm",
- "nvisy-ner",
- "nvisy-ocr",
- "nvisy-pattern",
- "nvisy-stt",
- "regex",
- "schemars",
- "serde",
- "serde_json",
- "sha2 0.11.0",
- "tokio",
- "tracing",
- "type-map",
- "unicode-normalization",
- "uuid",
-]
-
 [[package]]
 name = "once_cell"
 version = "1.21.4"
@@ -3237,27 +2739,12 @@ version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "44d11de466f4a3006fe8a5e7ec84e93b79c70cb992ae0aa0eb631ad2df8abfe2"
 
-[[package]]
-name = "opaque-debug"
-version = "0.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c08d65885ee38876c4f86fa503fb49d7b507c2b62552df7c70b2fce627e06381"
-
 [[package]]
 name = "openssl-probe"
 version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7c87def4c32ab89d880effc9e097653c8da5d6ef28e6b539d313baaacfbafcbe"
 
-[[package]]
-name = "ordered-float"
-version = "5.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7d950ca161dc355eaf28f82b11345ed76c6e1f6eb1f4f4479e0323b9e2fbd0e"
-dependencies = [
- "num-traits",
-]
-
 [[package]]
 name = "owned_ttf_parser"
 version = "0.25.1"
@@ -3311,32 +2798,6 @@ version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "35fb2e5f958ec131621fdd531e9fc186ed768cbe395337403ae56c17a74c68ec"
 
-[[package]]
-name = "pdfium-render"
-version = "0.9.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e06f0df3ca17554c1b8f31eb17bc77eedbbafa120d35b90b3096fa46b2fdc94c"
-dependencies = [
- "bitflags",
- "bytemuck",
- "bytes",
- "chrono",
- "console_error_panic_hook",
- "console_log",
- "image",
- "itertools",
- "js-sys",
- "libloading",
- "log",
- "maybe-owned",
- "once_cell",
- "utf16string",
- "vecmath",
- "wasm-bindgen",
- "wasm-bindgen-futures",
- "web-sys",
-]
-
 [[package]]
 name = "percent-encoding"
 version = "2.3.2"
@@ -3407,7 +2868,7 @@ dependencies = [
  "nom 7.1.3",
  "once_cell",
  "postcard",
- "quick-xml",
+ "quick-xml 0.38.4",
  "regex",
  "regex-cache",
  "serde",
@@ -3416,26 +2877,6 @@ dependencies = [
  "thiserror",
 ]
 
-[[package]]
-name = "pin-project"
-version = "1.1.13"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2466b2336ed02bcdca6b294417127b90ec92038d1d5c4fbeac971a922e0e0924"
-dependencies = [
- "pin-project-internal",
-]
-
-[[package]]
-name = "pin-project-internal"
-version = "1.1.13"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c96395f0a926bc13b1c17622aaddda1ecb55d49c8f1bf9777e4d877800a43f8b"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
-]
-
 [[package]]
 name = "pin-project-lite"
 version = "0.2.17"
@@ -3448,12 +2889,6 @@ version = "0.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8b870d8c151b6f2fb93e84a13146138f05d02ed11c7e7c54f8826aaaf7c9f184"
 
-[[package]]
-name = "piston-float"
-version = "1.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ad78bf43dcf80e8f950c92b84f938a0fc7590b7f6866fbcbeca781609c115590"
-
 [[package]]
 name = "pkg-config"
 version = "0.3.33"
@@ -3466,25 +2901,13 @@ version = "0.18.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "60769b8b31b2a9f263dae2776c37b1b28ae246943cf719eb6946a1db05128a61"
 dependencies = [
- "bitflags",
+ "bitflags 2.13.0",
  "crc32fast",
  "fdeflate",
  "flate2",
  "miniz_oxide",
 ]
 
-[[package]]
-name = "polyval"
-version = "0.6.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d1fe60d06143b2430aa532c94cfe9e29783047f06c0d7fd359a9a51b729fa25"
-dependencies = [
- "cfg-if",
- "cpufeatures 0.2.17",
- "opaque-debug",
- "universal-hash",
-]
-
 [[package]]
 name = "portable-atomic"
 version = "1.13.1"
@@ -3522,12 +2945,6 @@ dependencies = [
  "zerovec",
 ]
 
-[[package]]
-name = "powerfmt"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "439ee305def115ba05938db6eb1644ff94165c5ab5e9420d1c1bcedbba909391"
-
 [[package]]
 name = "ppv-lite86"
 version = "0.2.21"
@@ -3543,16 +2960,6 @@ version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "925383efa346730478fb4838dbe9137d2a47675ad789c546d150a6e1dd4ab31c"
 
-[[package]]
-name = "prettyplease"
-version = "0.2.37"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "479ca8adacdd7ce8f1fb39ce9ecccbfe93a3f1344b3d0d97f20bc0196208f62b"
-dependencies = [
- "proc-macro2",
- "syn",
-]
-
 [[package]]
 name = "primal-check"
 version = "0.3.4"
@@ -3562,25 +2969,6 @@ dependencies = [
  "num-integer",
 ]
 
-[[package]]
-name = "proc-macro-crate"
-version = "1.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7f4c021e1093a56626774e81216a4ce732a735e5bad4868a03f3ed65ca0c3919"
-dependencies = [
- "once_cell",
- "toml_edit 0.19.15",
-]
-
-[[package]]
-name = "proc-macro-crate"
-version = "3.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e67ba7e9b2b56446f1d419b1d807906278ffa1a658a8a5d8a39dcb1f5a78614f"
-dependencies = [
- "toml_edit 0.25.12+spec-1.1.0",
-]
-
 [[package]]
 name = "proc-macro-error-attr2"
 version = "2.0.0"
@@ -3661,11 +3049,20 @@ dependencies = [
  "memchr",
 ]
 
+[[package]]
+name = "quick-xml"
+version = "0.40.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2474bd2e5029e7ccb6abb2ba48cf2383a333851dedf495901544281590c7da7f"
+dependencies = [
+ "memchr",
+]
+
 [[package]]
 name = "quick_cache"
-version = "0.6.23"
+version = "0.6.24"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3a3db184a8b66cfe87f0263a1de147a6b554c864d1767c6f7fa4eb0e5497b565"
+checksum = "b9c6658afe513a3b484e3abfdaa0d03ef3c0bbf017542c178dd55f94eb3051f9"
 dependencies = [
  "equivalent",
  "hashbrown 0.16.1",
@@ -3673,9 +3070,9 @@ dependencies = [
 
 [[package]]
 name = "quinn"
-version = "0.11.9"
+version = "0.11.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9e20a958963c291dc322d98411f541009df2ced7b5a4f2bd52337638cfccf20"
+checksum = "0c1a41e437b6bbd489372cd4971de128e85c855f56c57f283d20ff016cf7c0a8"
 dependencies = [
  "bytes",
  "cfg_aliases",
@@ -3693,9 +3090,9 @@ dependencies = [
 
 [[package]]
 name = "quinn-proto"
-version = "0.11.14"
+version = "0.11.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "434b42fec591c96ef50e21e886936e66d3cc3f737104fdb9b737c40ffb94c098"
+checksum = "4fcb935c5bec503c2f0e306bdd3e58bb9029dcb14fa8d9ac76e3a5256ac0763e"
 dependencies = [
  "aws-lc-rs",
  "bytes",
@@ -3729,9 +3126,9 @@ dependencies = [
 
 [[package]]
 name = "quote"
-version = "1.0.45"
+version = "1.0.46"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "41f2619966050689382d2b44f664f4bc593e129785a36d6ee376ddf37259b924"
+checksum = "dfbc457d0c7a0759a614551b11a6409e5951f6c7537be1f1b7682b9ae9230368"
 dependencies = [
  "proc-macro2",
 ]
@@ -3748,24 +3145,13 @@ version = "6.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f8dcc9c7d52a811697d2151c701e0d08956f92b0e24136cf4cf27b57a6a0d9bf"
 
-[[package]]
-name = "rand"
-version = "0.8.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5ca0ecfa931c29007047d1bc58e623ab12e5590e8c7cc53200d5202b69266d8a"
-dependencies = [
- "libc",
- "rand_chacha 0.3.1",
- "rand_core 0.6.4",
-]
-
 [[package]]
 name = "rand"
 version = "0.9.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "44c5af06bb1b7d3216d91932aed5265164bf384dc89cd6ba05cf59a35f5f76ea"
 dependencies = [
- "rand_chacha 0.9.0",
+ "rand_chacha",
  "rand_core 0.9.5",
 ]
 
@@ -3776,20 +3162,10 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d2e8e8bcc7961af1fdac401278c6a831614941f6164ee3bf4ce61b7edb162207"
 dependencies = [
  "chacha20",
- "getrandom 0.4.2",
+ "getrandom 0.4.3",
  "rand_core 0.10.1",
 ]
 
-[[package]]
-name = "rand_chacha"
-version = "0.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e6c10a63a0fa32252be49d21e7709d4d4baf8d231c2dbce1eaa8141b9b127d88"
-dependencies = [
- "ppv-lite86",
- "rand_core 0.6.4",
-]
-
 [[package]]
 name = "rand_chacha"
 version = "0.9.0"
@@ -3800,15 +3176,6 @@ dependencies = [
  "rand_core 0.9.5",
 ]
 
-[[package]]
-name = "rand_core"
-version = "0.6.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ec0be4795e2f6a28069bec0b5ff3e2ac9bafc99e6a9a7dc3547996c5c816922c"
-dependencies = [
- "getrandom 0.2.17",
-]
-
 [[package]]
 name = "rand_core"
 version = "0.9.5"
@@ -3834,12 +3201,6 @@ dependencies = [
  "rand 0.10.1",
 ]
 
-[[package]]
-name = "rangemap"
-version = "1.7.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "973443cf09a9c8656b574a866ab68dfa19f0867d0340648c7d2f6a71b8a8ea68"
-
 [[package]]
 name = "rav1e"
 version = "0.8.1"
@@ -3868,7 +3229,7 @@ dependencies = [
  "paste",
  "profiling",
  "rand 0.9.4",
- "rand_chacha 0.9.0",
+ "rand_chacha",
  "simd_helpers",
  "thiserror",
  "v_frame",
@@ -3922,7 +3283,7 @@ version = "0.5.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ed2bf2547551a7053d6fdfafda3f938979645c44812fbfcda098faae3f1a362d"
 dependencies = [
- "bitflags",
+ "bitflags 2.13.0",
 ]
 
 [[package]]
@@ -4006,10 +3367,8 @@ checksum = "219c5811de6525e5416c7d5d53bb656d3afdbc6c5af816e0802bcfa42dbdc1c3"
 dependencies = [
  "base64",
  "bytes",
- "encoding_rs",
  "futures-core",
  "futures-util",
- "h2",
  "http",
  "http-body",
  "http-body-util",
@@ -4018,7 +3377,6 @@ dependencies = [
  "hyper-util",
  "js-sys",
  "log",
- "mime",
  "mime_guess",
  "percent-encoding",
  "pin-project-lite",
@@ -4031,14 +3389,12 @@ dependencies = [
  "sync_wrapper",
  "tokio",
  "tokio-rustls",
- "tokio-util",
  "tower",
  "tower-http",
  "tower-service",
  "url",
  "wasm-bindgen",
  "wasm-bindgen-futures",
- "wasm-streams",
  "web-sys",
 ]
 
@@ -4078,22 +3434,6 @@ dependencies = [
  "wasmtimer",
 ]
 
-[[package]]
-name = "reqwest-tracing"
-version = "0.7.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b5e5af0cd6fc3d3c8f703d597af70b6e4e62432c63157b49419fa1ffaf481702"
-dependencies = [
- "anyhow",
- "async-trait",
- "getrandom 0.2.17",
- "http",
- "matchit",
- "reqwest",
- "reqwest-middleware",
- "tracing",
-]
-
 [[package]]
 name = "retry-policies"
 version = "0.5.2"
@@ -4109,66 +3449,6 @@ version = "0.8.53"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "47b34b781b31e5d73e9fbc8689c70551fd1ade9a19e3e28cfec8580a79290cc4"
 
-[[package]]
-name = "rig"
-version = "0.38.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "95947817936a8a51866a8b7d1599b721cd7d1567dd49257dfaf55e13d59110bf"
-dependencies = [
- "rig-core",
-]
-
-[[package]]
-name = "rig-core"
-version = "0.38.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "557f11c26c2c2ea61d9cb843ce5adff138cc5785f58ff4b87d779de8acaa32ae"
-dependencies = [
- "as-any",
- "async-stream",
- "base64",
- "bytes",
- "eventsource-stream",
- "fastrand",
- "futures",
- "futures-timer",
- "glob",
- "http",
- "mime",
- "mime_guess",
- "nanoid",
- "ordered-float",
- "pin-project-lite",
- "reqwest",
- "reqwest-middleware",
- "rig-derive",
- "schemars",
- "serde",
- "serde_json",
- "thiserror",
- "tokio",
- "tokio-tungstenite",
- "tracing",
- "tracing-futures",
- "url",
-]
-
-[[package]]
-name = "rig-derive"
-version = "0.38.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "643eb495acbd4bd5976164cd068f186d534f741def8d5d052f860266b98d07f8"
-dependencies = [
- "convert_case 0.11.0",
- "deluxe",
- "indoc",
- "proc-macro-crate 3.5.0",
- "proc-macro2",
- "quote",
- "serde_json",
- "syn",
-]
-
 [[package]]
 name = "ring"
 version = "0.17.14"
@@ -4227,18 +3507,18 @@ version = "1.1.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6fe4565b9518b83ef4f91bb47ce29620ca828bd32cb7e408f0062e9930ba190"
 dependencies = [
- "bitflags",
+ "bitflags 2.13.0",
  "errno",
  "libc",
  "linux-raw-sys",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
 name = "rustls"
-version = "0.23.40"
+version = "0.23.41"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ef86cd5876211988985292b91c96a8f2d298df24e75989a43a3c73f2d4d8168b"
+checksum = "6b92b125634d9b795e7beca796cc790df15a7fb38323bf3196fda83292d06b1f"
 dependencies = [
  "aws-lc-rs",
  "once_cell",
@@ -4288,7 +3568,7 @@ dependencies = [
  "security-framework",
  "security-framework-sys",
  "webpki-root-certs",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -4409,7 +3689,7 @@ version = "3.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b7f4bc775c73d9a02cde8bf7b2ec4c9d12743edf609006c7facc23998404cd1d"
 dependencies = [
- "bitflags",
+ "bitflags 2.13.0",
  "core-foundation",
  "core-foundation-sys",
  "libc",
@@ -4432,7 +3712,7 @@ version = "0.38.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8adfa1c298912827b8a28b223b3b874357397ae706e6190acd9bf28cee99114d"
 dependencies = [
- "bitflags",
+ "bitflags 2.13.0",
  "cssparser",
  "derive_more",
  "log",
@@ -4471,17 +3751,6 @@ dependencies = [
  "serde_derive",
 ]
 
-[[package]]
-name = "serde-wasm-bindgen"
-version = "0.6.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8302e169f0eddcc139c70f139d19d6467353af16f9fce27e8c30158036a1e16b"
-dependencies = [
- "js-sys",
- "serde",
- "wasm-bindgen",
-]
-
 [[package]]
 name = "serde_core"
 version = "1.0.228"
@@ -4591,17 +3860,6 @@ dependencies = [
  "xxhash-rust",
 ]
 
-[[package]]
-name = "sha1"
-version = "0.10.6"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e3bf829a2d51ab4a5ddf1352d8470c140cadc8301b2ae1789db023f01cedd6ba"
-dependencies = [
- "cfg-if",
- "cpufeatures 0.2.17",
- "digest 0.10.7",
-]
-
 [[package]]
 name = "sha2"
 version = "0.10.9"
@@ -4706,9 +3964,9 @@ checksum = "0c790de23124f9ab44544d7ac05d60440adc586479ce501c1d6d7da3cd8c9cf5"
 
 [[package]]
 name = "smallvec"
-version = "1.15.1"
+version = "1.15.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "67b1b7a3b5fe4f1376887184045fcf45c69e92af734b7aaddc05fb777b6fbd03"
+checksum = "8ed6a63f02c8539c91a8685a86f4099661ba3da017932f6ebbea6de3f0fa7c90"
 
 [[package]]
 name = "socket2"
@@ -4717,7 +3975,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "52d1cfed4120b4d927bf7c0f86d2087a4a7d6027c906d9f9d525a80573b9be51"
 dependencies = [
  "libc",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -4765,23 +4023,6 @@ dependencies = [
  "quote",
 ]
 
-[[package]]
-name = "stringprep"
-version = "0.1.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7b4df3d392d81bd458a8a621b8bffbd2302a12ffe288a9d931670948749463b1"
-dependencies = [
- "unicode-bidi",
- "unicode-normalization",
- "unicode-properties",
-]
-
-[[package]]
-name = "strsim"
-version = "0.10.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "73473c0e59e6d5812c5dfe2a064a6444949f089e20eec9a2e5506596494e4623"
-
 [[package]]
 name = "strsim"
 version = "0.11.1"
@@ -4812,7 +4053,7 @@ version = "0.27.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7695ce3845ea4b33927c055a39dc438a45b059f7c1b3d91d38d10355fb8cbca7"
 dependencies = [
- "heck 0.5.0",
+ "heck",
  "proc-macro2",
  "quote",
  "syn",
@@ -4824,7 +4065,7 @@ version = "0.28.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ab85eea0270ee17587ed4156089e10b9e6880ee688791d45a905f5b1ca36f664"
 dependencies = [
- "heck 0.5.0",
+ "heck",
  "proc-macro2",
  "quote",
  "syn",
@@ -4877,7 +4118,7 @@ version = "0.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "95ec293b5f288383b72a7bffcade6b2860b642cf66f28b3bd5967349a49938b1"
 dependencies = [
- "bitflags",
+ "bitflags 2.13.0",
  "bytemuck",
  "lazy_static",
  "log",
@@ -4912,9 +4153,9 @@ dependencies = [
 
 [[package]]
 name = "syn"
-version = "2.0.117"
+version = "2.0.118"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e665b8803e7b1d2a727f4023456bbbbe74da67099c585258af0ad9c5013b9b99"
+checksum = "1b9ae57f904213ebb649ce6895b8a66c66f0203b9319718f69a5612a065b1422"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -4948,10 +4189,10 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
 dependencies = [
  "fastrand",
- "getrandom 0.4.2",
+ "getrandom 0.4.3",
  "once_cell",
  "rustix",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -5004,38 +4245,7 @@ dependencies = [
  "half",
  "quick-error",
  "weezl",
- "zune-jpeg",
-]
-
-[[package]]
-name = "time"
-version = "0.3.47"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "743bd48c283afc0388f9b8827b976905fb217ad9e647fae3a379a9283c4def2c"
-dependencies = [
- "deranged",
- "itoa",
- "num-conv",
- "powerfmt",
- "serde_core",
- "time-core",
- "time-macros",
-]
-
-[[package]]
-name = "time-core"
-version = "0.1.8"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7694e1cfe791f8d31026952abf09c69ca6f6fa4e1a1229e18988f06a04a12dca"
-
-[[package]]
-name = "time-macros"
-version = "0.2.27"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2e70e4c5a0e0a8a4823ad65dfe1a6930e4f4d756dcd9dd7939022b5e8c501215"
-dependencies = [
- "num-conv",
- "time-core",
+ "zune-jpeg",
 ]
 
 [[package]]
@@ -5100,22 +4310,6 @@ dependencies = [
  "tokio",
 ]
 
-[[package]]
-name = "tokio-tungstenite"
-version = "0.28.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d25a406cddcc431a75d3d9afc6a7c0f7428d4891dd973e4d54c56b46127bf857"
-dependencies = [
- "futures-util",
- "log",
- "rustls",
- "rustls-pki-types",
- "tokio",
- "tokio-rustls",
- "tungstenite",
- "webpki-roots 0.26.11",
-]
-
 [[package]]
 name = "tokio-util"
 version = "0.7.18"
@@ -5138,18 +4332,12 @@ dependencies = [
  "indexmap",
  "serde_core",
  "serde_spanned",
- "toml_datetime 1.1.1+spec-1.1.0",
+ "toml_datetime",
  "toml_parser",
  "toml_writer",
- "winnow 1.0.3",
+ "winnow",
 ]
 
-[[package]]
-name = "toml_datetime"
-version = "0.6.11"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "22cddaf88f4fbc13c51aebbf5f8eceb5c7c5a9da2ac40a13519eb5b0a0e8f11c"
-
 [[package]]
 name = "toml_datetime"
 version = "1.1.1+spec-1.1.0"
@@ -5159,36 +4347,13 @@ dependencies = [
  "serde_core",
 ]
 
-[[package]]
-name = "toml_edit"
-version = "0.19.15"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1b5bb770da30e5cbfde35a2d7b9b8a2c4b8ef89548a7a6aeab5c9a576e3e7421"
-dependencies = [
- "indexmap",
- "toml_datetime 0.6.11",
- "winnow 0.5.40",
-]
-
-[[package]]
-name = "toml_edit"
-version = "0.25.12+spec-1.1.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d2153edc6955a6c354fad8f5efd38b6a8769bdccf9fe50f8e1329f81b0baa5d7"
-dependencies = [
- "indexmap",
- "toml_datetime 1.1.1+spec-1.1.0",
- "toml_parser",
- "winnow 1.0.3",
-]
-
 [[package]]
 name = "toml_parser"
 version = "1.1.2+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a2abe9b86193656635d2411dc43050282ca48aa31c2451210f4202550afb7526"
 dependencies = [
- "winnow 1.0.3",
+ "winnow",
 ]
 
 [[package]]
@@ -5220,7 +4385,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4cfcf7e2740e6fc6d4d688b4ef00650406bb94adf4731e43c096c3a19fe40840"
 dependencies = [
  "async-compression",
- "bitflags",
+ "bitflags 2.13.0",
  "bytes",
  "futures-core",
  "futures-util",
@@ -5283,18 +4448,6 @@ dependencies = [
  "valuable",
 ]
 
-[[package]]
-name = "tracing-futures"
-version = "0.2.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "97d095ae15e245a057c8e8451bab9b3ee1e1f68e9ba2b4fbc18d0ac5237835f2"
-dependencies = [
- "futures",
- "futures-task",
- "pin-project",
- "tracing",
-]
-
 [[package]]
 name = "tracing-log"
 version = "0.2.0"
@@ -5359,25 +4512,6 @@ version = "0.25.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d2df906b07856748fa3f6e0ad0cbaa047052d4a7dd609e231c4f72cee8c36f31"
 
-[[package]]
-name = "tungstenite"
-version = "0.28.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8628dcc84e5a09eb3d8423d6cb682965dea9133204e8fb3efee74c2a0c259442"
-dependencies = [
- "bytes",
- "data-encoding",
- "http",
- "httparse",
- "log",
- "rand 0.9.4",
- "rustls",
- "rustls-pki-types",
- "sha1",
- "thiserror",
- "utf-8",
-]
-
 [[package]]
 name = "twox-hash"
 version = "2.1.2"
@@ -5393,6 +4527,12 @@ dependencies = [
  "rustc-hash",
 ]
 
+[[package]]
+name = "typeid"
+version = "1.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bc7d623258602320d5c55d1bc22793b57daff0ec7efc270ea7d55ce1d5f5471c"
+
 [[package]]
 name = "typenum"
 version = "1.20.1"
@@ -5405,12 +4545,6 @@ version = "2.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "dbc4bc3a9f746d862c45cb89d705aa10f187bb96c76001afab07a0d35ce60142"
 
-[[package]]
-name = "unicode-bidi"
-version = "0.3.18"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5c1cb5db39152898a79168971543b1cb5020dff7fe43c8dc468b0885f5e29df5"
-
 [[package]]
 name = "unicode-ident"
 version = "1.0.24"
@@ -5426,12 +4560,6 @@ dependencies = [
  "tinyvec",
 ]
 
-[[package]]
-name = "unicode-properties"
-version = "0.1.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7df058c713841ad818f1dc5d3fd88063241cc61f49f5fbea4b951e8cf5a8d71d"
-
 [[package]]
 name = "unicode-segmentation"
 version = "1.13.3"
@@ -5450,16 +4578,6 @@ version = "0.2.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ebc1c04c71510c7f702b52b7c350734c9ff1295c464a03335b00bb84fc54f853"
 
-[[package]]
-name = "universal-hash"
-version = "0.5.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fc1de2c688dc15305988b563c3854064043356019f97a4b46276fe734c4f07ea"
-dependencies = [
- "crypto-common 0.1.7",
- "subtle",
-]
-
 [[package]]
 name = "untrusted"
 version = "0.9.0"
@@ -5484,15 +4602,6 @@ version = "0.7.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "09cc8ee72d2a9becf2f2febe0205bbed8fc6615b7cb429ad062dc7b7ddd036a9"
 
-[[package]]
-name = "utf16string"
-version = "0.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0b62a1e85e12d5d712bf47a85f426b73d303e2d00a90de5f3004df3596e9d216"
-dependencies = [
- "byteorder",
-]
-
 [[package]]
 name = "utf8_iter"
 version = "1.0.4"
@@ -5507,11 +4616,11 @@ checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
 
 [[package]]
 name = "uuid"
-version = "1.23.3"
+version = "1.23.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "144d6b123cef80b301b8f72a9e2ca4370ddec21950d0a103dd22c437006d2db7"
+checksum = "bf80a72845275afea99e7f2b434723d3bc7e38470fcd1c7ed39a599c73319a53"
 dependencies = [
- "getrandom 0.4.2",
+ "getrandom 0.4.3",
  "js-sys",
  "serde_core",
  "wasm-bindgen",
@@ -5528,36 +4637,6 @@ dependencies = [
  "wasm-bindgen",
 ]
 
-[[package]]
-name = "validator"
-version = "0.20.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "43fb22e1a008ece370ce08a3e9e4447a910e92621bb49b85d6e48a45397e7cfa"
-dependencies = [
- "idna",
- "once_cell",
- "regex",
- "serde",
- "serde_derive",
- "serde_json",
- "url",
- "validator_derive",
-]
-
-[[package]]
-name = "validator_derive"
-version = "0.20.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7df16e474ef958526d1205f6dda359fdfab79d9aa6d54bafcb92dcd07673dca"
-dependencies = [
- "darling 0.20.11",
- "once_cell",
- "proc-macro-error2",
- "proc-macro2",
- "quote",
- "syn",
-]
-
 [[package]]
 name = "valuable"
 version = "0.1.1"
@@ -5566,18 +4645,9 @@ checksum = "ba73ea9cf16a25df0c8caa16c51acb937d5712a8429db78a3ee29d5dcacd3a65"
 
 [[package]]
 name = "varint-rs"
-version = "2.2.0"
+version = "2.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8f54a172d0620933a27a4360d3db3e2ae0dd6cceae9730751a036bbf182c4b23"
-
-[[package]]
-name = "vecmath"
-version = "1.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "956ae1e0d85bca567dee1dcf87fb1ca2e792792f66f87dced8381f99cd91156a"
-dependencies = [
- "piston-float",
-]
+checksum = "bfa6c38708f6257f1ec2ca7e5a11f9bbf58a27d7060078b6b333624968183d96"
 
 [[package]]
 name = "version_check"
@@ -5612,27 +4682,18 @@ checksum = "ccf3ec651a847eb01de73ccad15eb7d99f80485de043efb2f370cd654f4ea44b"
 
 [[package]]
 name = "wasip2"
-version = "1.0.3+wasi-0.2.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "20064672db26d7cdc89c7798c48a0fdfac8213434a1186e5ef29fd560ae223d6"
-dependencies = [
- "wit-bindgen 0.57.1",
-]
-
-[[package]]
-name = "wasip3"
-version = "0.4.0+wasi-0.3.0-rc-2026-01-06"
+version = "1.0.4+wasi-0.2.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5428f8bf88ea5ddc08faddef2ac4a67e390b88186c703ce6dbd955e1c145aca5"
+checksum = "b67efb37e106e55ce722a510d6b5f9c17f083e5fc79afc2badeb12cc313d9487"
 dependencies = [
- "wit-bindgen 0.51.0",
+ "wit-bindgen",
 ]
 
 [[package]]
 name = "wasm-bindgen"
-version = "0.2.122"
+version = "0.2.126"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3ed04576f974d2b2fba0f38c51dbc5518011e38c36bf1143164be765528fd409"
+checksum = "4b067c0c11094aef6b7a801c1e34a26affafdf3d051dba08456b868789aaf9a4"
 dependencies = [
  "cfg-if",
  "once_cell",
@@ -5643,9 +4704,9 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen-futures"
-version = "0.4.72"
+version = "0.4.76"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9473dbd2991ae90b6291c3c32c30c6187ac49aa32f9905d1cce280ec1e110b0f"
+checksum = "c62df1340f32221cb9c54d6a27b030e3dba64361d4a95bed55f9aacb44da291d"
 dependencies = [
  "js-sys",
  "wasm-bindgen",
@@ -5653,9 +4714,9 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen-macro"
-version = "0.2.122"
+version = "0.2.126"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "916151b09da36bd82f6615cbf3a419e2f0ba23a03c6160e8e92eb6bd4aa1dec6"
+checksum = "167ce5e579f6bcf889c4f7175a8a5a585de84e8ff93976ce393efa5f2837aab1"
 dependencies = [
  "quote",
  "wasm-bindgen-macro-support",
@@ -5663,9 +4724,9 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen-macro-support"
-version = "0.2.122"
+version = "0.2.126"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "299047362ccbfce148b67ab7e73349f77748e00c8296f9542adfad2ad82c5c5e"
+checksum = "f3997c7839262f4ef12cf90b818d6340c18e80f263f1a94bf157d0ec4420380e"
 dependencies = [
  "bumpalo",
  "proc-macro2",
@@ -5676,60 +4737,13 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen-shared"
-version = "0.2.122"
+version = "0.2.126"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9a929b2c61f11ba3e9bc35b50c1f25cb38e0e892c0c231ae2b8cf78d5dad4437"
+checksum = "dc1b4cb0cc549fcf58d7dfc081778139b3d283a081644e833e84682ad71cea24"
 dependencies = [
  "unicode-ident",
 ]
 
-[[package]]
-name = "wasm-encoder"
-version = "0.244.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "990065f2fe63003fe337b932cfb5e3b80e0b4d0f5ff650e6985b1048f62c8319"
-dependencies = [
- "leb128fmt",
- "wasmparser",
-]
-
-[[package]]
-name = "wasm-metadata"
-version = "0.244.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bb0e353e6a2fbdc176932bbaab493762eb1255a7900fe0fea1a2f96c296cc909"
-dependencies = [
- "anyhow",
- "indexmap",
- "wasm-encoder",
- "wasmparser",
-]
-
-[[package]]
-name = "wasm-streams"
-version = "0.5.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d1ec4f6517c9e11ae630e200b2b65d193279042e28edd4a2cda233e46670bbb"
-dependencies = [
- "futures-util",
- "js-sys",
- "wasm-bindgen",
- "wasm-bindgen-futures",
- "web-sys",
-]
-
-[[package]]
-name = "wasmparser"
-version = "0.244.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "47b807c72e1bac69382b3a6fb3dbe8ea4c0ed87ff5629b8685ae6b9a611028fe"
-dependencies = [
- "bitflags",
- "hashbrown 0.15.5",
- "indexmap",
- "semver",
-]
-
 [[package]]
 name = "wasmtimer"
 version = "0.4.3"
@@ -5746,9 +4760,9 @@ dependencies = [
 
 [[package]]
 name = "web-sys"
-version = "0.3.99"
+version = "0.3.103"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6d621441cfc37b84979402712047321980c178f299193a3589d05b99e8763436"
+checksum = "8622dcb61c0bcc9fffa6938bed81210af2da9a7e4a1a834b2e37a59b6dfb6141"
 dependencies = [
  "js-sys",
  "wasm-bindgen",
@@ -5766,9 +4780,9 @@ dependencies = [
 
 [[package]]
 name = "web_atoms"
-version = "0.2.4"
+version = "0.2.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d7cff6eef815df1834fd250e3a2ff436044d82a9f1bc1980ca1dbdf07effc538"
+checksum = "075474b12bcb3d2e3d4546580e9de478eeeead668a1761e2a8860c836b7ef297"
 dependencies = [
  "phf",
  "phf_codegen",
@@ -5778,27 +4792,9 @@ dependencies = [
 
 [[package]]
 name = "webpki-root-certs"
-version = "1.0.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f31141ce3fc3e300ae89b78c0dd67f9708061d1d2eda54b8209346fd6be9a92c"
-dependencies = [
- "rustls-pki-types",
-]
-
-[[package]]
-name = "webpki-roots"
-version = "0.26.11"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "521bc38abb08001b01866da9f51eb7c5d647a19260e00054a8c7fd5f9e57f7a9"
-dependencies = [
- "webpki-roots 1.0.7",
-]
-
-[[package]]
-name = "webpki-roots"
-version = "1.0.7"
+version = "1.0.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "52f5ee44c96cf55f1b349600768e3ece3a8f26010c05265ab73f945bb1a2eb9d"
+checksum = "0d46a5a140e6f7afeccd8eae97eff335163939eac8b929834875168b29b3d267"
 dependencies = [
  "rustls-pki-types",
 ]
@@ -5825,42 +4821,7 @@ version = "0.1.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
 dependencies = [
- "windows-sys 0.52.0",
-]
-
-[[package]]
-name = "windows-core"
-version = "0.62.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
-dependencies = [
- "windows-implement",
- "windows-interface",
- "windows-link",
- "windows-result",
- "windows-strings",
-]
-
-[[package]]
-name = "windows-implement"
-version = "0.60.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
-]
-
-[[package]]
-name = "windows-interface"
-version = "0.59.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -5878,15 +4839,6 @@ dependencies = [
  "windows-link",
 ]
 
-[[package]]
-name = "windows-strings"
-version = "0.5.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7837d08f69c77cf6b07689544538e017c1bfcf57e34b4c0ff58e6c2cd3b37091"
-dependencies = [
- "windows-link",
-]
-
 [[package]]
 name = "windows-sys"
 version = "0.52.0"
@@ -6043,32 +4995,11 @@ version = "0.53.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d6bbff5f0aada427a1e5a6da5f1f98158182f26556f345ac9e04d36d0ebed650"
 
-[[package]]
-name = "winnow"
-version = "0.5.40"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f593a95398737aeed53e489c785df13f3618e41dbcd6718c6addbf1395aa6876"
-dependencies = [
- "memchr",
-]
-
 [[package]]
 name = "winnow"
 version = "1.0.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0592e1c9d151f854e6fd382574c3a0855250e1d9b2f99d9281c6e6391af352f1"
-dependencies = [
- "memchr",
-]
-
-[[package]]
-name = "wit-bindgen"
-version = "0.51.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d7249219f66ced02969388cf2bb044a09756a083d0fab1e566056b04d9fbcaa5"
-dependencies = [
- "wit-bindgen-rust-macro",
-]
 
 [[package]]
 name = "wit-bindgen"
@@ -6076,85 +5007,6 @@ version = "0.57.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1ebf944e87a7c253233ad6766e082e3cd714b5d03812acc24c318f549614536e"
 
-[[package]]
-name = "wit-bindgen-core"
-version = "0.51.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ea61de684c3ea68cb082b7a88508a8b27fcc8b797d738bfc99a82facf1d752dc"
-dependencies = [
- "anyhow",
- "heck 0.5.0",
- "wit-parser",
-]
-
-[[package]]
-name = "wit-bindgen-rust"
-version = "0.51.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7c566e0f4b284dd6561c786d9cb0142da491f46a9fbed79ea69cdad5db17f21"
-dependencies = [
- "anyhow",
- "heck 0.5.0",
- "indexmap",
- "prettyplease",
- "syn",
- "wasm-metadata",
- "wit-bindgen-core",
- "wit-component",
-]
-
-[[package]]
-name = "wit-bindgen-rust-macro"
-version = "0.51.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0c0f9bfd77e6a48eccf51359e3ae77140a7f50b1e2ebfe62422d8afdaffab17a"
-dependencies = [
- "anyhow",
- "prettyplease",
- "proc-macro2",
- "quote",
- "syn",
- "wit-bindgen-core",
- "wit-bindgen-rust",
-]
-
-[[package]]
-name = "wit-component"
-version = "0.244.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d66ea20e9553b30172b5e831994e35fbde2d165325bec84fc43dbf6f4eb9cb2"
-dependencies = [
- "anyhow",
- "bitflags",
- "indexmap",
- "log",
- "serde",
- "serde_derive",
- "serde_json",
- "wasm-encoder",
- "wasm-metadata",
- "wasmparser",
- "wit-parser",
-]
-
-[[package]]
-name = "wit-parser"
-version = "0.244.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ecc8ac4bc1dc3381b7f59c34f00b67e18f910c2c0f50015669dde7def656a736"
-dependencies = [
- "anyhow",
- "id-arena",
- "indexmap",
- "log",
- "semver",
- "serde",
- "serde_derive",
- "serde_json",
- "unicode-xid",
- "wasmparser",
-]
-
 [[package]]
 name = "writeable"
 version = "0.6.3"
@@ -6198,18 +5050,18 @@ dependencies = [
 
 [[package]]
 name = "zerocopy"
-version = "0.8.50"
+version = "0.8.52"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3b065d4f0e55f82fae73202e189638116a87c55ab6b8e6c2721e13dd9d854ad1"
+checksum = "ce1022995ff5ff5d841ad7d994facc23098cd40152f2c1d11cd607c6f530653f"
 dependencies = [
  "zerocopy-derive",
 ]
 
 [[package]]
 name = "zerocopy-derive"
-version = "0.8.50"
+version = "0.8.52"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0b631b19d36a892ab55420c92dbc83ccd79274f25be714855d3074aa71cab639"
+checksum = "1ae7f38b72ec2a254e2b87ef277cf2cd4fb97cbebf944faa6f33354da0867930"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -6239,9 +5091,9 @@ dependencies = [
 
 [[package]]
 name = "zeroize"
-version = "1.8.2"
+version = "1.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b97154e67e32c85465826e8bcc1c59429aaaf107c1e4a9e53c8d8ccd5eff88d0"
+checksum = "e13c156562582aa81c60cb29407084cdb54c4164760106ab78e6c5b0858cf64e"
 
 [[package]]
 name = "zerotrie"
diff --git a/Cargo.toml b/Cargo.toml
index 8a6555cfc..2b0bd49e4 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -3,19 +3,11 @@
 [workspace]
 resolver = "3"
 members = [
+    "./crates/elide-bento",
     "./crates/nvisy-cli",
-    "./crates/nvisy-codec",
-    "./crates/nvisy-context",
     "./crates/nvisy-core",
     "./crates/nvisy-engine",
-    "./crates/nvisy-fake",
-    "./crates/nvisy-llm",
-    "./crates/nvisy-ner",
-    "./crates/nvisy-ocr",
-    "./crates/nvisy-pattern",
     "./crates/nvisy-server",
-    "./crates/nvisy-stt",
-    "./crates/nvisy-toolkit",
 ]
 
 [workspace.package]
@@ -36,19 +28,27 @@ documentation = "https://docs.rs/nvisy-runtime"
 #
 # See for more details: https://github.com/rust-lang/cargo/issues/11329
 
+# Elide toolkit (upstream)
+elide = { git = "https://github.com/nvisycom/elide", branch = "main", default-features = false }
+elide-core = { git = "https://github.com/nvisycom/elide", branch = "main" }
+# `elide-ner` + `elide-ocr` are pulled directly only by `elide-bento`, which
+# implements the per-backend traits these crates export. Engine and other
+# consumers reach the same types through `elide::recognition::{ner, ocr}`.
+elide-ner = { git = "https://github.com/nvisycom/elide", branch = "main", default-features = false }
+elide-ocr = { git = "https://github.com/nvisycom/elide", branch = "main", default-features = false }
+elide-stt = { git = "https://github.com/nvisycom/elide", branch = "main", default-features = false }
+
+# Runtime-owned elide extensions
+elide-bento = { path = "./crates/elide-bento", version = "0.1.0" }
+
 # Internal crates
-nvisy-codec = { path = "./crates/nvisy-codec", version = "0.1.0", default-features = false }
-nvisy-context = { path = "./crates/nvisy-context", version = "0.1.0" }
-nvisy-core = { path = "./crates/nvisy-core", version = "0.1.0" }
 nvisy-engine = { path = "./crates/nvisy-engine", version = "0.1.0" }
-nvisy-fake = { path = "./crates/nvisy-fake", version = "0.1.0" }
-nvisy-llm = { path = "./crates/nvisy-llm", version = "0.1.0" }
-nvisy-ner = { path = "./crates/nvisy-ner", version = "0.1.0" }
-nvisy-ocr = { path = "./crates/nvisy-ocr", version = "0.1.0" }
-nvisy-pattern = { path = "./crates/nvisy-pattern", version = "0.1.0" }
 nvisy-server = { path = "./crates/nvisy-server", version = "0.1.0" }
-nvisy-stt = { path = "./crates/nvisy-stt", version = "0.1.0" }
-nvisy-toolkit = { path = "./crates/nvisy-toolkit", version = "0.1.0" }
+
+# Internal crate being deleted: kept here as a path-dep so consumer
+# manifests (nvisy-engine/server/cli) parse during the migration.
+# Leaves the workspace once engine's source no longer imports it.
+nvisy-core = { path = "./crates/nvisy-core", version = "0.1.0" }
 
 # Serialization
 serde = { version = "1.0", features = ["derive"] }
diff --git a/Nvisy.example.toml b/Nvisy.example.toml
index 372b40332..423fe3613 100644
--- a/Nvisy.example.toml
+++ b/Nvisy.example.toml
@@ -1,13 +1,11 @@
 # Nvisy configuration
 #
-# Copy to Nvisy.toml (or run `make generate-config`), then edit as needed.
-# The server loads Nvisy.toml by default; override with --config or NVISY_CONFIG.
+# Copy to Nvisy.toml (or run `make generate-config`), then edit as
+# needed. The server loads Nvisy.toml by default; override with
+# --config or NVISY_CONFIG.
 #
-# All sections are optional: omitted sections use runtime defaults.
+# All sections are optional: omitted sections use server defaults.
 
-version = "0.1.0"
-
-# Server binding and storage.
 # CLI flags (--host, --port, --data-dir) override these.
 [server]
 host = "0.0.0.0"
@@ -24,107 +22,68 @@ body_limit_mb = 4
 request_timeout = "5m"
 cors = { allowed_origins = ["*"], max_age = "1h" }
 
-# Engine-level resource limits and shared infrastructure.
-[engine]
-
-# Run-level resource limits.
-[engine.limits]
-concurrency = 4              # max parallel documents in flight
-run_timeout = "60s"
-
-# Shared HTTP client for all downstream provider calls.
-[engine.http]
-max_retries = 3
-timeout = "120s"
-connect_timeout = "10s"
-idle_timeout = "90s"
-
-# Extraction registry.
-#
-# Each opted-in `[extraction.*]` section is built once at server
-# startup. Set `enabled = false` to keep the config but skip
-# construction and dispatch. The pipeline `Extraction` step carries
-# per-call flags only.
-
-# Extraction: OCR (optical character recognition, images only).
-[extraction.ocr]
-enabled = true
-
-[extraction.ocr.backend]
-kind = "bento"
-base_url = "http://localhost:3001"
-
-# Extraction: STT (speech-to-text, audio only).
-[extraction.stt]
-enabled = true
-
-[extraction.stt.backend]
-kind = "noop"
-
-# Detection: pattern-based (regex + dictionary + checksum).
-# Optional section — pattern detection runs by default. Include this
-# block to narrow it (named patterns, tag filters) or to disable it
-# entirely with `enabled = false`.
-[detection.pattern]
-enabled = true
-patterns = [] # empty = all built-in patterns
-
-# Detection: NER (named entity recognition).
-[detection.ner]
-enabled = true
-
-[detection.ner.backend]
-kind = "bento"
-base_url = "http://localhost:3000"
-
-# Detection: LLM (large language model).
-[detection.llm]
-enabled = true
-unresolved_policy = "drop" # drop | first-match
-
-[detection.llm.provider]
-kind = "open-ai"
-api_key = "sk-example-replace-me"
-model = "gpt-4o"
-
-[detection.llm.detect]
-enabled = true
-temperature = 0.1
-max_tokens = 4096
-max_retries = 3
-# context_window = { ... }   # optional, for chunking large inputs
-# preamble = "..."           # optional, overrides the agent's default system prompt
-
-[detection.llm.verify]
-enabled = true
-temperature = 0.1
-max_tokens = 4096
-max_retries = 3
-# context_window = { ... }
-# preamble = "..."
-
-# Detection: VLM (vision-language model, images only).
-[detection.vlm]
-enabled = true
-
-[detection.vlm.provider]
-kind = "open-ai"
-api_key = "sk-example-replace-me"
-model = "gpt-4o"
-
-[detection.vlm.detect]
-enabled = true
-temperature = 0.1
-max_tokens = 4096
-max_retries = 3
-
-[detection.vlm.verify]
-enabled = true
-temperature = 0.1
-max_tokens = 4096
-max_retries = 3
-
-# Redaction params.
-[redaction]
-confidence_threshold = 0.5
-process_metadata = false
+# The deployment default for AnalyzerParams. Requests carry
+# per-field `analyzer` overrides on top of this default (inherit
+# the slot, replace it, remove it from an optional slot, or patch
+# a list with extend + remove selectors). Omit the whole
+# [analyzer] section and the server falls back to the type-level
+# default — no recognizers, no enrichers, default dedup + scope,
+# empty catalog — useful for tests, not useful in production.
+
+# Caller-asserted scope threaded into every recognizer's context.
+# Add languages and country codes the analyzer should consider
+# (empty = no restriction).
+[analyzer.scope]
+languages = []
+jurisdictions = []
+
+# Deduplication pipeline applied after recognition:
+# calibrate → fuse → resolve → filter.
+[analyzer.deduplication]
+fusion = "max_confidence"        # max_confidence | mean | noisy_or
+resolution = "highest_confidence" # highest_confidence | longest_span
+min_confidence = 0.7              # filter threshold, 0.0..=1.0
+
+# Per-request label catalog. Entries are LabelSchema records:
+# { name, description?, tags?, metadata? }. The server unions
+# every submitted policy's labels with this catalog at request
+# time; entries here are the baseline.
+# [[analyzer.label_catalog]]
+# name = "email_address"
+# description = "RFC 5322 email"
+# tags = ["pii", "contact"]
+
+# Pattern recognizer: at-most-one per analyzer. Loads every
+# shipped pattern + dictionary and wraps the recognizer in
+# elide's context-boost layer.
+[analyzer.recognizers.pattern]
+builtins = true
+context_enhanced = true
+
+# NER recognizer list. Each entry needs a unique `name`; multiple
+# entries run in parallel (e.g. one English model, one Spanish).
+# [[analyzer.recognizers.ner]]
+# name = "default_ner"
+# backend = { kind = "mock" }
+# # backend = { kind = "bento", base_url = "http://localhost:3000", model = "..." }
+
+# LLM recognizer list. Same shape; backend chooses provider.
+# [[analyzer.recognizers.llm]]
+# name = "default_llm"
+# backend = { kind = "mock" }
+
+# Language enricher (writes the document's detected languages
+# into the recognizer context, driving jurisdiction-aware
+# dispatch). At-most-one per analyzer.
+# [analyzer.enrichers.language]
+# min_confidence = 0.6              # 0.0..=1.0; None lets the engine choose
+
+# OCR enricher (image modality only). At-most-one per analyzer.
+# [analyzer.enrichers.ocr]
+# backend = { kind = "mock" }
+# # backend = { kind = "bento", base_url = "http://localhost:3001", model = "..." }
+
+# STT enricher (audio modality only). At-most-one per analyzer.
+# [analyzer.enrichers.stt]
+# backend = { kind = "mock" }
+# # backend = { kind = "bento", base_url = "http://localhost:3002", model = "..." }
diff --git a/crates/elide-bento/Cargo.toml b/crates/elide-bento/Cargo.toml
new file mode 100644
index 000000000..f7b5d2cbd
--- /dev/null
+++ b/crates/elide-bento/Cargo.toml
@@ -0,0 +1,59 @@
+# https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[package]
+name = "elide-bento"
+description = "BentoML-backed NER and OCR backends for the elide toolkit"
+keywords = ["elide", "bento", "bentoml", "ner", "ocr"]
+categories = ["api-bindings"]
+readme = "README.md"
+
+version = { workspace = true }
+rust-version = { workspace = true }
+edition = { workspace = true }
+license = { workspace = true }
+publish = { workspace = true }
+
+authors = { workspace = true }
+repository = { workspace = true }
+homepage = { workspace = true }
+documentation = { workspace = true }
+
+[package.metadata.docs.rs]
+all-features = true
+rustdoc-args = ["--cfg", "docsrs"]
+
+[features]
+default = ["ner", "ocr"]
+## Ship the NER backend (implements `elide_ner::backend::NerBackend`
+## against the `nvisy-inference-ner` BentoML service).
+ner = ["dep:elide-ner"]
+## Ship the OCR backend (implements `elide_ocr::OcrBackend`
+## against the `nvisy-inference-ocr` BentoML service).
+ocr = ["dep:elide-ocr", "dep:base64", "elide-core/image"]
+
+[dependencies]
+# Elide toolkit (upstream)
+elide-core = { workspace = true, features = [] }
+elide-ner = { workspace = true, features = [], optional = true }
+elide-ocr = { workspace = true, features = [], optional = true }
+
+# Serialization
+serde = { workspace = true, features = ["derive"] }
+
+# Primitive datatypes
+hipstr = { workspace = true, features = [] }
+
+# Async runtime
+async-trait = { workspace = true, features = [] }
+
+# Error handling
+thiserror = { workspace = true, features = [] }
+
+# Image bytes → base64 (OCR only)
+base64 = { workspace = true, features = [], optional = true }
+
+# BentoML client
+bentoml = { workspace = true, default-features = false, features = ["rustls-tls", "tracing"] }
+
+[dev-dependencies]
+tokio = { workspace = true, features = ["rt", "macros"] }
diff --git a/crates/elide-bento/README.md b/crates/elide-bento/README.md
new file mode 100644
index 000000000..49b8e23e3
--- /dev/null
+++ b/crates/elide-bento/README.md
@@ -0,0 +1,7 @@
+# elide-bento
+
+Shared BentoML HTTP client wrapper for elide backends.
+
+Per-modality backends (NER, OCR, …) live in their consuming crates
+(`elide-ner`, `elide-ocr`) and pull this crate for the common HTTP
+client, params validation, and error translation.
diff --git a/crates/elide-bento/src/error.rs b/crates/elide-bento/src/error.rs
new file mode 100644
index 000000000..fcdd5caeb
--- /dev/null
+++ b/crates/elide-bento/src/error.rs
@@ -0,0 +1,39 @@
+//! Error translation: `bentoml` errors → [`elide_core::Error`].
+//!
+//! Crate-private — the public API of every backend reports
+//! [`elide_core::Error`]; this enum is the internal seam the
+//! per-route helpers use before bubbling up.
+
+use elide_core::{Error, ErrorKind};
+
+/// Errors surfaced internally by the bento backends.
+///
+/// Two structural categories the consuming crate maps onto
+/// [`ErrorKind`] when bubbling up: transport (HTTP / network /
+/// client construction) and protocol (service answered but the
+/// body did not match the contract — decode error, batch length
+/// mismatch, …).
+#[derive(Debug, thiserror::Error)]
+pub(crate) enum BentoError {
+    /// HTTP / transport failure — client construction, network
+    /// I/O, status-code rejections.
+    #[error("bento transport error: {0}")]
+    Transport(#[from] bentoml::Error),
+    /// Protocol failure — the service answered but the body did not
+    /// match the contract.
+    #[error("bento protocol error: {0}")]
+    Protocol(String),
+}
+
+impl From<BentoError> for Error {
+    /// Map transport to [`ErrorKind::Transport`] and protocol to
+    /// [`ErrorKind::Validation`], carrying the original error as the
+    /// source cause.
+    fn from(err: BentoError) -> Self {
+        let kind = match err {
+            BentoError::Transport(_) => ErrorKind::Transport,
+            BentoError::Protocol(_) => ErrorKind::Validation,
+        };
+        Error::new(kind, err)
+    }
+}
diff --git a/crates/elide-bento/src/lib.rs b/crates/elide-bento/src/lib.rs
new file mode 100644
index 000000000..ad4151ba3
--- /dev/null
+++ b/crates/elide-bento/src/lib.rs
@@ -0,0 +1,18 @@
+#![forbid(unsafe_code)]
+#![cfg_attr(docsrs, feature(doc_cfg))]
+#![doc = include_str!("../README.md")]
+
+mod error;
+
+#[cfg(feature = "ner")]
+#[cfg_attr(docsrs, doc(cfg(feature = "ner")))]
+pub mod ner;
+
+#[cfg(feature = "ocr")]
+#[cfg_attr(docsrs, doc(cfg(feature = "ocr")))]
+pub mod ocr;
+
+#[cfg(feature = "ner")]
+pub use self::ner::BentoNer;
+#[cfg(feature = "ocr")]
+pub use self::ocr::BentoOcr;
diff --git a/crates/elide-bento/src/ner/mod.rs b/crates/elide-bento/src/ner/mod.rs
new file mode 100644
index 000000000..f7d8545b5
--- /dev/null
+++ b/crates/elide-bento/src/ner/mod.rs
@@ -0,0 +1,140 @@
+//! [`BentoNer`]: an [`elide_ner::backend::NerBackend`] backed by the
+//! `nvisy-inference-ner` BentoML service.
+//!
+//! Wire contract: `POST /recognize` accepts a batched list of
+//! requests (one schema-driven entity-extraction call per item)
+//! and returns the matching list of responses. Each request carries
+//! a schema (entities + optional classifications + structures); this
+//! backend uses entities only and ignores the rest. Per-call
+//! correlation IDs propagate as `x-request-id` headers when set.
+//!
+//! Wire types live in the private `request` (outgoing) and
+//! `response` (incoming) submodules; only the public
+//! [`BentoNer`] backend is part of this crate's API.
+
+mod request;
+mod response;
+
+use bentoml::{Client, Endpoint};
+use elide_core::Result;
+use elide_core::entity::provenance::ModelEvent;
+use elide_ner::backend::{NerBackend, NerRequest, NerResponse};
+use hipstr::HipStr;
+
+use self::request::WireNerRequest;
+use self::response::WireNerResponse;
+use crate::error::BentoError;
+
+const ROUTE: &str = "recognize";
+
+/// BentoML NER backend.
+///
+/// Owns a cached [`Endpoint`] pointing at the `nvisy-inference-ner`
+/// `/recognize` route, plus the per-deployment model id (echoed
+/// into [`NerBackend::provenance`]) and a default per-label
+/// confidence threshold the service applies when a schema entry
+/// does not pin its own.
+#[derive(Debug, Clone)]
+pub struct BentoNer {
+    /// Pre-built endpoint at the `/recognize` route. Cloned per
+    /// call so per-request headers (`x-request-id`) layer onto a
+    /// fresh instance without rebuilding the route.
+    endpoint: Endpoint,
+    /// Service-side model identifier echoed in provenance.
+    model_id: HipStr<'static>,
+    /// Default per-label confidence cutoff sent on every request.
+    /// Per-label thresholds in the schema override it.
+    default_threshold: f32,
+}
+
+impl BentoNer {
+    /// Build from a service URL + the deployment's model id. The
+    /// default per-label threshold starts at `0.5` (matches the
+    /// service's own default); use [`with_default_threshold`] to
+    /// override.
+    ///
+    /// [`with_default_threshold`]: Self::with_default_threshold
+    pub fn new(base_url: impl Into<String>, model_id: impl Into<HipStr<'static>>) -> Result<Self> {
+        let client = Client::builder()
+            .with_base_url(base_url)
+            .build()
+            .map_err(BentoError::Transport)?;
+        Ok(Self {
+            endpoint: client.endpoint(ROUTE),
+            model_id: model_id.into(),
+            default_threshold: 0.5,
+        })
+    }
+
+    /// Override the per-request default confidence threshold (the
+    /// service applies it when a schema entity has no per-label
+    /// `threshold` of its own).
+    #[must_use]
+    pub fn with_default_threshold(mut self, threshold: f32) -> Self {
+        self.default_threshold = threshold;
+        self
+    }
+
+    /// Send one batched `/recognize` POST and parse the response
+    /// body. Clones the cached endpoint so per-request headers
+    /// layer on without touching the original.
+    async fn post_recognize(
+        &self,
+        requests: &[NerRequest<'_>],
+    ) -> Result<Vec<WireNerResponse>, BentoError> {
+        let body: Vec<WireNerRequest> = requests
+            .iter()
+            .map(|r| WireNerRequest::from_request(r, self.default_threshold))
+            .collect();
+        let mut endpoint = self.endpoint.clone();
+        if let Some(id) = requests.iter().find_map(|r| r.correlation_id) {
+            endpoint = endpoint.with_request_id(id.to_string());
+        }
+        endpoint
+            .invoke::<_, Vec<WireNerResponse>>(&body)
+            .await
+            .map_err(BentoError::Transport)
+    }
+}
+
+#[async_trait::async_trait]
+impl NerBackend for BentoNer {
+    fn provenance(&self) -> ModelEvent {
+        ModelEvent {
+            name: self.model_id.clone(),
+            version: None,
+            contextual: false,
+        }
+    }
+
+    async fn recognize(&self, request: NerRequest<'_>) -> Result<NerResponse> {
+        let responses = self.post_recognize(&[request]).await?;
+        let mut iter = responses.into_iter();
+        let response = iter
+            .next()
+            .ok_or_else(|| BentoError::Protocol("bento ner returned an empty batch".into()))?;
+        if iter.next().is_some() {
+            return Err(BentoError::Protocol(
+                "bento ner returned more responses than requests".into(),
+            )
+            .into());
+        }
+        Ok(response.decode())
+    }
+
+    async fn recognize_batch(&self, requests: &[NerRequest<'_>]) -> Result<Vec<NerResponse>> {
+        if requests.is_empty() {
+            return Ok(Vec::new());
+        }
+        let responses = self.post_recognize(requests).await?;
+        if responses.len() != requests.len() {
+            return Err(BentoError::Protocol(format!(
+                "bento ner returned {} responses for {} requests",
+                responses.len(),
+                requests.len(),
+            ))
+            .into());
+        }
+        Ok(responses.into_iter().map(WireNerResponse::decode).collect())
+    }
+}
diff --git a/crates/elide-bento/src/ner/request.rs b/crates/elide-bento/src/ner/request.rs
new file mode 100644
index 000000000..4fb6f86b9
--- /dev/null
+++ b/crates/elide-bento/src/ner/request.rs
@@ -0,0 +1,72 @@
+//! Outgoing wire types for the NER `/recognize` endpoint.
+//!
+//! Mirrors `nvisy_core.ner.v1.NerRequest` from the inference
+//! repository: a `(text, schema, threshold)` triple where the
+//! schema lists the entities to extract. Classifications and
+//! structured records (also part of the upstream schema) are
+//! omitted — this backend surfaces entity extraction only.
+
+use elide_ner::backend::NerRequest;
+use serde::Serialize;
+
+/// Outgoing per-call request body element.
+#[derive(Debug, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireNerRequest {
+    /// Source text to scan.
+    pub text: String,
+    /// Schema describing what to extract.
+    pub schema: WireSchema,
+    /// Default per-label confidence cutoff. Per-entity `threshold`
+    /// overrides this when present.
+    pub threshold: f32,
+}
+
+impl WireNerRequest {
+    /// Translate an elide [`NerRequest`] into the wire shape,
+    /// pinning the service-default threshold when the request has
+    /// no per-label thresholds of its own.
+    pub(super) fn from_request(request: &NerRequest<'_>, default_threshold: f32) -> Self {
+        let entities = request
+            .labels
+            .map(|labels| {
+                labels
+                    .iter()
+                    .map(|label| WireEntitySpec {
+                        label: label.name().to_owned(),
+                        description: label.description().map(str::to_owned),
+                        threshold: None,
+                    })
+                    .collect()
+            })
+            .unwrap_or_default();
+        Self {
+            text: request.text.to_owned(),
+            schema: WireSchema { entities },
+            threshold: default_threshold,
+        }
+    }
+}
+
+/// `Schema` group: the entities the call extracts.
+#[derive(Debug, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireSchema {
+    pub entities: Vec<WireEntitySpec>,
+}
+
+/// One entity to extract.
+#[derive(Debug, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireEntitySpec {
+    /// Stable label identifier (e.g. `"email_address"`).
+    pub label: String,
+    /// Optional natural-language description that steers zero-shot
+    /// extraction.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub description: Option<String>,
+    /// Per-label confidence cutoff. `None` falls through to the
+    /// request-level [`WireNerRequest::threshold`].
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub threshold: Option<f32>,
+}
diff --git a/crates/elide-bento/src/ner/response.rs b/crates/elide-bento/src/ner/response.rs
new file mode 100644
index 000000000..6250da111
--- /dev/null
+++ b/crates/elide-bento/src/ner/response.rs
@@ -0,0 +1,53 @@
+//! Incoming wire types for the NER `/recognize` endpoint.
+//!
+//! Mirrors `nvisy_core.ner.v1.NerResponse` from the inference
+//! repository. Classifications, structures, and the response-level
+//! `modelId` are deserialised-and-discarded — this backend surfaces
+//! entity-extraction results only.
+
+use elide_ner::backend::{NerResponse, NerSpan};
+use serde::Deserialize;
+
+/// Incoming per-call response body element.
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireNerResponse {
+    /// Extracted entities, in backend order.
+    #[serde(default)]
+    pub entities: Vec<WireEntity>,
+    // `classifications`, `structures`, `modelId` ignored.
+}
+
+/// One extracted entity span.
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireEntity {
+    /// Model-native label string.
+    pub label: String,
+    /// Confidence in `[0, 1]`.
+    pub score: f32,
+    /// Byte offset, inclusive.
+    pub start: usize,
+    /// Byte offset, exclusive.
+    pub end: usize,
+}
+
+impl WireNerResponse {
+    /// Translate into the elide [`NerResponse`] the backend trait
+    /// expects. Drops malformed (`end <= start`) spans defensively
+    /// — the wire validator already rejects them, but the guard
+    /// keeps a misbehaving service from poisoning the recognizer.
+    pub(super) fn decode(self) -> NerResponse {
+        let spans = self
+            .entities
+            .into_iter()
+            .filter_map(|e| {
+                if e.end <= e.start {
+                    return None;
+                }
+                Some(NerSpan::new(e.label, e.score, e.start..e.end))
+            })
+            .collect();
+        NerResponse::new(spans)
+    }
+}
diff --git a/crates/elide-bento/src/ocr/mod.rs b/crates/elide-bento/src/ocr/mod.rs
new file mode 100644
index 000000000..641f37ef3
--- /dev/null
+++ b/crates/elide-bento/src/ocr/mod.rs
@@ -0,0 +1,119 @@
+//! [`BentoOcr`]: an [`elide_ocr::OcrBackend`] backed by the
+//! `nvisy-inference-ocr` BentoML service.
+//!
+//! Wire contract: `POST /recognize` accepts a batched list of
+//! requests (each carrying base64-encoded image bytes + a
+//! confidence threshold) and returns the matching list of
+//! responses. Each response is a `Page -> Block -> Line -> Word`
+//! tree; this backend flattens it to elide's
+//! [`LayoutBlock`]/[`LayoutWord`] vocabulary: one [`LayoutBlock`]
+//! per inference `Block`, every per-block word lifted into a
+//! [`LayoutWord`] regardless of its parent `Line`. Per-call
+//! correlation IDs propagate as `x-request-id` headers when set.
+//!
+//! Wire types live in the private `request` (outgoing) and
+//! `response` (incoming) submodules; only the public
+//! [`BentoOcr`] backend is part of this crate's API.
+//!
+//! [`LayoutBlock`]: elide_core::modality::image::LayoutBlock
+//! [`LayoutWord`]: elide_core::modality::image::LayoutWord
+
+mod request;
+mod response;
+
+use bentoml::{Client, Endpoint};
+use elide_core::Result;
+use elide_core::entity::provenance::ModelEvent;
+use elide_ocr::{OcrBackend, OcrRequest, OcrResponse};
+use hipstr::HipStr;
+
+use self::request::WireOcrRequest;
+use self::response::WireOcrResponse;
+use crate::error::BentoError;
+
+const ROUTE: &str = "recognize";
+
+/// BentoML OCR backend.
+///
+/// Owns a cached [`Endpoint`] pointing at the `nvisy-inference-ocr`
+/// `/recognize` route, plus the per-deployment model id (echoed
+/// into [`OcrBackend::provenance`]) and a default per-word
+/// confidence threshold (the service drops anything weaker before
+/// returning).
+#[derive(Debug, Clone)]
+pub struct BentoOcr {
+    /// Pre-built endpoint at the `/recognize` route. Cloned per
+    /// call so per-request headers (`x-request-id`) layer onto a
+    /// fresh instance without rebuilding the route.
+    endpoint: Endpoint,
+    /// Service-side model identifier echoed in provenance.
+    model_id: HipStr<'static>,
+    /// Default confidence floor sent on every request; the service
+    /// drops weaker per-word recognitions before responding.
+    default_threshold: f32,
+}
+
+impl BentoOcr {
+    /// Build from a service URL + the deployment's model id.
+    /// Default per-word confidence threshold is `0.0` (no
+    /// filtering, matches the service's own default); use
+    /// [`with_default_threshold`] to override.
+    ///
+    /// [`with_default_threshold`]: Self::with_default_threshold
+    pub fn new(base_url: impl Into<String>, model_id: impl Into<HipStr<'static>>) -> Result<Self> {
+        let client = Client::builder()
+            .with_base_url(base_url)
+            .build()
+            .map_err(BentoError::Transport)?;
+        Ok(Self {
+            endpoint: client.endpoint(ROUTE),
+            model_id: model_id.into(),
+            default_threshold: 0.0,
+        })
+    }
+
+    /// Override the per-request default per-word confidence
+    /// threshold.
+    #[must_use]
+    pub fn with_default_threshold(mut self, threshold: f32) -> Self {
+        self.default_threshold = threshold;
+        self
+    }
+}
+
+#[async_trait::async_trait]
+impl OcrBackend for BentoOcr {
+    fn provenance(&self) -> ModelEvent {
+        ModelEvent {
+            name: self.model_id.clone(),
+            version: None,
+            contextual: false,
+        }
+    }
+
+    async fn recognize(&self, request: OcrRequest<'_>) -> Result<OcrResponse> {
+        let body = vec![WireOcrRequest::from_request(
+            &request,
+            self.default_threshold,
+        )];
+        let mut endpoint = self.endpoint.clone();
+        if let Some(id) = request.correlation_id {
+            endpoint = endpoint.with_request_id(id.to_string());
+        }
+        let responses: Vec<WireOcrResponse> = endpoint
+            .invoke(&body)
+            .await
+            .map_err(BentoError::Transport)?;
+        let mut iter = responses.into_iter();
+        let response = iter
+            .next()
+            .ok_or_else(|| BentoError::Protocol("bento ocr returned an empty batch".into()))?;
+        if iter.next().is_some() {
+            return Err(BentoError::Protocol(
+                "bento ocr returned more responses than requests".into(),
+            )
+            .into());
+        }
+        Ok(response.decode())
+    }
+}
diff --git a/crates/elide-bento/src/ocr/request.rs b/crates/elide-bento/src/ocr/request.rs
new file mode 100644
index 000000000..a15725762
--- /dev/null
+++ b/crates/elide-bento/src/ocr/request.rs
@@ -0,0 +1,29 @@
+//! Outgoing wire types for the OCR `/recognize` endpoint.
+//!
+//! Mirrors `nvisy_core.ocr.v1.OcrRequest` from the inference
+//! repository: base64-encoded image bytes plus a per-word
+//! confidence floor the service applies before returning.
+
+use base64::Engine;
+use base64::engine::general_purpose::STANDARD as BASE64;
+use elide_ocr::OcrRequest;
+use serde::Serialize;
+
+/// Outgoing per-call request body element.
+#[derive(Debug, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireOcrRequest {
+    /// Base64-encoded image bytes.
+    pub image: String,
+    /// Drop per-word recognitions weaker than this; `0.0` keeps all.
+    pub confidence_threshold: f32,
+}
+
+impl WireOcrRequest {
+    pub(super) fn from_request(request: &OcrRequest<'_>, default_threshold: f32) -> Self {
+        Self {
+            image: BASE64.encode(request.image),
+            confidence_threshold: default_threshold,
+        }
+    }
+}
diff --git a/crates/elide-bento/src/ocr/response.rs b/crates/elide-bento/src/ocr/response.rs
new file mode 100644
index 000000000..110fa6174
--- /dev/null
+++ b/crates/elide-bento/src/ocr/response.rs
@@ -0,0 +1,138 @@
+//! Incoming wire types for the OCR `/recognize` endpoint.
+//!
+//! Mirrors `nvisy_core.ocr.v1.OcrResponse` from the inference
+//! repository. The full upstream tree is
+//! `Page -> Block -> Line -> Word`; elide's vocabulary collapses
+//! lines into the parent block — [`WireOcrResponse::decode`]
+//! flattens every word under its grandparent block. The
+//! response-level `modelId`,
+//! per-page `width`/`height`, per-block `kind`, and any rotated
+//! polygons are deserialised-and-discarded for now.
+
+use elide_core::modality::image::{ImageLocation, LayoutBlock, LayoutWord};
+use elide_core::primitive::{BoundingBox, Confidence, Point};
+use elide_ocr::OcrResponse;
+use serde::Deserialize;
+
+/// Incoming per-call response body element.
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireOcrResponse {
+    #[serde(default)]
+    pub pages: Vec<WirePage>,
+    // `modelId` ignored: provenance comes from `BentoOcr::model_id`
+    // (the deployment-level id the operator wired at construction).
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WirePage {
+    /// 1-based page index; flows straight onto [`ImageLocation::page`].
+    pub page_number: Option<u32>,
+    #[serde(default)]
+    pub blocks: Vec<WireBlock>,
+    // `width`, `height` ignored: elide's layout does not carry page
+    // dimensions today.
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireBlock {
+    pub text: String,
+    pub bbox: WireBoundingBox,
+    #[serde(default)]
+    pub lines: Vec<WireLine>,
+    // `kind` (text / table / figure / other) ignored: elide's
+    // `LayoutBlock` does not yet model block kind. When upstream
+    // grows it, surface here.
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireLine {
+    #[serde(default)]
+    pub words: Vec<WireWord>,
+    // Per-line text + bbox are subsumed by the block's text + the
+    // per-word geometry; elide does not model the intermediate line
+    // layer today.
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireWord {
+    pub text: String,
+    pub confidence: Option<f32>,
+    pub bbox: WireBoundingBox,
+    // `polygon` ignored: rotated regions land on a future
+    // `LayoutWord::polygon` field once elide grows one.
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(super) struct WireBoundingBox {
+    pub x: f64,
+    pub y: f64,
+    pub width: f64,
+    pub height: f64,
+}
+
+impl From<WireBoundingBox> for BoundingBox {
+    fn from(b: WireBoundingBox) -> Self {
+        BoundingBox::new(
+            Point::new(b.x, b.y),
+            Point::new(b.x + b.width, b.y + b.height),
+        )
+    }
+}
+
+impl WireOcrResponse {
+    /// Translate into the elide [`OcrResponse`] the backend trait
+    /// expects. Flattens pages → blocks → words; every per-block
+    /// word becomes a [`LayoutWord`] on the resulting
+    /// [`LayoutBlock`].
+    pub(super) fn decode(self) -> OcrResponse {
+        let blocks = self
+            .pages
+            .into_iter()
+            .flat_map(|page| {
+                let page_number = page.page_number;
+                page.blocks
+                    .into_iter()
+                    .map(move |block| block.decode(page_number))
+            })
+            .collect();
+        OcrResponse::new(blocks)
+    }
+}
+
+impl WireBlock {
+    fn decode(self, page_number: Option<u32>) -> LayoutBlock {
+        let region = ImageLocation {
+            bounding_box: self.bbox.into(),
+            polygon: None,
+            page: page_number,
+        };
+        let words: Vec<LayoutWord> = self
+            .lines
+            .into_iter()
+            .flat_map(|line| line.words.into_iter())
+            .map(|word| word.decode(page_number))
+            .collect();
+        LayoutBlock::new(region, self.text).with_words(words)
+    }
+}
+
+impl WireWord {
+    fn decode(self, page_number: Option<u32>) -> LayoutWord {
+        let region = ImageLocation {
+            bounding_box: self.bbox.into(),
+            polygon: None,
+            page: page_number,
+        };
+        let mut layout = LayoutWord::new(region, self.text);
+        if let Some(c) = self.confidence {
+            layout = layout.with_confidence(Confidence::clamped(c));
+        }
+        layout
+    }
+}
diff --git a/crates/nvisy-cli/Cargo.toml b/crates/nvisy-cli/Cargo.toml
index c8fa9aec5..1bb5bd235 100644
--- a/crates/nvisy-cli/Cargo.toml
+++ b/crates/nvisy-cli/Cargo.toml
@@ -19,11 +19,7 @@ homepage = { workspace = true }
 documentation = { workspace = true }
 
 [features]
-default = [
-    "tabular", "image", "audio", "rich",
-    "openai", "anthropic", "google",
-    "bento",
-]
+default = ["tabular", "image", "audio"]
 
 ## Tabular modality — csv, xlsx.
 tabular = ["nvisy-server/tabular"]
@@ -31,19 +27,6 @@ tabular = ["nvisy-server/tabular"]
 image = ["nvisy-server/image"]
 ## Audio modality — wav, mp3.
 audio = ["nvisy-server/audio"]
-## Rich-document modality — pdf, docx.
-rich = ["nvisy-server/rich"]
-
-## Enable all OpenAI providers (GPT, Whisper STT).
-openai = ["nvisy-server/openai"]
-## Enable Anthropic Claude completion provider.
-anthropic = ["nvisy-server/anthropic"]
-## Enable Google Gemini.
-google = ["nvisy-server/google"]
-
-## Enable the externalized BentoML backends (NER + OCR) — forwards
-## to `nvisy-server/bento`.
-bento = ["nvisy-server/bento"]
 
 [package.metadata.docs.rs]
 all-features = true
@@ -55,7 +38,7 @@ path = "src/main.rs"
 
 [dependencies]
 # Internal crates
-nvisy-engine = { workspace = true, features = [] }
+nvisy-core = { workspace = true, features = [] }
 nvisy-server = { workspace = true, features = [] }
 
 # Serialization
diff --git a/crates/nvisy-cli/src/config/mod.rs b/crates/nvisy-cli/src/config/mod.rs
index c2fa93b92..dacf13d7e 100644
--- a/crates/nvisy-cli/src/config/mod.rs
+++ b/crates/nvisy-cli/src/config/mod.rs
@@ -1,8 +1,8 @@
 //! CLI configuration management.
 //!
-//! TOML file is the source of truth; CLI flags override a small set
-//! of network / lifecycle fields. The resolved [`AppConfig`] is what
-//! the rest of the binary consumes.
+//! TOML file is the source of truth; CLI flags override a small
+//! set of network / lifecycle fields. The resolved [`AppConfig`]
+//! is what the rest of the binary consumes.
 //!
 //! # Architecture
 //!
@@ -11,12 +11,20 @@
 //! └── Overrides     thin CLI overlay applied to ServerConfig
 //!
 //! AppConfig (TOML)
-//! ├── server: ServerConfig            [server] + nested .observability / .middleware
-//! └── runtime: RuntimeConfig          [engine], [extraction.*], [detection.*], [redaction]
+//! └── server: ServerConfig            [server] + nested .observability / .middleware
 //! ```
 //!
-//! [`Cli::load`] reads the file, merges CLI overrides, and returns
-//! the resolved [`AppConfig`].
+//! There is no engine config section. Recognizer / analyzer /
+//! anonymizer settings travel per-request inside
+//! [`AnalyzerParams`] on `POST /detections`; the server just
+//! orchestrates and persists. The only server-side
+//! engine-shaped setting is the data directory, which lives on
+//! [`ServerConfig`].
+//!
+//! [`Cli::load`] reads the file, merges CLI overrides, and
+//! returns the resolved [`AppConfig`].
+//!
+//! [`AnalyzerParams`]: nvisy_core::plan::AnalyzerParams
 
 pub mod middleware;
 pub mod observability;
@@ -29,7 +37,7 @@ use std::time::Duration;
 
 use anyhow::Context;
 use clap::{Args, Parser};
-use nvisy_engine::core::RuntimeConfig;
+use nvisy_core::plan::AnalyzerParams;
 use serde::Deserialize;
 
 pub use self::server::ServerConfig;
@@ -91,39 +99,32 @@ impl Overrides {
     }
 }
 
-/// Resolved top-level configuration: server settings + engine
-/// subsystem settings, all merged from TOML + CLI overrides.
-///
-/// `deny_unknown_fields` catches typos at the top level (e.g.
-/// `[serer]` instead of `[server]`). Each nested struct denies
-/// unknown fields too so typos inside a section also fail loudly.
+/// Resolved top-level configuration: server settings + the
+/// deployment's default [`AnalyzerParams`], merged from TOML +
+/// CLI overrides.
 #[derive(Debug, Clone, Default, Deserialize)]
 pub struct AppConfig {
     /// Server, observability, and middleware configuration.
     #[serde(default)]
     pub server: ServerConfig,
-    /// Engine and provider subsystem settings. Flattened so its
-    /// sections (`[engine]`, `[extraction.*]`, ...) sit at the
-    /// TOML root alongside `[server]`. The top-level `version`
-    /// key the TOML carries is consumed here by `RuntimeConfig`.
-    #[serde(flatten)]
-    pub runtime: RuntimeConfig,
+    /// Default analyzer spec the server applies to every
+    /// request that doesn't override its analyzer fields.
+    /// Empty when the `[analyzer]` section is absent —
+    /// requests then have to ship a complete spec or get
+    /// "nothing detects" semantics.
+    #[serde(default)]
+    pub analyzer: AnalyzerParams,
 }
 
 impl Cli {
-    /// Read the TOML file, apply CLI overrides, run runtime
-    /// validation, and return the resolved [`AppConfig`].
+    /// Read the TOML file, apply CLI overrides, return the
+    /// resolved [`AppConfig`].
     ///
-    /// Missing TOML file resolves to defaults (everything from CLI
-    /// + built-ins).
+    /// Missing TOML file resolves to defaults (everything from
+    /// CLI + built-ins).
     pub fn load(self) -> anyhow::Result<AppConfig> {
         let mut config = read_toml(&self.config)?;
         self.overrides.merge_into(&mut config.server);
-        config.runtime.resolve_env();
-        config
-            .runtime
-            .validate()
-            .map_err(|e| anyhow::anyhow!("invalid configuration: {}", e.message()))?;
         Ok(config)
     }
 }
@@ -145,8 +146,8 @@ mod tests {
 
     use super::*;
 
-    /// `Nvisy.example.toml` is the source of truth for the documented
-    /// schema. If it stops parsing, the docs lie.
+    /// `Nvisy.example.toml` is the source of truth for the
+    /// documented schema. If it stops parsing, the docs lie.
     #[test]
     fn example_toml_parses() {
         let path = PathBuf::from(env!("CARGO_MANIFEST_DIR"))
@@ -154,20 +155,7 @@ mod tests {
             .join("Nvisy.example.toml");
         let contents = fs::read_to_string(&path).expect("Nvisy.example.toml exists");
         let config: AppConfig = toml::from_str(&contents).expect("Nvisy.example.toml parses");
-
-        assert!(config.runtime.engine.is_some(), "[engine] should be set");
-        assert!(
-            config.runtime.extraction.is_some(),
-            "[extraction.*] should be set"
-        );
-        assert!(
-            config.runtime.detection.is_some(),
-            "[detection.*] should be set"
-        );
-        assert!(
-            config.runtime.redaction.is_some(),
-            "[redaction] should be set"
-        );
+        assert_eq!(config.server.port, 8080);
     }
 
     #[test]
diff --git a/crates/nvisy-cli/src/main.rs b/crates/nvisy-cli/src/main.rs
index 606b53c4c..2d914b545 100644
--- a/crates/nvisy-cli/src/main.rs
+++ b/crates/nvisy-cli/src/main.rs
@@ -39,13 +39,10 @@ async fn run() -> anyhow::Result<()> {
     tracing::info!(
         target: TARGET,
         binary = env!("CARGO_PKG_VERSION"),
-        config = %config.runtime.version,
         "starting nvisy",
     );
-    let AppConfig {
-        server, runtime, ..
-    } = config;
-    let state = ServiceState::new(runtime, server.data_dir.clone()).await?;
+    let AppConfig { server, analyzer } = config;
+    let state = ServiceState::new(server.data_dir.clone(), analyzer).await?;
     let router = create_router(&server, state);
     server::run(&server, router).await
 }
diff --git a/crates/nvisy-codec/Cargo.toml b/crates/nvisy-codec/Cargo.toml
deleted file mode 100644
index ff9752965..000000000
--- a/crates/nvisy-codec/Cargo.toml
+++ /dev/null
@@ -1,148 +0,0 @@
-# https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[package]
-name = "nvisy-codec"
-description = "Codec traits + format implementations (TXT, JSON, HTML, CSV, XLSX, PNG, JPEG, TIFF, WAV, MP3, PDF, DOCX) for the Nvisy multimodal redaction platform"
-keywords = ["nvisy", "codec", "loader", "pdf", "redaction"]
-categories = ["parser-implementations", "encoding"]
-readme = "README.md"
-
-version = { workspace = true }
-rust-version = { workspace = true }
-edition = { workspace = true }
-license = { workspace = true }
-publish = { workspace = true }
-
-authors = { workspace = true }
-repository = { workspace = true }
-homepage = { workspace = true }
-documentation = { workspace = true }
-
-[features]
-default = ["text", "tabular"]
-
-## Plain-text (`.txt`, `.log`) loader and handler.
-txt = ["internal_text"]
-## JSON loader and handler.
-json = ["internal_text"]
-## Markdown loader.
-markdown = ["internal_text"]
-## HTML loader and handler via `scraper`.
-html = ["internal_text", "dep:scraper", "dep:ego-tree"]
-
-## CSV loader and handler via the `csv` crate. Tabular cells reuse the
-## text redact helper, so this pulls `internal_text`.
-csv = ["internal_tabular", "internal_text", "dep:csv"]
-## Excel (`.xlsx`) loader and handler. Pulls `internal_text` for the
-## same reason as `csv`. The handler is a stub today (decode returns
-## empty); a parser dep will be wired back in when real extraction
-## lands.
-xlsx = ["internal_tabular", "internal_text"]
-
-## PNG loader and handler.
-png = ["internal_image"]
-## JPEG loader and handler.
-jpeg = ["internal_image"]
-## TIFF loader and handler.
-tiff = ["internal_image"]
-
-## WAV audio loader and handler. `hound` covers sample read + write;
-## `symphonia` probes container metadata for the clip duration.
-wav = ["internal_audio", "dep:hound", "dep:symphonia"]
-## MP3 audio loader and handler. `symphonia` covers the duration
-## probe and PCM decode; `mp3lame-encoder` (LGPL-3.0 via libmp3lame —
-## requires a C toolchain + autoconf/automake at build time) handles
-## re-encoding for the redaction round-trip.
-mp3 = ["internal_audio", "dep:symphonia", "dep:mp3lame-encoder"]
-
-## PDF loader, handler, and page-to-image rendering via `lopdf` +
-## `pdfium-render`. Pulls `internal_text` (page-text redact) +
-## `internal_image` (PngHandler for rendered pages + extracted
-## figures) + `png` (the concrete image format wrapping rendered
-## pages).
-pdf = ["internal_rich", "internal_text", "internal_image", "png", "dep:lopdf", "dep:pdfium-render", "dep:rayon"]
-## DOCX loader. Pulls `internal_text` only — DOCX text extraction is
-## stubbed today, no image rendering needed. Parser deps (`zip` +
-## `quick-xml`) will be wired back in when real extraction lands.
-docx = ["internal_rich", "internal_text"]
-
-## All text formats: `txt` + `json` + `markdown` + `html`.
-text = ["txt", "json", "markdown", "html"]
-## All tabular formats: `csv` + `xlsx`.
-tabular = ["csv", "xlsx"]
-## All image formats: `png` + `jpeg` + `tiff`.
-image = ["png", "jpeg", "tiff"]
-## All audio formats: `wav` + `mp3`.
-audio = ["wav", "mp3"]
-## All rich-document formats: `pdf` + `docx`.
-rich = ["pdf", "docx"]
-
-## In-memory decode helpers for tests in this and downstream crates.
-## Re-exported as `nvisy_codec::test_utils`.
-test-utils = []
-
-# Internal helpers — set automatically by any format feature above.
-# Library code uses `#[cfg(feature = "internal_text")]` to mean
-# "any text format is enabled", the right gate for shared
-# infrastructure like the modality's `impl Codable` block and
-# top-level `crate::text` module.
-internal_text = []
-internal_tabular = []
-internal_image = []
-internal_audio = []
-internal_rich = []
-
-[package.metadata.docs.rs]
-all-features = true
-rustdoc-args = ["--cfg", "docsrs"]
-
-[dependencies]
-# Internal crates
-nvisy-core = { workspace = true, features = [] }
-
-# Serialization
-serde = { workspace = true, features = [] }
-serde_json = { workspace = true, features = [] }
-schemars = { workspace = true, features = [] }
-
-# Derive macros and error handling
-derive_more = { workspace = true, features = ["as_ref", "deref", "from"] }
-
-# Primitive datatypes
-bytes = { workspace = true, features = [] }
-
-# Encoding and hashing
-hex = { workspace = true, features = [] }
-sha2 = { workspace = true, features = [] }
-
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-tokio = { workspace = true, features = ["sync"] }
-rayon = { workspace = true, optional = true, features = [] }
-
-# Observability
-tracing = { workspace = true, features = [] }
-
-# Tabular document parsing (feature-gated)
-csv = { workspace = true, optional = true, features = [] }
-
-# Rich-document parsing (feature-gated: HTML + PDF)
-scraper = { workspace = true, optional = true, features = [] }
-ego-tree = { workspace = true, optional = true, features = [] }
-lopdf = { workspace = true, optional = true, features = [] }
-pdfium-render = { workspace = true, optional = true, features = [] }
-
-# Image processing
-image = { workspace = true, features = [] }
-imageproc = { workspace = true, features = [] }
-
-# Audio processing (feature-gated)
-hound = { workspace = true, optional = true, features = [] }
-symphonia = { workspace = true, optional = true, features = [] }
-mp3lame-encoder = { workspace = true, optional = true, features = [] }
-
-# Storage and file-type detection
-infer = { workspace = true, features = [] }
-
-[dev-dependencies]
-tokio = { workspace = true, features = ["macros", "rt"] }
diff --git a/crates/nvisy-codec/README.md b/crates/nvisy-codec/README.md
deleted file mode 100644
index 693709c18..000000000
--- a/crates/nvisy-codec/README.md
+++ /dev/null
@@ -1,48 +0,0 @@
-# nvisy-codec
-
-[![Build](https://img.shields.io/github/actions/workflow/status/nvisycom/runtime/build.yml?branch=main&label=build%20%26%20test&style=flat-square)](https://github.com/nvisycom/runtime/actions/workflows/build.yml)
-
-Format handlers, the `Handler<M>` trait, and the `CodecRegistry`
-that drives ingest in the Nvisy runtime.
-
-## Overview
-
-Built-in handlers cover TXT, JSON, Markdown, HTML, CSV, XLSX, PNG,
-JPEG, TIFF, WAV, MP3, PDF, and DOCX. Each implements `Handler<M>`
-(streaming `next_chunk`, random-access `read` / `redact`,
-`lift_chunk` for offset translation) and pairs with a `Loader<M>`
-that decodes raw bytes into the handler. A `Format` descriptor
-built via `Format::new::<M, _>(id, loader)` plus chained
-`.with_extensions(...)` / `.with_content_types(...)` registers the
-pair into `CodecRegistry`.
-
-Consumers resolve a `Format` by extension, content-type, or id and
-get back an `UntypedDocumentHandle` they commit to a modality via
-`into_text` / `into_tabular` / `into_image` / `into_audio`. The
-typed `DocumentHandle<M>` implements `nvisy-core`'s `TextAt` /
-`DataAt` / `RedactAt` directly, so pipeline components read from
-and write to codec-backed sources through the same traits the
-engine bounds on. The `content` module (`Content`, `ContentData`,
-`ContentDescriptor`, `ContentDigest`, `ContentRecord`,
-`ContentSource`, `TextEncoding`) carries the raw-bytes side of the
-import surface. Each format is feature-gated (`txt`, `csv`,
-`png`, …) with umbrella features `text`, `tabular`, `image`,
-`audio`, `rich`. Depends only on `nvisy-core`.
-
-## Documentation
-
-See [`docs/`](../../docs/) for architecture, security, and API documentation.
-
-## Changelog
-
-See [CHANGELOG.md](../../CHANGELOG.md) for release notes and version history.
-
-## License
-
-Apache 2.0 License, see [LICENSE.txt](../../LICENSE.txt)
-
-## Support
-
-- **Documentation**: [docs.nvisy.com](https://docs.nvisy.com)
-- **Issues**: [GitHub Issues](https://github.com/nvisycom/runtime/issues)
-- **Email**: [support@nvisy.com](mailto:support@nvisy.com)
diff --git a/crates/nvisy-codec/src/content/bundle.rs b/crates/nvisy-codec/src/content/bundle.rs
deleted file mode 100644
index cc5246059..000000000
--- a/crates/nvisy-codec/src/content/bundle.rs
+++ /dev/null
@@ -1,114 +0,0 @@
-//! [`Content`]: data bytes optionally paired with a caller-supplied
-//! [`ContentDescriptor`].
-
-use std::path::Path;
-
-use derive_more::{AsRef, Deref};
-use nvisy_core::Result;
-use serde::{Deserialize, Serialize};
-
-use super::{ContentData, ContentDescriptor, ContentSource};
-
-/// Upload-shape carrier: raw bytes plus the caller's descriptive
-/// metadata.
-///
-/// [`ContentData`] holds the bytes and source identity.
-/// [`ContentDescriptor`] holds filename, MIME hint, and extras when
-/// the caller has them. The descriptor is optional because some
-/// import paths (raw byte uploads, generated content) have nothing
-/// to attach.
-///
-/// After `Registry::register_content` consumes a `Content`, the
-/// stored shape is a `ContentRecord` (descriptor + byte-derived
-/// digest), which is what registry reads return.
-#[derive(Debug, Clone, PartialEq)]
-#[derive(AsRef, Deref, Serialize, Deserialize)]
-pub struct Content {
-    /// Raw content bytes.
-    #[deref]
-    #[as_ref]
-    data: ContentData,
-    /// Caller-supplied descriptive metadata.
-    descriptor: Option<ContentDescriptor>,
-}
-
-impl From<ContentData> for Content {
-    fn from(data: ContentData) -> Self {
-        Self::new(data)
-    }
-}
-
-impl Content {
-    /// Create content from data without a descriptor.
-    pub fn new(data: ContentData) -> Self {
-        Self {
-            data,
-            descriptor: None,
-        }
-    }
-
-    /// Create content with a caller-supplied descriptor.
-    pub fn with_descriptor(data: ContentData, descriptor: ContentDescriptor) -> Self {
-        Self {
-            data,
-            descriptor: Some(descriptor),
-        }
-    }
-
-    /// Returns the raw content data.
-    pub fn data(&self) -> &ContentData {
-        &self.data
-    }
-
-    /// Returns the caller-supplied descriptor, if present.
-    pub fn descriptor(&self) -> Option<&ContentDescriptor> {
-        self.descriptor.as_ref()
-    }
-
-    /// Returns the content source identifier.
-    pub fn content_source(&self) -> ContentSource {
-        self.data.content_source
-    }
-
-    /// Returns the raw bytes.
-    pub fn as_bytes(&self) -> &[u8] {
-        self.data.as_bytes()
-    }
-
-    /// Returns `true` if the content appears to be text.
-    pub fn is_likely_text(&self) -> bool {
-        self.data.is_likely_text()
-    }
-
-    /// Try to get the content as a string slice.
-    ///
-    /// # Errors
-    ///
-    /// Returns an error if the content is not valid UTF-8.
-    pub fn as_str(&self) -> Result<&str> {
-        self.data.as_str()
-    }
-
-    /// Caller-supplied MIME type, if any. Detected MIME isn't
-    /// available pre-registration (the registry computes it).
-    pub fn content_type(&self) -> Option<&str> {
-        self.descriptor
-            .as_ref()
-            .and_then(|d| d.content_type.as_deref())
-    }
-
-    /// Original filename from the descriptor.
-    pub fn filename(&self) -> Option<&Path> {
-        self.descriptor.as_ref().and_then(|d| d.filename.as_deref())
-    }
-
-    /// File extension from the descriptor's source path.
-    pub fn file_extension(&self) -> Option<&str> {
-        self.descriptor.as_ref().and_then(|d| d.file_extension())
-    }
-
-    /// Consume and return both data and descriptor.
-    pub fn into_parts(self) -> (ContentData, Option<ContentDescriptor>) {
-        (self.data, self.descriptor)
-    }
-}
diff --git a/crates/nvisy-codec/src/content/content_data.rs b/crates/nvisy-codec/src/content/content_data.rs
deleted file mode 100644
index a31f5158e..000000000
--- a/crates/nvisy-codec/src/content/content_data.rs
+++ /dev/null
@@ -1,268 +0,0 @@
-//! Raw content bytes with source identity.
-//!
-//! [`ContentData`] is the pure data half of the content model. It holds
-//! the raw bytes and a [`ContentSource`] identifier. All descriptive
-//! attributes (MIME type, filename, arbitrary metadata) live on
-//! [`ContentMetadata`].
-//!
-//! [`ContentMetadata`]: super::ContentMetadata
-
-use std::{fmt, str};
-
-use bytes::Bytes;
-use nvisy_core::{Error, ErrorKind, Result};
-use serde::{Deserialize, Serialize};
-use sha2::{Digest, Sha256};
-
-use super::ContentSource;
-
-/// Raw content bytes with source identity.
-///
-/// This is the data-only half of the content model — it does not
-/// carry MIME type, filename, or other descriptive metadata.
-/// Pair with [`ContentDescriptor`] via [`Content`] for a complete
-/// representation.
-///
-/// [`ContentDescriptor`]: super::ContentDescriptor
-/// [`Content`]: super::Content
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
-pub struct ContentData {
-    /// Unique identifier for the content source.
-    pub content_source: ContentSource,
-    /// The actual content bytes.
-    data: Bytes,
-}
-
-impl ContentData {
-    /// Creates content data from raw bytes.
-    pub fn new(content_source: ContentSource, data: Bytes) -> Self {
-        Self {
-            content_source,
-            data,
-        }
-    }
-
-    /// Creates content data from a text string.
-    pub fn from_text(content_source: ContentSource, text: impl Into<String>) -> Self {
-        Self {
-            content_source,
-            data: Bytes::from(text.into().into_bytes()),
-        }
-    }
-
-    /// Returns the size of the content in bytes.
-    #[must_use]
-    pub fn size(&self) -> usize {
-        self.data.len()
-    }
-
-    /// Returns the content data as a byte slice.
-    #[must_use]
-    pub fn as_bytes(&self) -> &[u8] {
-        &self.data
-    }
-
-    /// Converts the content data to `Bytes`.
-    #[must_use]
-    pub fn to_bytes(&self) -> Bytes {
-        self.data.clone()
-    }
-
-    /// Consumes and converts into `Bytes`.
-    #[must_use]
-    pub fn into_bytes(self) -> Bytes {
-        self.data
-    }
-
-    /// Returns `true` if the content appears to be text.
-    ///
-    /// Checks that the content is valid UTF-8 and contains no control
-    /// characters other than common whitespace (tab, newline, carriage
-    /// return).
-    #[must_use]
-    pub fn is_likely_text(&self) -> bool {
-        let Ok(s) = str::from_utf8(&self.data) else {
-            return false;
-        };
-        s.chars()
-            .all(|c| !c.is_control() || matches!(c, '\t' | '\n' | '\r'))
-    }
-
-    /// Tries to convert the content data to a UTF-8 string slice.
-    ///
-    /// # Errors
-    ///
-    /// Returns an error if the content data contains invalid UTF-8 sequences.
-    pub fn as_str(&self) -> Result<&str> {
-        str::from_utf8(&self.data)
-            .map_err(|e| Error::new(ErrorKind::Serialization, format!("Invalid UTF-8: {e}")))
-    }
-
-    /// Computes and returns the SHA256 hash of the content.
-    #[must_use]
-    pub fn sha256(&self) -> Bytes {
-        let mut hasher = Sha256::new();
-        hasher.update(&self.data);
-        Bytes::from(hasher.finalize().to_vec())
-    }
-
-    /// Returns the SHA256 hash as a hex string.
-    #[must_use]
-    pub fn sha256_hex(&self) -> String {
-        hex::encode(self.sha256())
-    }
-
-    /// Verifies the content against a provided SHA256 hash.
-    ///
-    /// # Errors
-    ///
-    /// Returns an error if the computed hash does not match the expected hash.
-    pub fn verify_sha256(&self, expected_hash: impl AsRef<[u8]>) -> Result<()> {
-        let actual_hash = self.sha256();
-        let expected = expected_hash.as_ref();
-
-        if actual_hash.as_ref() == expected {
-            Ok(())
-        } else {
-            Err(Error::new(
-                ErrorKind::Validation,
-                format!(
-                    "Hash mismatch: expected {}, got {}",
-                    hex::encode(expected),
-                    hex::encode(actual_hash)
-                ),
-            ))
-        }
-    }
-
-    /// Returns a slice of the content data.
-    ///
-    /// # Errors
-    ///
-    /// Returns an error if the end index is beyond the content length
-    /// or if start is greater than end.
-    pub fn slice(&self, start: usize, end: usize) -> Result<Bytes> {
-        if end > self.data.len() {
-            return Err(Error::new(
-                ErrorKind::Validation,
-                format!(
-                    "Slice end {} exceeds content length {}",
-                    end,
-                    self.data.len()
-                ),
-            ));
-        }
-        if start > end {
-            return Err(Error::new(
-                ErrorKind::Validation,
-                format!("Slice start {start} is greater than end {end}"),
-            ));
-        }
-        Ok(Bytes::copy_from_slice(&self.data[start..end]))
-    }
-
-    /// Returns `true` if the content is empty.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.data.is_empty()
-    }
-
-    /// Detect MIME type from the raw bytes using magic-byte signatures.
-    ///
-    /// Returns `None` for content with no recognizable magic bytes
-    /// (e.g. plain text). Pairing the result with
-    /// `ContentMetadata::detected_content_type` is a caller-side
-    /// convention — no type-level link exists between the two.
-    #[must_use]
-    pub fn detect_mime(&self) -> Option<String> {
-        infer::get(&self.data).map(|t| t.mime_type().to_owned())
-    }
-}
-
-impl From<&str> for ContentData {
-    fn from(s: &str) -> Self {
-        Self::from_text(ContentSource::new(), s)
-    }
-}
-
-impl From<String> for ContentData {
-    fn from(s: String) -> Self {
-        Self::from_text(ContentSource::new(), s)
-    }
-}
-
-impl From<&[u8]> for ContentData {
-    fn from(bytes: &[u8]) -> Self {
-        Self::new(ContentSource::new(), Bytes::copy_from_slice(bytes))
-    }
-}
-
-impl From<Vec<u8>> for ContentData {
-    fn from(vec: Vec<u8>) -> Self {
-        Self::new(ContentSource::new(), Bytes::from(vec))
-    }
-}
-
-impl From<Bytes> for ContentData {
-    fn from(bytes: Bytes) -> Self {
-        Self::new(ContentSource::new(), bytes)
-    }
-}
-
-impl fmt::Display for ContentData {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        if let Ok(text) = self.as_str() {
-            write!(f, "{text}")
-        } else {
-            write!(f, "[Binary data: {} bytes]", self.size())
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn sha256_computation() {
-        let content = ContentData::from("Hello, world!");
-        let hash = content.sha256();
-        assert_eq!(hash.len(), 32);
-        assert_eq!(hash, content.sha256());
-    }
-
-    #[test]
-    fn sha256_verification() {
-        let content = ContentData::from("Hello, world!");
-        let hash = content.sha256().clone();
-        assert!(content.verify_sha256(&hash).is_ok());
-        assert!(content.verify_sha256([0u8; 32]).is_err());
-    }
-
-    #[test]
-    fn slice_operations() {
-        let content = ContentData::from("Hello, world!");
-        assert_eq!(content.slice(0, 5).unwrap(), Bytes::from("Hello"));
-        assert_eq!(content.slice(7, 12).unwrap(), Bytes::from("world"));
-        assert!(content.slice(0, 100).is_err());
-        assert!(content.slice(10, 5).is_err());
-    }
-
-    #[test]
-    fn detect_mime_png() {
-        let png = vec![
-            0x89, 0x50, 0x4E, 0x47, 0x0D, 0x0A, 0x1A, 0x0A, 0x00, 0x00, 0x00, 0x0D, 0x49, 0x48,
-            0x44, 0x52,
-        ];
-        let data = ContentData::from(png);
-        assert_eq!(data.detect_mime().as_deref(), Some("image/png"));
-    }
-
-    #[test]
-    fn is_likely_text() {
-        assert!(ContentData::from("ascii text").is_likely_text());
-        assert!(ContentData::from("").is_likely_text());
-        assert!(ContentData::from("café").is_likely_text());
-        assert!(!ContentData::from(vec![0x00]).is_likely_text());
-    }
-}
diff --git a/crates/nvisy-codec/src/content/content_metadata.rs b/crates/nvisy-codec/src/content/content_metadata.rs
deleted file mode 100644
index ca586515c..000000000
--- a/crates/nvisy-codec/src/content/content_metadata.rs
+++ /dev/null
@@ -1,176 +0,0 @@
-//! Two-layer content metadata:
-//!
-//! - [`ContentDescriptor`] holds caller-supplied descriptive bits
-//!   (filename, MIME hint, source path, arbitrary extras). All
-//!   optional — the caller might or might not have any of them.
-//! - [`ContentDigest`] holds facts the registry computes by looking
-//!   at the bytes (size, sha256, sniffed MIME). Required fields are
-//!   actually required.
-//! - [`ContentRecord`] bundles a descriptor with a digest. This is
-//!   what the registry persists and what read sites get back.
-
-use std::path::{Path, PathBuf};
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Caller-supplied descriptive metadata for an upload.
-///
-/// Built before the bytes have been written to the registry, so
-/// every field is optional — the caller knows whatever they know.
-/// The registry's `register_content` consumes this alongside the
-/// bytes to produce a [`ContentRecord`].
-#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct ContentDescriptor {
-    /// Optional path to the source file.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub source_path: Option<PathBuf>,
-    /// MIME type supplied by the caller (e.g. `"text/plain"` from
-    /// an HTTP `Content-Type` header or an explicit API call).
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub content_type: Option<String>,
-    /// Original filename, if known. Used by `CodecRegistry` for
-    /// extension-based format resolution.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub filename: Option<PathBuf>,
-    /// Caller-supplied key-value pairs that policy conditions
-    /// (`Condition::Metadata { key, value }`) match against.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub policy_metadata: Option<serde_json::Map<String, serde_json::Value>>,
-}
-
-impl ContentDescriptor {
-    /// Create an empty descriptor.
-    #[must_use]
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Create a descriptor with a source file path.
-    pub fn with_path(path: impl Into<PathBuf>) -> Self {
-        Self {
-            source_path: Some(path.into()),
-            ..Self::default()
-        }
-    }
-
-    /// Set the caller-supplied MIME type (builder pattern).
-    #[must_use]
-    pub fn with_content_type(mut self, mime: impl Into<String>) -> Self {
-        self.content_type = Some(mime.into());
-        self
-    }
-
-    /// Set the original filename (builder pattern).
-    #[must_use]
-    pub fn with_filename(mut self, name: impl Into<PathBuf>) -> Self {
-        self.filename = Some(name.into());
-        self
-    }
-
-    /// Get the file extension from the source path, if available.
-    #[must_use]
-    pub fn file_extension(&self) -> Option<&str> {
-        self.source_path
-            .as_ref()
-            .and_then(|path| path.extension())
-            .and_then(|ext| ext.to_str())
-    }
-
-    /// Get the full path if available.
-    #[must_use]
-    pub fn path(&self) -> Option<&Path> {
-        self.source_path.as_deref()
-    }
-
-    /// Get a single value from the policy metadata map.
-    #[must_use]
-    pub fn get_policy_metadata(&self, key: &str) -> Option<&serde_json::Value> {
-        self.policy_metadata.as_ref().and_then(|m| m.get(key))
-    }
-
-    /// Insert a key-value pair into the policy metadata map,
-    /// creating the map if it doesn't exist yet.
-    pub fn set_policy_metadata(&mut self, key: impl Into<String>, value: serde_json::Value) {
-        self.policy_metadata
-            .get_or_insert_with(serde_json::Map::new)
-            .insert(key.into(), value);
-    }
-
-    /// Remove a key from the policy metadata map. Returns the
-    /// removed value if the key existed.
-    pub fn remove_policy_metadata(&mut self, key: &str) -> Option<serde_json::Value> {
-        self.policy_metadata.as_mut().and_then(|m| m.remove(key))
-    }
-}
-
-/// Byte-derived facts about a piece of content.
-///
-/// Computed by `Registry::register_content` after the bytes are in
-/// hand. Required fields (`size`, `sha256`) are unconditional;
-/// `detected_content_type` is `Option` because magic-byte sniffing
-/// may legitimately fail (e.g. plain text).
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct ContentDigest {
-    /// Size in bytes.
-    pub size: u64,
-    /// SHA-256 hex digest of the raw bytes.
-    pub sha256: String,
-    /// MIME type sniffed from the bytes, if magic-byte detection
-    /// produced a result.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub detected_content_type: Option<String>,
-}
-
-/// Persisted, post-registration view of a piece of content.
-///
-/// Returned by registry read endpoints (`list_content_with_record`,
-/// `read_content`). The [`ContentDescriptor`] half is whatever the
-/// caller supplied at upload; the [`ContentDigest`] half is what
-/// the registry computed.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct ContentRecord {
-    /// Caller-supplied descriptor (filename, MIME hint, policy metadata).
-    pub descriptor: ContentDescriptor,
-    /// Registry-computed digest (size, sha256, detected MIME).
-    pub digest: ContentDigest,
-}
-
-impl ContentRecord {
-    /// Best-available MIME type: caller-supplied takes priority
-    /// over sniffed.
-    #[must_use]
-    pub fn content_type(&self) -> Option<&str> {
-        self.descriptor
-            .content_type
-            .as_deref()
-            .or(self.digest.detected_content_type.as_deref())
-    }
-
-    /// Original filename from the descriptor.
-    #[must_use]
-    pub fn filename(&self) -> Option<&Path> {
-        self.descriptor.filename.as_deref()
-    }
-
-    /// Original filename rendered as a UTF-8 string. Non-UTF-8 byte
-    /// sequences in the path are replaced with U+FFFD (lossy
-    /// conversion). Use [`filename`] when you need the raw `&Path`.
-    ///
-    /// [`filename`]: Self::filename
-    #[must_use]
-    pub fn filename_lossy(&self) -> Option<String> {
-        self.filename().map(|p| p.to_string_lossy().into_owned())
-    }
-
-    /// File extension from the descriptor's source path.
-    #[must_use]
-    pub fn file_extension(&self) -> Option<&str> {
-        self.descriptor.file_extension()
-    }
-}
diff --git a/crates/nvisy-codec/src/content/encoding.rs b/crates/nvisy-codec/src/content/encoding.rs
deleted file mode 100644
index 27f8d8d02..000000000
--- a/crates/nvisy-codec/src/content/encoding.rs
+++ /dev/null
@@ -1,23 +0,0 @@
-//! Character encoding for text-based loaders.
-
-use nvisy_core::{Error, Result};
-/// Character encoding used to decode raw bytes before parsing.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
-pub enum TextEncoding {
-    /// UTF-8 (the default and by far the most common encoding).
-    #[default]
-    Utf8,
-}
-
-impl TextEncoding {
-    /// Decode raw bytes to a UTF-8 string.
-    ///
-    /// `origin` identifies the caller for error messages
-    /// (e.g. `"json-loader"`).
-    pub fn decode_bytes(self, bytes: &[u8], origin: &str) -> Result<String> {
-        match self {
-            Self::Utf8 => String::from_utf8(bytes.to_vec())
-                .map_err(|e| Error::validation(format!("Invalid UTF-8: {e}"), origin.to_owned())),
-        }
-    }
-}
diff --git a/crates/nvisy-codec/src/content/mod.rs b/crates/nvisy-codec/src/content/mod.rs
deleted file mode 100644
index a0e4e4f0f..000000000
--- a/crates/nvisy-codec/src/content/mod.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-//! Content data containers, metadata, and source identity.
-//!
-//! - [`ContentData`]: raw content bytes with source identity.
-//! - [`ContentDescriptor`]: caller-supplied descriptive bits
-//!   (filename, MIME hint, extras) — built before bytes are
-//!   persisted.
-//! - [`ContentDigest`]: byte-derived facts (size, sha256, sniffed
-//!   MIME) — computed at registration time.
-//! - [`ContentRecord`]: persisted view (descriptor + digest), what
-//!   registry reads return.
-//! - [`Content`]: [`ContentData`] paired with an optional
-//!   [`ContentDescriptor`] — the upload-shape carrier.
-//! - [`ContentSource`]: UUIDv7-based content identity and lineage.
-//!
-//! Top-level format classification lives on [`FormatId`].
-//!
-//! [`FormatId`]: crate::FormatId
-
-mod bundle;
-mod content_data;
-mod content_metadata;
-mod encoding;
-
-pub use nvisy_core::entity::ContentSource;
-
-pub use self::bundle::Content;
-pub use self::content_data::ContentData;
-pub use self::content_metadata::{ContentDescriptor, ContentDigest, ContentRecord};
-pub use self::encoding::TextEncoding;
diff --git a/crates/nvisy-codec/src/core/format.rs b/crates/nvisy-codec/src/core/format.rs
deleted file mode 100644
index fe4015e98..000000000
--- a/crates/nvisy-codec/src/core/format.rs
+++ /dev/null
@@ -1,179 +0,0 @@
-//! Format identity: what kind of thing a registered codec is.
-//!
-//! - [`FormatId`] — stable string identifier (e.g.
-//!   `"nvisy.text.txt"`). Open namespace, no central enum.
-//! - [`Format`] — descriptor [`CodecRegistry`] indexes by id /
-//!   extension / content type. Bundles a `FormatId`, its
-//!   [`ModalityKind`], lookup keys, and an erased loader that
-//!   decodes bytes into a typed handle.
-//!
-//! [`CodecRegistry`]: super::CodecRegistry
-//! [`ModalityKind`]: nvisy_core::modality::ModalityKind
-
-use std::borrow::Cow;
-use std::fmt;
-use std::sync::Arc;
-
-use nvisy_core::modality::{Modality, ModalityKind};
-
-use super::{ErasedLoader, Loader, erase};
-use crate::document::{DocumentHandle, UntypedDocumentHandle};
-
-/// Stable identifier for a registered codec format. Open string
-/// namespace — downstream crates ship their own formats by
-/// registering a [`Format`] with a unique [`FormatId`].
-///
-/// Convention: dot-separated namespace. Built-in formats use the
-/// `nvisy.` prefix (e.g. `"nvisy.text.txt"`, `"nvisy.rich.pdf"`).
-/// Third-party formats use their own (e.g. `"acme.parquet.v2"`).
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-pub struct FormatId(Cow<'static, str>);
-
-impl FormatId {
-    /// Construct from a static string literal — no allocation.
-    pub const fn from_static(id: &'static str) -> Self {
-        Self(Cow::Borrowed(id))
-    }
-
-    /// Construct from an owned [`String`].
-    pub fn from_owned(id: String) -> Self {
-        Self(Cow::Owned(id))
-    }
-
-    /// Borrow as `&str`.
-    pub fn as_str(&self) -> &str {
-        &self.0
-    }
-}
-
-impl fmt::Display for FormatId {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.write_str(&self.0)
-    }
-}
-
-impl AsRef<str> for FormatId {
-    fn as_ref(&self) -> &str {
-        &self.0
-    }
-}
-
-/// Descriptor for one registered codec format. Indexed by
-/// [`CodecRegistry`] under its [`FormatId`], every extension in
-/// `extensions`, and every MIME in `content_types`.
-///
-/// Construct via [`Format::new`]; read the parts via the accessor
-/// methods. The fields are crate-private so the constructor stays
-/// the only path that produces a `Format` — that way the
-/// [`ModalityKind`] tag is always derived from the loader's
-/// modality and never hand-set.
-///
-/// [`CodecRegistry`]: super::CodecRegistry
-#[derive(Clone)]
-pub struct Format {
-    pub(crate) id: FormatId,
-    pub(crate) modality: ModalityKind,
-    pub(crate) extensions: Vec<Cow<'static, str>>,
-    pub(crate) content_types: Vec<Cow<'static, str>>,
-    pub(crate) loader: Arc<dyn ErasedLoader>,
-}
-
-impl Format {
-    /// Build a [`Format`] for modality `M`. The runtime
-    /// [`ModalityKind`] tag is taken from `M::KIND` and the loader
-    /// is erased internally — neither needs to be named at the call
-    /// site.
-    ///
-    /// Extensions and content types default to empty; chain
-    /// [`with_extensions`] / [`with_content_types`] to declare the
-    /// lookup keys the [`CodecRegistry`] indexes this format under.
-    ///
-    /// [`with_extensions`]: Self::with_extensions
-    /// [`with_content_types`]: Self::with_content_types
-    /// [`CodecRegistry`]: super::CodecRegistry
-    pub fn new<M, L>(id: FormatId, loader: L) -> Self
-    where
-        M: Modality,
-        L: Loader<M>,
-        DocumentHandle<M>: Into<UntypedDocumentHandle>,
-    {
-        Self {
-            id,
-            modality: M::KIND,
-            extensions: Vec::new(),
-            content_types: Vec::new(),
-            loader: erase(loader),
-        }
-    }
-
-    /// Declare the file extensions (lowercased, no leading dot) that
-    /// resolve to this format. Extends any previously-declared list.
-    #[must_use]
-    pub fn with_extensions<I, S>(mut self, extensions: I) -> Self
-    where
-        I: IntoIterator<Item = S>,
-        S: Into<Cow<'static, str>>,
-    {
-        self.extensions
-            .extend(extensions.into_iter().map(Into::into));
-        self
-    }
-
-    /// Declare the MIME content types (lowercased) that resolve to
-    /// this format. Extends any previously-declared list.
-    #[must_use]
-    pub fn with_content_types<I, S>(mut self, content_types: I) -> Self
-    where
-        I: IntoIterator<Item = S>,
-        S: Into<Cow<'static, str>>,
-    {
-        self.content_types
-            .extend(content_types.into_iter().map(Into::into));
-        self
-    }
-
-    /// Stable identifier of this format.
-    pub fn id(&self) -> &FormatId {
-        &self.id
-    }
-
-    /// Modality this format produces.
-    pub fn modality(&self) -> ModalityKind {
-        self.modality
-    }
-
-    /// File extensions (lowercased, no leading dot) that resolve to
-    /// this format.
-    pub fn extensions(&self) -> &[Cow<'static, str>] {
-        &self.extensions
-    }
-
-    /// MIME content types (lowercased) that resolve to this format.
-    pub fn content_types(&self) -> &[Cow<'static, str>] {
-        &self.content_types
-    }
-
-    /// Decode raw content through this format's loader, returning
-    /// the runtime-tagged handle. Equivalent to calling
-    /// [`CodecRegistry::decode`] after resolving the
-    /// format yourself.
-    ///
-    /// [`CodecRegistry::decode`]: super::CodecRegistry::decode
-    pub async fn decode(
-        &self,
-        content: crate::content::ContentData,
-    ) -> Result<crate::document::UntypedDocumentHandle, nvisy_core::Error> {
-        self.loader.decode(content).await
-    }
-}
-
-impl fmt::Debug for Format {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("Format")
-            .field("id", &self.id)
-            .field("modality", &self.modality)
-            .field("extensions", &self.extensions)
-            .field("content_types", &self.content_types)
-            .finish_non_exhaustive()
-    }
-}
diff --git a/crates/nvisy-codec/src/core/handler.rs b/crates/nvisy-codec/src/core/handler.rs
deleted file mode 100644
index 9ecd97b6c..000000000
--- a/crates/nvisy-codec/src/core/handler.rs
+++ /dev/null
@@ -1,127 +0,0 @@
-//! What a codec handler exposes — the trait surface every shipped
-//! format handler implements:
-//!
-//! - [`Handler<M>`] — per-modality capability trait. Identifies and
-//!   serialises the handler ([`format`], [`source`], [`encode`]),
-//!   streams chunks ([`next_chunk`]), supports random-access reads
-//!   and redactions ([`read`], [`redact`]), and lifts recognizer
-//!   offsets back to source coordinates ([`lift_chunk`]).
-//! - [`Chunk<M>`] — one decoded unit yielded by `next_chunk`.
-//!
-//! [`format`]: Handler::format
-//! [`source`]: Handler::source
-//! [`encode`]: Handler::encode
-//! [`next_chunk`]: Handler::next_chunk
-//! [`read`]: Handler::read
-//! [`redact`]: Handler::redact
-//! [`lift_chunk`]: Handler::lift_chunk
-
-use std::ops::Range;
-
-use nvisy_core::Error;
-use nvisy_core::modality::Modality;
-use nvisy_core::redaction::Redactions;
-
-use super::FormatId;
-use crate::content::{ContentData, ContentSource};
-
-/// One decoded unit yielded by [`Handler::next_chunk`].
-///
-/// `data` is the per-modality wire payload; `location` is the
-/// coordinate the handler will accept in [`Handler::read`] /
-/// [`Handler::redact`] to address the same chunk again. `hints`
-/// carries out-of-band context strings the chunk's structural
-/// neighbours surface — CSV/XLSX column headers, JSON object
-/// keys, HTML parent-element text — for downstream context-aware
-/// recognizers; handlers without such metadata leave it empty.
-#[derive(Debug, Clone, PartialEq)]
-pub struct Chunk<M: Modality> {
-    /// Coordinate addressing this chunk inside the handler.
-    pub location: M::Location,
-    /// Wire payload at the chunk's location.
-    pub data: M::Data,
-    /// Out-of-band context strings recognizers should treat as
-    /// in-context (column headers, parent element text, …).
-    /// Empty when the handler has no such metadata to surface.
-    pub hints: Vec<String>,
-}
-
-/// Per-modality capability trait every format handler implements.
-///
-/// Identifies and serialises the handler ([`format`], [`source`],
-/// [`encode`]), streams chunks ([`next_chunk`]), supports
-/// random-access reads and redactions ([`read`], [`redact`]), and
-/// lifts recognizer offsets back to source coordinates
-/// ([`lift_chunk`]).
-///
-/// The handler owns the streaming cursor — concurrent iteration
-/// of the same handle is not supported (only one `&mut self`).
-///
-/// [`format`]: Handler::format
-/// [`source`]: Handler::source
-/// [`encode`]: Handler::encode
-/// [`next_chunk`]: Handler::next_chunk
-/// [`read`]: Handler::read
-/// [`redact`]: Handler::redact
-/// [`lift_chunk`]: Handler::lift_chunk
-#[async_trait::async_trait]
-pub trait Handler<M: Modality>: Send + Sync + 'static {
-    /// Stable id of the format this handler represents (e.g.
-    /// `"nvisy.text.txt"`). Cheap to clone.
-    fn format(&self) -> FormatId;
-
-    /// Content source identity and lineage for this handler.
-    fn source(&self) -> ContentSource;
-
-    /// Serialize the current handler content back to [`ContentData`].
-    fn encode(&self) -> Result<ContentData, Error>;
-
-    /// Advance the cursor and yield the next chunk, or `None` at
-    /// end-of-stream.
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<M>>, Error>;
-
-    /// Read the wire payload at the given location. Used by
-    /// [`TextAt`] resolvers to fetch bytes for a coordinate already
-    /// known from somewhere else (an entity audit record, an
-    /// annotation). Extraction itself does not call this — it
-    /// drives [`next_chunk`] which returns `(location, data)`
-    /// together.
-    ///
-    /// [`next_chunk`]: Handler::next_chunk
-    /// [`TextAt`]: nvisy_core::extraction::TextAt
-    async fn read(&self, location: &M::Location) -> Result<Option<M::Data>, Error>;
-
-    /// Apply a batch of `(location, replacement)` pairs in whatever
-    /// order is correct for this format. Engine guarantees no two
-    /// locations overlap; handler decides ordering (right-to-left
-    /// for text/audio so deletions don't shift later indices, batch
-    /// per page for PDF, …). The first error aborts the batch.
-    ///
-    /// Use [`Redactions::single`] when only one replacement is needed.
-    async fn redact(&mut self, redactions: Redactions<M>) -> Result<(), Error>;
-
-    /// Translate a `value_range` expressed inside `chunk.data`'s
-    /// coordinate system into a source-coordinate `M::Location`.
-    ///
-    /// Recognizers see the unescaped, decoded chunk payload and
-    /// emit offsets into that. Downstream stages — dedup, redact,
-    /// audit — need locations that address the handler's source
-    /// bytes. `lift_chunk` is the bridge.
-    ///
-    /// For text-shaped handlers where `chunk.data` is byte-for-byte
-    /// a slice of source (TXT lines, HTML text nodes, PDF page
-    /// text, CSV cells, DOCX text runs), the mapping is the
-    /// identity offset add against `chunk.location.start`. Handlers
-    /// whose chunks decode escapes or otherwise transform the
-    /// payload (JSON `\"` / `\\`, future HTML entity refs) override
-    /// to walk their per-chunk escape map.
-    ///
-    /// Returns `None` when the range has no source pre-image — out
-    /// of bounds, lands inside an escape pair, or the modality
-    /// doesn't have a meaningful `usize` value-range concept (image
-    /// bounding boxes, audio time spans, tabular cell coords).
-    /// Non-text impls leave the default `None`.
-    fn lift_chunk(&self, _chunk: &Chunk<M>, _value_range: Range<usize>) -> Option<M::Location> {
-        None
-    }
-}
diff --git a/crates/nvisy-codec/src/core/loader.rs b/crates/nvisy-codec/src/core/loader.rs
deleted file mode 100644
index f0dc51200..000000000
--- a/crates/nvisy-codec/src/core/loader.rs
+++ /dev/null
@@ -1,129 +0,0 @@
-//! Decoding raw bytes into a typed handle.
-//!
-//! - [`Loader<M>`] — per-modality decoder format implementations
-//!   write. Returns a concrete handler that implements
-//!   [`Handler<M>`].
-//! - [`ErasedLoader`] — object-safe loader surface the
-//!   [`CodecRegistry`] stores behind `Arc`. Adapts a per-modality
-//!   `Loader<M>` into a uniform `decode` signature that returns
-//!   [`UntypedDocumentHandle`].
-//! - [`erase`] — bridge from typed `Loader<M>` to
-//!   `Arc<dyn ErasedLoader>` every shipped format uses when
-//!   populating [`Format::loader`].
-//!
-//! The handler's own [`Handler::format`] supplies the
-//! [`FormatId`] inside [`ErasedLoader::decode`]; [`erase`] only
-//! erases `M`.
-//!
-//! [`Handler<M>`]: super::Handler
-//! [`Handler::format`]: super::Handler::format
-//! [`CodecRegistry`]: super::CodecRegistry
-//! [`UntypedDocumentHandle`]: crate::document::UntypedDocumentHandle
-//! [`Format::loader`]: super::Format::loader
-//! [`FormatId`]: super::FormatId
-
-use std::marker::PhantomData;
-use std::sync::Arc;
-
-use nvisy_core::Error;
-use nvisy_core::modality::Modality;
-
-use super::Handler;
-use crate::content::ContentData;
-use crate::document::{DocumentHandle, UntypedDocumentHandle};
-
-/// Per-modality format loader.
-///
-/// A loader validates and parses raw content for modality `M`,
-/// producing a handler that implements [`Handler<M>`]. Loaders are
-/// the leaves the [`CodecRegistry`] composes — registering a
-/// format means registering its loader.
-///
-/// # Implementing a third-party format
-///
-/// 1. Implement [`Handler<M>`] for the per-format handler type that
-///    owns the parsed in-memory representation.
-/// 2. Implement `Loader<M>` for a stateless type whose [`decode`]
-///    validates raw [`ContentData`] and returns the handler.
-/// 3. Build a [`Format`] with [`Format::new`], chain
-///    [`with_extensions`] / [`with_content_types`] as needed, and
-///    register it on a [`CodecRegistry`] via
-///    [`CodecRegistry::add_format`].
-///
-/// The registry erases `M` internally; third-party callers never
-/// touch the object-safe loader surface.
-///
-/// [`Handler<M>`]: super::Handler
-/// [`CodecRegistry`]: super::CodecRegistry
-/// [`CodecRegistry::add_format`]: super::CodecRegistry::add_format
-/// [`Format`]: super::Format
-/// [`Format::new`]: super::Format::new
-/// [`with_extensions`]: super::Format::with_extensions
-/// [`with_content_types`]: super::Format::with_content_types
-/// [`decode`]: Loader::decode
-#[async_trait::async_trait]
-pub trait Loader<M: Modality>: Send + Sync + 'static {
-    /// The handler type this loader produces.
-    type Handler: Handler<M>;
-
-    /// Validate and parse the content, returning the loaded handler.
-    async fn decode(&self, content: ContentData) -> Result<Self::Handler, Error>;
-}
-
-/// Object-safe loader the [`CodecRegistry`] holds behind `Arc`.
-/// Adapts a per-modality [`Loader<M>`] into a uniform `decode`
-/// signature returning an [`UntypedDocumentHandle`].
-///
-/// Crate-internal: every consumer goes through [`Format::decode`]
-/// or [`CodecRegistry::decode`] instead of touching
-/// this trait directly.
-///
-/// [`CodecRegistry`]: super::CodecRegistry
-/// [`Format::decode`]: super::Format::decode
-/// [`CodecRegistry::decode`]: super::CodecRegistry::decode
-#[async_trait::async_trait]
-pub(crate) trait ErasedLoader: Send + Sync + 'static {
-    /// Decode raw content into an [`UntypedDocumentHandle`].
-    async fn decode(&self, content: ContentData) -> Result<UntypedDocumentHandle, Error>;
-}
-
-/// Erase a typed [`Loader<M>`] into an `Arc<dyn ErasedLoader>` the
-/// [`CodecRegistry`] can store. Called only by [`Format::new`] —
-/// not part of the public API.
-///
-/// [`CodecRegistry`]: super::CodecRegistry
-/// [`Format::new`]: super::Format::new
-pub(crate) fn erase<M, L>(loader: L) -> Arc<dyn ErasedLoader>
-where
-    M: Modality,
-    L: Loader<M>,
-    DocumentHandle<M>: Into<UntypedDocumentHandle>,
-{
-    Arc::new(LoaderAdapter {
-        loader,
-        _phantom: PhantomData,
-    })
-}
-
-/// Private wrapper that holds a typed [`Loader<M>`] and implements
-/// the object-safe [`ErasedLoader`] surface. Constructed only via
-/// [`erase`]; not part of the public API.
-struct LoaderAdapter<M: Modality, L: Loader<M>> {
-    loader: L,
-    _phantom: PhantomData<fn() -> M>,
-}
-
-#[async_trait::async_trait]
-impl<M, L> ErasedLoader for LoaderAdapter<M, L>
-where
-    M: Modality,
-    L: Loader<M>,
-    DocumentHandle<M>: Into<UntypedDocumentHandle>,
-{
-    async fn decode(&self, content: ContentData) -> Result<UntypedDocumentHandle, Error> {
-        let handler = self.loader.decode(content).await?;
-        let format = handler.format();
-        let handle: Box<dyn Handler<M>> = Box::new(handler);
-        Ok(DocumentHandle::new(format, handle).into())
-    }
-}
diff --git a/crates/nvisy-codec/src/core/mod.rs b/crates/nvisy-codec/src/core/mod.rs
deleted file mode 100644
index f013507d1..000000000
--- a/crates/nvisy-codec/src/core/mod.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-//! Codec core contracts, grouped by concern:
-//!
-//! - `format` — *what kind of thing a codec is*. [`FormatId`],
-//!   [`Format`] descriptor.
-//! - `handler` — *what a handler exposes*. [`Handler<M>`]
-//!   (per-modality capability surface — identify, encode, stream,
-//!   read, redact, lift), [`Chunk<M>`] payload.
-//! - `loader` — *how raw bytes become a handle*. [`Loader<M>`]
-//!   (per-modality decoder). The registry-side erasure machinery
-//!   (`ErasedLoader` trait, `erase` helper) is crate-internal and
-//!   wired through [`Format::new`] / [`Format::decode`].
-//! - `registry` — *the lookup engine*. [`CodecRegistry`] indexes
-//!   [`Format`]s by id, extension, and content type, and decodes
-//!   bytes through the matching loader.
-//!
-//! Concrete format implementations live in `crate::handler::*`.
-
-mod format;
-mod handler;
-mod loader;
-mod registry;
-
-pub use self::format::{Format, FormatId};
-pub use self::handler::{Chunk, Handler};
-pub use self::loader::Loader;
-pub(crate) use self::loader::{ErasedLoader, erase};
-pub use self::registry::CodecRegistry;
diff --git a/crates/nvisy-codec/src/core/registry.rs b/crates/nvisy-codec/src/core/registry.rs
deleted file mode 100644
index 9f65d12a6..000000000
--- a/crates/nvisy-codec/src/core/registry.rs
+++ /dev/null
@@ -1,173 +0,0 @@
-//! [`CodecRegistry`]: resolves an extension or content type to a
-//! registered [`Format`] and decodes content through its loader.
-//!
-//! Downstream crates register their own formats by calling
-//! [`CodecRegistry::add_format`] — there is no central enum to
-//! extend.
-
-use std::collections::HashMap;
-
-use nvisy_core::Error;
-
-use super::{Format, FormatId};
-use crate::content::ContentData;
-use crate::document::UntypedDocumentHandle;
-
-/// Codec registry — owns the set of registered [`Format`]s and
-/// resolves them by extension, content type, or id.
-#[derive(Debug, Default)]
-pub struct CodecRegistry {
-    formats: Vec<Format>,
-    by_id: HashMap<FormatId, usize>,
-    by_extension: HashMap<String, usize>,
-    by_content_type: HashMap<String, usize>,
-}
-
-impl CodecRegistry {
-    /// Empty registry. Use [`with_format`] / [`add_format`] to add
-    /// custom formats, or [`with_builtin`] to start from a pre-
-    /// populated set of every built-in format the active feature
-    /// set enables.
-    ///
-    /// [`with_format`]: Self::with_format
-    /// [`add_format`]: Self::add_format
-    /// [`with_builtin`]: Self::with_builtin
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Pre-populated registry containing every built-in format the
-    /// active feature set enables (TXT, JSON, HTML, CSV, PNG, JPEG,
-    /// WAV, PDF, …). Equivalent to [`new`] followed by registering
-    /// each built-in format.
-    ///
-    /// Add custom formats afterward with [`with_format`] (chainable)
-    /// or [`add_format`] (in-place); they take precedence on
-    /// extension / content-type collisions (last registration wins).
-    ///
-    /// [`new`]: Self::new
-    /// [`with_format`]: Self::with_format
-    /// [`add_format`]: Self::add_format
-    pub fn with_builtin() -> Self {
-        let mut registry = Self::new();
-        #[cfg(feature = "txt")]
-        registry.add_format(crate::handler::text::txt_format());
-        #[cfg(feature = "json")]
-        registry.add_format(crate::handler::text::json_format());
-        #[cfg(feature = "markdown")]
-        registry.add_format(crate::handler::text::markdown_format());
-        #[cfg(feature = "html")]
-        registry.add_format(crate::handler::text::html_format());
-        #[cfg(feature = "csv")]
-        registry.add_format(crate::handler::tabular::csv_format());
-        #[cfg(feature = "xlsx")]
-        registry.add_format(crate::handler::tabular::xlsx_format());
-        #[cfg(feature = "png")]
-        registry.add_format(crate::handler::image::png_format());
-        #[cfg(feature = "jpeg")]
-        registry.add_format(crate::handler::image::jpeg_format());
-        #[cfg(feature = "tiff")]
-        registry.add_format(crate::handler::image::tiff_format());
-        #[cfg(feature = "wav")]
-        registry.add_format(crate::handler::audio::wav_format());
-        #[cfg(feature = "mp3")]
-        registry.add_format(crate::handler::audio::mp3_format());
-        #[cfg(feature = "pdf")]
-        registry.add_format(crate::handler::rich::pdf_format());
-        #[cfg(feature = "docx")]
-        registry.add_format(crate::handler::rich::docx_format());
-        registry
-    }
-
-    /// Register a [`Format`] and return `self` for chained builder
-    /// calls. Delegates to [`add_format`] for the indexing body.
-    ///
-    /// # Panics
-    ///
-    /// Panics if the format's id is already registered. Extensions
-    /// and content types that conflict with an existing format are
-    /// overwritten (last registration wins) — register custom
-    /// formats *after* [`with_builtin`] if you want them to take
-    /// precedence.
-    ///
-    /// [`with_builtin`]: Self::with_builtin
-    /// [`add_format`]: Self::add_format
-    #[must_use]
-    pub fn with_format(mut self, format: Format) -> Self {
-        self.add_format(format);
-        self
-    }
-
-    /// In-place equivalent of [`with_format`]. Useful with an
-    /// already-mut binding (e.g. inside a cfg-stanza in
-    /// [`with_builtin`]) where the `let registry = registry.with_format(...)`
-    /// dance is just noise.
-    ///
-    /// # Panics
-    ///
-    /// Same conditions as [`with_format`].
-    ///
-    /// [`with_format`]: Self::with_format
-    /// [`with_builtin`]: Self::with_builtin
-    pub fn add_format(&mut self, format: Format) -> &mut Self {
-        assert!(
-            !self.by_id.contains_key(&format.id),
-            "format id already registered: {}",
-            format.id
-        );
-        let index = self.formats.len();
-        for ext in &format.extensions {
-            self.by_extension.insert(ext.to_ascii_lowercase(), index);
-        }
-        for ct in &format.content_types {
-            self.by_content_type.insert(ct.to_ascii_lowercase(), index);
-        }
-        self.by_id.insert(format.id.clone(), index);
-        self.formats.push(format);
-        self
-    }
-
-    /// Look up a registered format by id.
-    pub fn by_id(&self, id: &FormatId) -> Option<&Format> {
-        self.by_id.get(id).map(|&i| &self.formats[i])
-    }
-
-    /// Look up a registered format by file extension
-    /// (case-insensitive, no leading dot).
-    pub fn by_extension(&self, ext: &str) -> Option<&Format> {
-        self.by_extension
-            .get(&ext.to_ascii_lowercase())
-            .map(|&i| &self.formats[i])
-    }
-
-    /// Look up a registered format by MIME content type
-    /// (case-insensitive).
-    pub fn by_content_type(&self, mime: &str) -> Option<&Format> {
-        self.by_content_type
-            .get(&mime.to_ascii_lowercase())
-            .map(|&i| &self.formats[i])
-    }
-
-    /// Iterate over every registered format in registration order.
-    pub fn iter(&self) -> impl Iterator<Item = &Format> {
-        self.formats.iter()
-    }
-
-    /// Decode raw content using the format resolved from the
-    /// extension hint. Accepts anything convertible into
-    /// [`ContentData`] — `&str`, `&[u8]`, `Vec<u8>`, `Bytes`,
-    /// `String`.
-    pub async fn decode(
-        &self,
-        content: impl Into<ContentData>,
-        extension: &str,
-    ) -> Result<UntypedDocumentHandle, Error> {
-        let format = self.by_extension(extension).ok_or_else(|| {
-            Error::validation(
-                format!("no codec registered for extension `{extension}`"),
-                "nvisy_codec::handler::registry::decode",
-            )
-        })?;
-        format.loader.decode(content.into()).await
-    }
-}
diff --git a/crates/nvisy-codec/src/document/audio.rs b/crates/nvisy-codec/src/document/audio.rs
deleted file mode 100644
index 14b86c591..000000000
--- a/crates/nvisy-codec/src/document/audio.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-//! [`DocumentHandle<Audio>`] implements [`DataAt`] and
-//! [`RedactAt`] for the [`Audio`] modality by delegating to the
-//! underlying [`Handler<Audio>`].
-//!
-//! Audio does not implement [`TextAt`]: "text at this location" for
-//! audio is transcript text, which the extraction phase (in
-//! `nvisy-toolkit::extraction`) produces and the engine stamps onto
-//! document blocks. The codec layer has no visibility into STT.
-//!
-//! [`DocumentHandle<Audio>`]: crate::document::DocumentHandle
-//! [`Handler<Audio>`]: crate::core::Handler
-//! [`TextAt`]: nvisy_core::extraction::TextAt
-//! [`DataAt`]: nvisy_core::extraction::DataAt
-//! [`RedactAt`]: nvisy_core::redaction::RedactAt
-//! [`Audio`]: nvisy_core::modality::Audio
-
-use nvisy_core::Result;
-use nvisy_core::extraction::DataAt;
-use nvisy_core::modality::{Audio, AudioData, AudioLocation};
-use nvisy_core::redaction::{RedactAt, Redactions};
-
-use crate::document::DocumentHandle;
-
-#[async_trait::async_trait]
-impl DataAt<Audio> for DocumentHandle<Audio> {
-    async fn data_at(&self, location: &AudioLocation) -> Option<AudioData> {
-        self.handler().read(location).await.ok().flatten()
-    }
-}
-
-#[async_trait::async_trait]
-impl RedactAt<Audio> for DocumentHandle<Audio> {
-    async fn redact_at(&mut self, redactions: Redactions<Audio>) -> Result<()> {
-        self.handler_mut().redact(redactions).await
-    }
-}
diff --git a/crates/nvisy-codec/src/document/image.rs b/crates/nvisy-codec/src/document/image.rs
deleted file mode 100644
index e86fbc074..000000000
--- a/crates/nvisy-codec/src/document/image.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-//! [`DocumentHandle<Image>`] implements [`DataAt`] and
-//! [`RedactAt`] for the [`Image`] modality by delegating to the
-//! underlying [`Handler<Image>`].
-//!
-//! Image does not implement [`TextAt`]: "text at this location" for
-//! an image is OCR text, which the extraction phase (in
-//! `nvisy-toolkit::extraction`) produces and the engine stamps onto
-//! document blocks. The codec layer has no visibility into OCR.
-//!
-//! [`DocumentHandle<Image>`]: crate::document::DocumentHandle
-//! [`Handler<Image>`]: crate::core::Handler
-//! [`TextAt`]: nvisy_core::extraction::TextAt
-//! [`DataAt`]: nvisy_core::extraction::DataAt
-//! [`RedactAt`]: nvisy_core::redaction::RedactAt
-//! [`Image`]: nvisy_core::modality::Image
-
-use nvisy_core::Result;
-use nvisy_core::extraction::DataAt;
-use nvisy_core::modality::{Image, ImageData, ImageLocation};
-use nvisy_core::redaction::{RedactAt, Redactions};
-
-use crate::document::DocumentHandle;
-
-#[async_trait::async_trait]
-impl DataAt<Image> for DocumentHandle<Image> {
-    async fn data_at(&self, location: &ImageLocation) -> Option<ImageData> {
-        self.handler().read(location).await.ok().flatten()
-    }
-}
-
-#[async_trait::async_trait]
-impl RedactAt<Image> for DocumentHandle<Image> {
-    async fn redact_at(&mut self, redactions: Redactions<Image>) -> Result<()> {
-        self.handler_mut().redact(redactions).await
-    }
-}
diff --git a/crates/nvisy-codec/src/document/mod.rs b/crates/nvisy-codec/src/document/mod.rs
deleted file mode 100644
index 7d0a668f4..000000000
--- a/crates/nvisy-codec/src/document/mod.rs
+++ /dev/null
@@ -1,239 +0,0 @@
-//! [`UntypedDocumentHandle`] (runtime-tagged handle returned by the
-//! codec registry) and [`DocumentHandle<M>`] (typed view used
-//! downstream of decode).
-//!
-//! # Two-tier handle shape
-//!
-//! The registry can't know up front which modality a decoded format
-//! produces — that's a property of the format descriptor, resolved
-//! at decode time. [`UntypedDocumentHandle`] is the registry-level
-//! return: an enum with one variant per modality, each carrying an
-//! `Box<dyn Handler<M>>` + the [`FormatId`].
-//!
-//! Consumers downstream of decode commit to a modality via the
-//! consuming accessors ([`into_text`], [`into_image`], …),
-//! which yield a [`DocumentHandle<M>`] — a typed wrapper that owns
-//! the underlying [`Handler<M>`] and exposes the per-modality
-//! capability surface without further dispatch.
-//!
-//! [`Handler<M>`]: crate::core::Handler
-//! [`into_text`]: UntypedDocumentHandle::into_text
-//! [`into_image`]: UntypedDocumentHandle::into_image
-
-#[cfg(feature = "internal_audio")]
-mod audio;
-#[cfg(feature = "internal_image")]
-mod image;
-#[cfg(feature = "internal_tabular")]
-mod tabular;
-#[cfg(feature = "internal_text")]
-mod text;
-
-use std::any::type_name;
-use std::fmt;
-
-use derive_more::From;
-#[cfg(feature = "internal_audio")]
-use nvisy_core::modality::Audio;
-#[cfg(feature = "internal_image")]
-use nvisy_core::modality::Image;
-use nvisy_core::modality::Modality;
-#[cfg(feature = "internal_tabular")]
-use nvisy_core::modality::Tabular;
-#[cfg(feature = "internal_text")]
-use nvisy_core::modality::Text;
-
-use crate::core::{FormatId, Handler};
-
-/// Runtime-tagged handle returned by the codec registry, carrying the
-/// underlying [`Handler<M>`] for some `M` along with the
-/// [`FormatId`] of the producing loader.
-///
-/// Commit to a modality via [`into_text`] / [`into_image`] /
-/// [`into_audio`] / [`into_tabular`] to obtain a typed
-/// [`DocumentHandle<M>`]. The accessors are consuming — once you
-/// commit to a modality, the untyped form is gone.
-///
-/// [`Handler<M>`]: crate::core::Handler
-/// [`into_text`]: Self::into_text
-/// [`into_image`]: Self::into_image
-/// [`into_audio`]: Self::into_audio
-/// [`into_tabular`]: Self::into_tabular
-#[derive(Debug, From)]
-pub enum UntypedDocumentHandle {
-    #[cfg(feature = "internal_text")]
-    /// Text-modality handle.
-    Text(DocumentHandle<Text>),
-    #[cfg(feature = "internal_tabular")]
-    /// Tabular-modality handle.
-    Tabular(DocumentHandle<Tabular>),
-    #[cfg(feature = "internal_image")]
-    /// Image-modality handle.
-    Image(DocumentHandle<Image>),
-    #[cfg(feature = "internal_audio")]
-    /// Audio-modality handle.
-    Audio(DocumentHandle<Audio>),
-}
-
-impl UntypedDocumentHandle {
-    /// The [`FormatId`] of the loader that produced this handle.
-    pub fn format_id(&self) -> &FormatId {
-        match self {
-            #[cfg(feature = "internal_text")]
-            Self::Text(h) => h.format_id(),
-            #[cfg(feature = "internal_tabular")]
-            Self::Tabular(h) => h.format_id(),
-            #[cfg(feature = "internal_image")]
-            Self::Image(h) => h.format_id(),
-            #[cfg(feature = "internal_audio")]
-            Self::Audio(h) => h.format_id(),
-        }
-    }
-
-    /// Consume self, returning the inner [`DocumentHandle<Text>`] if
-    /// this handle carries text modality.
-    #[cfg(feature = "internal_text")]
-    pub fn into_text(self) -> Option<DocumentHandle<Text>> {
-        match self {
-            Self::Text(h) => Some(h),
-            #[cfg(feature = "internal_tabular")]
-            Self::Tabular(_) => None,
-            #[cfg(feature = "internal_image")]
-            Self::Image(_) => None,
-            #[cfg(feature = "internal_audio")]
-            Self::Audio(_) => None,
-        }
-    }
-
-    /// Consume self, returning the inner [`DocumentHandle<Tabular>`]
-    /// if this handle carries tabular modality.
-    #[cfg(feature = "internal_tabular")]
-    pub fn into_tabular(self) -> Option<DocumentHandle<Tabular>> {
-        match self {
-            Self::Tabular(h) => Some(h),
-            #[cfg(feature = "internal_text")]
-            Self::Text(_) => None,
-            #[cfg(feature = "internal_image")]
-            Self::Image(_) => None,
-            #[cfg(feature = "internal_audio")]
-            Self::Audio(_) => None,
-        }
-    }
-
-    /// Consume self, returning the inner [`DocumentHandle<Image>`] if
-    /// this handle carries image modality.
-    #[cfg(feature = "internal_image")]
-    pub fn into_image(self) -> Option<DocumentHandle<Image>> {
-        match self {
-            Self::Image(h) => Some(h),
-            #[cfg(feature = "internal_text")]
-            Self::Text(_) => None,
-            #[cfg(feature = "internal_tabular")]
-            Self::Tabular(_) => None,
-            #[cfg(feature = "internal_audio")]
-            Self::Audio(_) => None,
-        }
-    }
-
-    /// Consume self, returning the inner [`DocumentHandle<Audio>`] if
-    /// this handle carries audio modality.
-    #[cfg(feature = "internal_audio")]
-    pub fn into_audio(self) -> Option<DocumentHandle<Audio>> {
-        match self {
-            Self::Audio(h) => Some(h),
-            #[cfg(feature = "internal_text")]
-            Self::Text(_) => None,
-            #[cfg(feature = "internal_tabular")]
-            Self::Tabular(_) => None,
-            #[cfg(feature = "internal_image")]
-            Self::Image(_) => None,
-        }
-    }
-}
-
-/// Typed view of a single-modality handle. Carries the [`FormatId`]
-/// alongside the handler so audit and provenance can always answer
-/// "what format is this?" without re-decoding.
-///
-/// Constructed via the consuming accessors on
-/// [`UntypedDocumentHandle`]. The inner handler is held in a `Box`
-/// because the pipeline runs phases sequentially per document — there
-/// is no concurrent access to the handle within a single document's
-/// run, so reference counting buys nothing.
-///
-/// Implements the core `*At` trait surface
-/// ([`TextAt`] / [`DataAt`] / [`RedactAt`]) directly so any
-/// pipeline component can read from / write to a codec-backed
-/// source through the same traits the engine bounds on. The
-/// per-modality impls live in the `text` / `tabular` / `image` /
-/// `audio` sibling modules.
-///
-/// Modality coverage:
-///
-/// | Modality   | TextAt | DataAt | RedactAt |
-/// |------------|--------|--------|----------|
-/// | Text       |   ✓    |   ✓    |    ✓     |
-/// | Tabular    |   ✓    |   ✓    |    ✓     |
-/// | Image      |        |   ✓    |    ✓     |
-/// | Audio      |        |   ✓    |    ✓     |
-///
-/// Image and audio don't implement [`TextAt`] — "text at this
-/// location" for image means OCR text, for audio means transcript
-/// text, and both come from the extraction phase in
-/// `nvisy-toolkit::extraction`. The codec layer has no visibility
-/// into either.
-///
-/// [`TextAt`]: nvisy_core::extraction::TextAt
-/// [`DataAt`]: nvisy_core::extraction::DataAt
-/// [`RedactAt`]: nvisy_core::redaction::RedactAt
-pub struct DocumentHandle<M: Modality> {
-    format_id: FormatId,
-    handler: Box<dyn Handler<M>>,
-}
-
-impl<M: Modality> DocumentHandle<M> {
-    /// Wrap a handler and a format id into a typed handle. Used by
-    /// codec loaders to produce the typed handle, which is then
-    /// erased into an [`UntypedDocumentHandle`] variant for registry
-    /// return.
-    pub fn new(format_id: FormatId, handler: Box<dyn Handler<M>>) -> Self {
-        Self { format_id, handler }
-    }
-
-    /// The [`FormatId`] of the producing loader.
-    pub fn format_id(&self) -> &FormatId {
-        &self.format_id
-    }
-
-    /// Borrow the inner handler. Use this for read-only capability
-    /// methods ([`Handler::read`]).
-    ///
-    /// [`Handler::read`]: crate::core::Handler::read
-    pub fn handler(&self) -> &dyn Handler<M> {
-        &*self.handler
-    }
-
-    /// Mutably borrow the inner handler. Use this for cursor-advancing
-    /// methods ([`Handler::next_chunk`]) and the redaction batch
-    /// applicator ([`Handler::redact`]).
-    ///
-    /// [`Handler::next_chunk`]: crate::core::Handler::next_chunk
-    /// [`Handler::redact`]: crate::core::Handler::redact
-    pub fn handler_mut(&mut self) -> &mut dyn Handler<M> {
-        &mut *self.handler
-    }
-
-    /// Consume self, returning the inner handler.
-    pub fn into_handler(self) -> Box<dyn Handler<M>> {
-        self.handler
-    }
-}
-
-impl<M: Modality> fmt::Debug for DocumentHandle<M> {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("DocumentHandle")
-            .field("format_id", &self.format_id)
-            .field("modality", &type_name::<M>())
-            .finish()
-    }
-}
diff --git a/crates/nvisy-codec/src/document/tabular.rs b/crates/nvisy-codec/src/document/tabular.rs
deleted file mode 100644
index 974c3c064..000000000
--- a/crates/nvisy-codec/src/document/tabular.rs
+++ /dev/null
@@ -1,43 +0,0 @@
-//! [`DocumentHandle<Tabular>`] implements [`TextAt`], [`DataAt`],
-//! and [`RedactAt`] for the [`Tabular`] modality by delegating to
-//! the underlying [`Handler<Tabular>`].
-//!
-//! [`DocumentHandle<Tabular>`]: crate::document::DocumentHandle
-//! [`Handler<Tabular>`]: crate::core::Handler
-//! [`TextAt`]: nvisy_core::extraction::TextAt
-//! [`DataAt`]: nvisy_core::extraction::DataAt
-//! [`RedactAt`]: nvisy_core::redaction::RedactAt
-//! [`Tabular`]: nvisy_core::modality::Tabular
-
-use nvisy_core::Result;
-use nvisy_core::extraction::{DataAt, TextAt};
-use nvisy_core::modality::{Tabular, TabularLocation, TextData};
-use nvisy_core::redaction::{RedactAt, Redactions};
-
-use crate::document::DocumentHandle;
-
-#[async_trait::async_trait]
-impl TextAt<Tabular> for DocumentHandle<Tabular> {
-    async fn text_at(&self, location: &TabularLocation) -> Option<String> {
-        self.handler()
-            .read(location)
-            .await
-            .ok()
-            .flatten()
-            .map(|d| d.into_string())
-    }
-}
-
-#[async_trait::async_trait]
-impl DataAt<Tabular> for DocumentHandle<Tabular> {
-    async fn data_at(&self, location: &TabularLocation) -> Option<TextData> {
-        self.handler().read(location).await.ok().flatten()
-    }
-}
-
-#[async_trait::async_trait]
-impl RedactAt<Tabular> for DocumentHandle<Tabular> {
-    async fn redact_at(&mut self, redactions: Redactions<Tabular>) -> Result<()> {
-        self.handler_mut().redact(redactions).await
-    }
-}
diff --git a/crates/nvisy-codec/src/document/text.rs b/crates/nvisy-codec/src/document/text.rs
deleted file mode 100644
index 588cbe3f8..000000000
--- a/crates/nvisy-codec/src/document/text.rs
+++ /dev/null
@@ -1,43 +0,0 @@
-//! [`DocumentHandle<Text>`] implements [`TextAt`], [`DataAt`], and
-//! [`RedactAt`] for the [`Text`] modality by delegating to the
-//! underlying [`Handler<Text>`].
-//!
-//! [`DocumentHandle<Text>`]: crate::document::DocumentHandle
-//! [`Handler<Text>`]: crate::core::Handler
-//! [`TextAt`]: nvisy_core::extraction::TextAt
-//! [`DataAt`]: nvisy_core::extraction::DataAt
-//! [`RedactAt`]: nvisy_core::redaction::RedactAt
-//! [`Text`]: nvisy_core::modality::Text
-
-use nvisy_core::Result;
-use nvisy_core::extraction::{DataAt, TextAt};
-use nvisy_core::modality::{Text, TextData, TextLocation};
-use nvisy_core::redaction::{RedactAt, Redactions};
-
-use crate::document::DocumentHandle;
-
-#[async_trait::async_trait]
-impl TextAt<Text> for DocumentHandle<Text> {
-    async fn text_at(&self, location: &TextLocation) -> Option<String> {
-        self.handler()
-            .read(location)
-            .await
-            .ok()
-            .flatten()
-            .map(|d| d.into_string())
-    }
-}
-
-#[async_trait::async_trait]
-impl DataAt<Text> for DocumentHandle<Text> {
-    async fn data_at(&self, location: &TextLocation) -> Option<TextData> {
-        self.handler().read(location).await.ok().flatten()
-    }
-}
-
-#[async_trait::async_trait]
-impl RedactAt<Text> for DocumentHandle<Text> {
-    async fn redact_at(&mut self, redactions: Redactions<Text>) -> Result<()> {
-        self.handler_mut().redact(redactions).await
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/audio/duration.rs b/crates/nvisy-codec/src/handler/audio/duration.rs
deleted file mode 100644
index 421ccfa54..000000000
--- a/crates/nvisy-codec/src/handler/audio/duration.rs
+++ /dev/null
@@ -1,152 +0,0 @@
-//! Audio clip duration probing via Symphonia.
-//!
-//! Both [`WavHandler`] and [`Mp3Handler`] hand the whole clip back
-//! as a single [`Chunk<Audio>`] from `next_chunk`. The chunk's
-//! [`AudioLocation`] needs an honest [`TimeSpan`] so downstream
-//! consumers (most importantly the STT extractor) can anchor their
-//! per-segment timestamps to absolute clip coordinates.
-//!
-//! [`probe_duration_us`] runs Symphonia's format probe over the
-//! supplied bytes and reads the first audio track's container-level
-//! duration. No samples are decoded — the probe stops as soon as
-//! the demuxer has seen enough header to populate the track
-//! metadata.
-//!
-//! [`WavHandler`]: super::WavHandler
-//! [`Mp3Handler`]: super::Mp3Handler
-//! [`Chunk<Audio>`]: crate::Chunk
-//! [`AudioLocation`]: nvisy_core::modality::AudioLocation
-//! [`TimeSpan`]: nvisy_core::primitive::TimeSpan
-//! [`probe_duration_us`]: self::probe_duration_us
-
-use std::io::Cursor;
-
-use bytes::Bytes;
-use nvisy_core::Error;
-use symphonia::core::formats::FormatOptions;
-use symphonia::core::formats::probe::Hint;
-use symphonia::core::io::MediaSourceStream;
-use symphonia::core::meta::MetadataOptions;
-use symphonia::core::units::Timestamp;
-use symphonia::default::get_probe;
-
-const TARGET: &str = "nvisy_codec::handler::audio::duration";
-
-/// Probe `bytes` with Symphonia and return the clip duration in
-/// microseconds.
-///
-/// `extension_hint` is a Symphonia format hint (e.g. `"wav"`,
-/// `"mp3"`). The hint biases the probe toward the format the caller
-/// expects; an incorrect hint does not cause a failure as long as
-/// the bytes still match some registered format.
-///
-/// Errors when the probe cannot find a container format, when the
-/// first track lacks a timebase or a known duration, or when the
-/// computed duration would overflow `i64` microseconds.
-pub(super) fn probe_duration_us(bytes: &Bytes, extension_hint: &str) -> Result<i64, Error> {
-    let mss = MediaSourceStream::new(Box::new(Cursor::new(bytes.clone())), Default::default());
-
-    let mut hint = Hint::new();
-    hint.with_extension(extension_hint);
-
-    let reader = get_probe()
-        .probe(
-            &hint,
-            mss,
-            FormatOptions::default(),
-            MetadataOptions::default(),
-        )
-        .map_err(|e| Error::validation(format!("audio probe failed: {e}"), TARGET))?;
-
-    let track = reader
-        .tracks()
-        .first()
-        .ok_or_else(|| Error::validation("audio probe returned no tracks", TARGET))?;
-
-    let time_base = track
-        .time_base
-        .ok_or_else(|| Error::validation("audio track is missing a timebase", TARGET))?;
-    let duration = track.duration.ok_or_else(|| {
-        Error::validation("audio track is missing a container-level duration", TARGET)
-    })?;
-
-    // `Track::duration` is `Duration` (timebase ticks, u64). `TimeBase::calc_time`
-    // operates on `Timestamp` (timebase ticks, i64) — same unit, different signedness.
-    // Reinterpret as a timestamp anchored at zero.
-    let ticks = i64::try_from(duration.get()).map_err(|_| {
-        Error::validation(
-            format!(
-                "audio duration in timebase ticks overflowed i64: {}",
-                duration.get()
-            ),
-            TARGET,
-        )
-    })?;
-    let time = time_base.calc_time(Timestamp::new(ticks)).ok_or_else(|| {
-        Error::validation("audio duration overflowed when converted to time", TARGET)
-    })?;
-
-    let us = time.as_nanos() / 1_000;
-    if us < 0 {
-        return Err(Error::validation(
-            format!("audio duration is negative: {us} us"),
-            TARGET,
-        ));
-    }
-    i64::try_from(us).map_err(|_| {
-        Error::validation(
-            format!("audio duration overflowed i64 microseconds: {us}"),
-            TARGET,
-        )
-    })
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn write_silence_wav(duration_us: i64, sample_rate: u32) -> Bytes {
-        use hound::{SampleFormat, WavSpec, WavWriter};
-
-        let spec = WavSpec {
-            channels: 1,
-            sample_rate,
-            bits_per_sample: 16,
-            sample_format: SampleFormat::Int,
-        };
-        let total_samples = ((duration_us as i128) * (sample_rate as i128) / 1_000_000) as usize;
-
-        let mut buf = Cursor::new(Vec::<u8>::new());
-        {
-            let mut writer = WavWriter::new(&mut buf, spec).unwrap();
-            for _ in 0..total_samples {
-                writer.write_sample(0i16).unwrap();
-            }
-            writer.finalize().unwrap();
-        }
-        Bytes::from(buf.into_inner())
-    }
-
-    #[test]
-    fn probes_one_second_wav() {
-        let wav = write_silence_wav(1_000_000, 16_000);
-        let us = probe_duration_us(&wav, "wav").unwrap();
-        // 16_000 samples at 16_000 Hz = exactly 1_000_000 us.
-        assert_eq!(us, 1_000_000);
-    }
-
-    #[test]
-    fn probes_fractional_wav() {
-        let wav = write_silence_wav(2_500_000, 8_000);
-        let us = probe_duration_us(&wav, "wav").unwrap();
-        // 20_000 samples at 8_000 Hz = 2_500_000 us.
-        assert_eq!(us, 2_500_000);
-    }
-
-    #[test]
-    fn rejects_garbage_bytes() {
-        let garbage = Bytes::from_static(b"not an audio file");
-        let err = probe_duration_us(&garbage, "wav").unwrap_err();
-        assert!(err.to_string().contains("audio probe failed"));
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/audio/mod.rs b/crates/nvisy-codec/src/handler/audio/mod.rs
deleted file mode 100644
index 133edec24..000000000
--- a/crates/nvisy-codec/src/handler/audio/mod.rs
+++ /dev/null
@@ -1,37 +0,0 @@
-//! Audio modality: concrete audio format implementations (WAV, MP3).
-//!
-//! Audio handlers override [`Handler::redact`] to call
-//! [`Redactions::sort_descending`] so spans are applied right-to-left
-//! (an [`AudioReplacement::Remove`] shrinks the buffer and shifts
-//! every later sample index; right-to-left order keeps earlier
-//! indices valid). Replacements use [`AudioReplacement`].
-//!
-//! [`AudioReplacement`]: nvisy_core::redaction::AudioReplacement
-//! [`AudioReplacement::Remove`]: nvisy_core::redaction::AudioReplacement::Remove
-//! [`Redactions::sort_descending`]: nvisy_core::redaction::Redactions::sort_descending
-//! [`Handler::redact`]: crate::Handler::redact
-
-#[cfg(any(feature = "wav", feature = "mp3"))]
-pub(crate) mod duration;
-#[cfg(any(feature = "wav", feature = "mp3"))]
-pub(crate) mod redact;
-
-#[cfg(feature = "mp3")]
-mod mp3_codec;
-#[cfg(feature = "mp3")]
-mod mp3_handler;
-#[cfg(feature = "mp3")]
-mod mp3_loader;
-#[cfg(feature = "wav")]
-mod wav_handler;
-#[cfg(feature = "wav")]
-mod wav_loader;
-
-#[cfg(feature = "mp3")]
-pub use self::mp3_handler::{Mp3Handler, format as mp3_format};
-#[cfg(feature = "mp3")]
-pub use self::mp3_loader::Mp3Loader;
-#[cfg(feature = "wav")]
-pub use self::wav_handler::{WavHandler, format as wav_format};
-#[cfg(feature = "wav")]
-pub use self::wav_loader::WavLoader;
diff --git a/crates/nvisy-codec/src/handler/audio/mp3_codec.rs b/crates/nvisy-codec/src/handler/audio/mp3_codec.rs
deleted file mode 100644
index e8b6f7984..000000000
--- a/crates/nvisy-codec/src/handler/audio/mp3_codec.rs
+++ /dev/null
@@ -1,373 +0,0 @@
-//! MP3 decode + re-encode helpers for the redaction round-trip.
-//!
-//! Two helpers, each owning one side of the redact-via-PCM strategy
-//! the MP3 handler uses:
-//!
-//! - [`decode_to_pcm`] turns MP3 bytes into a flat, channel-
-//!   interleaved `Vec<f32>` plus the sample rate / channel count it
-//!   was decoded at. Sample-rate and channel-count come from the
-//!   first audio track's codec parameters and are assumed constant
-//!   across the clip (LAME requires them at builder time).
-//! - [`encode_from_pcm`] turns an interleaved `Vec<f32>` back into
-//!   MP3 bytes via `mp3lame-encoder` at a caller-supplied bitrate.
-//!
-//! Re-encoding is lossy in addition to whatever loss the input
-//! already had; the pair only exists to support sample-level
-//! redaction of MP3 streams. Callers wanting bit-perfect preservation
-//! of unredacted regions should not round-trip.
-
-use std::io::{Cursor, ErrorKind as IoErrorKind};
-
-use bytes::Bytes;
-use mp3lame_encoder::{Builder, FlushNoGap, InterleavedPcm, MonoPcm};
-use nvisy_core::Error;
-use symphonia::core::audio::conv::{ConvertibleSample, FromSample};
-use symphonia::core::audio::{Audio as _, AudioBuffer, GenericAudioBufferRef};
-use symphonia::core::codecs::audio::AudioDecoderOptions;
-use symphonia::core::errors::Error as SymError;
-use symphonia::core::formats::probe::Hint;
-use symphonia::core::formats::{FormatOptions, TrackType};
-use symphonia::core::io::MediaSourceStream;
-use symphonia::core::meta::MetadataOptions;
-use symphonia::default::{get_codecs, get_probe};
-
-const TARGET: &str = "nvisy_codec::handler::audio::mp3_codec";
-
-/// Decoded MP3 contents in interleaved `f32` PCM form.
-pub(super) struct DecodedMp3 {
-    pub samples: Vec<f32>,
-    pub sample_rate: u32,
-    pub channels: u16,
-}
-
-/// Probe `bytes` and return the channel count of the first audio
-/// track. Cheap — runs only the format probe + codec-params lookup,
-/// no packet decoding.
-///
-/// Used by the loader as a gate so >2-channel inputs are rejected
-/// before they ever reach the redact path (LAME's encoder side
-/// only supports mono and stereo, and silent downmixing would
-/// quietly edit the *unredacted* audio).
-pub(super) fn probe_channels(bytes: &Bytes) -> Result<u16, Error> {
-    let mss = MediaSourceStream::new(Box::new(Cursor::new(bytes.clone())), Default::default());
-    let mut hint = Hint::new();
-    hint.with_extension("mp3");
-
-    let reader = get_probe()
-        .probe(
-            &hint,
-            mss,
-            FormatOptions::default(),
-            MetadataOptions::default(),
-        )
-        .map_err(|e| Error::validation(format!("MP3 probe failed: {e}"), TARGET))?;
-
-    let track = reader
-        .default_track(TrackType::Audio)
-        .ok_or_else(|| Error::validation("MP3 stream has no audio track", TARGET))?;
-
-    let channels = track
-        .codec_params
-        .as_ref()
-        .and_then(|p| p.audio())
-        .and_then(|a| a.channels.as_ref())
-        .map(|c| c.count())
-        .ok_or_else(|| Error::validation("MP3 track is missing channel info", TARGET))?;
-
-    u16::try_from(channels).map_err(|_| {
-        Error::validation(
-            format!("MP3 track channel count {channels} exceeds u16"),
-            TARGET,
-        )
-    })
-}
-
-/// Decode `bytes` (an MP3 stream) into interleaved `f32` PCM samples.
-///
-/// Symphonia returns each decoded packet as a planar
-/// [`GenericAudioBufferRef`]; this helper interleaves the channels so
-/// the result is a single `Vec<f32>` suitable for both the shared
-/// [`super::redact::apply`] helper and for handing back to
-/// [`encode_from_pcm`].
-pub(super) fn decode_to_pcm(bytes: &Bytes) -> Result<DecodedMp3, Error> {
-    let mss = MediaSourceStream::new(Box::new(Cursor::new(bytes.clone())), Default::default());
-    let mut hint = Hint::new();
-    hint.with_extension("mp3");
-
-    let mut reader = get_probe()
-        .probe(
-            &hint,
-            mss,
-            FormatOptions::default(),
-            MetadataOptions::default(),
-        )
-        .map_err(|e| Error::validation(format!("MP3 probe failed: {e}"), TARGET))?;
-
-    let track = reader
-        .default_track(TrackType::Audio)
-        .ok_or_else(|| Error::validation("MP3 stream has no audio track", TARGET))?;
-    let track_id = track.id;
-
-    let audio_params = track
-        .codec_params
-        .as_ref()
-        .and_then(|p| p.audio())
-        .ok_or_else(|| Error::validation("MP3 track is missing audio codec params", TARGET))?
-        .clone();
-
-    let sample_rate = audio_params
-        .sample_rate
-        .ok_or_else(|| Error::validation("MP3 track is missing a sample rate", TARGET))?;
-    let channels = audio_params
-        .channels
-        .as_ref()
-        .map(|c| c.count())
-        .ok_or_else(|| Error::validation("MP3 track is missing channel info", TARGET))?;
-    let channels_u16 = u16::try_from(channels).map_err(|_| {
-        Error::validation(
-            format!("MP3 track channel count {channels} exceeds u16"),
-            TARGET,
-        )
-    })?;
-
-    let mut decoder = get_codecs()
-        .make_audio_decoder(&audio_params, &AudioDecoderOptions::default())
-        .map_err(|e| Error::validation(format!("MP3 decoder init failed: {e}"), TARGET))?;
-
-    let mut samples = Vec::<f32>::new();
-    let mut dropped_packets: u64 = 0;
-    loop {
-        let packet = match reader.next_packet() {
-            Ok(Some(p)) => p,
-            Ok(None) => break,
-            Err(SymError::IoError(io_err)) if io_err.kind() == IoErrorKind::UnexpectedEof => {
-                break;
-            }
-            Err(e) => {
-                return Err(Error::validation(
-                    format!("MP3 packet read failed: {e}"),
-                    TARGET,
-                ));
-            }
-        };
-        if packet.track_id != track_id {
-            continue;
-        }
-        match decoder.decode(&packet) {
-            Ok(buf_ref) => append_interleaved_f32(&buf_ref, channels, &mut samples),
-            Err(SymError::DecodeError(_)) => {
-                // Single-packet decode failures don't fatally end the
-                // stream; symphonia's reference player skips them and
-                // continues. Match that behaviour so a single corrupt
-                // frame doesn't abort the redact pass — but count the
-                // drops so the caller can correlate the shorter output
-                // with the input damage.
-                dropped_packets += 1;
-                continue;
-            }
-            Err(e) => {
-                return Err(Error::validation(format!("MP3 decode failed: {e}"), TARGET));
-            }
-        }
-    }
-
-    if dropped_packets > 0 {
-        // Each dropped packet removes ~1152 samples (MPEG layer 3
-        // frame). Surface this so support investigations can
-        // correlate a shorter-than-expected output with input damage
-        // and so callers can reject the redact result if drift is
-        // unacceptable.
-        tracing::warn!(
-            target: TARGET,
-            dropped_packets,
-            decoded_samples = samples.len(),
-            "decoded MP3 had corrupt packets; redacted output will be shorter than input",
-        );
-    }
-
-    Ok(DecodedMp3 {
-        samples,
-        sample_rate,
-        channels: channels_u16,
-    })
-}
-
-fn append_interleaved_f32(
-    buf_ref: &GenericAudioBufferRef<'_>,
-    channels: usize,
-    out: &mut Vec<f32>,
-) {
-    fn extend<S: ConvertibleSample + Copy>(
-        buf: &AudioBuffer<S>,
-        channels: usize,
-        out: &mut Vec<f32>,
-    ) where
-        f32: FromSample<S>,
-    {
-        let frames = buf.frames();
-        out.reserve(frames * channels);
-        for frame in 0..frames {
-            for ch in 0..channels {
-                let plane = buf.plane(ch).expect("plane for known channel index");
-                let sample = plane[frame];
-                out.push(<f32 as FromSample<S>>::from_sample(sample));
-            }
-        }
-    }
-
-    match buf_ref {
-        GenericAudioBufferRef::U8(buf) => extend(buf, channels, out),
-        GenericAudioBufferRef::U16(buf) => extend(buf, channels, out),
-        GenericAudioBufferRef::U32(buf) => extend(buf, channels, out),
-        GenericAudioBufferRef::S8(buf) => extend(buf, channels, out),
-        GenericAudioBufferRef::S16(buf) => extend(buf, channels, out),
-        GenericAudioBufferRef::S32(buf) => extend(buf, channels, out),
-        GenericAudioBufferRef::F32(buf) => extend(buf, channels, out),
-        GenericAudioBufferRef::F64(buf) => extend(buf, channels, out),
-        // MP3 decoders only output 8-/16-/32-bit integer or 32-/64-bit
-        // float samples — never 24-bit. If symphonia ever changes that
-        // we'd want to handle the new variants explicitly, not silently
-        // route them through a generic path that was never tested.
-        GenericAudioBufferRef::U24(_) | GenericAudioBufferRef::S24(_) => {
-            unreachable!("MP3 decoder does not emit 24-bit sample buffers");
-        }
-    }
-}
-
-/// Encode `samples` (interleaved `f32` PCM) back to MP3 bytes.
-///
-/// `target_bitrate_bps` is the desired average bitrate in bits/sec.
-/// It's snapped to the nearest [`mp3lame_encoder::Bitrate`] variant
-/// LAME accepts; passing the input file's measured average bitrate
-/// keeps the round-trip approximately size-stable.
-pub(super) fn encode_from_pcm(
-    samples: &[f32],
-    sample_rate: u32,
-    channels: u16,
-    target_bitrate_bps: u32,
-) -> Result<Vec<u8>, Error> {
-    let bitrate = snap_bitrate(target_bitrate_bps);
-
-    let mut encoder =
-        Builder::new().ok_or_else(|| Error::validation("LAME builder failed", TARGET))?;
-    encoder
-        .set_sample_rate(sample_rate)
-        .map_err(|e| Error::validation(format!("LAME sample-rate rejected: {e:?}"), TARGET))?;
-    encoder
-        .set_num_channels(channels as u8)
-        .map_err(|e| Error::validation(format!("LAME channel count rejected: {e:?}"), TARGET))?;
-    encoder
-        .set_brate(bitrate)
-        .map_err(|e| Error::validation(format!("LAME bitrate rejected: {e:?}"), TARGET))?;
-    encoder
-        .set_quality(mp3lame_encoder::Quality::Good)
-        .map_err(|e| Error::validation(format!("LAME quality rejected: {e:?}"), TARGET))?;
-
-    let mut encoder = encoder
-        .build()
-        .map_err(|e| Error::validation(format!("LAME init failed: {e:?}"), TARGET))?;
-
-    let frames = samples.len() / channels as usize;
-    let mut out = Vec::<u8>::with_capacity(mp3lame_encoder::max_required_buffer_size(frames));
-
-    match channels {
-        1 => {
-            encoder
-                .encode_to_vec(MonoPcm(samples), &mut out)
-                .map_err(|e| Error::validation(format!("LAME encode failed: {e:?}"), TARGET))?;
-        }
-        2 => {
-            encoder
-                .encode_to_vec(InterleavedPcm(samples), &mut out)
-                .map_err(|e| Error::validation(format!("LAME encode failed: {e:?}"), TARGET))?;
-        }
-        n => {
-            return Err(Error::validation(
-                format!("LAME supports 1 or 2 channels, got {n}"),
-                TARGET,
-            ));
-        }
-    }
-    encoder
-        .flush_to_vec::<FlushNoGap>(&mut out)
-        .map_err(|e| Error::validation(format!("LAME flush failed: {e:?}"), TARGET))?;
-
-    Ok(out)
-}
-
-/// Snap an arbitrary bits-per-second target to the nearest
-/// [`mp3lame_encoder::Bitrate`] variant LAME accepts.
-fn snap_bitrate(bps: u32) -> mp3lame_encoder::Bitrate {
-    use mp3lame_encoder::Bitrate::*;
-
-    // LAME's permissible bitrates. Variant tag value equals the kbps.
-    const ALL: &[mp3lame_encoder::Bitrate] = &[
-        Kbps8, Kbps16, Kbps24, Kbps32, Kbps40, Kbps48, Kbps64, Kbps80, Kbps96, Kbps112, Kbps128,
-        Kbps160, Kbps192, Kbps224, Kbps256, Kbps320,
-    ];
-
-    let kbps = (bps + 500) / 1_000; // round to nearest kbps
-    ALL.iter()
-        .copied()
-        .min_by_key(|b| ((*b as u32) as i64 - kbps as i64).abs())
-        .unwrap_or(Kbps128)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn snap_bitrate_picks_nearest() {
-        use mp3lame_encoder::Bitrate;
-
-        assert_eq!(snap_bitrate(127_000) as u32, Bitrate::Kbps128 as u32);
-        assert_eq!(snap_bitrate(96_000) as u32, Bitrate::Kbps96 as u32);
-        assert_eq!(snap_bitrate(0) as u32, Bitrate::Kbps8 as u32);
-        assert_eq!(snap_bitrate(1_000_000) as u32, Bitrate::Kbps320 as u32);
-    }
-
-    #[test]
-    fn snap_bitrate_resolves_ties_toward_lower() {
-        // 144_000 bps is exactly halfway between Kbps128 and Kbps160.
-        // `min_by_key` returns the first minimum seen in iteration
-        // order, so the lower variant wins. Locking this in so future
-        // refactors don't silently flip the tie-break direction.
-        assert_eq!(
-            snap_bitrate(144_000) as u32,
-            mp3lame_encoder::Bitrate::Kbps128 as u32
-        );
-    }
-
-    #[test]
-    fn snap_bitrate_clamps_below_floor() {
-        use mp3lame_encoder::Bitrate;
-
-        // Below the lowest valid bitrate (Kbps8 = 8 kbps) should still
-        // produce Kbps8 — degraded inputs (e.g. duration probe glitch)
-        // shouldn't crash the encoder.
-        assert_eq!(snap_bitrate(7_999) as u32, Bitrate::Kbps8 as u32);
-        assert_eq!(snap_bitrate(7_500) as u32, Bitrate::Kbps8 as u32);
-    }
-
-    #[test]
-    fn round_trips_silence() {
-        // Synthesise 0.5s of mono silence at 16 kHz, encode, decode, compare length.
-        let samples = vec![0f32; 8_000];
-        let encoded = encode_from_pcm(&samples, 16_000, 1, 64_000).unwrap();
-        assert!(!encoded.is_empty(), "encode produced no bytes");
-        let decoded = decode_to_pcm(&Bytes::from(encoded)).unwrap();
-        assert_eq!(decoded.sample_rate, 16_000);
-        assert_eq!(decoded.channels, 1);
-        // LAME adds encoder delay/padding so the decoded length isn't
-        // exactly equal to the input, but should land within roughly
-        // ±2 MP3 frames worth of samples (~2300 samples at 16 kHz).
-        let diff = (decoded.samples.len() as i64 - samples.len() as i64).abs();
-        assert!(
-            diff < 2_400,
-            "round-trip drift too large: input {}, decoded {}",
-            samples.len(),
-            decoded.samples.len()
-        );
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/audio/mp3_handler.rs b/crates/nvisy-codec/src/handler/audio/mp3_handler.rs
deleted file mode 100644
index c0b834229..000000000
--- a/crates/nvisy-codec/src/handler/audio/mp3_handler.rs
+++ /dev/null
@@ -1,282 +0,0 @@
-//! MP3 handler: holds raw MP3 audio bytes and exposes them as a
-//! single-track audio handle via [`Handler<Audio>`].
-//!
-//! [`Handler::redact`] decodes the stream via Symphonia, mutates an
-//! interleaved `f32` PCM buffer with the same
-//! [`super::redact::apply`] helper the WAV handler uses, then
-//! re-encodes via `mp3lame-encoder` at the input clip's measured
-//! average bitrate. The round-trip is therefore lossy in addition to
-//! whatever loss the input already had — operators wanting bit-exact
-//! preservation of unredacted regions should not redact MP3 inputs.
-//!
-//! The MP3 encoder dependency carries an LGPL-3.0 licence (via
-//! libmp3lame) and requires a C toolchain plus `autoconf` /
-//! `automake` at build time. Both prerequisites travel with the
-//! `mp3` crate feature.
-//!
-//! [`Handler<Audio>`]: crate::Handler
-//! [`Handler::redact`]: crate::Handler::redact
-//! [`super::redact::apply`]: super::redact::apply
-
-use bytes::Bytes;
-use nvisy_core::Error;
-use nvisy_core::modality::{Audio, AudioData, AudioLocation};
-use nvisy_core::primitive::TimeSpan;
-use nvisy_core::redaction::Redactions;
-
-use super::duration::probe_duration_us;
-use super::mp3_codec::{decode_to_pcm, encode_from_pcm};
-use super::{Mp3Loader, redact};
-use crate::content::{ContentData, ContentSource};
-use crate::{Chunk, Format, FormatId, Handler};
-
-/// Stable [`FormatId`] for the MP3 codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.audio.mp3");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Audio, _>(FORMAT_ID.clone(), Mp3Loader)
-        .with_extensions(["mp3"])
-        .with_content_types(["audio/mpeg"])
-}
-
-/// Handler for loaded MP3 content.
-#[derive(Debug)]
-pub struct Mp3Handler {
-    source: ContentSource,
-    bytes: Bytes,
-    filename: String,
-    yielded: bool,
-}
-
-impl Mp3Handler {
-    /// Create a handler from raw MP3 bytes.
-    pub fn new(bytes: Bytes) -> Self {
-        Self {
-            source: ContentSource::new(),
-            bytes,
-            filename: "audio.mp3".to_owned(),
-            yielded: false,
-        }
-    }
-
-    /// Attach a content source for lineage tracking.
-    pub fn with_source(mut self, source: ContentSource) -> Self {
-        self.source = source;
-        self
-    }
-
-    /// Attach a filename hint for downstream extractors.
-    pub fn with_filename(mut self, filename: impl Into<String>) -> Self {
-        self.filename = filename.into();
-        self
-    }
-
-    /// Reference to the raw audio bytes.
-    pub fn bytes(&self) -> &Bytes {
-        &self.bytes
-    }
-
-    /// Rewind the streaming cursor.
-    pub fn rewind(&mut self) {
-        self.yielded = false;
-    }
-}
-
-#[async_trait::async_trait]
-impl Handler<Audio> for Mp3Handler {
-    fn format(&self) -> FormatId {
-        FORMAT_ID.clone()
-    }
-
-    fn source(&self) -> ContentSource {
-        self.source
-    }
-
-    #[tracing::instrument(name = "mp3.encode", skip_all, fields(output_bytes))]
-    fn encode(&self) -> Result<ContentData, Error> {
-        tracing::Span::current().record("output_bytes", self.bytes.len());
-        let source = ContentSource::new().with_parent(&self.source);
-        Ok(ContentData::new(source, self.bytes.clone()))
-    }
-
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<Audio>>, Error> {
-        if self.yielded {
-            return Ok(None);
-        }
-        let duration_us = probe_duration_us(&self.bytes, "mp3")?;
-        let location = AudioLocation::new(TimeSpan::new(0, duration_us));
-        let data = AudioData::new(self.bytes.clone()).with_filename(self.filename.clone());
-        self.yielded = true;
-        Ok(Some(Chunk {
-            location,
-            data,
-            hints: Vec::new(),
-        }))
-    }
-
-    async fn read(&self, _location: &AudioLocation) -> Result<Option<AudioData>, Error> {
-        Ok(Some(
-            AudioData::new(self.bytes.clone()).with_filename(self.filename.clone()),
-        ))
-    }
-
-    /// Apply spans right-to-left so a [`AudioReplacement::Remove`]
-    /// doesn't invalidate earlier sample indices. The MP3 round-trip
-    /// decodes via `symphonia`, mutates an interleaved `f32` PCM
-    /// buffer, then re-encodes via `mp3lame-encoder` at the input's
-    /// average bitrate so the output's size stays in the same
-    /// ballpark.
-    ///
-    /// [`AudioReplacement::Remove`]: nvisy_core::redaction::AudioReplacement::Remove
-    async fn redact(&mut self, mut redactions: Redactions<Audio>) -> Result<(), Error> {
-        if redactions.is_empty() {
-            return Ok(());
-        }
-        let original_bytes_len = self.bytes.len();
-        let original_duration_us = probe_duration_us(&self.bytes, "mp3")?;
-
-        let mut decoded = decode_to_pcm(&self.bytes)?;
-
-        redactions.sort_descending();
-        for (location, replacement) in redactions.into_items() {
-            redact::apply(
-                &mut decoded.samples,
-                location.time_span,
-                &replacement,
-                decoded.sample_rate,
-                decoded.channels,
-            );
-        }
-
-        let target_bitrate_bps = average_bitrate_bps(original_bytes_len, original_duration_us);
-        let new_bytes = encode_from_pcm(
-            &decoded.samples,
-            decoded.sample_rate,
-            decoded.channels,
-            target_bitrate_bps,
-        )?;
-        self.bytes = Bytes::from(new_bytes);
-        Ok(())
-    }
-}
-
-/// Compute the average bitrate of the source file in bits/sec, so
-/// the re-encode lands roughly on the same target. Falls back to
-/// 128 kbps when the duration is non-positive (degenerate input).
-///
-/// `file_bytes` is the **whole** MP3 file size including ID3v1/v2
-/// tags and any embedded album art; this slightly biases the
-/// computed bitrate above the true audio-payload bitrate. The
-/// inflation is typically <5% (a few KB of tags against megabytes
-/// of audio) and gets snapped to a discrete LAME variant anyway,
-/// so the practical effect on output size is negligible. Worth
-/// revisiting only if we ever ship clips with very large embedded
-/// art relative to audio length.
-fn average_bitrate_bps(file_bytes: usize, duration_us: i64) -> u32 {
-    if duration_us <= 0 {
-        return 128_000;
-    }
-    // bps = file_bits * 1_000_000 / duration_us
-    let bits = (file_bytes as u128).saturating_mul(8);
-    let bps = bits
-        .saturating_mul(1_000_000)
-        .checked_div(duration_us as u128)
-        .unwrap_or(128_000);
-    u32::try_from(bps).unwrap_or(320_000)
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::redaction::AudioReplacement;
-
-    use super::*;
-
-    /// Mint a small mono MP3 fixture for round-trip tests: 1 second
-    /// of constant-amplitude tone at 16 kHz, encoded at 64 kbps.
-    fn fixture_mono_tone_mp3(seconds: usize) -> Bytes {
-        let total = 16_000 * seconds;
-        let samples: Vec<f32> = (0..total).map(|i| (i as f32 * 0.01).sin() * 0.5).collect();
-        let encoded = encode_from_pcm(&samples, 16_000, 1, 64_000).expect("encode fixture");
-        Bytes::from(encoded)
-    }
-
-    #[tokio::test]
-    async fn empty_redactions_is_noop() {
-        let bytes = fixture_mono_tone_mp3(1);
-        let original = bytes.clone();
-        let mut handler = Mp3Handler::new(bytes);
-        let rs: Redactions<Audio> = Redactions::default();
-        handler.redact(rs).await.unwrap();
-        assert_eq!(handler.bytes(), &original);
-    }
-
-    #[tokio::test]
-    async fn silence_redaction_round_trips_and_zeros_target_window() {
-        let bytes = fixture_mono_tone_mp3(2);
-        let mut handler = Mp3Handler::new(bytes);
-
-        let mut rs = Redactions::new();
-        rs.push(
-            AudioLocation::new(TimeSpan::new(500_000, 1_500_000)),
-            AudioReplacement::Silence,
-        );
-        handler.redact(rs).await.unwrap();
-
-        // Decode the redacted MP3 back and assert the middle second is
-        // (approximately) silence. We allow generous tolerance because
-        // MP3 is lossy and the encoder smears boundaries slightly.
-        let decoded = decode_to_pcm(handler.bytes()).unwrap();
-        let sr = decoded.sample_rate as usize;
-        // Sample a window inside the silenced region, away from the
-        // boundary where smear is biggest.
-        let start = sr * 3 / 4;
-        let end = sr * 5 / 4;
-        let mean_abs: f32 = decoded.samples[start..end]
-            .iter()
-            .map(|s| s.abs())
-            .sum::<f32>()
-            / (end - start) as f32;
-        assert!(
-            mean_abs < 0.05,
-            "silenced region should be near zero, mean |s| = {mean_abs}"
-        );
-    }
-
-    #[tokio::test]
-    async fn remove_redaction_shortens_clip_duration() {
-        let bytes = fixture_mono_tone_mp3(3);
-        let original_duration = probe_duration_us(&bytes, "mp3").unwrap();
-        let mut handler = Mp3Handler::new(bytes);
-
-        let mut rs = Redactions::new();
-        rs.push(
-            AudioLocation::new(TimeSpan::new(1_000_000, 2_000_000)),
-            AudioReplacement::Remove,
-        );
-        handler.redact(rs).await.unwrap();
-
-        let new_duration = probe_duration_us(handler.bytes(), "mp3").unwrap();
-        // Removed 1s out of 3s; LAME pads the edges, so expect roughly
-        // 2s ± 200ms.
-        let diff = (new_duration - 2_000_000).abs();
-        assert!(
-            diff < 200_000,
-            "expected ~2s after 1s removal, got {new_duration} us"
-        );
-        assert!(
-            new_duration < original_duration,
-            "redacted duration must be shorter than original"
-        );
-    }
-
-    #[tokio::test]
-    async fn next_chunk_propagates_probe_error_for_garbage_bytes() {
-        // Real MP3 fixtures are constructed inline above; here we only
-        // need to confirm the handler wires the probe in and surfaces
-        // failures rather than silently stamping (0, 0).
-        let mut handler = Mp3Handler::new(Bytes::from_static(b"definitely not an mp3"));
-        let err = handler.next_chunk().await.unwrap_err();
-        assert!(err.to_string().contains("audio probe failed"));
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/audio/mp3_loader.rs b/crates/nvisy-codec/src/handler/audio/mp3_loader.rs
deleted file mode 100644
index 1b6010ecd..000000000
--- a/crates/nvisy-codec/src/handler/audio/mp3_loader.rs
+++ /dev/null
@@ -1,86 +0,0 @@
-//! MP3 loader: wraps raw audio bytes into a [`Mp3Handler`].
-
-use nvisy_core::Error;
-use nvisy_core::modality::Audio;
-
-use super::Mp3Handler;
-use super::mp3_codec::probe_channels;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource};
-
-const TARGET: &str = "nvisy_codec::handler::audio::mp3_loader";
-
-/// Loader that wraps raw MP3 bytes. Produces one [`Mp3Handler`] per input.
-#[derive(Debug, Default)]
-pub struct Mp3Loader;
-
-#[async_trait::async_trait]
-impl Loader<Audio> for Mp3Loader {
-    type Handler = Mp3Handler;
-
-    /// Decode the loader content into a handler.
-    ///
-    /// MP3 redaction round-trips through `mp3lame-encoder`, which
-    /// only supports 1 or 2 channels. Reject anything wider here so
-    /// the failure is surfaced at load time (with a clear message)
-    /// rather than later from the redact path. Silent downmixing
-    /// would quietly edit the *unredacted* audio.
-    #[tracing::instrument(name = "mp3.decode", skip_all, fields(input_bytes))]
-    async fn decode(&self, content: ContentData) -> Result<Mp3Handler, Error> {
-        tracing::Span::current().record("input_bytes", content.to_bytes().len());
-        let parent = content.content_source;
-        let bytes = content.to_bytes();
-
-        let channels = probe_channels(&bytes)?;
-        if channels > 2 {
-            return Err(Error::validation(
-                format!(
-                    "MP3 has {channels} channels; only mono and stereo MP3s are supported \
-                     (downmixing would silently edit unredacted audio)"
-                ),
-                TARGET,
-            ));
-        }
-
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(Mp3Handler::new(bytes).with_source(source))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use bytes::Bytes;
-
-    use super::super::mp3_codec::encode_from_pcm;
-    use super::*;
-
-    /// Mint a tiny stereo MP3 from sample-level data so the loader's
-    /// channel-count gate has something real to probe. We can't easily
-    /// fixture the >2-channel rejection path here — LAME only encodes
-    /// mono/stereo, so producing a 6-channel MP3 in-test would need an
-    /// external encoder. That arm is covered by `probe_channels` unit
-    /// tests upstream of the loader.
-    fn fixture_stereo_mp3() -> Bytes {
-        let samples = vec![0f32; 16_000 * 2]; // 1s stereo silence
-        let encoded = encode_from_pcm(&samples, 16_000, 2, 64_000).expect("encode fixture");
-        Bytes::from(encoded)
-    }
-
-    #[tokio::test]
-    async fn accepts_stereo_mp3() {
-        let bytes = fixture_stereo_mp3();
-        let content = ContentData::new(ContentSource::new(), bytes);
-        let handler = Mp3Loader.decode(content).await;
-        handler.expect("stereo MP3 should load");
-    }
-
-    #[tokio::test]
-    async fn rejects_garbage_bytes() {
-        let content = ContentData::new(
-            ContentSource::new(),
-            Bytes::from_static(b"definitely not an mp3"),
-        );
-        let err = Mp3Loader.decode(content).await.unwrap_err();
-        assert!(err.to_string().contains("MP3 probe failed"));
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/audio/redact.rs b/crates/nvisy-codec/src/handler/audio/redact.rs
deleted file mode 100644
index b12475086..000000000
--- a/crates/nvisy-codec/src/handler/audio/redact.rs
+++ /dev/null
@@ -1,150 +0,0 @@
-//! Per-sample-buffer audio redaction helper shared by every audio
-//! handler (WAV today; MP3 doesn't redact and so doesn't use it).
-//!
-//! Applies one [`AudioReplacement`] to a flat, channel-interleaved
-//! `Vec<S>` sample buffer in place, given the replacement's containing
-//! time span and the buffer's sample rate + channel count.
-//!
-//! Ordering across multiple replacements is the caller's
-//! responsibility: an [`AudioReplacement::Remove`] shrinks the buffer,
-//! so later time spans must be applied first to keep earlier ones'
-//! indices valid. Audio handlers typically override
-//! [`Handler::redact`] to call [`Redactions::sort_descending`].
-//!
-//! [`Handler::redact`]: crate::Handler::redact
-//! [`Redactions::sort_descending`]: nvisy_core::redaction::Redactions::sort_descending
-
-use nvisy_core::primitive::TimeSpan;
-use nvisy_core::redaction::AudioReplacement;
-
-const TARGET: &str = "nvisy_codec::handler::audio::redact";
-
-/// Apply a single replacement to `samples` in place.
-pub(crate) fn apply<S>(
-    samples: &mut Vec<S>,
-    time_span: TimeSpan,
-    replacement: &AudioReplacement,
-    sample_rate: u32,
-    channels: u16,
-) where
-    S: Default + Clone,
-{
-    let (start_sample, end_sample) = time_span.sample_range(sample_rate, channels);
-    let start = start_sample.min(samples.len());
-    let end = end_sample.min(samples.len());
-    if start >= end {
-        return;
-    }
-    match replacement {
-        AudioReplacement::Silence => {
-            for s in &mut samples[start..end] {
-                *s = S::default();
-            }
-        }
-        AudioReplacement::Remove => {
-            samples.drain(start..end);
-        }
-        AudioReplacement::Replace { .. } => {
-            tracing::warn!(
-                target: TARGET,
-                start_us = time_span.start_us,
-                end_us = time_span.end_us,
-                "AudioReplacement::Replace is not yet implemented, skipping",
-            );
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use bytes::Bytes;
-
-    use super::*;
-
-    #[test]
-    fn silence_zeroes_range_mono() {
-        let mut samples: Vec<i16> = (1..=10).collect();
-        apply(
-            &mut samples,
-            TimeSpan::new(3_000, 6_000),
-            &AudioReplacement::Silence,
-            1000,
-            1,
-        );
-        assert_eq!(samples, vec![1, 2, 3, 0, 0, 0, 7, 8, 9, 10]);
-    }
-
-    #[test]
-    fn remove_shrinks_range_mono() {
-        let mut samples: Vec<i16> = (1..=10).collect();
-        apply(
-            &mut samples,
-            TimeSpan::new(3_000, 6_000),
-            &AudioReplacement::Remove,
-            1000,
-            1,
-        );
-        assert_eq!(samples, vec![1, 2, 3, 7, 8, 9, 10]);
-    }
-
-    #[test]
-    fn stereo_silence_aligns_to_frames() {
-        let mut samples: Vec<i16> = (1..=20).collect();
-        apply(
-            &mut samples,
-            TimeSpan::new(3_000, 6_000),
-            &AudioReplacement::Silence,
-            1000,
-            2,
-        );
-        assert_eq!(
-            samples,
-            vec![
-                1, 2, 3, 4, 5, 6, 0, 0, 0, 0, 0, 0, 13, 14, 15, 16, 17, 18, 19, 20
-            ],
-        );
-    }
-
-    #[test]
-    fn stereo_remove_drops_frames_not_samples() {
-        let mut samples: Vec<i16> = (1..=20).collect();
-        apply(
-            &mut samples,
-            TimeSpan::new(3_000, 6_000),
-            &AudioReplacement::Remove,
-            1000,
-            2,
-        );
-        assert_eq!(samples.len(), 14);
-        assert_eq!(
-            samples,
-            vec![1, 2, 3, 4, 5, 6, 13, 14, 15, 16, 17, 18, 19, 20]
-        );
-    }
-
-    #[test]
-    fn out_of_bounds_clipped() {
-        let mut samples: Vec<i16> = (1..=5).collect();
-        apply(
-            &mut samples,
-            TimeSpan::new(0, 999_999_000),
-            &AudioReplacement::Silence,
-            1000,
-            1,
-        );
-        assert_eq!(samples, vec![0, 0, 0, 0, 0]);
-    }
-
-    #[test]
-    fn replace_is_warned_and_skipped() {
-        let mut samples: Vec<i16> = (1..=5).collect();
-        apply(
-            &mut samples,
-            TimeSpan::new(0, 3_000),
-            &AudioReplacement::Replace { data: Bytes::new() },
-            1000,
-            1,
-        );
-        assert_eq!(samples, vec![1, 2, 3, 4, 5]);
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/audio/wav_handler.rs b/crates/nvisy-codec/src/handler/audio/wav_handler.rs
deleted file mode 100644
index 99297a0da..000000000
--- a/crates/nvisy-codec/src/handler/audio/wav_handler.rs
+++ /dev/null
@@ -1,326 +0,0 @@
-//! WAV handler: holds raw WAV audio bytes and exposes them as a
-//! single-track audio handle via [`Handler<Audio>`].
-//!
-//! Redaction decodes the WAV via [`hound`], applies sample-level
-//! mutations, and re-encodes back to bytes. Supported formats: `i8` /
-//! `i16` / `i32` PCM and `f32` IEEE float; other bit depths surface a
-//! clear error.
-//!
-//! Batched [`Handler::redact`] sorts right-to-left by
-//! `time_span.start_us` so [`AudioReplacement::Remove`] operations don't
-//! shift the indices of pending redactions.
-//!
-//! [`Handler<Audio>`]: crate::Handler
-//! [`Handler::redact`]: crate::Handler::redact
-//! [`AudioReplacement::Remove`]: nvisy_core::redaction::AudioReplacement::Remove
-
-use std::io::Cursor;
-
-use bytes::Bytes;
-use hound::{Sample, SampleFormat, WavReader, WavSpec, WavWriter};
-use nvisy_core::Error;
-use nvisy_core::modality::{Audio, AudioData, AudioLocation};
-use nvisy_core::primitive::TimeSpan;
-use nvisy_core::redaction::{AudioReplacement, Redactions};
-
-use super::duration::probe_duration_us;
-use super::{WavLoader, redact};
-use crate::content::{ContentData, ContentSource};
-use crate::{Chunk, Format, FormatId, Handler};
-
-const TARGET: &str = "nvisy_codec::handler::audio::wav";
-
-/// Stable [`FormatId`] for the WAV codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.audio.wav");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Audio, _>(FORMAT_ID.clone(), WavLoader)
-        .with_extensions(["wav"])
-        .with_content_types(["audio/wav", "audio/x-wav"])
-}
-
-/// Handler for loaded WAV content. Stores the encoded bytes; decode
-/// happens on demand inside [`Handler::redact`].
-///
-/// [`Handler::redact`]: crate::Handler::redact
-#[derive(Debug)]
-pub struct WavHandler {
-    source: ContentSource,
-    bytes: Bytes,
-    filename: String,
-    yielded: bool,
-}
-
-impl WavHandler {
-    /// Create a handler from raw WAV bytes.
-    pub fn new(bytes: Bytes) -> Self {
-        Self {
-            source: ContentSource::new(),
-            bytes,
-            filename: "audio.wav".to_owned(),
-            yielded: false,
-        }
-    }
-
-    /// Attach a content source for lineage tracking.
-    pub fn with_source(mut self, source: ContentSource) -> Self {
-        self.source = source;
-        self
-    }
-
-    /// Attach a filename hint for downstream extractors.
-    pub fn with_filename(mut self, filename: impl Into<String>) -> Self {
-        self.filename = filename.into();
-        self
-    }
-
-    /// Reference to the raw audio bytes.
-    pub fn bytes(&self) -> &Bytes {
-        &self.bytes
-    }
-
-    /// Rewind the streaming cursor.
-    pub fn rewind(&mut self) {
-        self.yielded = false;
-    }
-}
-
-#[async_trait::async_trait]
-impl Handler<Audio> for WavHandler {
-    fn format(&self) -> FormatId {
-        FORMAT_ID.clone()
-    }
-
-    fn source(&self) -> ContentSource {
-        self.source
-    }
-
-    #[tracing::instrument(name = "wav.encode", skip_all, fields(output_bytes))]
-    fn encode(&self) -> Result<ContentData, Error> {
-        tracing::Span::current().record("output_bytes", self.bytes.len());
-        let source = ContentSource::new().with_parent(&self.source);
-        Ok(ContentData::new(source, self.bytes.clone()))
-    }
-
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<Audio>>, Error> {
-        if self.yielded {
-            return Ok(None);
-        }
-        let duration_us = probe_duration_us(&self.bytes, "wav")?;
-        let location = AudioLocation::new(TimeSpan::new(0, duration_us));
-        let data = AudioData::new(self.bytes.clone()).with_filename(self.filename.clone());
-        self.yielded = true;
-        Ok(Some(Chunk {
-            location,
-            data,
-            hints: Vec::new(),
-        }))
-    }
-
-    async fn read(&self, _location: &AudioLocation) -> Result<Option<AudioData>, Error> {
-        Ok(Some(
-            AudioData::new(self.bytes.clone()).with_filename(self.filename.clone()),
-        ))
-    }
-
-    /// Apply spans right-to-left so a [`AudioReplacement::Remove`] doesn't
-    /// invalidate earlier sample indices.
-    async fn redact(&mut self, mut redactions: Redactions<Audio>) -> Result<(), Error> {
-        redactions.sort_descending();
-        for (location, replacement) in redactions.into_items() {
-            self.redact_one(&location, replacement)?;
-        }
-        Ok(())
-    }
-}
-
-impl WavHandler {
-    fn redact_one(
-        &mut self,
-        location: &AudioLocation,
-        replacement: AudioReplacement,
-    ) -> Result<(), Error> {
-        let spec = read_spec(&self.bytes)?;
-        let new_bytes = match (spec.sample_format, spec.bits_per_sample) {
-            (SampleFormat::Int, 8) => {
-                redact_typed::<i8>(&self.bytes, spec, location.time_span, &replacement)?
-            }
-            (SampleFormat::Int, 16) => {
-                redact_typed::<i16>(&self.bytes, spec, location.time_span, &replacement)?
-            }
-            (SampleFormat::Int, 24 | 32) => {
-                redact_typed::<i32>(&self.bytes, spec, location.time_span, &replacement)?
-            }
-            (SampleFormat::Float, 32) => {
-                redact_typed::<f32>(&self.bytes, spec, location.time_span, &replacement)?
-            }
-            _ => {
-                return Err(Error::validation(
-                    format!(
-                        "WAV format not yet supported: {:?}/{} bits",
-                        spec.sample_format, spec.bits_per_sample
-                    ),
-                    TARGET,
-                ));
-            }
-        };
-        self.bytes = Bytes::from(new_bytes);
-        Ok(())
-    }
-}
-
-fn read_spec(bytes: &Bytes) -> Result<WavSpec, Error> {
-    let reader = WavReader::new(Cursor::new(bytes.as_ref()))
-        .map_err(|e| Error::validation(format!("invalid WAV: {e}"), TARGET))?;
-    Ok(reader.spec())
-}
-
-fn redact_typed<S>(
-    bytes: &Bytes,
-    spec: WavSpec,
-    time_span: TimeSpan,
-    replacement: &AudioReplacement,
-) -> Result<Vec<u8>, Error>
-where
-    S: Sample + Default + Clone,
-{
-    let mut reader = WavReader::new(Cursor::new(bytes.as_ref()))
-        .map_err(|e| Error::validation(format!("invalid WAV: {e}"), TARGET))?;
-    let mut samples: Vec<S> = reader
-        .samples::<S>()
-        .collect::<Result<Vec<_>, _>>()
-        .map_err(|e| Error::validation(format!("WAV sample decode error: {e}"), TARGET))?;
-
-    redact::apply(
-        &mut samples,
-        time_span,
-        replacement,
-        spec.sample_rate,
-        spec.channels,
-    );
-
-    let mut buf = Cursor::new(Vec::<u8>::new());
-    {
-        let mut writer = WavWriter::new(&mut buf, spec)
-            .map_err(|e| Error::validation(format!("WAV writer init error: {e}"), TARGET))?;
-        for sample in samples {
-            writer
-                .write_sample(sample)
-                .map_err(|e| Error::validation(format!("WAV sample write error: {e}"), TARGET))?;
-        }
-        writer
-            .finalize()
-            .map_err(|e| Error::validation(format!("WAV finalize error: {e}"), TARGET))?;
-    }
-    Ok(buf.into_inner())
-}
-
-#[cfg(test)]
-mod tests {
-    use hound::SampleFormat;
-    use nvisy_core::redaction::AudioReplacement;
-
-    use super::*;
-
-    fn encode_wav_mono_i16(samples: &[i16]) -> Bytes {
-        let spec = WavSpec {
-            channels: 1,
-            sample_rate: 1000,
-            bits_per_sample: 16,
-            sample_format: SampleFormat::Int,
-        };
-        let mut buf = Cursor::new(Vec::<u8>::new());
-        {
-            let mut writer = WavWriter::new(&mut buf, spec).unwrap();
-            for &s in samples {
-                writer.write_sample(s).unwrap();
-            }
-            writer.finalize().unwrap();
-        }
-        Bytes::from(buf.into_inner())
-    }
-
-    fn decode_wav_mono_i16(bytes: &Bytes) -> Vec<i16> {
-        let mut reader = WavReader::new(Cursor::new(bytes.as_ref())).unwrap();
-        reader.samples::<i16>().map(Result::unwrap).collect()
-    }
-
-    #[tokio::test]
-    async fn silence_zeros_samples_in_range() {
-        let bytes = encode_wav_mono_i16(&[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
-        let mut handler = WavHandler::new(bytes);
-        let mut rs = Redactions::new();
-        rs.push(
-            AudioLocation::new(TimeSpan::new(3_000, 6_000)),
-            AudioReplacement::Silence,
-        );
-        handler.redact(rs).await.unwrap();
-        let samples = decode_wav_mono_i16(handler.bytes());
-        assert_eq!(samples, vec![1, 2, 3, 0, 0, 0, 7, 8, 9, 10]);
-    }
-
-    #[tokio::test]
-    async fn remove_shortens_file() {
-        let bytes = encode_wav_mono_i16(&[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
-        let mut handler = WavHandler::new(bytes);
-        let mut rs = Redactions::new();
-        rs.push(
-            AudioLocation::new(TimeSpan::new(3_000, 6_000)),
-            AudioReplacement::Remove,
-        );
-        handler.redact(rs).await.unwrap();
-        let samples = decode_wav_mono_i16(handler.bytes());
-        assert_eq!(samples, vec![1, 2, 3, 7, 8, 9, 10]);
-    }
-
-    #[tokio::test]
-    async fn multiple_removes_apply_right_to_left() {
-        let bytes = encode_wav_mono_i16(&[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]);
-        let mut handler = WavHandler::new(bytes);
-        let mut rs = Redactions::new();
-        rs.push(
-            AudioLocation::new(TimeSpan::new(1_000, 3_000)),
-            AudioReplacement::Remove,
-        );
-        rs.push(
-            AudioLocation::new(TimeSpan::new(6_000, 8_000)),
-            AudioReplacement::Remove,
-        );
-        handler.redact(rs).await.unwrap();
-        let samples = decode_wav_mono_i16(handler.bytes());
-        assert_eq!(samples, vec![1, 4, 5, 6, 9, 10]);
-    }
-
-    #[tokio::test]
-    async fn empty_redactions_is_noop() {
-        let bytes = encode_wav_mono_i16(&[1, 2, 3]);
-        let original = bytes.clone();
-        let mut handler = WavHandler::new(bytes);
-        let rs: Redactions<Audio> = Redactions::default();
-        handler.redact(rs).await.unwrap();
-        assert_eq!(handler.bytes(), &original);
-    }
-
-    #[tokio::test]
-    async fn unsupported_format_returns_error() {
-        let mut handler = WavHandler::new(Bytes::from_static(b"not-a-wav"));
-        let mut rs = Redactions::new();
-        rs.push(
-            AudioLocation::new(TimeSpan::new(0, 1_000)),
-            AudioReplacement::Silence,
-        );
-        let err = handler.redact(rs).await.unwrap_err();
-        assert!(err.to_string().contains("invalid WAV"));
-    }
-
-    #[tokio::test]
-    async fn next_chunk_reports_real_duration() {
-        // 1000 samples at 1000 Hz mono = exactly 1_000_000 us.
-        let bytes = encode_wav_mono_i16(&vec![0i16; 1000]);
-        let mut handler = WavHandler::new(bytes);
-        let chunk = handler.next_chunk().await.unwrap().unwrap();
-        assert_eq!(chunk.location.time_span.start_us, 0);
-        assert_eq!(chunk.location.time_span.end_us, 1_000_000);
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/audio/wav_loader.rs b/crates/nvisy-codec/src/handler/audio/wav_loader.rs
deleted file mode 100644
index 133f62905..000000000
--- a/crates/nvisy-codec/src/handler/audio/wav_loader.rs
+++ /dev/null
@@ -1,26 +0,0 @@
-//! WAV loader: wraps raw audio bytes into a [`WavHandler`].
-
-use nvisy_core::Error;
-use nvisy_core::modality::Audio;
-
-use super::WavHandler;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource};
-
-/// Loader that wraps raw WAV bytes. Produces one [`WavHandler`] per input.
-#[derive(Debug, Default)]
-pub struct WavLoader;
-
-#[async_trait::async_trait]
-impl Loader<Audio> for WavLoader {
-    type Handler = WavHandler;
-
-    #[tracing::instrument(name = "wav.decode", skip_all, fields(input_bytes))]
-    async fn decode(&self, content: ContentData) -> Result<WavHandler, Error> {
-        tracing::Span::current().record("input_bytes", content.to_bytes().len());
-        let parent = content.content_source;
-        let bytes = content.to_bytes();
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(WavHandler::new(bytes).with_source(source))
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/image/image_ops.rs b/crates/nvisy-codec/src/handler/image/image_ops.rs
deleted file mode 100644
index a5f73f1a2..000000000
--- a/crates/nvisy-codec/src/handler/image/image_ops.rs
+++ /dev/null
@@ -1,83 +0,0 @@
-//! Per-region pixel operations on [`DynamicImage`]: gaussian blur,
-//! solid-color overlay, and pixelation. Used by [`super::redact`] to
-//! render image-modality redaction outputs.
-
-use image::DynamicImage;
-use image::imageops::FilterType;
-use imageproc::filter::gaussian_blur_f32;
-use nvisy_core::primitive::{Color, IBoundingBox};
-
-/// Mutating image-transform operations on individual bounding-box
-/// regions.
-pub(super) trait ImageOps {
-    /// Apply a gaussian blur to `region` with the given `sigma`.
-    fn apply_gaussian_blur(&mut self, region: &IBoundingBox, sigma: f32);
-
-    /// Fill `region` with a solid `color`.
-    fn apply_block_overlay(&mut self, region: &IBoundingBox, color: Color);
-
-    /// Pixelate `region` with the given `block_size`.
-    fn apply_pixelate(&mut self, region: &IBoundingBox, block_size: u32);
-}
-
-impl ImageOps for DynamicImage {
-    fn apply_gaussian_blur(&mut self, region: &IBoundingBox, sigma: f32) {
-        let (x, y, w, h) = (region.x, region.y, region.width, region.height);
-
-        let img_w = self.width();
-        let img_h = self.height();
-        if x >= img_w || y >= img_h {
-            return;
-        }
-        let w = w.min(img_w - x);
-        let h = h.min(img_h - y);
-        if w == 0 || h == 0 {
-            return;
-        }
-
-        let sub = self.crop_imm(x, y, w, h);
-        let blurred = DynamicImage::ImageRgba8(gaussian_blur_f32(&sub.to_rgba8(), sigma));
-        image::imageops::overlay(self, &blurred, x as i64, y as i64);
-    }
-
-    fn apply_block_overlay(&mut self, region: &IBoundingBox, color: Color) {
-        let (x, y, w, h) = (region.x, region.y, region.width, region.height);
-
-        let img_w = self.width();
-        let img_h = self.height();
-        if x >= img_w || y >= img_h {
-            return;
-        }
-        let w = w.min(img_w - x);
-        let h = h.min(img_h - y);
-
-        let rgba = image::Rgba([color.r, color.g, color.b, 255]);
-        let block = image::RgbaImage::from_pixel(w, h, rgba);
-        image::imageops::overlay(self, &block, x as i64, y as i64);
-    }
-
-    fn apply_pixelate(&mut self, region: &IBoundingBox, block_size: u32) {
-        let block_size = block_size.max(1);
-        let (x, y, w, h) = (region.x, region.y, region.width, region.height);
-
-        let img_w = self.width();
-        let img_h = self.height();
-        if x >= img_w || y >= img_h {
-            return;
-        }
-        let w = w.min(img_w - x);
-        let h = h.min(img_h - y);
-        if w == 0 || h == 0 {
-            return;
-        }
-
-        let small_w = (w / block_size).max(1);
-        let small_h = (h / block_size).max(1);
-
-        let sub = self.crop_imm(x, y, w, h);
-        let small = sub.resize_exact(small_w, small_h, FilterType::Nearest);
-        let pixelated = small.resize_exact(w, h, FilterType::Nearest);
-
-        image::imageops::overlay(self, &pixelated, x as i64, y as i64);
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/image/jpeg_handler.rs b/crates/nvisy-codec/src/handler/image/jpeg_handler.rs
deleted file mode 100644
index 50f92ac5d..000000000
--- a/crates/nvisy-codec/src/handler/image/jpeg_handler.rs
+++ /dev/null
@@ -1,25 +0,0 @@
-//! JPEG handler: see [`super::png_handler`] for the shared shape.
-
-use super::JpegLoader;
-use crate::content::ContentSource;
-
-/// Handler for loaded JPEG content. Holds the decoded
-/// [`image::DynamicImage`] and a streaming cursor; the macro-generated
-/// `Handler<Image>` impl encodes back to JPEG on demand.
-#[derive(Debug)]
-pub struct JpegHandler {
-    source: ContentSource,
-    image: image::DynamicImage,
-    yielded: bool,
-}
-
-impl_image_handler!(
-    handler = JpegHandler,
-    loader = JpegLoader,
-    format_id = "nvisy.image.jpeg",
-    extensions = ["jpg", "jpeg"],
-    content_types = ["image/jpeg"],
-    image_format = image::ImageFormat::Jpeg,
-    origin = "jpeg-handler",
-    encode_span = "jpeg.encode",
-);
diff --git a/crates/nvisy-codec/src/handler/image/jpeg_loader.rs b/crates/nvisy-codec/src/handler/image/jpeg_loader.rs
deleted file mode 100644
index 6fbae6f62..000000000
--- a/crates/nvisy-codec/src/handler/image/jpeg_loader.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! JPEG loader: validates and decodes raw JPEG bytes into a
-//! [`JpegHandler`].
-
-use nvisy_core::Error;
-use nvisy_core::modality::Image;
-
-use super::JpegHandler;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource};
-
-/// Loader for JPEG files. Produces one [`JpegHandler`] per input.
-#[derive(Debug, Default)]
-pub struct JpegLoader;
-
-#[async_trait::async_trait]
-impl Loader<Image> for JpegLoader {
-    type Handler = JpegHandler;
-
-    #[tracing::instrument(name = "jpeg.decode", skip_all, fields(input_bytes, width, height))]
-    async fn decode(&self, content: ContentData) -> Result<JpegHandler, Error> {
-        tracing::Span::current().record("input_bytes", content.to_bytes().len());
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        let image = image::load_from_memory(&raw)
-            .map_err(|e| Error::validation(format!("JPEG decode failed: {e}"), "jpeg-loader"))?;
-        tracing::Span::current().record("width", image.width());
-        tracing::Span::current().record("height", image.height());
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(JpegHandler::new(image).with_source(source))
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/image/macros.rs b/crates/nvisy-codec/src/handler/image/macros.rs
deleted file mode 100644
index e32d3e484..000000000
--- a/crates/nvisy-codec/src/handler/image/macros.rs
+++ /dev/null
@@ -1,204 +0,0 @@
-//! [`impl_image_handler!`]: shared macro for single-image handler
-//! structs (PNG, JPEG, TIFF — anything backed by one
-//! `image::DynamicImage` + a `ContentSource`).
-//!
-//! Generates the [`Handler<Image>`] impl plus inherent constructors
-//! and a `pub fn format() -> Format` descriptor that registers with
-//! [`crate::CodecRegistry`].
-//!
-//! [`Handler<Image>`]: crate::Handler
-
-/// Implement [`Handler<Image>`] and the shared inherent methods for
-/// an image handler struct that holds a single `DynamicImage`, a
-/// `ContentSource`, and a streaming cursor.
-///
-/// Each invocation also emits:
-/// - a `pub const FORMAT_ID: FormatId` for the handler's stable id
-/// - a `pub fn format() -> Format` descriptor that wires the
-///   matching loader through [`Format::new`] into the registry
-///
-/// [`Handler<Image>`]: crate::Handler
-/// [`Format::new`]: crate::Format::new
-#[macro_export]
-macro_rules! impl_image_handler {
-    (
-        handler = $handler:ident,
-        loader = $loader:ident,
-        format_id = $format_id:literal,
-        extensions = [$($ext:literal),+ $(,)?],
-        content_types = [$($mime:literal),+ $(,)?],
-        image_format = $img_fmt:expr,
-        origin = $origin:literal,
-        encode_span = $encode_name:literal $(,)?
-    ) => {
-        pub const FORMAT_ID: $crate::FormatId =
-            $crate::FormatId::from_static($format_id);
-
-        /// [`Format`] descriptor registered into
-        /// [`crate::CodecRegistry`].
-        ///
-        /// [`Format`]: $crate::Format
-        pub fn format() -> $crate::Format {
-            $crate::Format::new::<::nvisy_core::modality::Image, _>(
-                FORMAT_ID.clone(),
-                $loader::default(),
-            )
-            .with_extensions([$($ext),+])
-            .with_content_types([$($mime),+])
-        }
-
-        #[::async_trait::async_trait]
-        impl $crate::Handler<::nvisy_core::modality::Image> for $handler {
-            fn format(&self) -> $crate::FormatId {
-                FORMAT_ID.clone()
-            }
-
-            fn source(&self) -> $crate::content::ContentSource {
-                self.source
-            }
-
-            #[::tracing::instrument(name = $encode_name, skip_all, fields(output_bytes))]
-            fn encode(
-                &self,
-            ) -> ::std::result::Result<$crate::content::ContentData, ::nvisy_core::Error>
-            {
-                let out = $crate::handler::image::macros::encode_image(&self.image, $img_fmt, $origin)?;
-                ::tracing::Span::current().record("output_bytes", out.len());
-                let source = $crate::content::ContentSource::new()
-                    .with_parent(&self.source);
-                Ok($crate::content::ContentData::new(source, out.into()))
-            }
-
-            async fn next_chunk(
-                &mut self,
-            ) -> ::std::result::Result<
-                ::std::option::Option<
-                    $crate::Chunk<::nvisy_core::modality::Image>,
-                >,
-                ::nvisy_core::Error,
-            > {
-                if self.yielded {
-                    return Ok(None);
-                }
-                let (w, h) = (self.image.width(), self.image.height());
-                let location = ::nvisy_core::modality::ImageLocation {
-                    bounding_box: ::nvisy_core::primitive::BoundingBox {
-                        x: 0.0,
-                        y: 0.0,
-                        width: w as f64,
-                        height: h as f64,
-                    },
-                    polygon: None,
-                    image_id: None,
-                    page_number: None,
-                };
-                let bytes = $crate::handler::image::macros::encode_image(
-                    &self.image,
-                    $img_fmt,
-                    $origin,
-                )?;
-                let data = ::nvisy_core::modality::ImageData::new(
-                    bytes,
-                    ::nvisy_core::primitive::Dimensions::new(w, h),
-                );
-                self.yielded = true;
-                Ok(Some($crate::Chunk {
-                    location,
-                    data,
-                    hints: ::std::vec::Vec::new(),
-                }))
-            }
-
-            async fn read(
-                &self,
-                location: &::nvisy_core::modality::ImageLocation,
-            ) -> ::std::result::Result<
-                ::std::option::Option<::nvisy_core::modality::ImageData>,
-                ::nvisy_core::Error,
-            > {
-                let bb = &location.bounding_box;
-                let x = bb.x.max(0.0) as u32;
-                let y = bb.y.max(0.0) as u32;
-                let w = (bb.width as u32)
-                    .min(self.image.width().saturating_sub(x));
-                let h = (bb.height as u32)
-                    .min(self.image.height().saturating_sub(y));
-                if w == 0 || h == 0 {
-                    return Ok(None);
-                }
-                let cropped = self.image.crop_imm(x, y, w, h);
-                let bytes = $crate::handler::image::macros::encode_image(
-                    &cropped,
-                    $img_fmt,
-                    $origin,
-                )?;
-                Ok(Some(::nvisy_core::modality::ImageData::new(
-                    bytes,
-                    ::nvisy_core::primitive::Dimensions::new(w, h),
-                )))
-            }
-
-            async fn redact(
-                &mut self,
-                redactions: ::nvisy_core::redaction::Redactions<
-                    ::nvisy_core::modality::Image,
-                >,
-            ) -> ::std::result::Result<(), ::nvisy_core::Error> {
-                for (location, replacement) in redactions.into_items() {
-                    $crate::handler::image::redact::apply(
-                        &mut self.image,
-                        &replacement,
-                        location.bounding_box,
-                    );
-                }
-                Ok(())
-            }
-        }
-
-        impl $handler {
-            /// Create a handler from an already-decoded image.
-            pub fn new(image: ::image::DynamicImage) -> Self {
-                Self {
-                    source: $crate::content::ContentSource::new(),
-                    image,
-                    yielded: false,
-                }
-            }
-
-            /// Attach a content source for lineage tracking.
-            pub fn with_source(
-                mut self,
-                source: $crate::content::ContentSource,
-            ) -> Self {
-                self.source = source;
-                self
-            }
-
-            /// Reference to the decoded image.
-            pub fn image(&self) -> &::image::DynamicImage {
-                &self.image
-            }
-
-            /// Rewind the streaming cursor so [`next_chunk`] yields
-            /// the full-image chunk again.
-            ///
-            /// [`next_chunk`]: $crate::Handler::next_chunk
-            pub fn rewind(&mut self) {
-                self.yielded = false;
-            }
-        }
-    };
-}
-
-/// Encode a [`DynamicImage`] into bytes via the given [`ImageFormat`].
-pub(crate) fn encode_image(
-    img: &::image::DynamicImage,
-    fmt: ::image::ImageFormat,
-    origin: &'static str,
-) -> ::std::result::Result<::std::vec::Vec<u8>, ::nvisy_core::Error> {
-    use ::std::io::Cursor;
-    let mut buf = Cursor::new(Vec::new());
-    img.write_to(&mut buf, fmt)
-        .map_err(|e| ::nvisy_core::Error::validation(format!("encode failed: {e}"), origin))?;
-    Ok(buf.into_inner())
-}
diff --git a/crates/nvisy-codec/src/handler/image/mod.rs b/crates/nvisy-codec/src/handler/image/mod.rs
deleted file mode 100644
index a2496f219..000000000
--- a/crates/nvisy-codec/src/handler/image/mod.rs
+++ /dev/null
@@ -1,39 +0,0 @@
-//! Image modality: concrete image format implementations (PNG,
-//! JPEG, TIFF) and pixel-decode helpers.
-//!
-//! Replacements written during [`Handler::redact`] use
-//! [`ImageReplacement`].
-//!
-//! [`ImageReplacement`]: nvisy_core::redaction::ImageReplacement
-//! [`Handler::redact`]: crate::Handler::redact
-
-#[macro_use]
-pub(crate) mod macros;
-mod image_ops;
-pub(crate) mod redact;
-
-#[cfg(feature = "jpeg")]
-mod jpeg_handler;
-#[cfg(feature = "jpeg")]
-mod jpeg_loader;
-#[cfg(feature = "png")]
-mod png_handler;
-#[cfg(feature = "png")]
-mod png_loader;
-#[cfg(feature = "tiff")]
-mod tiff_handler;
-#[cfg(feature = "tiff")]
-mod tiff_loader;
-
-#[cfg(feature = "jpeg")]
-pub use self::jpeg_handler::{JpegHandler, format as jpeg_format};
-#[cfg(feature = "jpeg")]
-pub use self::jpeg_loader::JpegLoader;
-#[cfg(feature = "png")]
-pub use self::png_handler::{PngHandler, format as png_format};
-#[cfg(feature = "png")]
-pub use self::png_loader::PngLoader;
-#[cfg(feature = "tiff")]
-pub use self::tiff_handler::{TiffHandler, format as tiff_format};
-#[cfg(feature = "tiff")]
-pub use self::tiff_loader::TiffLoader;
diff --git a/crates/nvisy-codec/src/handler/image/png_handler.rs b/crates/nvisy-codec/src/handler/image/png_handler.rs
deleted file mode 100644
index 7edb61acc..000000000
--- a/crates/nvisy-codec/src/handler/image/png_handler.rs
+++ /dev/null
@@ -1,33 +0,0 @@
-//! PNG handler: holds a decoded image and exposes it as a single
-//! full-image chunk via [`Handler<Image>`], with random-access region
-//! reads / pixel redactions via [`Handler<Image>`].
-//!
-//! [`Handler<Image>`]: crate::Handler
-//! [`Handler<Image>`]: crate::Handler
-
-use super::PngLoader;
-use crate::content::ContentSource;
-
-/// Handler for loaded PNG content. Stores the decoded
-/// [`DynamicImage`] directly; raw PNG bytes are produced on
-/// demand by [`Handler::encode`].
-///
-/// [`DynamicImage`]: image::DynamicImage
-/// [`Handler::encode`]: crate::Handler::encode
-#[derive(Debug)]
-pub struct PngHandler {
-    source: ContentSource,
-    image: image::DynamicImage,
-    yielded: bool,
-}
-
-impl_image_handler!(
-    handler = PngHandler,
-    loader = PngLoader,
-    format_id = "nvisy.image.png",
-    extensions = ["png"],
-    content_types = ["image/png"],
-    image_format = image::ImageFormat::Png,
-    origin = "png-handler",
-    encode_span = "png.encode",
-);
diff --git a/crates/nvisy-codec/src/handler/image/png_loader.rs b/crates/nvisy-codec/src/handler/image/png_loader.rs
deleted file mode 100644
index df117b190..000000000
--- a/crates/nvisy-codec/src/handler/image/png_loader.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! PNG loader: validates and decodes raw PNG bytes into a
-//! [`PngHandler`].
-
-use nvisy_core::Error;
-use nvisy_core::modality::Image;
-
-use super::PngHandler;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource};
-
-/// Loader for PNG files. Produces one [`PngHandler`] per input.
-#[derive(Debug, Default)]
-pub struct PngLoader;
-
-#[async_trait::async_trait]
-impl Loader<Image> for PngLoader {
-    type Handler = PngHandler;
-
-    #[tracing::instrument(name = "png.decode", skip_all, fields(input_bytes, width, height))]
-    async fn decode(&self, content: ContentData) -> Result<PngHandler, Error> {
-        tracing::Span::current().record("input_bytes", content.to_bytes().len());
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        let image = image::load_from_memory(&raw)
-            .map_err(|e| Error::validation(format!("PNG decode failed: {e}"), "png-loader"))?;
-        tracing::Span::current().record("width", image.width());
-        tracing::Span::current().record("height", image.height());
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(PngHandler::new(image).with_source(source))
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/image/redact.rs b/crates/nvisy-codec/src/handler/image/redact.rs
deleted file mode 100644
index 848699138..000000000
--- a/crates/nvisy-codec/src/handler/image/redact.rs
+++ /dev/null
@@ -1,56 +0,0 @@
-//! Per-region image redaction helper shared by every image handler
-//! (PNG, JPEG, TIFF — all reduce to single-image `DynamicImage`s).
-//!
-//! Applies one [`ImageReplacement`] in place against a bounding box.
-//! The per-output pixel transforms (blur, block, pixelate) live on
-//! the [`ImageOps`] trait in [`super::image_ops`].
-
-use image::DynamicImage;
-use image::imageops::FilterType;
-use nvisy_core::primitive::BoundingBox;
-use nvisy_core::redaction::ImageReplacement;
-
-use super::image_ops::ImageOps;
-
-const TARGET: &str = "nvisy_codec::handler::image::redact";
-
-/// Apply a single replacement to `img` in place at the given bounding
-/// box.
-///
-/// `Replace` variants whose embedded image data fails to decode are
-/// skipped with a warning.
-pub(crate) fn apply(
-    img: &mut DynamicImage,
-    replacement: &ImageReplacement,
-    bounding_box: BoundingBox,
-) {
-    let region = bounding_box.to_pixel();
-    match replacement {
-        ImageReplacement::Blur { sigma } => {
-            img.apply_gaussian_blur(&region, *sigma);
-        }
-        ImageReplacement::Block { color } => {
-            img.apply_block_overlay(&region, *color);
-        }
-        ImageReplacement::Pixelate { block_size } => {
-            img.apply_pixelate(&region, *block_size);
-        }
-        ImageReplacement::Replace { data } => {
-            let replacement_img = match image::load_from_memory(data) {
-                Ok(r) => r,
-                Err(e) => {
-                    tracing::warn!(
-                        target: TARGET,
-                        region = ?region,
-                        error = %e,
-                        "failed to decode replacement image data, skipping region"
-                    );
-                    return;
-                }
-            };
-            let resized =
-                replacement_img.resize_exact(region.width, region.height, FilterType::Lanczos3);
-            image::imageops::overlay(img, &resized, region.x as i64, region.y as i64);
-        }
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/image/tiff_handler.rs b/crates/nvisy-codec/src/handler/image/tiff_handler.rs
deleted file mode 100644
index 702e32045..000000000
--- a/crates/nvisy-codec/src/handler/image/tiff_handler.rs
+++ /dev/null
@@ -1,25 +0,0 @@
-//! TIFF handler: see [`super::png_handler`] for the shared shape.
-
-use super::TiffLoader;
-use crate::content::ContentSource;
-
-/// Handler for loaded TIFF content. Holds the decoded
-/// [`image::DynamicImage`] and a streaming cursor; the macro-generated
-/// `Handler<Image>` impl encodes back to TIFF on demand.
-#[derive(Debug)]
-pub struct TiffHandler {
-    source: ContentSource,
-    image: image::DynamicImage,
-    yielded: bool,
-}
-
-impl_image_handler!(
-    handler = TiffHandler,
-    loader = TiffLoader,
-    format_id = "nvisy.image.tiff",
-    extensions = ["tif", "tiff"],
-    content_types = ["image/tiff"],
-    image_format = image::ImageFormat::Tiff,
-    origin = "tiff-handler",
-    encode_span = "tiff.encode",
-);
diff --git a/crates/nvisy-codec/src/handler/image/tiff_loader.rs b/crates/nvisy-codec/src/handler/image/tiff_loader.rs
deleted file mode 100644
index 20a01bf64..000000000
--- a/crates/nvisy-codec/src/handler/image/tiff_loader.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! TIFF loader: validates and decodes raw TIFF bytes into a
-//! [`TiffHandler`].
-
-use nvisy_core::Error;
-use nvisy_core::modality::Image;
-
-use super::TiffHandler;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource};
-
-/// Loader for TIFF files. Produces one [`TiffHandler`] per input.
-#[derive(Debug, Default)]
-pub struct TiffLoader;
-
-#[async_trait::async_trait]
-impl Loader<Image> for TiffLoader {
-    type Handler = TiffHandler;
-
-    #[tracing::instrument(name = "tiff.decode", skip_all, fields(input_bytes, width, height))]
-    async fn decode(&self, content: ContentData) -> Result<TiffHandler, Error> {
-        tracing::Span::current().record("input_bytes", content.to_bytes().len());
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        let image = image::load_from_memory(&raw)
-            .map_err(|e| Error::validation(format!("TIFF decode failed: {e}"), "tiff-loader"))?;
-        tracing::Span::current().record("width", image.width());
-        tracing::Span::current().record("height", image.height());
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(TiffHandler::new(image).with_source(source))
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/mod.rs b/crates/nvisy-codec/src/handler/mod.rs
deleted file mode 100644
index 96cd53751..000000000
--- a/crates/nvisy-codec/src/handler/mod.rs
+++ /dev/null
@@ -1,18 +0,0 @@
-//! Concrete per-modality format implementations.
-//!
-//! Each modality module owns the concrete handlers + loaders for
-//! every format that produces that modality (TXT/JSON/HTML for
-//! text, PNG/JPEG/TIFF for image, …). Rich-document formats (PDF,
-//! DOCX) live under [`rich`] because they produce text handles
-//! with embedded image children.
-
-#[cfg(feature = "internal_audio")]
-pub mod audio;
-#[cfg(feature = "internal_image")]
-pub mod image;
-#[cfg(feature = "internal_rich")]
-pub mod rich;
-#[cfg(feature = "internal_tabular")]
-pub mod tabular;
-#[cfg(feature = "internal_text")]
-pub mod text;
diff --git a/crates/nvisy-codec/src/handler/rich/docx_handler.rs b/crates/nvisy-codec/src/handler/rich/docx_handler.rs
deleted file mode 100644
index ad0e3d464..000000000
--- a/crates/nvisy-codec/src/handler/rich/docx_handler.rs
+++ /dev/null
@@ -1,99 +0,0 @@
-//! DOCX handler (stub: text extraction awaiting implementation).
-//!
-//! Preserves the raw bytes for round-trip encoding and exposes no
-//! text chunks until per-paragraph extraction lands.
-
-use std::ops::Range;
-
-use bytes::Bytes;
-use nvisy_core::Error;
-use nvisy_core::modality::{Text, TextData, TextLocation};
-use nvisy_core::redaction::Redactions;
-
-use super::DocxLoader;
-use crate::content::{ContentData, ContentSource};
-use crate::{Chunk, Format, FormatId, Handler};
-
-const TARGET: &str = "nvisy_codec::handler::rich::docx";
-
-/// Stable [`FormatId`] for the DOCX codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.rich.docx");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Text, _>(FORMAT_ID.clone(), DocxLoader)
-        .with_extensions(["docx"])
-        .with_content_types([
-            "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
-        ])
-}
-
-/// Handler for loaded DOCX content. Owns the raw document bytes;
-/// chunking and read/redact operate on the parsed paragraph stream.
-#[derive(Debug)]
-pub struct DocxHandler {
-    source: ContentSource,
-    raw: Bytes,
-}
-
-impl DocxHandler {
-    /// Build a handler from raw DOCX bytes; the parsed-paragraph
-    /// stream is reconstructed lazily inside the `Handler<Text>` impl.
-    pub fn new(raw: impl Into<Bytes>) -> Self {
-        Self {
-            source: ContentSource::new(),
-            raw: raw.into(),
-        }
-    }
-
-    /// Attach a content source for lineage tracking.
-    pub fn with_source(mut self, source: ContentSource) -> Self {
-        self.source = source;
-        self
-    }
-
-    /// Borrow the raw DOCX bytes.
-    pub fn raw(&self) -> &[u8] {
-        &self.raw
-    }
-}
-
-#[async_trait::async_trait]
-impl Handler<Text> for DocxHandler {
-    fn format(&self) -> FormatId {
-        FORMAT_ID.clone()
-    }
-
-    fn source(&self) -> ContentSource {
-        self.source
-    }
-
-    #[tracing::instrument(name = "docx.encode", skip_all, fields(output_bytes))]
-    fn encode(&self) -> Result<ContentData, Error> {
-        tracing::Span::current().record("output_bytes", self.raw.len());
-        let source = ContentSource::new().with_parent(&self.source);
-        Ok(ContentData::new(source, self.raw.clone()))
-    }
-
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<Text>>, Error> {
-        Ok(None)
-    }
-
-    fn lift_chunk(&self, chunk: &Chunk<Text>, value_range: Range<usize>) -> Option<TextLocation> {
-        chunk.location.subslice(value_range)
-    }
-
-    async fn read(&self, _location: &TextLocation) -> Result<Option<TextData>, Error> {
-        Ok(None)
-    }
-
-    async fn redact(&mut self, redactions: Redactions<Text>) -> Result<(), Error> {
-        if redactions.is_empty() {
-            return Ok(());
-        }
-        Err(Error::validation(
-            "DOCX redaction is not yet supported",
-            TARGET,
-        ))
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/rich/docx_loader.rs b/crates/nvisy-codec/src/handler/rich/docx_loader.rs
deleted file mode 100644
index bf296e609..000000000
--- a/crates/nvisy-codec/src/handler/rich/docx_loader.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-//! DOCX loader (stub: awaiting text extraction implementation).
-//!
-//! Wraps the raw bytes into a [`DocxHandler`] that preserves the
-//! input for round-trip encoding but exposes no text chunks.
-
-use nvisy_core::Error;
-use nvisy_core::modality::Text;
-
-use super::DocxHandler;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource};
-
-/// Loader that wraps raw DOCX bytes.
-#[derive(Debug, Default)]
-pub struct DocxLoader;
-
-#[async_trait::async_trait]
-impl Loader<Text> for DocxLoader {
-    type Handler = DocxHandler;
-
-    #[tracing::instrument(name = "docx.decode", skip_all, fields(input_bytes))]
-    async fn decode(&self, content: ContentData) -> Result<DocxHandler, Error> {
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        tracing::Span::current().record("input_bytes", raw.len());
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(DocxHandler::new(raw).with_source(source))
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/rich/mod.rs b/crates/nvisy-codec/src/handler/rich/mod.rs
deleted file mode 100644
index 6353583ea..000000000
--- a/crates/nvisy-codec/src/handler/rich/mod.rs
+++ /dev/null
@@ -1,21 +0,0 @@
-//! Rich-document format implementations: PDF, DOCX.
-
-#[cfg(feature = "docx")]
-mod docx_handler;
-#[cfg(feature = "docx")]
-mod docx_loader;
-#[cfg(feature = "pdf")]
-mod pdf_handler;
-#[cfg(feature = "pdf")]
-mod pdf_loader;
-#[cfg(feature = "pdf")]
-mod pdf_render;
-
-#[cfg(feature = "docx")]
-pub use self::docx_handler::{DocxHandler, format as docx_format};
-#[cfg(feature = "docx")]
-pub use self::docx_loader::DocxLoader;
-#[cfg(feature = "pdf")]
-pub use self::pdf_handler::{PdfHandler, format as pdf_format};
-#[cfg(feature = "pdf")]
-pub use self::pdf_loader::PdfLoader;
diff --git a/crates/nvisy-codec/src/handler/rich/pdf_handler.rs b/crates/nvisy-codec/src/handler/rich/pdf_handler.rs
deleted file mode 100644
index f67c283d0..000000000
--- a/crates/nvisy-codec/src/handler/rich/pdf_handler.rs
+++ /dev/null
@@ -1,339 +0,0 @@
-//! PDF handler: holds per-page extracted text and raw document bytes,
-//! exposing per-page chunks via [`Handler<Text>`].
-//!
-//! Text offsets are cumulative over the per-page text sequence in
-//! document order. Each chunk carries the page number on its
-//! [`TextLocation`] for downstream provenance.
-//!
-//! Embedded image extraction (figures, page rasterization for OCR)
-//! lives on inherent methods rather than `Handler<Image>` — see
-//! [`render_pages`] and [`extract_embedded_images`]. These
-//! return [`DocumentHandle<Image>`] values backed by PNG bytes so
-//! downstream extractors can route them through the standard image
-//! pipeline.
-//!
-//! [`Handler<Text>`]: crate::Handler
-//! [`render_pages`]: PdfHandler::render_pages
-//! [`extract_embedded_images`]: PdfHandler::extract_embedded_images
-
-use std::ops::Range;
-
-use bytes::Bytes;
-use nvisy_core::Error;
-use nvisy_core::modality::{Image, Text, TextData, TextLocation};
-use nvisy_core::primitive::Dpi;
-use nvisy_core::redaction::{Redactions, TextReplacement};
-
-use super::PdfLoader;
-use super::pdf_render::PdfRenderer;
-use crate::content::{ContentData, ContentSource};
-use crate::handler::image::PngHandler;
-use crate::handler::text::redact;
-use crate::{Chunk, DocumentHandle, Format, FormatId, Handler};
-
-const TARGET: &str = "nvisy_codec::handler::rich::pdf";
-
-/// Stable [`FormatId`] for the PDF codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.rich.pdf");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Text, _>(FORMAT_ID.clone(), PdfLoader::default())
-        .with_extensions(["pdf"])
-        .with_content_types(["application/pdf"])
-}
-
-/// Handler for loaded PDF content.
-///
-/// `page_starts` is a cumulative-offset index over the per-page text
-/// sequence: `page_starts[i]` is the byte position of page `i+1`'s
-/// text. Maintained on every redaction; resolves `(byte offset →
-/// page index)` in `O(log N)`.
-#[derive(Debug)]
-pub struct PdfHandler {
-    source: ContentSource,
-    pages: Vec<String>,
-    page_starts: Vec<usize>,
-    raw: Bytes,
-    cursor: usize,
-}
-
-impl PdfHandler {
-    /// Create a handler from per-page text and raw PDF bytes.
-    pub fn new(pages: Vec<String>, raw: impl Into<Bytes>) -> Self {
-        let page_starts = compute_page_starts(&pages);
-        Self {
-            source: ContentSource::new(),
-            pages,
-            page_starts,
-            raw: raw.into(),
-            cursor: 0,
-        }
-    }
-
-    /// Attach a content source for lineage tracking.
-    pub fn with_source(mut self, source: ContentSource) -> Self {
-        self.source = source;
-        self
-    }
-
-    /// Parse raw PDF bytes, extract per-page text, and return a new handler.
-    pub fn from_pdf(raw: impl Into<Bytes>, password: Option<&str>) -> Result<Self, Error> {
-        let raw: Bytes = raw.into();
-        let mut doc = lopdf::Document::load_mem(&raw).map_err(|e| {
-            Error::runtime(
-                format!("failed to extract text from PDF: {e}"),
-                TARGET,
-                false,
-            )
-        })?;
-        if doc.is_encrypted() {
-            doc.decrypt(password.unwrap_or("")).map_err(|e| {
-                Error::runtime(
-                    format!("failed to extract text from PDF: {e}"),
-                    TARGET,
-                    false,
-                )
-            })?;
-        }
-        let page_count = doc.get_pages().len();
-        let mut pages = Vec::with_capacity(page_count);
-        for page_num in 1..=(page_count as u32) {
-            let chunks = doc.extract_text_chunks(&[page_num]);
-            let text: String = chunks.into_iter().filter_map(|r| r.ok()).collect();
-            pages.push(text);
-        }
-        Ok(Self::new(pages, raw))
-    }
-
-    /// The raw PDF bytes.
-    pub fn raw(&self) -> &[u8] {
-        &self.raw
-    }
-
-    /// Number of pages.
-    pub fn len(&self) -> usize {
-        self.pages.len()
-    }
-
-    /// Whether the document has no pages.
-    pub fn is_empty(&self) -> bool {
-        self.pages.is_empty()
-    }
-
-    /// Rewind the streaming cursor to the first page.
-    pub fn rewind(&mut self) {
-        self.cursor = 0;
-    }
-
-    /// Render all pages as PNG-backed [`DocumentHandle<Image>`]s at
-    /// the given DPI. Used by the OCR fallback when text extraction
-    /// yields nothing.
-    pub fn render_pages(&self, dpi: Dpi) -> Result<Vec<DocumentHandle<Image>>, Error> {
-        let images = PdfRenderer::parallel_render(&self.raw, dpi)?;
-        Ok(images.into_iter().map(wrap_as_image_handle).collect())
-    }
-
-    /// Extract embedded image objects (figures, photos) from the PDF
-    /// content streams. Returns each as a PNG-backed
-    /// [`DocumentHandle<Image>`].
-    pub fn extract_embedded_images(&self) -> Result<Vec<DocumentHandle<Image>>, Error> {
-        let images = PdfRenderer::extract_images(&self.raw)?;
-        Ok(images.into_iter().map(wrap_as_image_handle).collect())
-    }
-
-    fn page_for(&self, byte_offset: usize) -> Option<usize> {
-        match self.page_starts.binary_search(&byte_offset) {
-            Ok(i) if i < self.pages.len() => Some(i),
-            Ok(_) => None,
-            Err(i) if i > 0 && i <= self.pages.len() => Some(i - 1),
-            _ => None,
-        }
-    }
-
-    fn shift_starts_after(&mut self, i: usize, delta: isize) {
-        if delta == 0 {
-            return;
-        }
-        for s in &mut self.page_starts[i + 1..] {
-            *s = (*s as isize + delta) as usize;
-        }
-    }
-}
-
-#[async_trait::async_trait]
-impl Handler<Text> for PdfHandler {
-    fn format(&self) -> FormatId {
-        FORMAT_ID.clone()
-    }
-
-    fn source(&self) -> ContentSource {
-        self.source
-    }
-
-    #[tracing::instrument(name = "pdf.encode", skip_all, fields(output_bytes))]
-    fn encode(&self) -> Result<ContentData, Error> {
-        tracing::Span::current().record("output_bytes", self.raw.len());
-        let source = ContentSource::new().with_parent(&self.source);
-        Ok(ContentData::new(source, self.raw.clone()))
-    }
-
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<Text>>, Error> {
-        if self.cursor >= self.pages.len() {
-            return Ok(None);
-        }
-        let i = self.cursor;
-        let start = self.page_starts[i];
-        let end = self.page_starts[i + 1];
-        let text = &self.pages[i];
-        self.cursor += 1;
-        Ok(Some(Chunk {
-            location: TextLocation {
-                start,
-                end,
-                page_number: Some((i + 1) as u32),
-                ..Default::default()
-            },
-            data: TextData::from(text.as_str()),
-            hints: Vec::new(),
-        }))
-    }
-
-    fn lift_chunk(&self, chunk: &Chunk<Text>, value_range: Range<usize>) -> Option<TextLocation> {
-        chunk.location.subslice(value_range)
-    }
-
-    async fn read(&self, location: &TextLocation) -> Result<Option<TextData>, Error> {
-        let Some(i) = self.page_for(location.start) else {
-            return Ok(None);
-        };
-        let page_start = self.page_starts[i];
-        let page_end = self.page_starts[i + 1];
-        if location.end > page_end {
-            return Ok(None);
-        }
-        let local_start = location.start - page_start;
-        let local_end = location.end - page_start;
-        Ok(self.pages[i]
-            .get(local_start..local_end)
-            .map(TextData::from))
-    }
-
-    async fn redact(&mut self, mut redactions: Redactions<Text>) -> Result<(), Error> {
-        // Right-to-left so each page's length delta doesn't invalidate
-        // earlier byte offsets.
-        redactions.sort_descending();
-        for (location, replacement) in redactions.into_items() {
-            self.redact_one(&location, replacement)?;
-        }
-        Ok(())
-    }
-}
-
-impl PdfHandler {
-    fn redact_one(
-        &mut self,
-        location: &TextLocation,
-        replacement: TextReplacement,
-    ) -> Result<(), Error> {
-        let Some(i) = self.page_for(location.start) else {
-            return Ok(());
-        };
-        let page_start = self.page_starts[i];
-        let page_end = self.page_starts[i + 1];
-        if location.end > page_end {
-            return Ok(());
-        }
-        let local_start = location.start - page_start;
-        let local_end = location.end - page_start;
-
-        let mut content = self.pages[i].clone();
-        let value = replacement.replacement_value().unwrap_or_default();
-        let before_len = self.pages[i].len();
-        redact::replace_range(&mut content, value, local_start..local_end, TARGET)?;
-
-        // Bake the edit into the raw PDF content stream so the encoded
-        // bytes round-trip with the redaction.
-        let old_text = &self.pages[i];
-        if !old_text.is_empty() && old_text != &content {
-            let mut doc = lopdf::Document::load_mem(&self.raw).map_err(|e| {
-                Error::runtime(
-                    format!("failed to load PDF for editing: {e}"),
-                    TARGET,
-                    false,
-                )
-            })?;
-            let _ = doc.replace_text((i as u32) + 1, old_text, &content, None);
-            let mut buf = Vec::new();
-            doc.save_to(&mut buf).map_err(|e| {
-                Error::runtime(format!("failed to save edited PDF: {e}"), TARGET, false)
-            })?;
-            self.raw = Bytes::from(buf);
-        }
-
-        self.pages[i] = content;
-        let delta = self.pages[i].len() as isize - before_len as isize;
-        self.shift_starts_after(i, delta);
-        Ok(())
-    }
-}
-
-fn compute_page_starts(pages: &[String]) -> Vec<usize> {
-    let mut starts = Vec::with_capacity(pages.len() + 1);
-    let mut offset = 0usize;
-    for page in pages {
-        starts.push(offset);
-        offset += page.len();
-    }
-    starts.push(offset);
-    starts
-}
-
-/// Wrap a decoded [`image::DynamicImage`] into a PNG-backed
-/// [`DocumentHandle<Image>`].
-fn wrap_as_image_handle(img: image::DynamicImage) -> DocumentHandle<Image> {
-    let handler = PngHandler::new(img);
-    let format = handler.format();
-    DocumentHandle::new(format, Box::new(handler))
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn handler(pages: &[&str]) -> PdfHandler {
-        PdfHandler::new(pages.iter().map(|s| s.to_string()).collect(), Vec::new())
-    }
-
-    #[tokio::test]
-    async fn stream_yields_one_chunk_per_page() -> Result<(), Error> {
-        let mut h = handler(&["page one", "page two", "page three"]);
-        let mut pages = Vec::new();
-        while let Some(chunk) = h.next_chunk().await? {
-            pages.push(chunk.location.page_number);
-        }
-        assert_eq!(pages, vec![Some(1), Some(2), Some(3)]);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn read_returns_page_text() -> Result<(), Error> {
-        let mut h = handler(&["page one", "page two"]);
-        let chunk = h.next_chunk().await?.unwrap();
-        let data = h.read(&chunk.location).await?.unwrap();
-        assert_eq!(data.as_str(), "page one");
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn read_cross_page_returns_none() -> Result<(), Error> {
-        let h = handler(&["page one", "page two"]);
-        let bogus = TextLocation {
-            start: 5,
-            end: 14,
-            ..Default::default()
-        };
-        assert!(h.read(&bogus).await?.is_none());
-        Ok(())
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/rich/pdf_loader.rs b/crates/nvisy-codec/src/handler/rich/pdf_loader.rs
deleted file mode 100644
index cf2189b37..000000000
--- a/crates/nvisy-codec/src/handler/rich/pdf_loader.rs
+++ /dev/null
@@ -1,96 +0,0 @@
-//! PDF loader: parses raw PDF content into a [`PdfHandler`]. Text is
-//! extracted per page via [`lopdf`]; the raw bytes are preserved for
-//! encoding and rendering.
-//!
-//! [`lopdf`]: https://docs.rs/lopdf
-
-use nvisy_core::Error;
-use nvisy_core::modality::Text;
-
-use super::PdfHandler;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource};
-
-/// Loader for PDF files. Produces one [`PdfHandler`] per input.
-#[derive(Debug, Default)]
-pub struct PdfLoader {
-    /// Optional password for encrypted PDFs.
-    pub password: Option<String>,
-}
-
-#[async_trait::async_trait]
-impl Loader<Text> for PdfLoader {
-    type Handler = PdfHandler;
-
-    #[tracing::instrument(name = "pdf.decode", skip_all, fields(input_bytes))]
-    async fn decode(&self, content: ContentData) -> Result<PdfHandler, Error> {
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        tracing::Span::current().record("input_bytes", raw.len());
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(PdfHandler::from_pdf(raw, self.password.as_deref())?.with_source(source))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use bytes::Bytes;
-    use lopdf::{Dictionary, Document, Object, Stream, dictionary};
-
-    use super::*;
-    use crate::Handler;
-    use crate::content::ContentSource;
-
-    fn content_from_bytes(bytes: &[u8]) -> ContentData {
-        ContentData::new(ContentSource::new(), Bytes::from(bytes.to_vec()))
-    }
-
-    fn minimal_pdf() -> Vec<u8> {
-        let mut doc = Document::with_version("1.5");
-        let pages_id = doc.new_object_id();
-        let page_id = doc.new_object_id();
-        let content_id = doc.add_object(Stream::new(Dictionary::new(), Vec::new()));
-        doc.set_object(
-            page_id,
-            Object::Dictionary(dictionary! {
-                "Type" => "Page",
-                "Parent" => pages_id,
-                "MediaBox" => vec![0.into(), 0.into(), 612.into(), 792.into()],
-                "Contents" => content_id,
-            }),
-        );
-        doc.set_object(
-            pages_id,
-            Object::Dictionary(dictionary! {
-                "Type" => "Pages",
-                "Kids" => vec![page_id.into()],
-                "Count" => 1,
-            }),
-        );
-        let catalog_id = doc.add_object(dictionary! {
-            "Type" => "Catalog",
-            "Pages" => pages_id,
-        });
-        doc.trailer.set("Root", catalog_id);
-        let mut buf = Vec::new();
-        doc.save_to(&mut buf).expect("failed to save minimal PDF");
-        buf
-    }
-
-    #[tokio::test]
-    async fn load_invalid_pdf_returns_error() {
-        let content = content_from_bytes(b"not a pdf");
-        let err = PdfLoader::default().decode(content).await.unwrap_err();
-        assert!(err.to_string().contains("failed to extract text from PDF"));
-    }
-
-    #[tokio::test]
-    async fn load_minimal_pdf() {
-        let raw = minimal_pdf();
-        let content = content_from_bytes(&raw);
-        let mut doc = PdfLoader::default().decode(content).await.unwrap();
-        assert_eq!(doc.format().as_str(), "nvisy.rich.pdf");
-        let chunk = doc.next_chunk().await.unwrap().expect("one page");
-        assert!(chunk.data.as_str().trim().is_empty());
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/rich/pdf_render.rs b/crates/nvisy-codec/src/handler/rich/pdf_render.rs
deleted file mode 100644
index d630d52be..000000000
--- a/crates/nvisy-codec/src/handler/rich/pdf_render.rs
+++ /dev/null
@@ -1,152 +0,0 @@
-//! PDF-to-image rendering via PDFium.
-//!
-//! PDFium is not thread-safe, so all rendering is serialised on a
-//! dedicated single-thread [`ThreadPool`]. The [`PdfRenderer`]
-//! binding is created once on first use via a `thread_local!` and
-//! reused for all subsequent calls.
-//!
-//! [`ThreadPool`]: rayon::ThreadPool
-
-use std::cell::RefCell;
-use std::sync::LazyLock;
-
-use image::DynamicImage;
-use nvisy_core::Error;
-use nvisy_core::primitive::Dpi;
-use pdfium_render::prelude::*;
-
-/// Dedicated single-thread pool for PDFium operations.
-static PDF_POOL: LazyLock<rayon::ThreadPool> = LazyLock::new(|| {
-    rayon::ThreadPoolBuilder::new()
-        .num_threads(1)
-        .thread_name(|_| "pdfium".into())
-        .build()
-        .expect("failed to create PDFium thread pool")
-});
-
-thread_local! {
-    static RENDERER: RefCell<Option<PdfRenderer>> = const { RefCell::new(None) };
-}
-
-/// Renders PDF pages to images for OCR processing.
-///
-/// Binding to the PDFium shared library is expensive, so the renderer
-/// is lazily initialised on a dedicated thread and reused across calls.
-/// Requires the PDFium shared library to be available at runtime
-/// (bundled in the Docker image or installed on the host).
-pub(super) struct PdfRenderer {
-    pdfium: Pdfium,
-}
-
-impl PdfRenderer {
-    /// Create a new renderer by binding to a system-provided PDFium library.
-    fn new() -> Result<Self, Error> {
-        let bindings = Pdfium::bind_to_system_library()
-            .or_else(|_| Pdfium::bind_to_library("libpdfium"))
-            .map_err(|e| {
-                Error::runtime(
-                    format!("failed to load PDFium library: {e}"),
-                    "pdf_render",
-                    false,
-                )
-            })?;
-        Ok(Self {
-            pdfium: Pdfium::new(bindings),
-        })
-    }
-
-    /// Render all pages of a PDF to images at the given DPI.
-    ///
-    /// Dispatches work to a dedicated single-thread pool where the
-    /// PDFium binding is lazily initialised and reused. A typical DPI
-    /// value for OCR is [`Dpi::OCR`] (300).
-    pub fn parallel_render(pdf_bytes: &[u8], dpi: Dpi) -> Result<Vec<DynamicImage>, Error> {
-        let bytes = pdf_bytes.to_vec();
-
-        PDF_POOL.install(|| {
-            RENDERER.with_borrow_mut(|slot| {
-                if slot.is_none() {
-                    *slot = Some(PdfRenderer::new()?);
-                }
-                slot.as_ref().unwrap().render(&bytes, dpi)
-            })
-        })
-    }
-
-    /// Extract all embedded images from a PDF.
-    ///
-    /// Iterates over all page objects across every page, collecting
-    /// image objects via [`PdfPageImageObject::get_raw_image`]. Each
-    /// image is paired with its 0-based sequential index.
-    pub fn extract_images(pdf_bytes: &[u8]) -> Result<Vec<DynamicImage>, Error> {
-        let bytes = pdf_bytes.to_vec();
-
-        PDF_POOL.install(|| {
-            RENDERER.with_borrow_mut(|slot| {
-                if slot.is_none() {
-                    *slot = Some(PdfRenderer::new()?);
-                }
-                slot.as_ref().unwrap().extract(&bytes)
-            })
-        })
-    }
-
-    /// Extract embedded images using the bound PDFium instance.
-    fn extract(&self, pdf_bytes: &[u8]) -> Result<Vec<DynamicImage>, Error> {
-        let document = self
-            .pdfium
-            .load_pdf_from_byte_slice(pdf_bytes, None)
-            .map_err(|e| Error::runtime(format!("failed to load PDF: {e}"), "pdf_render", false))?;
-
-        let mut images = Vec::new();
-        for page in document.pages().iter() {
-            for object in page.objects().iter() {
-                if let Some(image_object) = object.as_image_object() {
-                    match image_object.get_raw_image() {
-                        Ok(img) => images.push(img),
-                        Err(e) => {
-                            tracing::warn!(
-                                target: "nvisy_codec::pdf_render",
-                                error = %e,
-                                "failed to extract embedded image, skipping",
-                            );
-                        }
-                    }
-                }
-            }
-        }
-
-        Ok(images)
-    }
-
-    /// Render all pages using the bound PDFium instance.
-    fn render(&self, pdf_bytes: &[u8], dpi: Dpi) -> Result<Vec<DynamicImage>, Error> {
-        let document = self
-            .pdfium
-            .load_pdf_from_byte_slice(pdf_bytes, None)
-            .map_err(|e| Error::runtime(format!("failed to load PDF: {e}"), "pdf_render", false))?;
-
-        let config = PdfRenderConfig::new().scale_page_by_factor(dpi.scale_factor());
-
-        let mut images = Vec::new();
-        for page in document.pages().iter() {
-            let bitmap = page.render_with_config(&config).map_err(|e| {
-                Error::runtime(
-                    format!("failed to render PDF page: {e}"),
-                    "pdf_render",
-                    false,
-                )
-            })?;
-            let image = bitmap.as_image().map_err(|e| {
-                Error::runtime(
-                    format!("failed to convert PDF page bitmap: {e}"),
-                    "pdf_render",
-                    false,
-                )
-            })?;
-            images.push(image);
-        }
-
-        Ok(images)
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/tabular/csv_handler.rs b/crates/nvisy-codec/src/handler/tabular/csv_handler.rs
deleted file mode 100644
index 43f114fd6..000000000
--- a/crates/nvisy-codec/src/handler/tabular/csv_handler.rs
+++ /dev/null
@@ -1,571 +0,0 @@
-//! CSV handler: holds parsed CSV content and streams cell coordinates
-//! via [`Handler<Tabular>`], with random-access reads / redactions via
-//! [`Handler<Tabular>`].
-//!
-//! Cell coordinates are `(row, col)`. Row 0 is the header row (if
-//! present); row 1 is the first data row when headers exist, else row
-//! 0 is the first data row. Intra-cell byte offsets on a
-//! [`TabularLocation`] address sub-strings within a cell value;
-//! omitting them redacts the whole cell.
-
-use std::ops::Range;
-
-use nvisy_core::Error;
-use nvisy_core::modality::{Tabular, TabularLocation, TextData};
-use nvisy_core::redaction::{Redactions, TabularReplacement};
-
-use super::CsvLoader;
-use crate::content::{ContentData, ContentSource};
-use crate::handler::tabular::TabularHandle;
-use crate::handler::text::redact;
-use crate::{Chunk, Format, FormatId, Handler};
-
-const TARGET: &str = "nvisy_codec::handler::tabular::csv";
-
-/// Stable [`FormatId`] for the CSV codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.tabular.csv");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Tabular, _>(FORMAT_ID.clone(), CsvLoader::default())
-        .with_extensions(["csv"])
-        .with_content_types(["text/csv"])
-}
-
-/// Parsed CSV content.
-#[derive(Debug, Clone)]
-pub struct CsvData {
-    /// Column headers, if present.
-    pub headers: Option<Vec<String>>,
-    /// Data rows (excluding the header row).
-    pub rows: Vec<Vec<String>>,
-    /// Field delimiter byte (e.g. `b','`, `b'\t'`, `b';'`).
-    pub delimiter: u8,
-    /// Whether the original source had a trailing newline.
-    pub trailing_newline: bool,
-}
-
-/// Handler for loaded CSV content. Cell coordinates double as the
-/// index — no derived offset table to maintain.
-#[derive(Debug)]
-pub struct CsvHandler {
-    source: ContentSource,
-    data: CsvData,
-    cursor: CsvCursor,
-}
-
-/// Streaming cursor over a CSV: walks header row (if present) then
-/// data rows, one cell at a time in row-major order.
-#[derive(Debug, Default)]
-struct CsvCursor {
-    row: u32,
-    col: u32,
-}
-
-#[async_trait::async_trait]
-impl Handler<Tabular> for CsvHandler {
-    fn format(&self) -> FormatId {
-        FORMAT_ID.clone()
-    }
-
-    fn source(&self) -> ContentSource {
-        self.source
-    }
-
-    #[tracing::instrument(name = "csv.encode", skip_all, fields(output_bytes))]
-    fn encode(&self) -> Result<ContentData, Error> {
-        let bytes = self.serialize_bytes()?;
-        tracing::Span::current().record("output_bytes", bytes.len());
-        let source = ContentSource::new().with_parent(&self.source);
-        Ok(ContentData::new(source, bytes.into()))
-    }
-
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<Tabular>>, Error> {
-        let total_rows = if self.data.headers.is_some() {
-            self.data.rows.len() as u32 + 1
-        } else {
-            self.data.rows.len() as u32
-        };
-
-        if self.cursor.row >= total_rows {
-            return Ok(None);
-        }
-        let row_len = self.row_len(self.cursor.row).unwrap_or(0);
-        if (self.cursor.col as usize) >= row_len {
-            self.cursor.row += 1;
-            self.cursor.col = 0;
-            return Box::pin(self.next_chunk()).await;
-        }
-
-        let row = self.cursor.row;
-        let col = self.cursor.col;
-        let location = TabularLocation {
-            row_index: row,
-            column_index: col,
-            start_offset: None,
-            end_offset: None,
-            column_name: self
-                .data
-                .headers
-                .as_ref()
-                .and_then(|h| h.get(col as usize).cloned()),
-            sheet_name: None,
-        };
-        let cell = self.cell_at(row, col).expect("bounds checked above");
-        let data = TextData::from(cell.to_owned());
-        let hints = location.column_name.iter().cloned().collect();
-
-        self.cursor.col += 1;
-        Ok(Some(Chunk {
-            location,
-            data,
-            hints,
-        }))
-    }
-
-    fn lift_chunk(
-        &self,
-        chunk: &Chunk<Tabular>,
-        value_range: Range<usize>,
-    ) -> Option<TabularLocation> {
-        let cell = self.cell_at(chunk.location.row_index, chunk.location.column_index)?;
-        if value_range.start > value_range.end || value_range.end > cell.len() {
-            return None;
-        }
-        Some(TabularLocation {
-            row_index: chunk.location.row_index,
-            column_index: chunk.location.column_index,
-            start_offset: Some(value_range.start),
-            end_offset: Some(value_range.end),
-            column_name: chunk.location.column_name.clone(),
-            sheet_name: chunk.location.sheet_name.clone(),
-        })
-    }
-
-    async fn read(&self, location: &TabularLocation) -> Result<Option<TextData>, Error> {
-        Ok(self
-            .cell_at(location.row_index, location.column_index)
-            .map(|s| TextData::from(s.to_owned())))
-    }
-
-    async fn redact(&mut self, mut redactions: Redactions<Tabular>) -> Result<(), Error> {
-        // Multiple redactions can target intra-cell byte ranges within
-        // the same cell; apply right-to-left over byte offsets so an
-        // earlier shrink doesn't invalidate later offsets.
-        redactions.sort_descending();
-        for (location, replacement) in redactions.into_items() {
-            self.redact_one(&location, replacement)?;
-        }
-        Ok(())
-    }
-}
-
-impl TabularHandle for CsvHandler {
-    fn has_header(&self) -> bool {
-        self.data.headers.is_some()
-    }
-}
-
-impl CsvHandler {
-    /// Create a new handler from parsed CSV data.
-    pub fn new(data: CsvData) -> Self {
-        Self {
-            source: ContentSource::new(),
-            data,
-            cursor: CsvCursor::default(),
-        }
-    }
-
-    /// Attach a content source for lineage tracking.
-    pub fn with_source(mut self, source: ContentSource) -> Self {
-        self.source = source;
-        self
-    }
-
-    /// Column headers, if present.
-    pub fn headers(&self) -> Option<&[String]> {
-        self.data.headers.as_deref()
-    }
-
-    /// All data rows (excluding the header row).
-    pub fn rows(&self) -> &[Vec<String>] {
-        &self.data.rows
-    }
-
-    /// Mutable access to all data rows.
-    pub fn rows_mut(&mut self) -> &mut Vec<Vec<String>> {
-        &mut self.data.rows
-    }
-
-    /// A specific cell by `(data_row, col)`. `data_row` is 0-based
-    /// against the data rows; the header is *not* data row 0. Use
-    /// [`Handler::read`] with [`TabularLocation`] if you need
-    /// to address the header.
-    pub fn cell(&self, data_row: usize, col: usize) -> Option<&str> {
-        self.data
-            .rows
-            .get(data_row)
-            .and_then(|r| r.get(col))
-            .map(String::as_str)
-    }
-
-    /// Number of data rows (excluding the header).
-    pub fn len(&self) -> usize {
-        self.data.rows.len()
-    }
-
-    /// Whether the document has no data rows.
-    pub fn is_empty(&self) -> bool {
-        self.data.rows.is_empty()
-    }
-
-    /// Detected field delimiter.
-    pub fn delimiter(&self) -> u8 {
-        self.data.delimiter
-    }
-
-    /// Whether the original source had a trailing newline.
-    pub fn trailing_newline(&self) -> bool {
-        self.data.trailing_newline
-    }
-
-    /// Rewind the streaming cursor to the start of the document.
-    pub fn rewind(&mut self) {
-        self.cursor = CsvCursor::default();
-    }
-
-    /// Consume the handler and return the inner [`CsvData`].
-    pub fn into_data(self) -> CsvData {
-        self.data
-    }
-
-    fn cell_at(&self, row_index: u32, column_index: u32) -> Option<&str> {
-        let col = column_index as usize;
-        if self.data.headers.is_some() && row_index == 0 {
-            return self.data.headers.as_ref()?.get(col).map(String::as_str);
-        }
-        let data_row = if self.data.headers.is_some() {
-            (row_index - 1) as usize
-        } else {
-            row_index as usize
-        };
-        self.data.rows.get(data_row)?.get(col).map(String::as_str)
-    }
-
-    fn cell_at_mut(&mut self, row_index: u32, column_index: u32) -> Option<&mut String> {
-        let col = column_index as usize;
-        if self.data.headers.is_some() && row_index == 0 {
-            return self.data.headers.as_mut()?.get_mut(col);
-        }
-        let data_row = if self.data.headers.is_some() {
-            (row_index - 1) as usize
-        } else {
-            row_index as usize
-        };
-        self.data.rows.get_mut(data_row)?.get_mut(col)
-    }
-
-    fn row_len(&self, row_index: u32) -> Option<usize> {
-        if self.data.headers.is_some() && row_index == 0 {
-            return self.data.headers.as_ref().map(|h| h.len());
-        }
-        let data_row = if self.data.headers.is_some() {
-            (row_index - 1) as usize
-        } else {
-            row_index as usize
-        };
-        self.data.rows.get(data_row).map(|r| r.len())
-    }
-
-    fn redact_one(
-        &mut self,
-        location: &TabularLocation,
-        replacement: TabularReplacement,
-    ) -> Result<(), Error> {
-        let Some(cell) = self.cell_at_mut(location.row_index, location.column_index) else {
-            return Ok(());
-        };
-        let start = location.start_offset.unwrap_or(0);
-        let end = location.end_offset.unwrap_or(cell.len());
-        let value = replacement.replacement_value().unwrap_or_default();
-        redact::replace_range(cell, value, start..end, TARGET)
-    }
-
-    fn serialize_bytes(&self) -> Result<Vec<u8>, Error> {
-        let mut wtr = csv::WriterBuilder::new()
-            .delimiter(self.data.delimiter)
-            .has_headers(false)
-            .from_writer(Vec::new());
-
-        if let Some(headers) = &self.data.headers {
-            wtr.write_record(headers)
-                .map_err(|e| Error::validation(format!("CSV encode error: {e}"), TARGET))?;
-        }
-        for row in &self.data.rows {
-            wtr.write_record(row)
-                .map_err(|e| Error::validation(format!("CSV encode error: {e}"), TARGET))?;
-        }
-        let mut bytes = wtr
-            .into_inner()
-            .map_err(|e| Error::validation(format!("CSV encode error: {e}"), TARGET))?;
-        bytes.retain(|&b| b != b'\r');
-        if !self.data.trailing_newline && bytes.last() == Some(&b'\n') {
-            bytes.pop();
-        }
-        Ok(bytes)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::Error;
-
-    use super::*;
-
-    fn handler_with_headers(headers: Vec<&str>, rows: Vec<Vec<&str>>) -> CsvHandler {
-        CsvHandler::new(CsvData {
-            headers: Some(headers.into_iter().map(String::from).collect()),
-            rows: rows
-                .into_iter()
-                .map(|r| r.into_iter().map(String::from).collect())
-                .collect(),
-            delimiter: b',',
-            trailing_newline: true,
-        })
-    }
-
-    fn handler_no_headers(rows: Vec<Vec<&str>>) -> CsvHandler {
-        CsvHandler::new(CsvData {
-            headers: None,
-            rows: rows
-                .into_iter()
-                .map(|r| r.into_iter().map(String::from).collect())
-                .collect(),
-            delimiter: b',',
-            trailing_newline: true,
-        })
-    }
-
-    fn cell_range(row: u32, col: u32, start: usize, end: usize) -> TabularLocation {
-        TabularLocation {
-            start_offset: Some(start),
-            end_offset: Some(end),
-            ..TabularLocation::new(row, col)
-        }
-    }
-
-    #[tokio::test]
-    async fn stream_walks_headers_then_rows() -> Result<(), Error> {
-        let mut h = handler_with_headers(vec!["name", "age"], vec![vec!["Alice", "30"]]);
-        let first = h.next_chunk().await?.unwrap();
-        assert_eq!(first.location.row_index, 0);
-        assert_eq!(first.location.column_index, 0);
-        assert_eq!(first.location.column_name.as_deref(), Some("name"));
-        // Drain the rest and count.
-        let mut count = 1;
-        while h.next_chunk().await?.is_some() {
-            count += 1;
-        }
-        assert_eq!(count, 4); // 2 header + 2 data
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn stream_no_headers_starts_at_row_zero() -> Result<(), Error> {
-        let mut h = handler_no_headers(vec![vec!["a", "b"], vec!["c", "d"]]);
-        let first = h.next_chunk().await?.unwrap();
-        assert_eq!(first.location.row_index, 0);
-        assert_eq!(first.location.column_index, 0);
-        assert!(first.location.column_name.is_none());
-        let mut count = 1;
-        while h.next_chunk().await?.is_some() {
-            count += 1;
-        }
-        assert_eq!(count, 4);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn read_returns_cell_value() -> Result<(), Error> {
-        let h = handler_with_headers(vec!["name"], vec![vec!["Alice"]]);
-        assert_eq!(
-            h.read(&TabularLocation::new(0, 0)).await?.unwrap().as_str(),
-            "name"
-        );
-        assert_eq!(
-            h.read(&TabularLocation::new(1, 0)).await?.unwrap().as_str(),
-            "Alice"
-        );
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn read_out_of_bounds_returns_none() -> Result<(), Error> {
-        let h = handler_with_headers(vec!["a"], vec![vec!["1"]]);
-        assert!(h.read(&TabularLocation::new(99, 0)).await?.is_none());
-        assert!(h.read(&TabularLocation::new(0, 99)).await?.is_none());
-        Ok(())
-    }
-
-    /// Lifting a recognizer-emitted intra-cell range turns into a
-    /// TabularLocation whose row/col match the chunk and whose
-    /// start_offset/end_offset point at the substring. Round-trips
-    /// through `read` (whole-cell) and a partial-cell redact.
-    #[tokio::test]
-    async fn lift_chunk_addresses_intra_cell_range() -> Result<(), Error> {
-        // Data row 1 col 1 is the email cell.
-        let mut h = handler_with_headers(
-            vec!["name", "email"],
-            vec![vec!["Alice", "alice@example.com"]],
-        );
-        // Advance the cursor to the email cell.
-        let chunk = loop {
-            let c = h.next_chunk().await?.expect("chunk");
-            if c.data.as_str() == "alice@example.com" {
-                break c;
-            }
-        };
-        // Recognizer says `alice` starts at byte 0 within the cell.
-        let lifted = h
-            .lift_chunk(&chunk, 0.."alice".len())
-            .expect("range in bounds");
-        assert_eq!(lifted.row_index, 1);
-        assert_eq!(lifted.column_index, 1);
-        assert_eq!(lifted.column_name.as_deref(), Some("email"));
-        assert_eq!(lifted.start_offset, Some(0));
-        assert_eq!(lifted.end_offset, Some(5));
-
-        // Out of bounds returns None.
-        assert!(h.lift_chunk(&chunk, 0..9999).is_none());
-        assert!(h.lift_chunk(&chunk, 99..100).is_none());
-        Ok(())
-    }
-
-    /// Pipeline: lift a recognizer-emitted intra-cell range, push it
-    /// through `redact`, and confirm only the matched substring
-    /// changes (not the whole cell).
-    #[tokio::test]
-    async fn lift_chunk_into_partial_cell_redact() -> Result<(), Error> {
-        let mut h = handler_with_headers(vec!["email"], vec![vec!["alice@example.com"]]);
-        let chunk = loop {
-            let c = h.next_chunk().await?.expect("chunk");
-            if c.data.as_str() == "alice@example.com" {
-                break c;
-            }
-        };
-        let lifted = h
-            .lift_chunk(&chunk, 0.."alice".len())
-            .expect("range in bounds");
-        let mut rs = Redactions::new();
-        rs.push(lifted, TabularReplacement::substituted("[USER]"));
-        h.redact(rs).await?;
-        let encoded = h.encode()?.as_str().unwrap().to_owned();
-        assert!(
-            encoded.contains("[USER]@example.com"),
-            "partial cell redaction lost: {encoded}",
-        );
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_full_cell() -> Result<(), Error> {
-        let mut h = handler_with_headers(vec!["ssn"], vec![vec!["123-45-6789"]]);
-        let mut rs = Redactions::new();
-        rs.push(
-            cell_range(1, 0, 0, 11),
-            TabularReplacement::substituted("[REDACTED]"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.cell(0, 0), Some("[REDACTED]"));
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_partial_cell() -> Result<(), Error> {
-        let mut h = handler_with_headers(vec!["bio"], vec![vec!["Alice Smith"]]);
-        let mut rs = Redactions::new();
-        rs.push(
-            cell_range(1, 0, 0, 5),
-            TabularReplacement::substituted("[NAME]"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.cell(0, 0), Some("[NAME] Smith"));
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_header() -> Result<(), Error> {
-        let mut h = handler_with_headers(vec!["secret_field"], vec![vec!["v"]]);
-        let mut rs = Redactions::new();
-        rs.push(
-            cell_range(0, 0, 0, 12),
-            TabularReplacement::substituted("redacted"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.headers(), Some(["redacted".to_string()].as_slice()));
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_two_ranges_same_cell() -> Result<(), Error> {
-        // Two intra-cell ranges within "Alice Smith"; expected right-to-left.
-        let mut h = handler_with_headers(vec!["bio"], vec![vec!["Alice Smith"]]);
-        let mut rs = Redactions::new();
-        rs.push(
-            cell_range(1, 0, 0, 5),
-            TabularReplacement::substituted("[A]"),
-        );
-        rs.push(
-            cell_range(1, 0, 6, 11),
-            TabularReplacement::substituted("[B]"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.cell(0, 0), Some("[A] [B]"));
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_unknown_row_skipped() -> Result<(), Error> {
-        let mut h = handler_with_headers(vec!["a"], vec![vec!["one"]]);
-        let mut rs = Redactions::new();
-        rs.push(
-            cell_range(99, 0, 0, 1),
-            TabularReplacement::substituted("X"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.cell(0, 0), Some("one"));
-        Ok(())
-    }
-
-    #[test]
-    fn encode_with_headers() -> Result<(), Error> {
-        let h = handler_with_headers(
-            vec!["name", "age"],
-            vec![vec!["Alice", "30"], vec!["Bob", "25"]],
-        );
-        let content = h.encode()?;
-        assert_eq!(
-            content.as_str().expect("valid utf-8"),
-            "name,age\nAlice,30\nBob,25\n"
-        );
-        Ok(())
-    }
-
-    #[test]
-    fn encode_with_quoting() -> Result<(), Error> {
-        let h = handler_with_headers(vec!["name", "bio"], vec![vec!["Alice", "Has a, comma"]]);
-        let content = h.encode()?;
-        let text = content.as_str().expect("valid utf-8");
-        assert!(text.contains("\"Has a, comma\""));
-        Ok(())
-    }
-
-    #[test]
-    fn encode_without_trailing_newline() -> Result<(), Error> {
-        let mut h = handler_with_headers(vec!["a"], vec![vec!["1"]]);
-        h.data.trailing_newline = false;
-        let content = h.encode()?;
-        assert_eq!(content.as_str().expect("valid utf-8"), "a\n1");
-        Ok(())
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/tabular/csv_loader.rs b/crates/nvisy-codec/src/handler/tabular/csv_loader.rs
deleted file mode 100644
index 2f31e16d1..000000000
--- a/crates/nvisy-codec/src/handler/tabular/csv_loader.rs
+++ /dev/null
@@ -1,204 +0,0 @@
-//! CSV loader: validates and parses raw CSV content into a
-//! [`CsvHandler`]. Auto-detects the field delimiter (comma, tab,
-//! semicolon, pipe) by inspecting the first few lines.
-
-use nvisy_core::Error;
-use nvisy_core::modality::Tabular;
-
-use super::{CsvData, CsvHandler};
-use crate::Loader;
-use crate::content::{ContentData, ContentSource, TextEncoding};
-
-const TARGET: &str = "nvisy_codec::handler::tabular::csv";
-
-/// Loader for CSV files. Produces one [`CsvHandler`] per input.
-#[derive(Debug)]
-pub struct CsvLoader {
-    /// Character encoding of the input bytes. Defaults to UTF-8.
-    pub encoding: TextEncoding,
-    /// Whether the first row contains column headers. Defaults to `true`.
-    pub has_headers: bool,
-    /// Override the field delimiter. `None` triggers auto-detection.
-    pub delimiter: Option<u8>,
-}
-
-impl Default for CsvLoader {
-    fn default() -> Self {
-        Self {
-            encoding: TextEncoding::Utf8,
-            has_headers: true,
-            delimiter: None,
-        }
-    }
-}
-
-#[async_trait::async_trait]
-impl Loader<Tabular> for CsvLoader {
-    type Handler = CsvHandler;
-
-    #[tracing::instrument(name = "csv.decode", skip_all, fields(input_bytes, rows, delimiter))]
-    async fn decode(&self, content: ContentData) -> Result<CsvHandler, Error> {
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        tracing::Span::current().record("input_bytes", raw.len());
-        let text = self
-            .encoding
-            .decode_bytes(&raw, "nvisy_codec::handler::tabular::csv")?;
-        let trailing_newline = text.ends_with('\n');
-        let delimiter = self.delimiter.unwrap_or_else(|| detect_delimiter(&text));
-        tracing::Span::current().record("delimiter", tracing::field::display(delimiter as char));
-
-        let mut reader = csv::ReaderBuilder::new()
-            .has_headers(self.has_headers)
-            .delimiter(delimiter)
-            .flexible(true)
-            .from_reader(text.as_bytes());
-
-        let headers = if self.has_headers {
-            let hdr = reader.headers().map_err(|e| {
-                Error::validation(
-                    format!("CSV header error: {e}"),
-                    "nvisy_codec::handler::tabular::csv",
-                )
-            })?;
-            Some(hdr.iter().map(String::from).collect())
-        } else {
-            None
-        };
-
-        let mut rows = Vec::new();
-        for result in reader.records() {
-            let record = result.map_err(|e| {
-                Error::validation(
-                    format!("CSV parse error: {e}"),
-                    "nvisy_codec::handler::tabular::csv",
-                )
-            })?;
-            rows.push(record.iter().map(String::from).collect());
-        }
-        tracing::Span::current().record("rows", rows.len());
-
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(CsvHandler::new(CsvData {
-            headers,
-            rows,
-            delimiter,
-            trailing_newline,
-        })
-        .with_source(source))
-    }
-}
-
-/// Auto-detect the CSV delimiter by sampling up to 5 lines and picking
-/// the candidate with the highest, most consistent count. Tie-break:
-/// prefer comma.
-fn detect_delimiter(text: &str) -> u8 {
-    let candidates: &[(u8, char)] = &[(b',', ','), (b'\t', '\t'), (b';', ';'), (b'|', '|')];
-    let sample_lines: Vec<&str> = text.lines().take(5).collect();
-    if sample_lines.is_empty() {
-        return b',';
-    }
-
-    let mut best_byte = b',';
-    let mut best_score = (0usize, 0usize);
-
-    for &(byte, ch) in candidates {
-        let counts: Vec<usize> = sample_lines
-            .iter()
-            .map(|line| line.matches(ch).count())
-            .collect();
-        let total: usize = counts.iter().sum();
-        let min = counts.iter().copied().min().unwrap_or(0);
-        let score = (min, total);
-        if score > best_score || (score == best_score && byte == b',') {
-            best_score = score;
-            best_byte = byte;
-        }
-    }
-
-    if best_byte != b',' {
-        tracing::debug!(target: TARGET, delimiter = %char::from(best_byte), "detected non-comma CSV delimiter");
-    }
-    best_byte
-}
-
-#[cfg(test)]
-mod tests {
-    use bytes::Bytes;
-    use nvisy_core::Error;
-
-    use super::*;
-    use crate::Handler;
-    use crate::content::ContentSource;
-
-    fn content_from_str(s: &str) -> ContentData {
-        ContentData::new(ContentSource::new(), Bytes::from(s.to_owned()))
-    }
-
-    #[tokio::test]
-    async fn load_with_headers() -> Result<(), Error> {
-        let content = content_from_str("name,age\nAlice,30\nBob,25\n");
-        let doc = CsvLoader::default().decode(content).await?;
-        assert_eq!(doc.format().as_str(), "nvisy.tabular.csv");
-        assert_eq!(
-            doc.headers(),
-            Some(["name", "age"].map(String::from).as_slice())
-        );
-        assert_eq!(doc.len(), 2);
-        assert_eq!(doc.cell(0, 0), Some("Alice"));
-        assert_eq!(doc.cell(1, 1), Some("25"));
-        assert!(doc.trailing_newline());
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn load_without_headers() -> Result<(), Error> {
-        let loader = CsvLoader {
-            has_headers: false,
-            ..CsvLoader::default()
-        };
-        let content = content_from_str("x,y\n1,2\n");
-        let doc = loader.decode(content).await?;
-        assert!(doc.headers().is_none());
-        assert_eq!(doc.len(), 2);
-        assert_eq!(doc.cell(0, 0), Some("x"));
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn load_tab_delimited() -> Result<(), Error> {
-        let content = content_from_str("a\tb\n1\t2\n");
-        let doc = CsvLoader::default().decode(content).await?;
-        assert_eq!(doc.delimiter(), b'\t');
-        assert_eq!(doc.headers(), Some(["a", "b"].map(String::from).as_slice()));
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn load_invalid_utf8() {
-        let content = ContentData::new(
-            ContentSource::new(),
-            Bytes::from_static(&[0xFF, 0xFE, 0x00]),
-        );
-        let err = CsvLoader::default().decode(content).await.unwrap_err();
-        assert!(err.to_string().contains("UTF-8"));
-    }
-
-    #[test]
-    fn detect_tab_delimited() {
-        let text = "a\tb\tc\n1\t2\t3\n4\t5\t6\n";
-        assert_eq!(detect_delimiter(text), b'\t');
-    }
-
-    #[test]
-    fn detect_semicolons_with_commas_in_content() {
-        let text = "\"a,b\";c;d\n\"e,f\";g;h\n";
-        assert_eq!(detect_delimiter(text), b';');
-    }
-
-    #[test]
-    fn detect_no_delimiters_defaults_to_comma() {
-        let text = "just plain text\nno delimiters here\n";
-        assert_eq!(detect_delimiter(text), b',');
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/tabular/mod.rs b/crates/nvisy-codec/src/handler/tabular/mod.rs
deleted file mode 100644
index cf42b0922..000000000
--- a/crates/nvisy-codec/src/handler/tabular/mod.rs
+++ /dev/null
@@ -1,57 +0,0 @@
-//! Tabular modality: the [`TabularHandle`] extension trait plus
-//! concrete tabular format implementations (CSV, XLSX).
-//!
-//! Tabular handlers address content by cell coordinate
-//! ([`Tabular`] = row + column, optionally with intra-cell byte
-//! offsets) and return [`TextData`] from `read` — the cell's string
-//! value — so [`Modality::Data`] aliases [`TextData`] for the
-//! [`Tabular`] modality. Replacements written during
-//! [`Handler::redact`] use [`TabularReplacement`]; cells are
-//! strings, so the per-format handlers share the text crate's redact
-//! helper.
-//!
-//! [`Handler::redact`]: crate::Handler::redact
-//! [`Tabular`]: nvisy_core::modality::Tabular
-//! [`TabularReplacement`]: nvisy_core::redaction::TabularReplacement
-//! [`TextData`]: nvisy_core::modality::TextData
-//! [`Modality::Data`]: nvisy_core::modality::Modality::Data
-
-use nvisy_core::modality::Tabular;
-
-use crate::Handler;
-
-/// Extension trait implemented by every tabular handler exposing the
-/// "do I have a header row?" signal.
-///
-/// The engine importer reads this to pick
-/// [`TabularExtraction::SchemaTyped`] when headers are known from the
-/// source (header row in CSV, schema in Parquet/XLSX) vs.
-/// [`TabularExtraction::SchemaInferred`] when column semantics must be
-/// inferred from data.
-///
-/// [`TabularExtraction::SchemaTyped`]: nvisy_core::modality::TabularExtraction::SchemaTyped
-/// [`TabularExtraction::SchemaInferred`]: nvisy_core::modality::TabularExtraction::SchemaInferred
-pub trait TabularHandle: Handler<Tabular> {
-    /// `true` when the source format carries explicit column headers
-    /// or typed schema (CSV with header row, Parquet, XLSX); `false`
-    /// when column semantics have to be inferred from the data.
-    fn has_header(&self) -> bool;
-}
-
-#[cfg(feature = "csv")]
-mod csv_handler;
-#[cfg(feature = "csv")]
-mod csv_loader;
-#[cfg(feature = "xlsx")]
-mod xlsx_handler;
-#[cfg(feature = "xlsx")]
-mod xlsx_loader;
-
-#[cfg(feature = "csv")]
-pub use self::csv_handler::{CsvData, CsvHandler, format as csv_format};
-#[cfg(feature = "csv")]
-pub use self::csv_loader::CsvLoader;
-#[cfg(feature = "xlsx")]
-pub use self::xlsx_handler::{XlsxHandler, format as xlsx_format};
-#[cfg(feature = "xlsx")]
-pub use self::xlsx_loader::XlsxLoader;
diff --git a/crates/nvisy-codec/src/handler/tabular/xlsx_handler.rs b/crates/nvisy-codec/src/handler/tabular/xlsx_handler.rs
deleted file mode 100644
index 69c4778e1..000000000
--- a/crates/nvisy-codec/src/handler/tabular/xlsx_handler.rs
+++ /dev/null
@@ -1,85 +0,0 @@
-//! XLSX handler (stub: awaiting full spreadsheet support).
-//!
-//! Decodes to an empty handler — emits no chunks, reads return `None`,
-//! redactions are no-ops, encode fails. Wired into the registry so
-//! downstream code can resolve XLSX without a runtime panic, while
-//! signalling that meaningful tabular access isn't implemented yet.
-
-use nvisy_core::Error;
-use nvisy_core::modality::{Tabular, TabularLocation, TextData};
-use nvisy_core::redaction::Redactions;
-
-use super::XlsxLoader;
-use crate::content::{ContentData, ContentSource};
-use crate::handler::tabular::TabularHandle;
-use crate::{Chunk, Format, FormatId, Handler};
-
-/// Stable [`FormatId`] for the XLSX codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.tabular.xlsx");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Tabular, _>(FORMAT_ID.clone(), XlsxLoader)
-        .with_extensions(["xlsx"])
-        .with_content_types(["application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"])
-}
-
-/// Handler for loaded XLSX content. Placeholder: the current XLSX
-/// pipeline only carries source identity through to downstream phases;
-/// chunking and redact return empty / `None` until the full
-/// spreadsheet parser lands.
-#[derive(Debug, Default)]
-pub struct XlsxHandler {
-    source: ContentSource,
-}
-
-impl XlsxHandler {
-    /// Build an empty handler with no attached source.
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Attach a content source for lineage tracking.
-    pub fn with_source(mut self, source: ContentSource) -> Self {
-        self.source = source;
-        self
-    }
-}
-
-#[async_trait::async_trait]
-impl Handler<Tabular> for XlsxHandler {
-    fn format(&self) -> FormatId {
-        FORMAT_ID.clone()
-    }
-
-    fn source(&self) -> ContentSource {
-        self.source
-    }
-
-    #[tracing::instrument(name = "xlsx.encode", skip_all)]
-    fn encode(&self) -> Result<ContentData, Error> {
-        Err(Error::validation(
-            "encode not supported for XLSX",
-            "xlsx-handler",
-        ))
-    }
-
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<Tabular>>, Error> {
-        Ok(None)
-    }
-
-    async fn read(&self, _location: &TabularLocation) -> Result<Option<TextData>, Error> {
-        Ok(None)
-    }
-
-    async fn redact(&mut self, _redactions: Redactions<Tabular>) -> Result<(), Error> {
-        Ok(())
-    }
-}
-
-impl TabularHandle for XlsxHandler {
-    fn has_header(&self) -> bool {
-        // XLSX always carries a typed schema.
-        true
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/tabular/xlsx_loader.rs b/crates/nvisy-codec/src/handler/tabular/xlsx_loader.rs
deleted file mode 100644
index 6d2c0a098..000000000
--- a/crates/nvisy-codec/src/handler/tabular/xlsx_loader.rs
+++ /dev/null
@@ -1,25 +0,0 @@
-//! XLSX loader (stub: awaiting full spreadsheet support).
-
-use nvisy_core::Error;
-use nvisy_core::modality::Tabular;
-
-use super::XlsxHandler;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource};
-
-/// Loader for XLSX spreadsheets. Produces a single (stub) [`XlsxHandler`].
-#[derive(Debug, Default)]
-pub struct XlsxLoader;
-
-#[async_trait::async_trait]
-impl Loader<Tabular> for XlsxLoader {
-    type Handler = XlsxHandler;
-
-    #[tracing::instrument(name = "xlsx.decode", skip_all, fields(input_bytes))]
-    async fn decode(&self, content: ContentData) -> Result<XlsxHandler, Error> {
-        tracing::Span::current().record("input_bytes", content.to_bytes().len());
-        let parent = content.content_source;
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(XlsxHandler::new().with_source(source))
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/text/html_encode.rs b/crates/nvisy-codec/src/handler/text/html_encode.rs
deleted file mode 100644
index 36f4184cd..000000000
--- a/crates/nvisy-codec/src/handler/text/html_encode.rs
+++ /dev/null
@@ -1,172 +0,0 @@
-//! HTML encode-side machinery: the [`EncodePlan`] that buckets the
-//! handler's [`RedactableItem`] stream by kind so encode's single
-//! DOM walk can answer "what does item-of-kind-X at index N look
-//! like?" in O(1), plus the per-kind DOM mutation helpers.
-
-use ego_tree::NodeId;
-use scraper::Html;
-use scraper::node::Node;
-
-use super::html_handler::{ElementTarget, RedactableItem, RedactableKind};
-
-/// Pre-bucketed view of the loader's items, indexed by kind so
-/// encode's single DOM walk can answer "what does item-of-kind-X at
-/// index N look like?" in O(1).
-pub(super) struct EncodePlan<'a> {
-    /// Indexed by text-node ordinal in document order. `Some(value)`
-    /// when the loader emitted an item for that text node; `None`
-    /// for skipped text nodes (e.g. script / style children when
-    /// the policy is `Skip`).
-    text_values: Vec<Option<&'a str>>,
-    /// Indexed by comment ordinal. Same shape as `text_values`.
-    comment_values: Vec<Option<&'a str>>,
-    /// Indexed by element ordinal. Each entry is the (possibly
-    /// empty) list of element-bound items at that element.
-    element_targets: Vec<Vec<ElementTargetPatch<'a>>>,
-}
-
-/// A single element-bound patch staged for encode.
-struct ElementTargetPatch<'a> {
-    target: &'a ElementTarget,
-    value: &'a str,
-}
-
-impl<'a> EncodePlan<'a> {
-    /// Build a plan from the handler's items list.
-    pub(super) fn from_items(items: &'a [RedactableItem]) -> Self {
-        let mut text_values: Vec<Option<&'a str>> = Vec::new();
-        let mut comment_values: Vec<Option<&'a str>> = Vec::new();
-        let mut element_targets: Vec<Vec<ElementTargetPatch<'a>>> = Vec::new();
-
-        for item in items {
-            match &item.kind {
-                RedactableKind::TextNode { index } => {
-                    grow_to(&mut text_values, *index);
-                    text_values[*index] = Some(item.value.as_str());
-                }
-                RedactableKind::Comment { index } => {
-                    grow_to(&mut comment_values, *index);
-                    comment_values[*index] = Some(item.value.as_str());
-                }
-                RedactableKind::Element {
-                    element_index,
-                    target,
-                } => {
-                    while element_targets.len() <= *element_index {
-                        element_targets.push(Vec::new());
-                    }
-                    element_targets[*element_index].push(ElementTargetPatch {
-                        target,
-                        value: item.value.as_str(),
-                    });
-                }
-            }
-        }
-
-        Self {
-            text_values,
-            comment_values,
-            element_targets,
-        }
-    }
-
-    /// Walk `dom` in document order, applying the per-kind patches
-    /// from this plan against the matching ordinals. Mutates `dom`
-    /// in place.
-    pub(super) fn apply(&self, dom: &mut Html) {
-        let mut text_seen = 0usize;
-        let mut comment_seen = 0usize;
-        let mut element_seen = 0usize;
-
-        let node_ids: Vec<_> = dom.tree.nodes().map(|n| n.id()).collect();
-        for node_id in node_ids {
-            let Some(node) = dom.tree.get(node_id) else {
-                continue;
-            };
-            match node.value() {
-                Node::Text(_) => {
-                    if let Some(value) = self.text_values.get(text_seen).copied().flatten() {
-                        write_text(dom, node_id, value);
-                    }
-                    text_seen += 1;
-                }
-                Node::Comment(_) => {
-                    if let Some(value) = self.comment_values.get(comment_seen).copied().flatten() {
-                        write_comment(dom, node_id, value);
-                    }
-                    comment_seen += 1;
-                }
-                Node::Element(_) => {
-                    if let Some(targets) = self.element_targets.get(element_seen) {
-                        for et in targets {
-                            apply_element_target(dom, node_id, et);
-                        }
-                    }
-                    element_seen += 1;
-                }
-                _ => {}
-            }
-        }
-    }
-}
-
-fn grow_to(v: &mut Vec<Option<&str>>, index: usize) {
-    while v.len() <= index {
-        v.push(None);
-    }
-}
-
-fn write_text(dom: &mut Html, node_id: NodeId, value: &str) {
-    if let Some(mut n) = dom.tree.get_mut(node_id)
-        && let Node::Text(t) = n.value()
-        && t.text.as_ref() != value
-    {
-        t.text = value.into();
-    }
-}
-
-fn write_comment(dom: &mut Html, node_id: NodeId, value: &str) {
-    if let Some(mut n) = dom.tree.get_mut(node_id)
-        && let Node::Comment(c) = n.value()
-        && c.comment.as_ref() != value
-    {
-        c.comment = value.into();
-    }
-}
-
-fn apply_element_target(dom: &mut Html, node_id: NodeId, patch: &ElementTargetPatch<'_>) {
-    match patch.target {
-        ElementTarget::Attribute { attr_name } => {
-            write_attribute(dom, node_id, attr_name, patch.value);
-        }
-        ElementTarget::ScriptText | ElementTarget::StyleText => {
-            write_text_child(dom, node_id, patch.value);
-        }
-    }
-}
-
-fn write_attribute(dom: &mut Html, element_id: NodeId, attr_name: &str, value: &str) {
-    if let Some(mut n) = dom.tree.get_mut(element_id)
-        && let Node::Element(e) = n.value()
-    {
-        for (qn, val) in &mut e.attrs {
-            if qn.local.as_ref() == attr_name {
-                if val.as_ref() != value {
-                    *val = value.into();
-                }
-                return;
-            }
-        }
-    }
-}
-
-fn write_text_child(dom: &mut Html, element_id: NodeId, value: &str) {
-    let child_id = dom
-        .tree
-        .get(element_id)
-        .and_then(|n| n.first_child().map(|c| c.id()));
-    let Some(child_id) = child_id else {
-        return;
-    };
-    write_text(dom, child_id, value);
-}
diff --git a/crates/nvisy-codec/src/handler/text/html_handler.rs b/crates/nvisy-codec/src/handler/text/html_handler.rs
deleted file mode 100644
index c47ecad51..000000000
--- a/crates/nvisy-codec/src/handler/text/html_handler.rs
+++ /dev/null
@@ -1,437 +0,0 @@
-//! HTML handler: holds a parsed HTML document as a stream of
-//! [`RedactableItem`]s — text nodes, comments, element
-//! attributes, and (per [`HtmlLoader::script_policy`] /
-//! [`HtmlLoader::style_policy`]) `<script>` / `<style>` bodies.
-//!
-//! Offsets are cumulative over the redactable-item sequence in
-//! document order, not raw HTML source bytes. [`Handler::encode`]
-//! reconstructs the HTML by re-parsing the original source into a
-//! DOM, splicing each mutated value back in document order, and
-//! serializing with [`Html::html`].
-//!
-//! [`Html::html`]: scraper::Html::html
-//! [`HtmlLoader`]: super::HtmlLoader
-//! [`HtmlLoader::script_policy`]: super::HtmlLoader::script_policy
-//! [`HtmlLoader::style_policy`]: super::HtmlLoader::style_policy
-
-use std::ops::Range;
-
-use nvisy_core::Error;
-use nvisy_core::modality::{Text, TextData, TextLocation};
-use nvisy_core::redaction::{Redactions, TextReplacement};
-use scraper::Html;
-
-use super::html_encode::EncodePlan;
-use super::{HtmlLoader, redact};
-use crate::content::{ContentData, ContentSource};
-use crate::{Chunk, Format, FormatId, Handler};
-
-const TARGET: &str = "nvisy_codec::handler::text::html";
-
-/// Stable [`FormatId`] for the HTML codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.text.html");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Text, _>(FORMAT_ID.clone(), HtmlLoader::default())
-        .with_extensions(["html", "htm"])
-        .with_content_types(["text/html"])
-}
-
-/// Parsed HTML content: the redactable-item stream plus the raw
-/// source bytes (kept for round-trip reconstruction).
-#[derive(Debug, Clone)]
-pub struct HtmlData {
-    /// Redactable items in document order. Decoded once by the
-    /// [`HtmlLoader`] and never re-derived afterwards.
-    pub items: Vec<RedactableItem>,
-    /// The raw HTML source. Re-parsed at encode time so the
-    /// serialiser can splice mutated values back into a fresh
-    /// DOM and emit the result.
-    pub raw: String,
-}
-
-/// One redactable unit yielded by [`HtmlHandler::next_chunk`].
-///
-/// `value` is the text the recognizer scans and that
-/// [`HtmlHandler::redact`] mutates in place. `kind` tells encode
-/// where to splice the mutated value back into the document.
-#[derive(Debug, Clone)]
-pub struct RedactableItem {
-    /// Where this item lives in the document.
-    pub kind: RedactableKind,
-    /// Text-node text, comment body, attribute value, or script /
-    /// style element text.
-    pub value: String,
-    /// Out-of-band context strings surfaced from the item's
-    /// structural neighbours — currently the parent element's
-    /// concatenated text content minus this item's own text.
-    /// Empty for items without useful surrounding context (e.g.
-    /// items whose parent element only contains the item itself).
-    pub hints: Vec<String>,
-}
-
-/// Where a [`RedactableItem`] lives inside the parsed HTML
-/// document. Used by encode to splice the mutated `value` back into
-/// a freshly-parsed DOM.
-#[derive(Debug, Clone)]
-pub enum RedactableKind {
-    /// A text node, addressed by its 0-based index in the
-    /// document-order text-node sequence.
-    TextNode {
-        /// Document-order index among text nodes.
-        index: usize,
-    },
-    /// An HTML comment, addressed by its 0-based index in the
-    /// document-order comment sequence.
-    Comment {
-        /// Document-order index among comments.
-        index: usize,
-    },
-    /// Any element-bound item: an attribute value or a script /
-    /// style element's text body.
-    Element {
-        /// Document-order index among elements.
-        element_index: usize,
-        /// Which part of the element this item addresses.
-        target: ElementTarget,
-    },
-}
-
-/// The element-bound location a [`RedactableKind::Element`] item
-/// points at.
-#[derive(Debug, Clone)]
-pub enum ElementTarget {
-    /// The value of `attr_name` on this element.
-    Attribute {
-        /// Attribute local name.
-        attr_name: String,
-    },
-    /// The text body of a `<script>` element scanned as plain text.
-    ScriptText,
-    /// The text body of a `<style>` element scanned as plain text.
-    StyleText,
-}
-
-/// Handler for loaded HTML content.
-///
-/// `item_starts` is a cumulative-offset index over the redactable
-/// items: `item_starts[i]` is the byte position of item `i`, and
-/// `item_starts[items.len()]` is the total length sentinel.
-/// Maintained on every redaction so random-access reads run in
-/// `O(log N)`.
-#[derive(Debug)]
-pub struct HtmlHandler {
-    source: ContentSource,
-    data: HtmlData,
-    item_starts: Vec<usize>,
-    cursor: usize,
-}
-
-#[async_trait::async_trait]
-impl Handler<Text> for HtmlHandler {
-    fn format(&self) -> FormatId {
-        FORMAT_ID.clone()
-    }
-
-    fn source(&self) -> ContentSource {
-        self.source
-    }
-
-    #[tracing::instrument(name = "html.encode", skip_all, fields(output_bytes))]
-    fn encode(&self) -> Result<ContentData, Error> {
-        let mut dom = Html::parse_document(&self.data.raw);
-        EncodePlan::from_items(&self.data.items).apply(&mut dom);
-
-        let bytes = dom.html().into_bytes();
-        tracing::Span::current().record("output_bytes", bytes.len());
-        let source = ContentSource::new().with_parent(&self.source);
-        Ok(ContentData::new(source, bytes.into()))
-    }
-
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<Text>>, Error> {
-        if self.cursor >= self.data.items.len() {
-            return Ok(None);
-        }
-        let i = self.cursor;
-        let start = self.item_starts[i];
-        let end = self.item_starts[i + 1];
-        let item = &self.data.items[i];
-        let value = &item.value;
-        let hints = item.hints.clone();
-        self.cursor += 1;
-        Ok(Some(Chunk {
-            location: TextLocation {
-                start,
-                end,
-                ..Default::default()
-            },
-            data: TextData::from(value.as_str()),
-            hints,
-        }))
-    }
-
-    fn lift_chunk(&self, chunk: &Chunk<Text>, value_range: Range<usize>) -> Option<TextLocation> {
-        chunk.location.subslice(value_range)
-    }
-
-    async fn read(&self, location: &TextLocation) -> Result<Option<TextData>, Error> {
-        let Some(i) = self.item_for(location.start) else {
-            return Ok(None);
-        };
-        let item_start = self.item_starts[i];
-        let item_end = self.item_starts[i + 1];
-        if location.end > item_end {
-            return Ok(None);
-        }
-        let local_start = location.start - item_start;
-        let local_end = location.end - item_start;
-        Ok(self.data.items[i]
-            .value
-            .get(local_start..local_end)
-            .map(TextData::from))
-    }
-
-    async fn redact(&mut self, mut redactions: Redactions<Text>) -> Result<(), Error> {
-        redactions.sort_descending();
-        for (location, replacement) in redactions.into_items() {
-            self.redact_one(&location, replacement)?;
-        }
-        Ok(())
-    }
-}
-
-impl HtmlHandler {
-    /// Build a handler from a populated [`HtmlData`].
-    pub fn new(data: HtmlData) -> Self {
-        let item_starts = compute_item_starts(&data.items);
-        Self {
-            source: ContentSource::new(),
-            data,
-            item_starts,
-            cursor: 0,
-        }
-    }
-
-    /// Attach a content source for lineage tracking.
-    pub fn with_source(mut self, source: ContentSource) -> Self {
-        self.source = source;
-        self
-    }
-
-    /// All redactable items in document order.
-    pub fn items(&self) -> &[RedactableItem] {
-        &self.data.items
-    }
-
-    /// Total number of redactable items.
-    pub fn len(&self) -> usize {
-        self.data.items.len()
-    }
-
-    /// Whether the document has no redactable items.
-    pub fn is_empty(&self) -> bool {
-        self.data.items.is_empty()
-    }
-
-    /// The raw HTML source.
-    pub fn raw(&self) -> &str {
-        &self.data.raw
-    }
-
-    /// Consume the handler and return the inner [`HtmlData`].
-    pub fn into_data(self) -> HtmlData {
-        self.data
-    }
-
-    /// Rewind the streaming cursor to the start of the document.
-    pub fn rewind(&mut self) {
-        self.cursor = 0;
-    }
-
-    fn item_for(&self, byte_offset: usize) -> Option<usize> {
-        match self.item_starts.binary_search(&byte_offset) {
-            Ok(i) if i < self.data.items.len() => Some(i),
-            Ok(_) => None,
-            Err(i) if i > 0 && i <= self.data.items.len() => Some(i - 1),
-            _ => None,
-        }
-    }
-
-    fn shift_starts_after(&mut self, i: usize, delta: isize) {
-        if delta == 0 {
-            return;
-        }
-        for s in &mut self.item_starts[i + 1..] {
-            *s = (*s as isize + delta) as usize;
-        }
-    }
-
-    fn redact_one(
-        &mut self,
-        location: &TextLocation,
-        replacement: TextReplacement,
-    ) -> Result<(), Error> {
-        let Some(i) = self.item_for(location.start) else {
-            return Ok(());
-        };
-        let item_start = self.item_starts[i];
-        let item_end = self.item_starts[i + 1];
-        if location.end > item_end {
-            return Ok(());
-        }
-        let local_start = location.start - item_start;
-        let local_end = location.end - item_start;
-        let value = replacement.replacement_value().unwrap_or_default();
-        let before_len = self.data.items[i].value.len();
-        redact::replace_range(
-            &mut self.data.items[i].value,
-            value,
-            local_start..local_end,
-            TARGET,
-        )?;
-        let delta = self.data.items[i].value.len() as isize - before_len as isize;
-        self.shift_starts_after(i, delta);
-        Ok(())
-    }
-}
-
-/// Cumulative byte-offset table over the redactable items: `[0, len(item[0]),
-/// len(item[0]) + len(item[1]), …, total]`.
-fn compute_item_starts(items: &[RedactableItem]) -> Vec<usize> {
-    let mut starts = Vec::with_capacity(items.len() + 1);
-    let mut offset = 0usize;
-    for item in items {
-        starts.push(offset);
-        offset += item.value.len();
-    }
-    starts.push(offset);
-    starts
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::Error;
-    use nvisy_core::redaction::TextReplacement;
-
-    use super::super::HtmlLoader;
-    use super::*;
-    use crate::Loader;
-    use crate::content::{ContentData, ContentSource};
-
-    async fn load(raw: &str) -> HtmlHandler {
-        load_with(raw, HtmlLoader::default()).await
-    }
-
-    async fn load_with(raw: &str, loader: HtmlLoader) -> HtmlHandler {
-        let content = ContentData::new(ContentSource::new(), bytes::Bytes::from(raw.to_owned()));
-        loader.decode(content).await.expect("html decode succeeds")
-    }
-
-    #[tokio::test]
-    async fn encode_unchanged_round_trips() -> Result<(), Error> {
-        let raw = "<html><head></head><body><p>Hello</p></body></html>";
-        let h = load(raw).await;
-        let content = h.encode()?;
-        assert_eq!(content.as_str().unwrap(), raw);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_replaces_text_node() -> Result<(), Error> {
-        let raw = "<html><head></head><body><p>Hello</p><p>World</p></body></html>";
-        let mut h = load(raw).await;
-        let first = h.next_chunk().await?.unwrap();
-        let mut rs = Redactions::new();
-        rs.push(first.location, TextReplacement::substituted("[REDACTED]"));
-        h.redact(rs).await?;
-        let out = h.encode()?.as_str().unwrap().to_owned();
-        assert!(out.contains("[REDACTED]"));
-        assert!(out.contains("World"));
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn stream_yields_text_attribute_and_comment() -> Result<(), Error> {
-        let raw = r#"<html><head></head><body><!-- secret 1 --><img alt="hello" title="alt"></body></html>"#;
-        let mut h = load(raw).await;
-        let mut values: Vec<String> = Vec::new();
-        while let Some(chunk) = h.next_chunk().await? {
-            values.push(chunk.data.as_str().to_owned());
-        }
-        assert!(values.iter().any(|v| v == " secret 1 "));
-        assert!(values.iter().any(|v| v == "hello"));
-        assert!(values.iter().any(|v| v == "alt"));
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn attribute_redact_round_trips() -> Result<(), Error> {
-        let raw = r#"<html><head></head><body><img alt="alice@example.com"></body></html>"#;
-        let mut h = load(raw).await;
-        // Find the attribute chunk.
-        let mut loc = None;
-        while let Some(chunk) = h.next_chunk().await? {
-            if chunk.data.as_str() == "alice@example.com" {
-                loc = Some(chunk.location);
-                break;
-            }
-        }
-        let loc = loc.expect("alt attribute yields a chunk");
-        let mut rs = Redactions::new();
-        rs.push(loc, TextReplacement::substituted("[email]"));
-        h.redact(rs).await?;
-        let out = h.encode()?.as_str().unwrap().to_owned();
-        assert!(
-            out.contains(r#"alt="[email]""#),
-            "alt attribute not rewritten: {out}"
-        );
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn comment_redact_round_trips() -> Result<(), Error> {
-        let raw = "<html><head></head><body><!-- alice@example.com --></body></html>";
-        let mut h = load(raw).await;
-        let mut loc = None;
-        while let Some(chunk) = h.next_chunk().await? {
-            if chunk.data.as_str().contains("alice@example.com") {
-                loc = Some(TextLocation {
-                    start: chunk.location.start + chunk.data.as_str().find("alice").unwrap(),
-                    end: chunk.location.start
-                        + chunk.data.as_str().find("alice").unwrap()
-                        + "alice@example.com".len(),
-                    ..Default::default()
-                });
-                break;
-            }
-        }
-        let loc = loc.expect("comment yields a chunk containing the email");
-        let mut rs = Redactions::new();
-        rs.push(loc, TextReplacement::substituted("[email]"));
-        h.redact(rs).await?;
-        let out = h.encode()?.as_str().unwrap().to_owned();
-        assert!(
-            out.contains("<!-- [email] -->"),
-            "comment not rewritten: {out}"
-        );
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn href_attribute_carries_url_verbatim() -> Result<(), Error> {
-        let raw = r#"<html><head></head><body><a href="mailto:alice@example.com?subject=Hi">contact</a></body></html>"#;
-        let mut h = load(raw).await;
-        let mut emitted = Vec::new();
-        while let Some(chunk) = h.next_chunk().await? {
-            emitted.push(chunk.data.as_str().to_owned());
-        }
-        assert!(
-            emitted
-                .iter()
-                .any(|v| v == "mailto:alice@example.com?subject=Hi"),
-            "href emitted as one verbatim attribute item: {emitted:?}"
-        );
-        Ok(())
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/text/html_loader.rs b/crates/nvisy-codec/src/handler/text/html_loader.rs
deleted file mode 100644
index 1667de012..000000000
--- a/crates/nvisy-codec/src/handler/text/html_loader.rs
+++ /dev/null
@@ -1,282 +0,0 @@
-//! HTML loader: parses raw HTML into an [`HtmlHandler`] populated
-//! with a heterogeneous [`RedactableItem`] stream.
-//!
-//! Text nodes, every element attribute, and HTML comments are
-//! emitted as items. Attribute values pass through verbatim;
-//! percent-encoded URL values are a known gap, tracked at
-//! <https://github.com/nvisycom/runtime/issues/267>. `<script>`
-//! and `<style>` bodies follow [`HtmlLoader::script_policy`] /
-//! [`HtmlLoader::style_policy`].
-//!
-//! [`RedactableItem`]: super::RedactableItem
-
-use ego_tree::NodeRef;
-use nvisy_core::Error;
-use nvisy_core::modality::Text;
-use scraper::Html;
-use scraper::node::Node;
-
-use super::html_handler::{ElementTarget, HtmlData, HtmlHandler, RedactableItem, RedactableKind};
-use crate::Loader;
-use crate::content::{ContentData, ContentSource, TextEncoding};
-
-const TARGET: &str = "nvisy_codec::handler::text::html";
-
-/// Loader for HTML files. Produces one [`HtmlHandler`] per input.
-#[derive(Debug, Clone)]
-pub struct HtmlLoader {
-    /// Character encoding of the input bytes.
-    pub encoding: TextEncoding,
-    /// How `<script>` element bodies enter the detection stream.
-    pub script_policy: ScriptPolicy,
-    /// How `<style>` element bodies enter the detection stream.
-    pub style_policy: ScriptPolicy,
-}
-
-impl Default for HtmlLoader {
-    fn default() -> Self {
-        Self {
-            encoding: TextEncoding::Utf8,
-            script_policy: ScriptPolicy::Skip,
-            style_policy: ScriptPolicy::Skip,
-        }
-    }
-}
-
-/// How the loader handles `<script>` or `<style>` element bodies.
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub enum ScriptPolicy {
-    /// Skip the element entirely — its body never enters the
-    /// detection stream.
-    Skip,
-    /// Treat the element body as plain text and scan it like a
-    /// regular text node.
-    ScanText,
-    /// Parse the element body as JSON and route it through the
-    /// JSON handler's slot model.
-    ///
-    /// **Not yet implemented** — a loader configured with this
-    /// policy returns an error at decode time. Tracked in
-    /// <https://github.com/nvisycom/runtime/issues/266>.
-    ScanJson,
-}
-
-#[async_trait::async_trait]
-impl Loader<Text> for HtmlLoader {
-    type Handler = HtmlHandler;
-
-    #[tracing::instrument(name = "html.decode", skip_all, fields(input_bytes, items))]
-    async fn decode(&self, content: ContentData) -> Result<HtmlHandler, Error> {
-        if matches!(self.script_policy, ScriptPolicy::ScanJson)
-            || matches!(self.style_policy, ScriptPolicy::ScanJson)
-        {
-            return Err(Error::validation(
-                "ScriptPolicy::ScanJson is not yet implemented; see https://github.com/nvisycom/runtime/issues/266",
-                TARGET,
-            ));
-        }
-
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        tracing::Span::current().record("input_bytes", raw.len());
-        let text = self.encoding.decode_bytes(&raw, TARGET)?;
-        let dom = Html::parse_document(&text);
-        let items = build_items(&dom, self);
-        tracing::Span::current().record("items", items.len());
-
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(HtmlHandler::new(HtmlData { items, raw: text }).with_source(source))
-    }
-}
-
-fn build_items(dom: &Html, loader: &HtmlLoader) -> Vec<RedactableItem> {
-    let mut items = Vec::new();
-    let mut text_index: usize = 0;
-    let mut comment_index: usize = 0;
-    let mut element_index: usize = 0;
-
-    for node in dom.tree.nodes() {
-        match node.value() {
-            Node::Text(t) => {
-                if !skip_text_under(node) {
-                    let hints = sibling_text_hint(node, &t.text);
-                    items.push(RedactableItem {
-                        kind: RedactableKind::TextNode { index: text_index },
-                        value: t.text.to_string(),
-                        hints,
-                    });
-                }
-                text_index += 1;
-            }
-            Node::Comment(c) => {
-                items.push(RedactableItem {
-                    kind: RedactableKind::Comment {
-                        index: comment_index,
-                    },
-                    value: c.comment.to_string(),
-                    hints: Vec::new(),
-                });
-                comment_index += 1;
-            }
-            Node::Element(e) => {
-                let element_name = e.name.local.as_ref();
-
-                // Every attribute on this element. Values pass
-                // through verbatim — URLs like `mailto:alice@x.com`
-                // have the email matched in place by the
-                // recognizer.
-                for (qn, val) in &e.attrs {
-                    items.push(RedactableItem {
-                        kind: RedactableKind::Element {
-                            element_index,
-                            target: ElementTarget::Attribute {
-                                attr_name: qn.local.as_ref().to_owned(),
-                            },
-                        },
-                        value: val.to_string(),
-                        hints: Vec::new(),
-                    });
-                }
-
-                // `<script>` / `<style>` body, when policy says ScanText.
-                let policy = match element_name {
-                    "script" => Some(loader.script_policy),
-                    "style" => Some(loader.style_policy),
-                    _ => None,
-                };
-                if let Some(ScriptPolicy::ScanText) = policy
-                    && let Some(body) = first_child_text(node)
-                {
-                    let target = if element_name == "script" {
-                        ElementTarget::ScriptText
-                    } else {
-                        ElementTarget::StyleText
-                    };
-                    items.push(RedactableItem {
-                        kind: RedactableKind::Element {
-                            element_index,
-                            target,
-                        },
-                        value: body,
-                        hints: Vec::new(),
-                    });
-                }
-
-                element_index += 1;
-            }
-            _ => {}
-        }
-    }
-
-    items
-}
-
-/// Collect the surrounding-text content of the text node's
-/// nearest block-level ancestor as a single hint string.
-///
-/// Used by [`build_items`] to surface the surrounding sentence
-/// (`"the payment card 4111… is on file"`) as an out-of-band
-/// hint when a text node sits inside an inline wrapper
-/// (`<code>4111…</code>`) that splits the prose into multiple
-/// chunks. The walk targets the nearest *block* ancestor
-/// (`<p>`, `<div>`, `<li>`, `<td>`, `<th>`, `<h1>`–`<h6>`,
-/// `<blockquote>`, `<dt>`, `<dd>`) — stopping at the immediate
-/// inline parent would yield only the chunk's own text.
-///
-/// `own_text` is excluded so the hint doesn't echo the node's
-/// own bytes. Returns an empty `Vec` when no useful surrounding
-/// text exists (no block ancestor, or the ancestor contains
-/// only this text).
-fn sibling_text_hint(text_node: NodeRef<'_, Node>, own_text: &str) -> Vec<String> {
-    let Some(ancestor) = nearest_block_ancestor(text_node) else {
-        return Vec::new();
-    };
-    let mut buf = String::new();
-    for descendant in ancestor.descendants() {
-        if let Node::Text(t) = descendant.value() {
-            let chunk = t.text.as_ref();
-            if chunk == own_text {
-                continue;
-            }
-            if !buf.is_empty() {
-                buf.push(' ');
-            }
-            buf.push_str(chunk);
-        }
-    }
-    let trimmed = buf.trim();
-    if trimmed.is_empty() {
-        Vec::new()
-    } else {
-        vec![trimmed.to_owned()]
-    }
-}
-
-/// Walk parents until we hit a block-level element (or root).
-/// Used to find the "sentence boundary" around an inline text
-/// node so the hint covers the full prose around an inline
-/// wrapper like `<code>` or `<span>`.
-fn nearest_block_ancestor(text_node: NodeRef<'_, Node>) -> Option<NodeRef<'_, Node>> {
-    let mut current = text_node.parent();
-    while let Some(node) = current {
-        if let Node::Element(e) = node.value()
-            && is_block_element(e.name.local.as_ref())
-        {
-            return Some(node);
-        }
-        current = node.parent();
-    }
-    None
-}
-
-fn is_block_element(name: &str) -> bool {
-    matches!(
-        name,
-        "p" | "div"
-            | "li"
-            | "td"
-            | "th"
-            | "h1"
-            | "h2"
-            | "h3"
-            | "h4"
-            | "h5"
-            | "h6"
-            | "blockquote"
-            | "dt"
-            | "dd"
-            | "section"
-            | "article"
-            | "aside"
-            | "header"
-            | "footer"
-            | "main"
-            | "nav"
-            | "figcaption"
-            | "caption"
-    )
-}
-
-/// Don't emit text-node items for text that lives directly inside
-/// a `<script>` or `<style>` element — those bodies are handled by
-/// the script / style policy on the parent element instead. The
-/// `text_index` counter still advances so encode's
-/// document-order index lines up with decode.
-fn skip_text_under(text_node: NodeRef<'_, Node>) -> bool {
-    text_node
-        .parent()
-        .and_then(|p| p.value().as_element())
-        .map(|e| {
-            let name = e.name.local.as_ref();
-            matches!(name, "script" | "style")
-        })
-        .unwrap_or(false)
-}
-
-fn first_child_text(node: NodeRef<'_, Node>) -> Option<String> {
-    let child = node.first_child()?;
-    match child.value() {
-        Node::Text(t) => Some(t.text.to_string()),
-        _ => None,
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/text/json_handler.rs b/crates/nvisy-codec/src/handler/text/json_handler.rs
deleted file mode 100644
index f99ff666d..000000000
--- a/crates/nvisy-codec/src/handler/text/json_handler.rs
+++ /dev/null
@@ -1,953 +0,0 @@
-//! JSON handler: a flat ordered sequence of source slots.
-//!
-//! The loader lexes the source once into [`Slot`]s — either
-//! [`Slot::Passthrough`] (whitespace + structural punctuation,
-//! kept verbatim) or [`Slot::Leaf`] (a key, string value, or
-//! scalar). Leaves carry both the original source bytes
-//! (`serialized`) and the unescaped UTF-8 value the recognizer
-//! sees (`value`). [`Handler::next_chunk`] yields leaves in
-//! document order; [`Handler::redact`] mutates the leaf's
-//! value and re-renders its serialized form; [`Handler::encode`]
-//! concatenates every slot.
-//!
-//! This keeps formatting (indentation, key order, comment-free
-//! whitespace) byte-identical to the source for any slot the
-//! caller didn't touch, and reduces the partial-leaf offset
-//! translation to a single per-leaf walk of its escape table.
-
-use std::ops::Range;
-
-use nvisy_core::Error;
-use nvisy_core::modality::{Text, TextData, TextLocation};
-use nvisy_core::redaction::Redactions;
-
-use super::redact;
-use crate::content::{ContentData, ContentSource};
-use crate::{Chunk, Format, FormatId, Handler};
-
-const TARGET: &str = "nvisy_codec::handler::text::json";
-
-/// Stable [`FormatId`] for the JSON codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.text.json");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Text, _>(FORMAT_ID.clone(), super::JsonLoader::default())
-        .with_extensions(["json"])
-        .with_content_types(["application/json"])
-}
-
-/// One element of the parsed source.
-#[derive(Debug, Clone)]
-pub(super) enum Slot {
-    /// Whitespace or structural punctuation (`{ } [ ] : ,` and
-    /// surrounding whitespace). Held verbatim and emitted back
-    /// unchanged.
-    Passthrough(String),
-    /// A key, string value, or scalar (number/bool/null) — every
-    /// position a recognizer is allowed to address.
-    Leaf(Leaf),
-}
-
-/// An addressable position in the document.
-#[derive(Debug, Clone)]
-pub(super) struct Leaf {
-    pub kind: LeafKind,
-    /// Current unescaped UTF-8 value — what the recognizer sees
-    /// in [`Chunk::data`] and what redactions edit.
-    pub value: String,
-    /// Current source bytes — what `encode` emits and what
-    /// [`TextLocation`] offsets address. For [`LeafKind::Key`]
-    /// and [`LeafKind::StringValue`] this is the quoted form
-    /// `"…"` with `\\` / `\"` escapes; for [`LeafKind::Scalar`]
-    /// it is the bare literal.
-    pub serialized: String,
-    /// Out-of-band context strings (currently the enclosing
-    /// object key) surfaced to recognizers as hints; empty for
-    /// keys and for value leaves outside any object (e.g. a
-    /// top-level scalar).
-    pub hints: Vec<String>,
-}
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub(super) enum LeafKind {
-    Key,
-    StringValue,
-    Scalar,
-}
-
-impl Leaf {
-    fn is_quoted(&self) -> bool {
-        matches!(self.kind, LeafKind::Key | LeafKind::StringValue)
-    }
-
-    fn render(&mut self) {
-        self.serialized = match self.kind {
-            LeafKind::Key | LeafKind::StringValue => format!("\"{}\"", json_escape(&self.value)),
-            LeafKind::Scalar => self.value.clone(),
-        };
-    }
-}
-
-/// Handler for loaded JSON content.
-#[derive(Debug)]
-pub struct JsonHandler {
-    source: ContentSource,
-    slots: Vec<Slot>,
-    cursor: usize,
-}
-
-#[async_trait::async_trait]
-impl Handler<Text> for JsonHandler {
-    fn format(&self) -> FormatId {
-        FORMAT_ID.clone()
-    }
-
-    fn source(&self) -> ContentSource {
-        self.source
-    }
-
-    #[tracing::instrument(name = "json.encode", skip_all, fields(output_bytes))]
-    fn encode(&self) -> Result<ContentData, Error> {
-        let mut out = String::new();
-        for slot in &self.slots {
-            match slot {
-                Slot::Passthrough(text) => out.push_str(text),
-                Slot::Leaf(leaf) => out.push_str(&leaf.serialized),
-            }
-        }
-        tracing::Span::current().record("output_bytes", out.len());
-        let source = ContentSource::new().with_parent(&self.source);
-        Ok(ContentData::new(source, out.into_bytes().into()))
-    }
-
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<Text>>, Error> {
-        while self.cursor < self.slots.len() {
-            let start = self.offset_of(self.cursor);
-            let slot = &self.slots[self.cursor];
-            self.cursor += 1;
-            if let Slot::Leaf(leaf) = slot {
-                return Ok(Some(Chunk {
-                    location: TextLocation {
-                        start,
-                        end: start + leaf.serialized.len(),
-                        ..Default::default()
-                    },
-                    data: TextData::from(leaf.value.as_str()),
-                    hints: leaf.hints.clone(),
-                }));
-            }
-        }
-        Ok(None)
-    }
-
-    fn lift_chunk(&self, chunk: &Chunk<Text>, value_range: Range<usize>) -> Option<TextLocation> {
-        let (idx, leaf) = self.find_leaf(&chunk.location)?;
-        let slot_start = self.offset_of(idx);
-        let source_start = value_to_source_offset(leaf, slot_start, value_range.start)?;
-        let source_end = value_to_source_offset(leaf, slot_start, value_range.end)?;
-        Some(TextLocation {
-            start: source_start,
-            end: source_end,
-            context: chunk.location.context,
-            page_number: chunk.location.page_number,
-        })
-    }
-
-    async fn read(&self, location: &TextLocation) -> Result<Option<TextData>, Error> {
-        Ok(self
-            .find_leaf(location)
-            .map(|(_, leaf)| TextData::from(leaf.value.as_str())))
-    }
-
-    async fn redact(&mut self, redactions: Redactions<Text>) -> Result<(), Error> {
-        // Resolve every redaction against the **pre-mutation** slot
-        // offsets first. Mutating a leaf shifts every later slot's
-        // source-byte offset, so resolving inline would mismatch
-        // later locations against the live (already-shifted) slot
-        // table. The plan stores per-leaf value-byte ranges, which
-        // stay valid regardless of how other slots change length.
-        let mut plan: Vec<(usize, usize, usize, String)> = Vec::new();
-        let mut slot_offset = 0usize;
-        let mut slot_iter = self.slots.iter().enumerate().peekable();
-        let mut items = redactions.into_items();
-        items.sort_by_key(|(loc, _)| loc.start);
-        for (loc, replacement) in items {
-            // Advance the slot cursor to the slot containing `loc`.
-            // Slot offsets are monotonic, so a single forward sweep
-            // resolves every redaction in O(slots + redactions).
-            while let Some(&(idx, slot)) = slot_iter.peek() {
-                let len = match slot {
-                    Slot::Passthrough(t) => t.len(),
-                    Slot::Leaf(l) => l.serialized.len(),
-                };
-                let slot_end = slot_offset + len;
-                if loc.start < slot_end {
-                    if let Slot::Leaf(leaf) = slot
-                        && let Some((value_start, value_end)) =
-                            translate_to_value(leaf, slot_offset, loc.start, loc.end)
-                    {
-                        let value = replacement
-                            .replacement_value()
-                            .unwrap_or_default()
-                            .to_owned();
-                        plan.push((idx, value_start, value_end, value));
-                    }
-                    break;
-                }
-                slot_offset = slot_end;
-                slot_iter.next();
-            }
-        }
-        // Apply per-leaf edits right-to-left within each leaf so
-        // earlier edits in the same leaf don't shift later ones.
-        plan.sort_by(|a, b| a.0.cmp(&b.0).then(b.1.cmp(&a.1)));
-        for (idx, value_start, value_end, value) in plan {
-            let Slot::Leaf(leaf) = &mut self.slots[idx] else {
-                continue;
-            };
-            redact::replace_range(&mut leaf.value, &value, value_start..value_end, TARGET)?;
-            leaf.render();
-        }
-        self.cursor = 0;
-        Ok(())
-    }
-}
-
-impl JsonHandler {
-    /// Build a handler from a synthetic [`serde_json::Value`].
-    /// Synthetic documents always re-emit compact JSON — loaded
-    /// documents preserve their source formatting via the slot
-    /// model.
-    pub fn from_value(value: serde_json::Value) -> Self {
-        let serialized = serde_json::to_string(&value).unwrap_or_default();
-        Self::from_source_string(serialized)
-    }
-
-    /// Build a handler directly from JSON source bytes. Used by
-    /// the loader; preserves the source formatting verbatim.
-    pub(super) fn from_source_string(source: String) -> Self {
-        let slots = parse_slots(&source).unwrap_or_else(|_| vec![Slot::Passthrough(source)]);
-        Self {
-            source: ContentSource::new(),
-            slots,
-            cursor: 0,
-        }
-    }
-
-    /// Attach a content source for lineage tracking.
-    pub fn with_source(mut self, source: ContentSource) -> Self {
-        self.source = source;
-        self
-    }
-
-    /// Rewind the streaming cursor to the start of the document.
-    pub fn rewind(&mut self) {
-        self.cursor = 0;
-    }
-
-    /// Byte offset where the slot at `idx` starts in the current
-    /// encoded output.
-    fn offset_of(&self, idx: usize) -> usize {
-        self.slots[..idx]
-            .iter()
-            .map(|s| match s {
-                Slot::Passthrough(t) => t.len(),
-                Slot::Leaf(l) => l.serialized.len(),
-            })
-            .sum()
-    }
-
-    /// Locate the leaf slot whose source range contains
-    /// `location`. Returns its index and a borrow.
-    fn find_leaf(&self, location: &TextLocation) -> Option<(usize, &Leaf)> {
-        let mut offset = 0usize;
-        for (idx, slot) in self.slots.iter().enumerate() {
-            let len = match slot {
-                Slot::Passthrough(t) => t.len(),
-                Slot::Leaf(l) => l.serialized.len(),
-            };
-            let slot_end = offset + len;
-            if let Slot::Leaf(leaf) = slot
-                && location.start >= offset
-                && location.end <= slot_end
-            {
-                return Some((idx, leaf));
-            }
-            offset = slot_end;
-        }
-        None
-    }
-}
-
-/// Escape a string for JSON matching (backslash and quote only —
-/// other control characters in keys/values are unsupported in
-/// this codec's redaction path and round-trip as-is).
-fn json_escape(s: &str) -> String {
-    s.replace('\\', "\\\\").replace('"', "\\\"")
-}
-
-/// Translate a `(source_start, source_end)` range expressed in
-/// the current encoded output into the corresponding
-/// `(value_start, value_end)` range inside `leaf.value`.
-///
-/// `slot_start` is the leaf's byte offset in the current output.
-/// For scalars the mapping is identity. For quoted leaves the
-/// boundary positions (opening/closing quote) map to the full
-/// value range; interior positions are translated by walking
-/// the `\\`/`\"` escape pairs.
-///
-/// Returns `None` when the requested boundary lands on a quote
-/// byte (outside the whole-leaf case) or in the middle of an
-/// escape pair.
-fn translate_to_value(
-    leaf: &Leaf,
-    slot_start: usize,
-    source_start: usize,
-    source_end: usize,
-) -> Option<(usize, usize)> {
-    let slot_end = slot_start + leaf.serialized.len();
-    if !leaf.is_quoted() {
-        return Some((source_start - slot_start, source_end - slot_start));
-    }
-    if source_start == slot_start && source_end == slot_end {
-        return Some((0, leaf.value.len()));
-    }
-    let escaped_start = slot_start + 1;
-    let escaped_end = slot_end - 1;
-    if source_start < escaped_start || source_end > escaped_end || source_start > source_end {
-        return None;
-    }
-    let bytes = leaf.serialized.as_bytes();
-    let mut src = escaped_start;
-    let mut val = 0usize;
-    let mut start = None;
-    let mut end = None;
-    while src <= escaped_end {
-        if src == source_start {
-            start = Some(val);
-        }
-        if src == source_end {
-            end = Some(val);
-            break;
-        }
-        // Index into leaf.serialized: `src - slot_start`.
-        let local = src - slot_start;
-        let is_escape = bytes.get(local) == Some(&b'\\');
-        if is_escape && src + 1 > escaped_end {
-            return None;
-        }
-        src += if is_escape { 2 } else { 1 };
-        val += 1;
-    }
-    Some((start?, end?))
-}
-
-/// Inverse of [`translate_to_value`]: map a value-byte offset
-/// inside `leaf.value` to the source byte offset inside the
-/// current encoded output.
-///
-/// `slot_start` is the leaf's source byte offset. For scalars the
-/// mapping is identity. For quoted leaves the value-byte cursor
-/// advances one for each interior source byte (or two source
-/// bytes when the next source byte is a `\` escape prefix).
-///
-/// Returns `None` if `value_offset` is past the end of the value.
-fn value_to_source_offset(leaf: &Leaf, slot_start: usize, value_offset: usize) -> Option<usize> {
-    if !leaf.is_quoted() {
-        if value_offset > leaf.value.len() {
-            return None;
-        }
-        return Some(slot_start + value_offset);
-    }
-    let escaped_start = slot_start + 1;
-    let escaped_end = slot_start + leaf.serialized.len() - 1;
-    let bytes = leaf.serialized.as_bytes();
-    let mut src = escaped_start;
-    let mut val = 0usize;
-    while val < value_offset {
-        if src >= escaped_end {
-            return None;
-        }
-        let local = src - slot_start;
-        let is_escape = bytes.get(local) == Some(&b'\\');
-        src += if is_escape { 2 } else { 1 };
-        val += 1;
-    }
-    Some(src)
-}
-
-/// Lex JSON source into a flat ordered slot list.
-///
-/// Whitespace and structural punctuation collapse into
-/// [`Slot::Passthrough`]; keys, string values and scalars
-/// become [`Slot::Leaf`]. Returns an error if the source is not
-/// well-formed JSON.
-pub(super) fn parse_slots(src: &str) -> Result<Vec<Slot>, Error> {
-    let mut p = SlotParser::new(src);
-    p.parse_value(None)?;
-    p.flush_passthrough();
-    p.consume_whitespace();
-    p.flush_passthrough();
-    if p.pos != src.len() {
-        return Err(Error::validation(
-            format!("trailing bytes after JSON value at offset {}", p.pos),
-            TARGET,
-        ));
-    }
-    Ok(p.slots)
-}
-
-struct SlotParser<'a> {
-    src: &'a str,
-    bytes: &'a [u8],
-    pos: usize,
-    /// Pending whitespace + structural bytes we haven't flushed
-    /// into a [`Slot::Passthrough`] yet.
-    pending: String,
-    slots: Vec<Slot>,
-}
-
-impl<'a> SlotParser<'a> {
-    fn new(src: &'a str) -> Self {
-        Self {
-            src,
-            bytes: src.as_bytes(),
-            pos: 0,
-            pending: String::new(),
-            slots: Vec::new(),
-        }
-    }
-
-    fn flush_passthrough(&mut self) {
-        if !self.pending.is_empty() {
-            self.slots
-                .push(Slot::Passthrough(std::mem::take(&mut self.pending)));
-        }
-    }
-
-    fn push_leaf(&mut self, leaf: Leaf) {
-        self.flush_passthrough();
-        self.slots.push(Slot::Leaf(leaf));
-    }
-
-    fn peek(&self) -> Option<u8> {
-        self.bytes.get(self.pos).copied()
-    }
-
-    fn consume_whitespace(&mut self) {
-        while let Some(b) = self.peek() {
-            if b == b' ' || b == b'\t' || b == b'\n' || b == b'\r' {
-                self.pending.push(b as char);
-                self.pos += 1;
-            } else {
-                break;
-            }
-        }
-    }
-
-    fn consume_punct(&mut self, c: u8) -> Result<(), Error> {
-        if self.peek() != Some(c) {
-            return Err(Error::validation(
-                format!("expected {:?} at offset {}", c as char, self.pos),
-                TARGET,
-            ));
-        }
-        self.pending.push(c as char);
-        self.pos += 1;
-        Ok(())
-    }
-
-    fn parse_value(&mut self, key_context: Option<&str>) -> Result<(), Error> {
-        self.consume_whitespace();
-        match self.peek() {
-            Some(b'{') => self.parse_object(),
-            Some(b'[') => self.parse_array(key_context),
-            Some(b'"') => {
-                let mut leaf = self.parse_string_leaf(LeafKind::StringValue)?;
-                if let Some(k) = key_context {
-                    leaf.hints.push(k.to_owned());
-                }
-                self.push_leaf(leaf);
-                Ok(())
-            }
-            Some(b't') | Some(b'f') | Some(b'n') | Some(b'-') | Some(b'0'..=b'9') => {
-                let mut leaf = self.parse_scalar()?;
-                if let Some(k) = key_context {
-                    leaf.hints.push(k.to_owned());
-                }
-                self.push_leaf(leaf);
-                Ok(())
-            }
-            Some(b) => Err(Error::validation(
-                format!("unexpected byte {b:#x} at offset {}", self.pos),
-                TARGET,
-            )),
-            None => Err(Error::validation(
-                "unexpected end of input".to_string(),
-                TARGET,
-            )),
-        }
-    }
-
-    fn parse_object(&mut self) -> Result<(), Error> {
-        self.consume_punct(b'{')?;
-        self.consume_whitespace();
-        if self.peek() == Some(b'}') {
-            self.consume_punct(b'}')?;
-            return Ok(());
-        }
-        loop {
-            self.consume_whitespace();
-            let key = self.parse_string_leaf(LeafKind::Key)?;
-            let key_value = key.value.clone();
-            self.push_leaf(key);
-            self.consume_whitespace();
-            self.consume_punct(b':')?;
-            self.parse_value(Some(&key_value))?;
-            self.consume_whitespace();
-            match self.peek() {
-                Some(b',') => {
-                    self.consume_punct(b',')?;
-                }
-                Some(b'}') => {
-                    self.consume_punct(b'}')?;
-                    return Ok(());
-                }
-                _ => {
-                    return Err(Error::validation(
-                        format!("expected ',' or '}}' at offset {}", self.pos),
-                        TARGET,
-                    ));
-                }
-            }
-        }
-    }
-
-    fn parse_array(&mut self, key_context: Option<&str>) -> Result<(), Error> {
-        self.consume_punct(b'[')?;
-        self.consume_whitespace();
-        if self.peek() == Some(b']') {
-            self.consume_punct(b']')?;
-            return Ok(());
-        }
-        loop {
-            // Array elements inherit the containing object key as
-            // their hint — `{"cards": ["4111…", "5555…"]}` should
-            // treat both PANs as living under `cards`.
-            self.parse_value(key_context)?;
-            self.consume_whitespace();
-            match self.peek() {
-                Some(b',') => {
-                    self.consume_punct(b',')?;
-                }
-                Some(b']') => {
-                    self.consume_punct(b']')?;
-                    return Ok(());
-                }
-                _ => {
-                    return Err(Error::validation(
-                        format!("expected ',' or ']' at offset {}", self.pos),
-                        TARGET,
-                    ));
-                }
-            }
-        }
-    }
-
-    fn parse_string_leaf(&mut self, kind: LeafKind) -> Result<Leaf, Error> {
-        if self.peek() != Some(b'"') {
-            return Err(Error::validation(
-                format!("expected '\"' at offset {}", self.pos),
-                TARGET,
-            ));
-        }
-        let start = self.pos;
-        self.pos += 1;
-        let mut value = String::new();
-        loop {
-            match self.peek() {
-                Some(b'"') => {
-                    self.pos += 1;
-                    let serialized = self.src[start..self.pos].to_string();
-                    return Ok(Leaf {
-                        kind,
-                        value,
-                        serialized,
-                        hints: Vec::new(),
-                    });
-                }
-                Some(b'\\') => {
-                    let next = self.bytes.get(self.pos + 1).copied();
-                    match next {
-                        Some(b'"') => value.push('"'),
-                        Some(b'\\') => value.push('\\'),
-                        Some(b'/') => value.push('/'),
-                        Some(b'n') => value.push('\n'),
-                        Some(b'r') => value.push('\r'),
-                        Some(b't') => value.push('\t'),
-                        Some(b'b') => value.push('\u{0008}'),
-                        Some(b'f') => value.push('\u{000c}'),
-                        Some(b'u') => {
-                            return Err(Error::validation(
-                                format!(
-                                    "JSON \\u escapes are not supported in redaction codec \
-                                     (offset {})",
-                                    self.pos
-                                ),
-                                TARGET,
-                            ));
-                        }
-                        Some(other) => {
-                            return Err(Error::validation(
-                                format!(
-                                    "invalid escape \\{} at offset {}",
-                                    other as char, self.pos
-                                ),
-                                TARGET,
-                            ));
-                        }
-                        None => {
-                            return Err(Error::validation(
-                                "unterminated escape at end of input".to_string(),
-                                TARGET,
-                            ));
-                        }
-                    }
-                    self.pos += 2;
-                }
-                Some(_) => {
-                    let ch_start = self.pos;
-                    // Advance one UTF-8 codepoint without reading the
-                    // escape table.
-                    let rest = &self.src[ch_start..];
-                    let ch = rest.chars().next().ok_or_else(|| {
-                        Error::validation("unterminated string".to_string(), TARGET)
-                    })?;
-                    value.push(ch);
-                    self.pos += ch.len_utf8();
-                }
-                None => {
-                    return Err(Error::validation("unterminated string".to_string(), TARGET));
-                }
-            }
-        }
-    }
-
-    fn parse_scalar(&mut self) -> Result<Leaf, Error> {
-        let start = self.pos;
-        while let Some(b) = self.peek() {
-            let is_scalar_byte =
-                b.is_ascii_alphanumeric() || matches!(b, b'-' | b'+' | b'.' | b'_');
-            if is_scalar_byte {
-                self.pos += 1;
-            } else {
-                break;
-            }
-        }
-        if start == self.pos {
-            return Err(Error::validation(
-                format!("expected scalar at offset {start}"),
-                TARGET,
-            ));
-        }
-        let literal = self.src[start..self.pos].to_string();
-        Ok(Leaf {
-            kind: LeafKind::Scalar,
-            value: literal.clone(),
-            serialized: literal,
-            hints: Vec::new(),
-        })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::Error;
-    use nvisy_core::redaction::TextReplacement;
-
-    use super::*;
-
-    fn handler(src: &str) -> JsonHandler {
-        JsonHandler::from_source_string(src.to_string())
-    }
-
-    #[tokio::test]
-    async fn stream_yields_keys_and_values_in_order() -> Result<(), Error> {
-        let mut h = handler(r#"{"name":"Alice","age":30}"#);
-        let mut chunks = Vec::new();
-        while let Some(c) = h.next_chunk().await? {
-            chunks.push(c.data.as_str().to_owned());
-        }
-        assert_eq!(chunks, vec!["name", "Alice", "age", "30"]);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn duplicate_values_get_distinct_offsets() -> Result<(), Error> {
-        let mut h = handler(r#"{"a":"same","b":"same"}"#);
-        let mut offsets = Vec::new();
-        while let Some(c) = h.next_chunk().await? {
-            if c.data.as_str() == "same" {
-                offsets.push(c.location.start);
-            }
-        }
-        assert_eq!(offsets.len(), 2);
-        assert_ne!(offsets[0], offsets[1]);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn read_returns_string() -> Result<(), Error> {
-        let mut h = handler(r#"{"name":"Alice"}"#);
-        let mut found = false;
-        while let Some(chunk) = h.next_chunk().await? {
-            if h.read(&chunk.location)
-                .await?
-                .map(|d| d.as_str().to_owned())
-                == Some("Alice".to_owned())
-            {
-                found = true;
-            }
-        }
-        assert!(found);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn read_rejects_arbitrary_offsets() -> Result<(), Error> {
-        let h = handler(r#"{"name":"Alice"}"#);
-        let bogus = TextLocation {
-            start: 3,
-            end: 7,
-            ..Default::default()
-        };
-        // 3..7 covers `name` mid-token plus a closing quote — not a
-        // whole-leaf hit, but it does fall inside the key span. read
-        // returns the key's unescaped value.
-        let result = h.read(&bogus).await?.map(|d| d.as_str().to_owned());
-        assert_eq!(result, Some("name".to_owned()));
-        Ok(())
-    }
-
-    #[test]
-    fn encode_preserves_source_compact() -> Result<(), Error> {
-        let src = r#"{"a":1}"#;
-        let h = handler(src);
-        assert_eq!(h.encode()?.as_str().unwrap(), src);
-        Ok(())
-    }
-
-    #[test]
-    fn encode_preserves_source_pretty() -> Result<(), Error> {
-        let src = "{\n  \"a\": 1\n}\n";
-        let h = handler(src);
-        assert_eq!(h.encode()?.as_str().unwrap(), src);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_whole_string_value() -> Result<(), Error> {
-        let mut h = handler(r#"{"name":"Alice"}"#);
-        let chunk = loop {
-            let c = h.next_chunk().await?.expect("expected chunk");
-            if c.data.as_str() == "Alice" {
-                break c;
-            }
-        };
-        let mut rs = Redactions::new();
-        rs.push(chunk.location.clone(), TextReplacement::substituted("Bob"));
-        h.redact(rs).await?;
-        let encoded = h.encode()?.as_str().unwrap().to_owned();
-        assert_eq!(encoded, r#"{"name":"Bob"}"#);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_partial_leaf_in_compact_source() -> Result<(), Error> {
-        let src = r#"{"email":"alice@example.com"}"#;
-        let mut h = handler(src);
-        let _ = loop {
-            let c = h.next_chunk().await?.expect("chunk");
-            if c.data.as_str() == "alice@example.com" {
-                break c;
-            }
-        };
-        let local_start = src.find("alice").unwrap();
-        let local_end = local_start + "alice".len();
-        let mut rs = Redactions::new();
-        rs.push(
-            TextLocation::new(local_start, local_end),
-            TextReplacement::substituted("[USER]"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(
-            h.encode()?.as_str().unwrap(),
-            r#"{"email":"[USER]@example.com"}"#
-        );
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_partial_leaf_with_escapes() -> Result<(), Error> {
-        // unescaped value: foo"bar — source: "foo\"bar"
-        let src = r#"{"msg":"foo\"bar"}"#;
-        let mut h = handler(src);
-        let _ = loop {
-            let c = h.next_chunk().await?.expect("chunk");
-            if c.data.as_str() == r#"foo"bar"# {
-                break c;
-            }
-        };
-        let local_start = src.find("bar").unwrap();
-        let local_end = local_start + "bar".len();
-        let mut rs = Redactions::new();
-        rs.push(
-            TextLocation::new(local_start, local_end),
-            TextReplacement::substituted("XXX"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.encode()?.as_str().unwrap(), r#"{"msg":"foo\"XXX"}"#);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_key() -> Result<(), Error> {
-        let mut h = handler(r#"{"email":"a@b.c"}"#);
-        let chunk = loop {
-            let c = h.next_chunk().await?.expect("chunk");
-            if c.data.as_str() == "email" {
-                break c;
-            }
-        };
-        let mut rs = Redactions::new();
-        rs.push(
-            chunk.location.clone(),
-            TextReplacement::substituted("contact"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.encode()?.as_str().unwrap(), r#"{"contact":"a@b.c"}"#);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_scalar() -> Result<(), Error> {
-        let mut h = handler(r#"{"n":42}"#);
-        let chunk = loop {
-            let c = h.next_chunk().await?.expect("chunk");
-            if c.data.as_str() == "42" {
-                break c;
-            }
-        };
-        let mut rs = Redactions::new();
-        rs.push(chunk.location.clone(), TextReplacement::substituted("0"));
-        h.redact(rs).await?;
-        assert_eq!(h.encode()?.as_str().unwrap(), r#"{"n":0}"#);
-        Ok(())
-    }
-
-    /// Multiple redactions in a single batch, each targeting a
-    /// different leaf, with length deltas that shift later slot
-    /// offsets. Regression test for the "only the first redaction
-    /// lands" bug: locations must resolve against pre-mutation
-    /// slot offsets so later locations still find their slot after
-    /// earlier slots changed length.
-    #[tokio::test]
-    async fn redact_multiple_leaves_with_shifting_offsets() -> Result<(), Error> {
-        let src = r#"{"a":"first","b":"second","c":"third"}"#;
-        let mut h = handler(src);
-        let mut locs = Vec::new();
-        while let Some(c) = h.next_chunk().await? {
-            let v = c.data.as_str();
-            if v == "first" || v == "second" || v == "third" {
-                locs.push(c.location);
-            }
-        }
-        assert_eq!(locs.len(), 3, "expected three string values");
-        let mut rs = Redactions::new();
-        for loc in locs {
-            rs.push(loc, TextReplacement::substituted("X"));
-        }
-        h.redact(rs).await?;
-        assert_eq!(
-            h.encode()?.as_str().unwrap(),
-            r#"{"a":"X","b":"X","c":"X"}"#
-        );
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn lift_chunk_simple_string() -> Result<(), Error> {
-        let src = r#"{"email":"alice@example.com"}"#;
-        let mut h = handler(src);
-        let chunk = loop {
-            let c = h.next_chunk().await?.expect("chunk");
-            if c.data.as_str() == "alice@example.com" {
-                break c;
-            }
-        };
-        // Recognizer sees the unescaped value "alice@example.com" and
-        // emits an offset against it. lift_chunk must turn that into
-        // a source-byte TextLocation.
-        let value_start = "alice@example.com".find("alice").unwrap();
-        let value_end = value_start + "alice".len();
-        let source_loc = h
-            .lift_chunk(&chunk, value_start..value_end)
-            .expect("range is in bounds");
-        let expected_start = src.find("alice").unwrap();
-        assert_eq!(source_loc.start, expected_start);
-        assert_eq!(source_loc.end, expected_start + "alice".len());
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn lift_chunk_walks_escapes() -> Result<(), Error> {
-        // unescaped value: foo"bar — source: "foo\"bar"
-        let src = r#"{"msg":"foo\"bar"}"#;
-        let mut h = handler(src);
-        let chunk = loop {
-            let c = h.next_chunk().await?.expect("chunk");
-            if c.data.as_str() == r#"foo"bar"# {
-                break c;
-            }
-        };
-        // Target the 'bar' substring as the recognizer would see it
-        // — in unescaped-value coords its start is 4 (after foo").
-        let value = chunk.data.as_str();
-        let value_start = value.find("bar").unwrap();
-        let value_end = value_start + "bar".len();
-        let source_loc = h
-            .lift_chunk(&chunk, value_start..value_end)
-            .expect("range is in bounds");
-        let expected_start = src.find("bar").unwrap();
-        assert_eq!(source_loc.start, expected_start);
-        assert_eq!(source_loc.end, expected_start + "bar".len());
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn lift_chunk_redact_roundtrip() -> Result<(), Error> {
-        // End-to-end: pull a chunk, look up a sub-range in value
-        // coords, lift to source coords, push redaction through.
-        // Asserts that lift_chunk and redact agree on coordinates.
-        let src = r#"{"msg":"foo\"bar"}"#;
-        let mut h = handler(src);
-        let chunk = loop {
-            let c = h.next_chunk().await?.expect("chunk");
-            if c.data.as_str() == r#"foo"bar"# {
-                break c;
-            }
-        };
-        let value = chunk.data.as_str();
-        let value_start = value.find("bar").unwrap();
-        let value_end = value_start + "bar".len();
-        let source_loc = h
-            .lift_chunk(&chunk, value_start..value_end)
-            .expect("range is in bounds");
-        let mut rs = Redactions::new();
-        rs.push(source_loc, TextReplacement::substituted("XXX"));
-        h.redact(rs).await?;
-        assert_eq!(h.encode()?.as_str().unwrap(), r#"{"msg":"foo\"XXX"}"#);
-        Ok(())
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/text/json_loader.rs b/crates/nvisy-codec/src/handler/text/json_loader.rs
deleted file mode 100644
index 8296cfc8f..000000000
--- a/crates/nvisy-codec/src/handler/text/json_loader.rs
+++ /dev/null
@@ -1,38 +0,0 @@
-//! JSON loader: decode source bytes and hand them to
-//! [`JsonHandler`] verbatim. Formatting (indentation, key order,
-//! trailing whitespace) is preserved by the handler's slot model;
-//! the loader only does encoding + well-formedness checks.
-
-use nvisy_core::Error;
-use nvisy_core::modality::Text;
-
-use super::JsonHandler;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource, TextEncoding};
-
-/// Loader for JSON files. Produces one [`JsonHandler`] per input.
-#[derive(Debug, Default)]
-pub struct JsonLoader {
-    /// Character encoding of the input bytes. Defaults to UTF-8.
-    pub encoding: TextEncoding,
-}
-
-#[async_trait::async_trait]
-impl Loader<Text> for JsonLoader {
-    type Handler = JsonHandler;
-
-    #[tracing::instrument(name = "json.decode", skip_all, fields(input_bytes))]
-    async fn decode(&self, content: ContentData) -> Result<JsonHandler, Error> {
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        tracing::Span::current().record("input_bytes", raw.len());
-        let text = self.encoding.decode_bytes(&raw, "json-loader")?;
-        // Validate well-formedness eagerly; the handler's lexer
-        // re-parses but with a friendlier error path. Reject here so
-        // callers get a single decode-time validation point.
-        serde_json::from_str::<serde_json::Value>(&text)
-            .map_err(|e| Error::validation(format!("Invalid JSON: {e}"), "json-loader"))?;
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(JsonHandler::from_source_string(text).with_source(source))
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/text/markdown_loader.rs b/crates/nvisy-codec/src/handler/text/markdown_loader.rs
deleted file mode 100644
index b0f486ec5..000000000
--- a/crates/nvisy-codec/src/handler/text/markdown_loader.rs
+++ /dev/null
@@ -1,50 +0,0 @@
-//! Markdown loader: validates and parses raw Markdown content into a
-//! [`TxtHandler`].
-//!
-//! Markdown is structurally line-based text. The loader reuses
-//! [`TxtHandler`] for span access and editing; the format distinction
-//! is carried by [`FORMAT_ID`] on the [`Format`] descriptor so
-//! downstream code can apply Markdown-aware processing when needed.
-
-use nvisy_core::Error;
-use nvisy_core::modality::Text;
-
-use super::TxtHandler;
-use crate::content::{ContentData, ContentSource, TextEncoding};
-use crate::{Format, FormatId, Loader};
-
-/// Stable [`FormatId`] for the Markdown codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.text.markdown");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Text, _>(FORMAT_ID.clone(), MdLoader::default())
-        .with_extensions(["md", "markdown"])
-        .with_content_types(["text/markdown"])
-}
-
-/// Loader for Markdown files. Produces a [`TxtHandler`] per input.
-#[derive(Debug, Default)]
-pub struct MdLoader {
-    /// Character encoding of the input bytes. Defaults to UTF-8.
-    pub encoding: TextEncoding,
-}
-
-#[async_trait::async_trait]
-impl Loader<Text> for MdLoader {
-    type Handler = TxtHandler;
-
-    #[tracing::instrument(name = "markdown.decode", skip_all, fields(input_bytes, lines))]
-    async fn decode(&self, content: ContentData) -> Result<TxtHandler, Error> {
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        tracing::Span::current().record("input_bytes", raw.len());
-        let text = self.encoding.decode_bytes(&raw, "markdown-loader")?;
-        let trailing_newline = text.ends_with('\n');
-        let lines: Vec<String> = text.lines().map(String::from).collect();
-        tracing::Span::current().record("lines", lines.len());
-
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(TxtHandler::new(lines, trailing_newline).with_source(source))
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/text/mod.rs b/crates/nvisy-codec/src/handler/text/mod.rs
deleted file mode 100644
index f40293347..000000000
--- a/crates/nvisy-codec/src/handler/text/mod.rs
+++ /dev/null
@@ -1,45 +0,0 @@
-//! Text modality: concrete format implementations that produce
-//! text handles (TXT, JSON, Markdown, HTML). The per-modality
-//! capability surface lives on the generic [`Handler<Text>`] trait
-//! re-exported at the crate root; replacements written during
-//! [`Handler::redact`] use [`TextReplacement`].
-//!
-//! [`Handler<Text>`]: crate::Handler
-//! [`Handler::redact`]: crate::Handler::redact
-//! [`TextReplacement`]: nvisy_core::redaction::TextReplacement
-
-pub(crate) mod redact;
-
-#[cfg(feature = "html")]
-mod html_encode;
-#[cfg(feature = "html")]
-mod html_handler;
-#[cfg(feature = "html")]
-mod html_loader;
-#[cfg(feature = "json")]
-mod json_handler;
-#[cfg(feature = "json")]
-mod json_loader;
-#[cfg(feature = "markdown")]
-mod markdown_loader;
-#[cfg(feature = "txt")]
-mod txt_handler;
-#[cfg(feature = "txt")]
-mod txt_loader;
-
-#[cfg(feature = "html")]
-pub use self::html_handler::{
-    ElementTarget, HtmlData, HtmlHandler, RedactableItem, RedactableKind, format as html_format,
-};
-#[cfg(feature = "html")]
-pub use self::html_loader::{HtmlLoader, ScriptPolicy};
-#[cfg(feature = "json")]
-pub use self::json_handler::{JsonHandler, format as json_format};
-#[cfg(feature = "json")]
-pub use self::json_loader::JsonLoader;
-#[cfg(feature = "markdown")]
-pub use self::markdown_loader::{MdLoader, format as markdown_format};
-#[cfg(feature = "txt")]
-pub use self::txt_handler::{TxtHandler, format as txt_format};
-#[cfg(feature = "txt")]
-pub use self::txt_loader::TxtLoader;
diff --git a/crates/nvisy-codec/src/handler/text/redact.rs b/crates/nvisy-codec/src/handler/text/redact.rs
deleted file mode 100644
index 897d50846..000000000
--- a/crates/nvisy-codec/src/handler/text/redact.rs
+++ /dev/null
@@ -1,76 +0,0 @@
-//! Shared byte-level string redaction helper used by every
-//! text-family handler (TXT, JSON, HTML, the per-page text inside
-//! the PDF rich handler) plus the tabular cell handlers (CSV, XLSX
-//! cells are flat strings).
-//!
-//! Replaces `buf[range]` with `value`. Clamps the range endpoints
-//! to `buf.len()`; errors when either endpoint falls mid-character.
-
-use std::ops::Range;
-
-use nvisy_core::Error;
-
-/// Replace `buf[range]` with `value` in place.
-///
-/// Returns an error if either endpoint falls mid-character. The
-/// range endpoints are clamped to `buf.len()`; an empty range is
-/// a no-op.
-pub(crate) fn replace_range(
-    buf: &mut String,
-    value: &str,
-    range: Range<usize>,
-    target: &'static str,
-) -> Result<(), Error> {
-    let s = range.start.min(buf.len());
-    let e = range.end.min(buf.len());
-    if s >= e {
-        return Ok(());
-    }
-    if !buf.is_char_boundary(s) || !buf.is_char_boundary(e) {
-        return Err(Error::validation(
-            format!(
-                "redaction offset falls mid-character \
-                 (start={}, end={}, len={})",
-                range.start,
-                range.end,
-                buf.len()
-            ),
-            target,
-        ));
-    }
-    buf.replace_range(s..e, value);
-    Ok(())
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn single_replacement() {
-        let mut s = String::from("hello world");
-        replace_range(&mut s, "[X]", 0..5, "test").unwrap();
-        assert_eq!(s, "[X] world");
-    }
-
-    #[test]
-    fn remove_empty_value() {
-        let mut s = String::from("hello world");
-        replace_range(&mut s, "", 5..11, "test").unwrap();
-        assert_eq!(s, "hello");
-    }
-
-    #[test]
-    fn out_of_bounds_clipped() {
-        let mut s = String::from("short");
-        replace_range(&mut s, "[X]", 0..999, "test").unwrap();
-        assert_eq!(s, "[X]");
-    }
-
-    #[test]
-    fn mid_character_rejected() {
-        let mut s = String::from("héllo"); // 'é' is 2 bytes
-        let err = replace_range(&mut s, "[X]", 0..2, "test").unwrap_err();
-        assert!(err.to_string().contains("mid-character"));
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/text/txt_handler.rs b/crates/nvisy-codec/src/handler/text/txt_handler.rs
deleted file mode 100644
index ed5b6e2e6..000000000
--- a/crates/nvisy-codec/src/handler/text/txt_handler.rs
+++ /dev/null
@@ -1,446 +0,0 @@
-//! Plain-text handler: holds loaded text content and streams it
-//! line-by-line via [`Handler<Text>`], with random-access reads /
-//! redactions via [`Handler<Text>`].
-//!
-//! The handler stores the text as a vector of lines together with a
-//! trailing-newline flag so the original file can be reconstructed
-//! byte-for-byte after edits.
-
-use std::ops::Range;
-
-use nvisy_core::Error;
-use nvisy_core::modality::{Text, TextData, TextLocation};
-use nvisy_core::redaction::{Redactions, TextReplacement};
-
-use super::{TxtLoader, redact};
-use crate::content::{ContentData, ContentSource};
-use crate::{Chunk, Format, FormatId, Handler};
-
-const TARGET: &str = "nvisy_codec::handler::text::txt";
-
-/// Stable [`FormatId`] for the plain-text codec.
-pub const FORMAT_ID: FormatId = FormatId::from_static("nvisy.text.txt");
-
-/// [`Format`] descriptor registered into [`crate::CodecRegistry`].
-pub fn format() -> Format {
-    Format::new::<Text, _>(FORMAT_ID.clone(), TxtLoader::default())
-        .with_extensions(["txt", "log"])
-        .with_content_types(["text/plain"])
-}
-
-/// Handler for loaded plain-text content. Each line is independently
-/// addressable via [`TextLocation`].
-///
-/// `line_starts` is a cumulative-offset index maintained alongside
-/// `lines`: `line_starts[i]` is the byte position of line `i` in the
-/// serialized output, and `line_starts[lines.len()]` is the total
-/// length sentinel. Random-access [`Handler::read`] and
-/// [`Handler::redact`] resolve a byte offset to a line in
-/// `O(log N)` instead of rebuilding the table on every call.
-#[derive(Debug)]
-pub struct TxtHandler {
-    source: ContentSource,
-    lines: Vec<String>,
-    line_starts: Vec<usize>,
-    trailing_newline: bool,
-    cursor: usize,
-}
-
-#[async_trait::async_trait]
-impl Handler<Text> for TxtHandler {
-    fn format(&self) -> FormatId {
-        FORMAT_ID.clone()
-    }
-
-    fn source(&self) -> ContentSource {
-        self.source
-    }
-
-    #[tracing::instrument(name = "txt.encode", skip_all, fields(output_bytes))]
-    fn encode(&self) -> Result<ContentData, Error> {
-        let mut out = self.lines.join("\n");
-        if self.trailing_newline && !self.lines.is_empty() {
-            out.push('\n');
-        }
-        let bytes = out.into_bytes();
-        tracing::Span::current().record("output_bytes", bytes.len());
-        let source = ContentSource::new().with_parent(&self.source);
-        Ok(ContentData::new(source, bytes.into()))
-    }
-
-    async fn next_chunk(&mut self) -> Result<Option<Chunk<Text>>, Error> {
-        if self.cursor >= self.lines.len() {
-            return Ok(None);
-        }
-        let i = self.cursor;
-        let start = self.line_starts[i];
-        let end = self.line_starts[i + 1] - 1; // strip the implicit '\n' separator
-        let line = &self.lines[i];
-        self.cursor += 1;
-        Ok(Some(Chunk {
-            location: TextLocation {
-                start,
-                end,
-                ..Default::default()
-            },
-            data: TextData::from(line.as_str()),
-            hints: Vec::new(),
-        }))
-    }
-
-    fn lift_chunk(&self, chunk: &Chunk<Text>, value_range: Range<usize>) -> Option<TextLocation> {
-        chunk.location.subslice(value_range)
-    }
-
-    async fn read(&self, location: &TextLocation) -> Result<Option<TextData>, Error> {
-        let Some(i) = self.line_for(location.start) else {
-            return Ok(None);
-        };
-        let line_start = self.line_starts[i];
-        let line_end = self.line_starts[i + 1] - 1;
-        if location.end > line_end {
-            return Ok(None); // crosses a line boundary
-        }
-        let local_start = location.start - line_start;
-        let local_end = location.end - line_start;
-        Ok(self.lines[i]
-            .get(local_start..local_end)
-            .map(TextData::from))
-    }
-
-    async fn redact(&mut self, mut redactions: Redactions<Text>) -> Result<(), Error> {
-        // Apply right-to-left so each edit's length delta doesn't
-        // invalidate earlier locations.
-        redactions.sort_descending();
-        for (location, replacement) in redactions.into_items() {
-            self.redact_one(&location, replacement)?;
-        }
-        Ok(())
-    }
-}
-
-impl TxtHandler {
-    /// Create a new handler from lines and a trailing-newline flag.
-    pub fn new(lines: Vec<String>, trailing_newline: bool) -> Self {
-        let line_starts = compute_line_starts(&lines);
-        Self {
-            source: ContentSource::new(),
-            lines,
-            line_starts,
-            trailing_newline,
-            cursor: 0,
-        }
-    }
-
-    /// Attach a content source for lineage tracking.
-    pub fn with_source(mut self, source: ContentSource) -> Self {
-        self.source = source;
-        self
-    }
-
-    /// All lines in the document.
-    pub fn lines(&self) -> &[String] {
-        &self.lines
-    }
-
-    /// A specific line by 0-based index.
-    pub fn line(&self, index: usize) -> Option<&str> {
-        self.lines.get(index).map(String::as_str)
-    }
-
-    /// Whether the original source had a trailing newline.
-    pub fn trailing_newline(&self) -> bool {
-        self.trailing_newline
-    }
-
-    /// Total number of lines.
-    pub fn len(&self) -> usize {
-        self.lines.len()
-    }
-
-    /// Whether the document has no lines.
-    pub fn is_empty(&self) -> bool {
-        self.lines.is_empty()
-    }
-
-    /// Rewind the streaming cursor to the start of the document.
-    pub fn rewind(&mut self) {
-        self.cursor = 0;
-    }
-
-    /// Line index containing `byte_offset`, or `None` if the offset
-    /// is past the end of the document.
-    fn line_for(&self, byte_offset: usize) -> Option<usize> {
-        match self.line_starts.binary_search(&byte_offset) {
-            Ok(i) if i < self.lines.len() => Some(i),
-            Ok(_) => None, // landed on the trailing sentinel
-            Err(i) if i > 0 && i <= self.lines.len() => Some(i - 1),
-            _ => None,
-        }
-    }
-
-    /// Shift every `line_starts[j]` for `j > i` by `delta`. Called
-    /// after a redaction changes the length of line `i`.
-    fn shift_starts_after(&mut self, i: usize, delta: isize) {
-        if delta == 0 {
-            return;
-        }
-        for s in &mut self.line_starts[i + 1..] {
-            *s = (*s as isize + delta) as usize;
-        }
-    }
-
-    fn redact_one(
-        &mut self,
-        location: &TextLocation,
-        replacement: TextReplacement,
-    ) -> Result<(), Error> {
-        let Some(i) = self.line_for(location.start) else {
-            return Ok(());
-        };
-        let line_start = self.line_starts[i];
-        let line_end = self.line_starts[i + 1] - 1;
-        if location.end > line_end {
-            return Ok(());
-        }
-        let local_start = location.start - line_start;
-        let local_end = location.end - line_start;
-        let value = replacement.replacement_value().unwrap_or_default();
-        let before_len = self.lines[i].len();
-        redact::replace_range(&mut self.lines[i], value, local_start..local_end, TARGET)?;
-        let after_len = self.lines[i].len();
-        self.shift_starts_after(i, after_len as isize - before_len as isize);
-        Ok(())
-    }
-}
-
-fn compute_line_starts(lines: &[String]) -> Vec<usize> {
-    let mut starts = Vec::with_capacity(lines.len() + 1);
-    let mut offset = 0usize;
-    for line in lines {
-        starts.push(offset);
-        offset += line.len() + 1; // +1 for the implicit '\n' separator
-    }
-    starts.push(offset);
-    starts
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::Error;
-
-    use super::*;
-
-    fn handler(text: &str) -> TxtHandler {
-        let trailing_newline = text.ends_with('\n');
-        let lines = text.lines().map(String::from).collect();
-        TxtHandler::new(lines, trailing_newline)
-    }
-
-    #[tokio::test]
-    async fn lift_chunk_is_identity_on_second_line() -> Result<(), Error> {
-        // line "world" starts at byte 6. A value-byte range 1..4
-        // ("orl") lifts to source bytes 7..10.
-        let mut h = handler("hello\nworld\n");
-        let _first = h.next_chunk().await?.unwrap();
-        let second = h.next_chunk().await?.unwrap();
-        let lifted = h.lift_chunk(&second, 1..4).expect("in bounds");
-        assert_eq!(lifted.start, 7);
-        assert_eq!(lifted.end, 10);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn stream_yields_each_line() -> Result<(), Error> {
-        let mut h = handler("hello\nworld\n");
-        let first = h.next_chunk().await?.unwrap();
-        assert_eq!(first.location.start, 0);
-        assert_eq!(first.location.end, 5);
-        assert_eq!(first.data.as_str(), "hello");
-        let second = h.next_chunk().await?.unwrap();
-        assert_eq!(second.location.start, 6);
-        assert_eq!(second.location.end, 11);
-        assert_eq!(second.data.as_str(), "world");
-        assert!(h.next_chunk().await?.is_none());
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn stream_single_line_no_newline() -> Result<(), Error> {
-        let mut h = handler("no newline");
-        let chunk = h.next_chunk().await?.unwrap();
-        assert_eq!(chunk.location.start, 0);
-        assert_eq!(chunk.location.end, 10);
-        assert!(!h.trailing_newline());
-        assert!(h.next_chunk().await?.is_none());
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn read_returns_line() -> Result<(), Error> {
-        let h = handler("hello\nworld\n");
-        let loc = TextLocation {
-            start: 6,
-            end: 11,
-            ..Default::default()
-        };
-        assert_eq!(h.read(&loc).await?.unwrap().as_str(), "world");
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn read_cross_line_returns_none() -> Result<(), Error> {
-        let h = handler("hello\nworld\n");
-        let loc = TextLocation {
-            start: 3,
-            end: 8,
-            ..Default::default()
-        };
-        assert!(h.read(&loc).await?.is_none());
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_replaces_whole_line() -> Result<(), Error> {
-        let mut h = handler("hello\nworld\n");
-        let mut rs = Redactions::new();
-        rs.push(
-            TextLocation {
-                start: 6,
-                end: 11,
-                ..Default::default()
-            },
-            TextReplacement::substituted("[REDACTED]"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.lines(), &["hello", "[REDACTED]"]);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_substring_within_line() -> Result<(), Error> {
-        let mut h = handler("hello world");
-        let mut rs = Redactions::new();
-        rs.push(
-            TextLocation {
-                start: 6,
-                end: 11,
-                ..Default::default()
-            },
-            TextReplacement::substituted("[X]"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.lines(), &["hello [X]"]);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_multiple_lines_left_to_right_input() -> Result<(), Error> {
-        // Batch is right-to-left-sorted internally; client passes any order.
-        let mut h = handler("alpha\nbravo\ncharlie\n");
-        let mut rs = Redactions::new();
-        rs.push(
-            TextLocation {
-                start: 0,
-                end: 5,
-                ..Default::default()
-            },
-            TextReplacement::substituted("[A]"),
-        );
-        rs.push(
-            TextLocation {
-                start: 12,
-                end: 19,
-                ..Default::default()
-            },
-            TextReplacement::substituted("[C]"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.lines(), &["[A]", "bravo", "[C]"]);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_unknown_location_skipped() -> Result<(), Error> {
-        let mut h = handler("one line");
-        let mut rs = Redactions::new();
-        rs.push(
-            TextLocation {
-                start: 999,
-                end: 1000,
-                ..Default::default()
-            },
-            TextReplacement::substituted("nope"),
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.lines(), &["one line"]);
-        Ok(())
-    }
-
-    #[test]
-    fn encode_with_trailing_newline() -> Result<(), Error> {
-        let h = handler("hello\nworld\n");
-        let content = h.encode()?;
-        assert_eq!(content.as_bytes(), b"hello\nworld\n");
-        Ok(())
-    }
-
-    #[test]
-    fn encode_without_trailing_newline() -> Result<(), Error> {
-        let h = handler("no newline");
-        let content = h.encode()?;
-        assert_eq!(content.as_bytes(), b"no newline");
-        Ok(())
-    }
-
-    #[test]
-    fn line_starts_initialised_correctly() {
-        let h = handler("hello\nworld\n");
-        // hello [0..5], world [6..11], total = 12 (sentinel)
-        assert_eq!(h.line_starts, vec![0, 6, 12]);
-    }
-
-    #[tokio::test]
-    async fn line_starts_shift_after_shrink() -> Result<(), Error> {
-        let mut h = handler("hello\nworld\n");
-        let mut rs = Redactions::new();
-        rs.push(
-            TextLocation {
-                start: 0,
-                end: 5,
-                ..Default::default()
-            },
-            TextReplacement::substituted("[X]"),
-        );
-        h.redact(rs).await?;
-        // line 0 shrank from 5 → 3, so line 1's start shifted by -2.
-        assert_eq!(h.line_starts, vec![0, 4, 10]);
-        assert_eq!(h.lines(), &["[X]", "world"]);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn redact_mixes_substituted_and_removed() -> Result<(), Error> {
-        let mut h = handler("alice@example.test and bob@example.test");
-        let mut rs = Redactions::new();
-        rs.push(
-            TextLocation {
-                start: 0,
-                end: 18,
-                ..Default::default()
-            },
-            TextReplacement::substituted("[EMAIL]"),
-        );
-        rs.push(
-            TextLocation {
-                start: 23,
-                end: 39,
-                ..Default::default()
-            },
-            TextReplacement::Removed,
-        );
-        h.redact(rs).await?;
-        assert_eq!(h.lines(), &["[EMAIL] and "]);
-        Ok(())
-    }
-}
diff --git a/crates/nvisy-codec/src/handler/text/txt_loader.rs b/crates/nvisy-codec/src/handler/text/txt_loader.rs
deleted file mode 100644
index b921f17dc..000000000
--- a/crates/nvisy-codec/src/handler/text/txt_loader.rs
+++ /dev/null
@@ -1,80 +0,0 @@
-//! Plain-text loader: validates and parses raw text content into a
-//! [`TxtHandler`].
-
-use nvisy_core::Error;
-use nvisy_core::modality::Text;
-
-use super::TxtHandler;
-use crate::Loader;
-use crate::content::{ContentData, ContentSource, TextEncoding};
-
-/// Loader that validates and parses plain-text files. Produces one
-/// [`TxtHandler`] per input.
-#[derive(Debug, Default)]
-pub struct TxtLoader {
-    /// Character encoding of the input bytes. Defaults to UTF-8.
-    pub encoding: TextEncoding,
-}
-
-#[async_trait::async_trait]
-impl Loader<Text> for TxtLoader {
-    type Handler = TxtHandler;
-
-    #[tracing::instrument(name = "txt.decode", skip_all, fields(input_bytes, lines))]
-    async fn decode(&self, content: ContentData) -> Result<TxtHandler, Error> {
-        let parent = content.content_source;
-        let raw = content.to_bytes();
-        tracing::Span::current().record("input_bytes", raw.len());
-        let text = self.encoding.decode_bytes(&raw, "txt-loader")?;
-        let trailing_newline = text.ends_with('\n');
-        let lines: Vec<String> = text.lines().map(String::from).collect();
-        tracing::Span::current().record("lines", lines.len());
-
-        let source = ContentSource::new().with_parent(&parent);
-        Ok(TxtHandler::new(lines, trailing_newline).with_source(source))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use bytes::Bytes;
-    use nvisy_core::Error;
-
-    use super::*;
-    use crate::Handler;
-    use crate::content::ContentSource;
-
-    fn content_from_str(s: &str) -> ContentData {
-        ContentData::new(ContentSource::new(), Bytes::from(s.to_owned()))
-    }
-
-    #[tokio::test]
-    async fn load_multiline() -> Result<(), Error> {
-        let content = content_from_str("hello\nworld\n");
-        let doc = TxtLoader::default().decode(content).await?;
-        assert_eq!(doc.format().as_str(), "nvisy.text.txt");
-        assert_eq!(doc.lines(), &["hello", "world"]);
-        assert!(doc.trailing_newline());
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn load_no_trailing_newline() -> Result<(), Error> {
-        let content = content_from_str("single line");
-        let doc = TxtLoader::default().decode(content).await?;
-        assert_eq!(doc.len(), 1);
-        assert_eq!(doc.line(0), Some("single line"));
-        assert!(!doc.trailing_newline());
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn load_invalid_utf8() {
-        let content = ContentData::new(
-            ContentSource::new(),
-            Bytes::from_static(&[0xFF, 0xFE, 0x00]),
-        );
-        let err = TxtLoader::default().decode(content).await.unwrap_err();
-        assert!(err.to_string().contains("UTF-8"));
-    }
-}
diff --git a/crates/nvisy-codec/src/lib.rs b/crates/nvisy-codec/src/lib.rs
deleted file mode 100644
index 0e9ed9d12..000000000
--- a/crates/nvisy-codec/src/lib.rs
+++ /dev/null
@@ -1,78 +0,0 @@
-#![forbid(unsafe_code)]
-#![cfg_attr(docsrs, feature(doc_cfg))]
-#![doc = include_str!("../README.md")]
-
-#[cfg(not(any(
-    feature = "internal_text",
-    feature = "internal_tabular",
-    feature = "internal_image",
-    feature = "internal_audio",
-    feature = "internal_rich",
-)))]
-compile_error!(
-    "nvisy-codec requires at least one format feature \
-     (txt/json/markdown/html/csv/xlsx/png/jpeg/tiff/wav/mp3/pdf/docx) \
-     — or the umbrella `text`/`tabular`/`image`/`audio`/`rich`"
-);
-
-pub mod content;
-mod core;
-mod document;
-pub mod handler;
-
-pub use self::core::{Chunk, CodecRegistry, Format, FormatId, Handler, Loader};
-pub use self::document::{DocumentHandle, UntypedDocumentHandle};
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[cfg(feature = "txt")]
-    #[tokio::test]
-    async fn registry_decodes_txt_from_extension() {
-        let reg = CodecRegistry::with_builtin();
-        let handle = reg
-            .decode("hello\nworld\n".to_owned().into_bytes(), "txt")
-            .await
-            .expect("txt decoded");
-        let typed = handle.into_text().expect("text variant");
-        assert_eq!(typed.format_id().as_str(), "nvisy.text.txt");
-    }
-
-    #[cfg(feature = "csv")]
-    #[tokio::test]
-    async fn registry_decodes_csv_from_extension() {
-        let reg = CodecRegistry::with_builtin();
-        let handle = reg
-            .decode("a,b\n1,2\n".to_owned().into_bytes(), "csv")
-            .await
-            .expect("csv decoded");
-        let typed = handle.into_tabular().expect("tabular variant");
-        assert_eq!(typed.format_id().as_str(), "nvisy.tabular.csv");
-    }
-
-    #[cfg(all(feature = "txt", feature = "csv"))]
-    #[test]
-    fn registry_resolves_by_extension_and_content_type() {
-        let reg = CodecRegistry::with_builtin();
-        assert_eq!(
-            reg.by_extension("txt").map(|f| f.id.as_str()),
-            Some("nvisy.text.txt"),
-        );
-        assert_eq!(
-            reg.by_extension("log").map(|f| f.id.as_str()),
-            Some("nvisy.text.txt"),
-        );
-        assert_eq!(
-            reg.by_content_type("text/csv").map(|f| f.id.as_str()),
-            Some("nvisy.tabular.csv"),
-        );
-    }
-
-    #[cfg(feature = "txt")]
-    #[test]
-    fn empty_registry_skipped_when_user_wants_no_builtins() {
-        let reg = CodecRegistry::new();
-        assert!(reg.by_extension("txt").is_none());
-    }
-}
diff --git a/crates/nvisy-context/Cargo.toml b/crates/nvisy-context/Cargo.toml
deleted file mode 100644
index 3f5642436..000000000
--- a/crates/nvisy-context/Cargo.toml
+++ /dev/null
@@ -1,39 +0,0 @@
-# https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[package]
-name = "nvisy-context"
-description = "Post-recognition keyword-boost enhancer for Nvisy entities"
-keywords = ["nvisy", "context", "enhancer", "pii"]
-categories = ["text-processing"]
-readme = "README.md"
-
-version = { workspace = true }
-rust-version = { workspace = true }
-edition = { workspace = true }
-license = { workspace = true }
-publish = { workspace = true }
-
-authors = { workspace = true }
-repository = { workspace = true }
-homepage = { workspace = true }
-documentation = { workspace = true }
-
-[package.metadata.docs.rs]
-all-features = true
-rustdoc-args = ["--cfg", "docsrs"]
-
-[dependencies]
-# Internal crates
-nvisy-core = { workspace = true, features = [] }
-
-# Primitive datatypes (cheap-clone surface form on `Token`)
-hipstr = { workspace = true, features = [] }
-
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-
-# Text processing (word-window walk for the substring fallback)
-unicode-segmentation = { workspace = true, features = [] }
-
-[dev-dependencies]
-nvisy-core = { workspace = true, features = ["test-utils"] }
diff --git a/crates/nvisy-context/README.md b/crates/nvisy-context/README.md
deleted file mode 100644
index cf59218a9..000000000
--- a/crates/nvisy-context/README.md
+++ /dev/null
@@ -1,44 +0,0 @@
-# nvisy-context
-
-[![Build](https://img.shields.io/github/actions/workflow/status/nvisycom/runtime/build.yml?branch=main&label=build%20%26%20test&style=flat-square)](https://github.com/nvisycom/runtime/actions/workflows/build.yml)
-
-Post-recognition keyword-boost enhancer for the Nvisy runtime.
-
-## Overview
-
-Context-aware confidence boosting. Every recognizer that wants
-score boosting declares a `Context` (a list of keywords
-plus optional window / boost overrides), registered against the
-recognizer's name. After recognition, `ContextEnhancer` walks each
-detected `Entity<Text>`, looks the recognizer name up in the
-`ContextRegistry`, scans the surrounding window for any declared
-keyword via the configured `KeywordMatcher`, and bumps the entity's
-confidence on a hit.
-
-`Tokens` is the optional NLP artifact (surface + lemma per token)
-that a tokenizing NLP engine stashes on `RecognizerInput.artifacts`
-so `LemmaMatcher` can match morphological variants (`running` →
-`run`). The `SubstringMatcher` fallback runs whenever no `Tokens`
-artifact is present.
-
-The crate depends only on `nvisy-core` for `Entity<Text>`,
-`TrailStep`, and `Confidence` — recognizer crates and the engine
-each depend on `nvisy-context` to participate.
-
-## Documentation
-
-See [`docs/`](../../docs/) for architecture, security, and API documentation.
-
-## Changelog
-
-See [CHANGELOG.md](../../CHANGELOG.md) for release notes and version history.
-
-## License
-
-Apache 2.0 License, see [LICENSE.txt](../../LICENSE.txt)
-
-## Support
-
-- **Documentation**: [docs.nvisy.com](https://docs.nvisy.com)
-- **Issues**: [GitHub Issues](https://github.com/nvisycom/runtime/issues)
-- **Email**: [support@nvisy.com](mailto:support@nvisy.com)
diff --git a/crates/nvisy-context/src/enhancer/context.rs b/crates/nvisy-context/src/enhancer/context.rs
deleted file mode 100644
index 85278b1c4..000000000
--- a/crates/nvisy-context/src/enhancer/context.rs
+++ /dev/null
@@ -1,68 +0,0 @@
-//! [`Context`]: per-call inputs bundled for [`Enhancer::enhance`].
-//!
-//! [`Enhancer::enhance`]: super::Enhancer::enhance
-
-use nvisy_core::primitive::LanguageTag;
-
-use crate::io::Token;
-
-/// Per-call inputs bundled together so the enhancer's internal
-/// methods don't drag a long argument list through every layer.
-///
-/// All fields borrow; the value lives for the duration of one
-/// [`Enhancer::enhance`] call.
-///
-/// [`Enhancer::enhance`]: super::Enhancer::enhance
-#[derive(Clone, Copy)]
-pub struct Context<'a> {
-    /// Full text the entities' byte offsets index into.
-    pub text: &'a str,
-    /// Optional token artifact produced by an upstream NLP
-    /// engine. When present, word-window counting walks the token
-    /// stream; when absent, words are derived from `text` via
-    /// Unicode word segmentation.
-    pub tokens: Option<&'a [Token]>,
-    /// Per-call language hint. `None` means "unknown" — every
-    /// per-language rule applies as a permissive fallback.
-    pub language: Option<&'a LanguageTag>,
-    /// Out-of-band context strings (CSV column headers, JSON
-    /// object keys, log field names) the caller wants treated as
-    /// in-context. Each hint is fed to the matcher as its own
-    /// one-string window; a hit boosts the entity exactly as an
-    /// in-text keyword would.
-    pub hints: &'a [String],
-}
-
-impl<'a> Context<'a> {
-    /// Construct a context with just the source text; every
-    /// other field defaults to empty.
-    pub fn new(text: &'a str) -> Self {
-        Self {
-            text,
-            tokens: None,
-            language: None,
-            hints: &[],
-        }
-    }
-
-    /// Attach a token artifact.
-    #[must_use]
-    pub fn with_tokens(mut self, tokens: &'a [Token]) -> Self {
-        self.tokens = Some(tokens);
-        self
-    }
-
-    /// Attach a language hint.
-    #[must_use]
-    pub fn with_language(mut self, language: &'a LanguageTag) -> Self {
-        self.language = Some(language);
-        self
-    }
-
-    /// Attach out-of-band hint strings.
-    #[must_use]
-    pub fn with_hints(mut self, hints: &'a [String]) -> Self {
-        self.hints = hints;
-        self
-    }
-}
diff --git a/crates/nvisy-context/src/enhancer/mod.rs b/crates/nvisy-context/src/enhancer/mod.rs
deleted file mode 100644
index 1c94dcfbc..000000000
--- a/crates/nvisy-context/src/enhancer/mod.rs
+++ /dev/null
@@ -1,506 +0,0 @@
-//! [`Enhancer`]: post-recognition keyword-boost pass for any
-//! [`Entity<Text>`] regardless of which recognizer produced it.
-
-use std::collections::HashMap;
-
-use nvisy_core::entity::{Entity, EntityLabelRef, TrailStep};
-use nvisy_core::modality::Text;
-
-use crate::io::Token;
-use crate::matching::KeywordMatcher;
-use crate::rule::BoostRule;
-
-mod context;
-mod window;
-
-pub use self::context::Context;
-use self::window::{slice_tokens_around, token_span, word_window};
-
-/// Source name stamped onto refinement [`TrailStep`]s the
-/// enhancer appends when the in-text word window fires.
-const TRAIL_SOURCE_WINDOW: &str = "context";
-
-/// Source name stamped onto refinement [`TrailStep`]s the
-/// enhancer appends when an out-of-band hint fires.
-const TRAIL_SOURCE_HINT: &str = "context-hint";
-
-/// Post-recognition enhancer. Holds a label-keyed [`BoostRule`]
-/// map plus the keyword-matching strategy, and lifts the
-/// confidence of each text entity whose label has a rule and
-/// whose surrounding word window contains one of the rule's
-/// keywords.
-///
-/// Construct via [`Enhancer::new`]. Rules are passed in by value;
-/// duplicates for the same label are merged via
-/// [`BoostRule::merge`] (union of keywords; window radii / `boost`
-/// kept from the first-seen rule).
-///
-/// The matcher defaults are picked by the engine that constructs
-/// the enhancer: [`SubstringMatcher`] when no upstream NLP engine
-/// produces tokens, [`LemmaMatcher`] when one does.
-///
-/// [`SubstringMatcher`]: crate::SubstringMatcher
-/// [`LemmaMatcher`]: crate::LemmaMatcher
-pub struct Enhancer {
-    /// Rules bucketed by label. Within one bucket, each entry is
-    /// a distinct `(language)` scope; rules sharing the same
-    /// `(label, language)` are pre-merged via [`BoostRule::merge`]
-    /// at construction. Per-entity application looks up the
-    /// bucket once by label, then walks the small inner vec
-    /// filtering on the per-call language hint.
-    rules: HashMap<EntityLabelRef, Vec<BoostRule>>,
-    matcher: Box<dyn KeywordMatcher>,
-}
-
-impl Enhancer {
-    /// Construct from a rule iterator and matcher. Rules sharing
-    /// the same `(label, language)` are merged via
-    /// [`BoostRule::merge`]; rules with the same label but
-    /// distinct languages live as separate entries inside the
-    /// label's bucket.
-    pub fn new(
-        rules: impl IntoIterator<Item = BoostRule>,
-        matcher: Box<dyn KeywordMatcher>,
-    ) -> Self {
-        let mut buckets: HashMap<EntityLabelRef, Vec<BoostRule>> = HashMap::new();
-        for rule in rules {
-            let bucket = buckets.entry(rule.label.clone()).or_default();
-            if let Some(existing) = bucket.iter_mut().find(|r| r.language == rule.language) {
-                existing.merge(rule);
-            } else {
-                bucket.push(rule);
-            }
-        }
-        Self {
-            rules: buckets,
-            matcher,
-        }
-    }
-
-    /// `true` when no rules are registered. Engine code uses this
-    /// to short-circuit calls to [`enhance`] entirely.
-    ///
-    /// [`enhance`]: Self::enhance
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.rules.is_empty()
-    }
-
-    /// Number of distinct labels with rules.
-    #[must_use]
-    pub fn len(&self) -> usize {
-        self.rules.len()
-    }
-
-    /// Apply boost rules to `entities` in place. For each entity:
-    /// walk every rule registered for its label whose language
-    /// scope applies under `ctx.language`, walk a window of
-    /// `prefix_words` words before and `suffix_words` words after
-    /// the entity's location, ask the matcher whether any keyword
-    /// fires, and on a hit lift confidence by the rule's `boost`
-    /// (saturating at the [`Confidence`] ceiling) plus append a
-    /// [`Refinement`] trail step.
-    ///
-    /// The in-text and hint paths are independent — at most one
-    /// boost per rule fires per entity (window first, hint as
-    /// fallback) so a rule with a long keyword list can't
-    /// double-dip.
-    ///
-    /// [`Confidence`]: nvisy_core::primitive::Confidence
-    /// [`Refinement`]: nvisy_core::entity::TrailStepKind::Refinement
-    pub fn enhance(&self, entities: &mut [Entity<Text>], ctx: &Context<'_>) {
-        if self.rules.is_empty() {
-            return;
-        }
-        for entity in entities {
-            self.enhance_one(entity, ctx);
-        }
-    }
-
-    fn enhance_one(&self, entity: &mut Entity<Text>, ctx: &Context<'_>) {
-        let Some(bucket) = self.rules.get(&entity.label) else {
-            return;
-        };
-        for rule in bucket {
-            if !rule.applies_to_language(ctx.language) {
-                continue;
-            }
-            if rule.keywords.is_empty() {
-                continue;
-            }
-            self.apply_rule(entity, rule, ctx);
-        }
-    }
-
-    fn apply_rule(&self, entity: &mut Entity<Text>, rule: &BoostRule, ctx: &Context<'_>) {
-        let start = entity.location.start;
-        let end = entity.location.end;
-
-        // Prefer the token stream when the producer reached this
-        // entity. Fall back to the word-segmented substring window
-        // whenever the token slice would be empty — that covers
-        // `tokens: None`, `tokens: Some(&[])`, and the "tokens
-        // present but none overlap the entity" case (e.g. NLP
-        // engine only tokenized part of the document).
-        let token_slice = ctx
-            .tokens
-            .map(|toks| slice_tokens_around(toks, start, end, rule.prefix_words, rule.suffix_words))
-            .unwrap_or(&[]);
-        let (snippet, tokens_in_window): (&str, &[Token]) = if token_slice.is_empty() {
-            let snippet = word_window(ctx.text, start, end, rule.prefix_words, rule.suffix_words);
-            (snippet, &[])
-        } else {
-            let snippet = token_span(ctx.text, token_slice, start, end);
-            (snippet, token_slice)
-        };
-
-        let source = if self
-            .matcher
-            .any_match(snippet, tokens_in_window, &rule.keywords)
-        {
-            TRAIL_SOURCE_WINDOW
-        } else if ctx
-            .hints
-            .iter()
-            .any(|h| self.matcher.any_match(h, &[], &rule.keywords))
-        {
-            TRAIL_SOURCE_HINT
-        } else {
-            return;
-        };
-
-        let original = entity.confidence;
-        let adjusted = original.saturating_add(rule.boost.get());
-        if adjusted == original {
-            return;
-        }
-        entity.confidence = adjusted;
-
-        entity.trail.push(TrailStep::refinement(
-            source,
-            original,
-            adjusted,
-            format!(
-                "context keyword near `{}` (+{:.3})",
-                entity.label.as_str(),
-                rule.boost.get(),
-            ),
-        ));
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{
-        EntityLabelRef, PatternProvenance, TrailProvenance, TrailStepKind, builtins,
-    };
-    use nvisy_core::modality::{Text, TextLocation};
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-    use crate::SubstringMatcher;
-
-    fn govid_label() -> EntityLabelRef {
-        builtins::GOVERNMENT_ID.label_ref()
-    }
-
-    fn person_label() -> EntityLabelRef {
-        builtins::PERSON_NAME.label_ref()
-    }
-
-    fn entity(label: EntityLabelRef, start: usize, end: usize, score: f64) -> Entity<Text> {
-        let confidence = Confidence::new(score).unwrap();
-        let step = TrailStep::recognition(
-            "test",
-            confidence,
-            TrailProvenance::Pattern(PatternProvenance::DenyList),
-            "test fixture",
-        );
-        Entity::builder()
-            .with_label(label)
-            .with_trail(vec![step])
-            .with_confidence(confidence)
-            .with_location(TextLocation::new(start, end))
-            .build()
-            .expect("entity builds")
-    }
-
-    fn enhancer(rules: Vec<BoostRule>) -> Enhancer {
-        Enhancer::new(rules, Box::new(SubstringMatcher))
-    }
-
-    fn rule(
-        label: EntityLabelRef,
-        keywords: &[&'static str],
-        prefix: usize,
-        suffix: usize,
-        boost: f64,
-    ) -> BoostRule {
-        BoostRule::new(
-            label,
-            keywords.iter().copied(),
-            prefix,
-            suffix,
-            Confidence::clamped(boost),
-        )
-    }
-
-    #[test]
-    fn boosts_entity_when_keyword_in_word_window() {
-        let enhancer = enhancer(vec![rule(
-            govid_label(),
-            &["ssn", "social security"],
-            5,
-            5,
-            0.2,
-        )]);
-        let text = "Your SSN: 123-45-6789";
-        let mut entities = vec![entity(govid_label(), 10, 21, 0.6)];
-        enhancer.enhance(&mut entities, &Context::new(text));
-        assert!(entities[0].confidence.get() > 0.6);
-        assert!(
-            entities[0]
-                .trail
-                .iter()
-                .any(|s| matches!(s.kind, TrailStepKind::Refinement)),
-        );
-    }
-
-    #[test]
-    fn suffix_zero_ignores_trailing_keyword() {
-        // Prefix-only: trailing keyword must not boost.
-        let enhancer = enhancer(vec![rule(govid_label(), &["social"], 5, 0, 0.2)]);
-        let text = "123-45-6789 (social security number)";
-        let mut entities = vec![entity(govid_label(), 0, 11, 0.6)];
-        let before = entities[0].confidence.get();
-        enhancer.enhance(&mut entities, &Context::new(text));
-        assert_eq!(entities[0].confidence.get(), before);
-    }
-
-    #[test]
-    fn skips_entity_with_no_rule_for_label() {
-        let enhancer = enhancer(vec![rule(govid_label(), &["ssn"], 5, 5, 0.2)]);
-        let text = "Mr. Smith is named in the report.";
-        let mut entities = vec![entity(person_label(), 4, 9, 0.5)];
-        let before = entities[0].confidence.get();
-        enhancer.enhance(&mut entities, &Context::new(text));
-        assert_eq!(entities[0].confidence.get(), before);
-    }
-
-    #[test]
-    fn window_bounds_the_search() {
-        // 2-word prefix / 2-word suffix: "far_keyword" is at the
-        // start; the entity is after many filler words.
-        let enhancer = enhancer(vec![rule(govid_label(), &["far_keyword"], 2, 2, 0.2)]);
-        let text = "far_keyword here is some filler between the keyword and XYZ here";
-        let xyz_start = text.find("XYZ").unwrap();
-        let xyz_end = xyz_start + "XYZ".len();
-        let mut entities = vec![entity(govid_label(), xyz_start, xyz_end, 0.6)];
-        let before = entities[0].confidence.get();
-        enhancer.enhance(&mut entities, &Context::new(text));
-        assert_eq!(entities[0].confidence.get(), before);
-    }
-
-    #[test]
-    fn boost_saturates_at_one() {
-        let enhancer = enhancer(vec![rule(govid_label(), &["here"], 5, 5, 0.9)]);
-        let text = "the value is right here in plain sight";
-        let mut entities = vec![entity(govid_label(), 16, 21, 0.95)];
-        enhancer.enhance(&mut entities, &Context::new(text));
-        assert!((entities[0].confidence.get() - 1.0).abs() < f64::EPSILON);
-    }
-
-    #[test]
-    fn duplicate_label_rules_merge_keywords() {
-        // Two rules for the same label, each contributing a
-        // distinct keyword. The merged rule must trigger boosts
-        // for matches near keywords from *either* original source,
-        // proving the keyword union survived the merge (not just
-        // last-write-wins).
-        let make_enhancer = || {
-            enhancer(vec![
-                rule(govid_label(), &["ssn"], 5, 5, 0.2),
-                rule(govid_label(), &["tax id"], 5, 5, 0.2),
-            ])
-        };
-        assert_eq!(make_enhancer().len(), 1);
-
-        // Keyword only from the first rule.
-        let ssn_only = "ssn: 123-45-6789";
-        let ssn_entity_start = ssn_only.find("123").unwrap();
-        let ssn_entity_end = ssn_entity_start + "123-45-6789".len();
-        let mut from_first = vec![entity(govid_label(), ssn_entity_start, ssn_entity_end, 0.6)];
-        make_enhancer().enhance(&mut from_first, &Context::new(ssn_only));
-        assert!(
-            from_first[0].confidence.get() > 0.6,
-            "keyword `ssn` from the first rule must still boost after merge",
-        );
-
-        // Keyword only from the second rule.
-        let taxid_only = "tax id: 987-65-4329";
-        let tax_entity_start = taxid_only.find("987").unwrap();
-        let tax_entity_end = tax_entity_start + "987-65-4329".len();
-        let mut from_second = vec![entity(govid_label(), tax_entity_start, tax_entity_end, 0.6)];
-        make_enhancer().enhance(&mut from_second, &Context::new(taxid_only));
-        assert!(
-            from_second[0].confidence.get() > 0.6,
-            "keyword `tax id` from the second rule must still boost after merge",
-        );
-    }
-
-    #[test]
-    fn word_window_handles_unicode() {
-        // 3-word prefix reaches "café" past "naïve" and "resume".
-        let enhancer = enhancer(vec![rule(govid_label(), &["café"], 3, 0, 0.2)]);
-        let text = "café naïve resume — 123-45-6789";
-        let entity_start = text.find("123").unwrap();
-        let entity_end = entity_start + "123-45-6789".len();
-        let mut entities = vec![entity(govid_label(), entity_start, entity_end, 0.6)];
-        enhancer.enhance(&mut entities, &Context::new(text));
-        assert!(
-            entities[0].confidence.get() > 0.6,
-            "unicode word should be reachable within 3-word prefix",
-        );
-    }
-
-    #[test]
-    fn empty_tokens_slice_matches_none_behaviour() {
-        // `Some(&[])` must not collapse the snippet to entity
-        // bytes — it should fall back to the word-window path
-        // just like `None`.
-        let enhancer = enhancer(vec![rule(govid_label(), &["ssn"], 5, 5, 0.2)]);
-        let text = "Your SSN: 123-45-6789";
-        let mut from_none = vec![entity(govid_label(), 10, 21, 0.6)];
-        let mut from_empty = vec![entity(govid_label(), 10, 21, 0.6)];
-        enhancer.enhance(&mut from_none, &Context::new(text));
-        enhancer.enhance(&mut from_empty, &Context::new(text).with_tokens(&[]));
-        assert_eq!(
-            from_none[0].confidence.get(),
-            from_empty[0].confidence.get(),
-            "Some(&[]) must behave identically to None",
-        );
-        assert!(
-            from_empty[0].confidence.get() > 0.6,
-            "empty tokens slice must still allow the word-window fallback to boost",
-        );
-    }
-
-    #[test]
-    fn token_path_counts_words_against_token_stream() {
-        // 1-word prefix, 0-word suffix: the only word the
-        // prefix reaches is the immediate predecessor token
-        // "Your". The tokenizer here treats "social security"
-        // as a single compound token outside the window, so the
-        // keyword "social security" must NOT fire.
-        let enhancer = enhancer(vec![rule(govid_label(), &["social security"], 1, 0, 0.2)]);
-        let text = "social security: Your 123-45-6789";
-        let entity_start = text.find("123").unwrap();
-        let entity_end = entity_start + "123-45-6789".len();
-        let tokens: Vec<Token> = vec![
-            Token::from_text("social security", 0..15),
-            Token::from_text("Your", 17..21),
-            Token::from_text("123-45-6789", 22..33),
-        ];
-        let mut entities = vec![entity(govid_label(), entity_start, entity_end, 0.6)];
-        let before = entities[0].confidence.get();
-        enhancer.enhance(&mut entities, &Context::new(text).with_tokens(&tokens));
-        assert_eq!(
-            entities[0].confidence.get(),
-            before,
-            "1-word prefix should not reach the `social security` token two positions back",
-        );
-    }
-
-    #[test]
-    fn lemma_matcher_boosts_on_morphological_variant() {
-        // Substring matcher would miss `running` for keyword
-        // `run`. Lemma matcher reads the lemma directly off the
-        // token and boosts.
-        let enhancer = Enhancer::new(
-            vec![rule(govid_label(), &["run"], 5, 5, 0.2)],
-            Box::new(crate::LemmaMatcher),
-        );
-        let text = "They were running 123-45-6789 across the system";
-        let entity_start = text.find("123").unwrap();
-        let entity_end = entity_start + "123-45-6789".len();
-        let tokens: Vec<Token> = vec![
-            Token::from_text("They", 0..4),
-            Token::from_text("were", 5..9),
-            Token::from_text("running", 10..17).with_lemma("run"),
-            Token::from_text("123-45-6789", 18..29),
-            Token::from_text("across", 30..36),
-            Token::from_text("the", 37..40),
-            Token::from_text("system", 41..47),
-        ];
-        let mut entities = vec![entity(govid_label(), entity_start, entity_end, 0.6)];
-        enhancer.enhance(&mut entities, &Context::new(text).with_tokens(&tokens));
-        assert!(
-            entities[0].confidence.get() > 0.6,
-            "lemma matcher should match `run` against the `running` token's lemma",
-        );
-    }
-
-    #[test]
-    fn tokens_with_no_overlap_fall_back_to_word_window() {
-        // Tokens cover the first half of the document; the entity
-        // is in the second half, outside any token's range. The
-        // word-window path must still reach the keyword.
-        let enhancer = enhancer(vec![rule(govid_label(), &["ssn"], 5, 5, 0.2)]);
-        let text = "First half of the document. Your SSN: 123-45-6789";
-        let entity_start = text.find("123").unwrap();
-        let entity_end = entity_start + "123-45-6789".len();
-        let tokens: Vec<Token> = vec![
-            Token::from_text("First", 0..5),
-            Token::from_text("half", 6..10),
-            Token::from_text("of", 11..13),
-            Token::from_text("the", 14..17),
-            Token::from_text("document", 18..26),
-        ];
-        let mut entities = vec![entity(govid_label(), entity_start, entity_end, 0.6)];
-        enhancer.enhance(&mut entities, &Context::new(text).with_tokens(&tokens));
-        assert!(
-            entities[0].confidence.get() > 0.6,
-            "tokens that don't overlap the entity must fall back to the word window",
-        );
-    }
-
-    #[test]
-    fn out_of_band_hint_boosts_when_window_is_empty() {
-        // Cell-only text has no surrounding context — the word
-        // window walk finds nothing — but the caller supplies the
-        // CSV column header as an out-of-band hint that contains
-        // a rule keyword. Confidence must lift, and the trail
-        // step must mark the source as `context-hint`.
-        let enhancer = enhancer(vec![rule(govid_label(), &["ssn"], 5, 5, 0.2)]);
-        let text = "123-45-6789";
-        let hints = ["ssn".to_owned()];
-        let mut entities = vec![entity(govid_label(), 0, 11, 0.6)];
-        enhancer.enhance(&mut entities, &Context::new(text).with_hints(&hints));
-        assert!(
-            entities[0].confidence.get() > 0.6,
-            "out-of-band hint matching a rule keyword must boost",
-        );
-        assert!(
-            entities[0].trail.iter().any(|s| s.source == "context-hint"),
-            "trail step must record the hint-source provenance",
-        );
-    }
-
-    #[test]
-    fn hint_path_is_independent_of_window_path() {
-        // The in-text window already fires, so the hint path
-        // shouldn't double-boost. Exactly one refinement step
-        // appears on the entity.
-        let enhancer = enhancer(vec![rule(govid_label(), &["ssn"], 5, 5, 0.2)]);
-        let text = "Your SSN: 123-45-6789";
-        let hints = ["ssn".to_owned()];
-        let mut entities = vec![entity(govid_label(), 10, 21, 0.6)];
-        enhancer.enhance(&mut entities, &Context::new(text).with_hints(&hints));
-        let refinements = entities[0]
-            .trail
-            .iter()
-            .filter(|s| matches!(s.kind, TrailStepKind::Refinement))
-            .count();
-        assert_eq!(refinements, 1, "rule must boost at most once per entity");
-    }
-}
diff --git a/crates/nvisy-context/src/enhancer/window.rs b/crates/nvisy-context/src/enhancer/window.rs
deleted file mode 100644
index 85252fc26..000000000
--- a/crates/nvisy-context/src/enhancer/window.rs
+++ /dev/null
@@ -1,118 +0,0 @@
-//! Window-slicing helpers shared by [`Enhancer::apply_rule`].
-//!
-//! Two coordinate systems matter here:
-//!
-//! - **Bytes**: source-text offsets. `word_window` walks Unicode
-//!   word segments to expand an entity's `[start, end)` to
-//!   `prefix`/`suffix` words on either side.
-//! - **Tokens**: pre-tokenized stream from an upstream NLP engine.
-//!   `slice_tokens_around` takes a `prefix`/`suffix` count and
-//!   returns the contiguous token slice that covers the entity
-//!   plus that many neighbours.
-//!
-//! Both paths feed the same downstream [`KeywordMatcher`] —
-//! [`token_span`] reduces a non-empty token slice back to its
-//! spanning substring for matchers that operate on raw text.
-//!
-//! [`Enhancer::apply_rule`]: super::Enhancer
-//! [`KeywordMatcher`]: crate::KeywordMatcher
-
-use unicode_segmentation::UnicodeSegmentation;
-
-use crate::io::Token;
-
-/// Walk `prefix` words before `[start, end)` and `suffix` words
-/// after, via Unicode word segmentation, and return the spanning
-/// substring (including any non-word whitespace and punctuation
-/// between words). The returned slice covers `[start, end)` itself
-/// plus the prefix / suffix words; the entity's own bytes are
-/// always inside.
-pub(super) fn word_window(
-    text: &str,
-    start: usize,
-    end: usize,
-    prefix: usize,
-    suffix: usize,
-) -> &str {
-    let prefix_text = &text[..start.min(text.len())];
-    let suffix_text = &text[end.min(text.len())..];
-
-    // `unicode_word_indices` yields `(byte_offset, word_str)` for
-    // every "word" (alphanumeric run) in source order. Take the
-    // last `prefix` on the prefix side, the first `suffix` on the
-    // suffix side, and compute the spanning byte range.
-    let prefix_words: Vec<(usize, &str)> = prefix_text.unicode_word_indices().collect();
-    let prefix_take = prefix_words.len().saturating_sub(prefix);
-    let prefix_byte = prefix_words
-        .get(prefix_take)
-        .map(|(idx, _)| *idx)
-        .unwrap_or(start.min(text.len()));
-
-    let suffix_byte = if suffix == 0 {
-        end.min(text.len())
-    } else {
-        suffix_text
-            .unicode_word_indices()
-            .nth(suffix - 1)
-            .map(|(idx, word)| end + idx + word.len())
-            .unwrap_or(text.len())
-    };
-
-    let lo = floor_char_boundary(text, prefix_byte);
-    let hi = ceil_char_boundary(text, suffix_byte.min(text.len()));
-    &text[lo..hi]
-}
-
-/// Slice tokens by *count*: take `prefix` tokens before the first
-/// token overlapping `[start, end)` and `suffix` tokens after the
-/// last. The returned slice is contiguous.
-pub(super) fn slice_tokens_around(
-    tokens: &[Token],
-    start: usize,
-    end: usize,
-    prefix: usize,
-    suffix: usize,
-) -> &[Token] {
-    if tokens.is_empty() {
-        return &[];
-    }
-    // First token whose `offset.end > start` overlaps or follows the entity.
-    let first_overlap = tokens.partition_point(|t| t.offset.end <= start);
-    // One past the last token whose `offset.start < end` overlaps the entity.
-    let last_overlap = tokens.partition_point(|t| t.offset.start < end);
-    let lo = first_overlap.saturating_sub(prefix);
-    let hi = (last_overlap + suffix).min(tokens.len());
-    if lo >= hi {
-        return &[];
-    }
-    &tokens[lo..hi]
-}
-
-/// Spanning substring covering `tokens` plus the entity itself.
-/// Used to give the matcher a contiguous text window when slicing
-/// against the token stream.
-///
-/// Precondition: `tokens` is non-empty. Callers must take the
-/// [`word_window`] fallback path when their token slice is empty.
-pub(super) fn token_span<'a>(text: &'a str, tokens: &[Token], start: usize, end: usize) -> &'a str {
-    debug_assert!(!tokens.is_empty(), "token_span requires non-empty slice");
-    let lo = tokens[0].offset.start.min(start);
-    let hi = tokens[tokens.len() - 1].offset.end.max(end);
-    let lo = floor_char_boundary(text, lo.min(text.len()));
-    let hi = ceil_char_boundary(text, hi.min(text.len()));
-    &text[lo..hi]
-}
-
-fn floor_char_boundary(s: &str, mut pos: usize) -> usize {
-    while pos > 0 && !s.is_char_boundary(pos) {
-        pos -= 1;
-    }
-    pos
-}
-
-fn ceil_char_boundary(s: &str, mut pos: usize) -> usize {
-    while pos < s.len() && !s.is_char_boundary(pos) {
-        pos += 1;
-    }
-    pos
-}
diff --git a/crates/nvisy-context/src/io/mod.rs b/crates/nvisy-context/src/io/mod.rs
deleted file mode 100644
index df3b9fb0b..000000000
--- a/crates/nvisy-context/src/io/mod.rs
+++ /dev/null
@@ -1,20 +0,0 @@
-//! Wiring between the [`Enhancer`] and the [`EntityRecognizer`]
-//! pipeline.
-//!
-//! - [`Token`] / [`Tokens`] is the shared NLP token artifact the
-//!   enhancer reads off `RecognizerInput.artifacts`.
-//! - [`ContextEnhanced`] wraps any [`EntityRecognizer<Text>`] so
-//!   the enhancer runs automatically after the inner recognizer's
-//!   pass.
-//!
-//! All three types are re-exported at the crate root.
-//!
-//! [`Enhancer`]: crate::Enhancer
-//! [`EntityRecognizer`]: nvisy_core::recognition::EntityRecognizer
-//! [`EntityRecognizer<Text>`]: nvisy_core::recognition::EntityRecognizer
-
-mod tokens;
-mod wrapper;
-
-pub use self::tokens::{Token, Tokens};
-pub use self::wrapper::ContextEnhanced;
diff --git a/crates/nvisy-context/src/io/tokens.rs b/crates/nvisy-context/src/io/tokens.rs
deleted file mode 100644
index 57262d6be..000000000
--- a/crates/nvisy-context/src/io/tokens.rs
+++ /dev/null
@@ -1,166 +0,0 @@
-//! [`Token`] and the [`Tokens`] collection.
-//!
-//! A [`Token`] is the engine's view of one lexical unit: the
-//! surface text, its lemma when the engine has a lemmatizer
-//! (otherwise == `text`), the byte range it occupies in the source
-//! text, and two precomputed predicates the enhancer reads
-//! (`is_stop`, `is_punct`).
-//!
-//! [`Tokens`] is the owning collection — a `Vec<Token>` newtype
-//! exposing iteration and length. The [`Enhancer`] slices the
-//! stream by *count* (prefix/suffix word radii) using its own
-//! internal helpers; the byte range carried on each [`Token`] is
-//! there for consumers that want to map a token back to its
-//! source-text substring.
-//!
-//! [`Enhancer`]: super::Enhancer
-//!
-//! Tokens live next to the [`Enhancer`] because that's the only
-//! consumer: the enhancer reads them off
-//! `RecognizerInput::artifacts` to drive lemma-aware keyword
-//! matching. The producer (a tokenizer in some upstream NLP
-//! backend) only needs to know the type by name; the type itself
-//! belongs in the consumer's neighbourhood.
-//!
-//! The shape is intentionally minimal. POS tags, morphology,
-//! dependency trees, and other heavier features are not part of
-//! the v1 surface; they get added as fields when a downstream
-//! consumer needs them. This keeps the artifact cheap for engines
-//! that don't produce them — `text == lemma`, `is_stop == false`,
-//! `is_punct == false` are the defaults for a tokenizer-only
-//! engine.
-
-use std::ops::Range;
-use std::{slice, vec};
-
-use hipstr::HipStr;
-
-/// One token produced by an upstream tokenizer.
-///
-/// `lemma` falls back to `text` when the producer has no
-/// lemmatizer — callers that want lemma-aware matching can read
-/// `token.lemma` uniformly without checking which engine produced
-/// the artifact.
-#[derive(Debug, Clone, PartialEq, Eq)]
-pub struct Token {
-    /// Surface form as it appears in the source text.
-    pub text: HipStr<'static>,
-    /// Lemma when the producer emitted one; otherwise == [`text`].
-    ///
-    /// [`text`]: Self::text
-    pub lemma: HipStr<'static>,
-    /// Byte range this token occupies in the source text. Use this
-    /// to map back to substrings of the original input.
-    pub offset: Range<usize>,
-    /// Producer-asserted stopword flag (e.g. "the", "a", "of" for
-    /// English). Producers without a stopword list set this to
-    /// `false`.
-    pub is_stop: bool,
-    /// Producer-asserted punctuation flag.
-    pub is_punct: bool,
-}
-
-impl Token {
-    /// Construct a token with the surface form mirrored into
-    /// `lemma`. Use this from tokenizer-only producers.
-    pub fn from_text(text: impl Into<HipStr<'static>>, offset: Range<usize>) -> Self {
-        let text = text.into();
-        Self {
-            lemma: text.clone(),
-            text,
-            offset,
-            is_stop: false,
-            is_punct: false,
-        }
-    }
-
-    /// Override the lemma.
-    #[must_use]
-    pub fn with_lemma(mut self, lemma: impl Into<HipStr<'static>>) -> Self {
-        self.lemma = lemma.into();
-        self
-    }
-
-    /// Override the stopword flag.
-    #[must_use]
-    pub fn with_is_stop(mut self, is_stop: bool) -> Self {
-        self.is_stop = is_stop;
-        self
-    }
-
-    /// Override the punctuation flag.
-    #[must_use]
-    pub fn with_is_punct(mut self, is_punct: bool) -> Self {
-        self.is_punct = is_punct;
-        self
-    }
-}
-
-/// Owning token sequence stamped on a
-/// [`RecognizerInput::artifacts`] bundle by an upstream NLP engine.
-///
-/// [`RecognizerInput::artifacts`]: nvisy_core::recognition::RecognizerInput::artifacts
-///
-/// Tokens are sorted by `offset.start` (producers should emit them
-/// in order; consumer-side code assumes this). The [`Enhancer`]
-/// borrows the underlying slice via [`as_slice`] and walks it by
-/// count when scoring the entity's neighbourhood.
-///
-/// [`Enhancer`]: crate::Enhancer
-/// [`as_slice`]: Tokens::as_slice
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-pub struct Tokens(Vec<Token>);
-
-impl Tokens {
-    /// Construct an empty token sequence. Use this when the
-    /// producer has no tokenizer (language-only engines).
-    #[must_use]
-    pub fn empty() -> Self {
-        Self::default()
-    }
-
-    /// Construct from an owned vector. The caller is responsible
-    /// for ensuring tokens are sorted by `offset.start`.
-    #[must_use]
-    pub fn new(tokens: Vec<Token>) -> Self {
-        Self(tokens)
-    }
-
-    /// Borrow the underlying slice.
-    #[must_use]
-    pub fn as_slice(&self) -> &[Token] {
-        &self.0
-    }
-
-    /// Number of tokens.
-    #[must_use]
-    pub fn len(&self) -> usize {
-        self.0.len()
-    }
-
-    /// Whether the sequence is empty.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.0.is_empty()
-    }
-
-    /// Iterate tokens in source order.
-    pub fn iter(&self) -> slice::Iter<'_, Token> {
-        self.0.iter()
-    }
-}
-
-impl FromIterator<Token> for Tokens {
-    fn from_iter<I: IntoIterator<Item = Token>>(iter: I) -> Self {
-        Self(iter.into_iter().collect())
-    }
-}
-
-impl IntoIterator for Tokens {
-    type IntoIter = vec::IntoIter<Token>;
-    type Item = Token;
-
-    fn into_iter(self) -> Self::IntoIter {
-        self.0.into_iter()
-    }
-}
diff --git a/crates/nvisy-context/src/io/wrapper.rs b/crates/nvisy-context/src/io/wrapper.rs
deleted file mode 100644
index 6329bf5d6..000000000
--- a/crates/nvisy-context/src/io/wrapper.rs
+++ /dev/null
@@ -1,82 +0,0 @@
-//! [`ContextEnhanced`]: post-recognition keyword-boost wrapper for any
-//! [`EntityRecognizer<Text>`].
-//!
-//! Composes an inner recognizer with an [`Enhancer`]: the wrapper
-//! delegates `recognize` to the inner, then runs the enhancer
-//! over the produced entities. Equivalent to "the recognizer
-//! owns its boosting" without each recognizer reimplementing the
-//! enhancement step.
-//!
-//! Typical use:
-//!
-//! ```ignore
-//! let inner = MyRecognizer::new(...);
-//! let enhancer = Enhancer::new(rules, Box::new(SubstringMatcher));
-//! let recognizer = ContextEnhanced::new(inner, enhancer);
-//! ```
-//!
-//! The wrapper implements [`EntityRecognizer<Text>`] so the engine
-//! never has to know boosting happened.
-
-use nvisy_core::Result;
-use nvisy_core::modality::Text;
-use nvisy_core::recognition::{EntityRecognizer, RecognizerInput, RecognizerOutput};
-
-use super::Tokens;
-use crate::{Context, Enhancer};
-
-/// Wraps an [`EntityRecognizer<Text>`] with a post-recognition
-/// [`Enhancer`] pass. Implements [`EntityRecognizer<Text>`] so
-/// the wrapped recognizer is a drop-in replacement.
-///
-/// Assumes the inner recognizer emits entities whose byte offsets
-/// index into `input.data.text` (the standard
-/// [`EntityRecognizer<Text>`] contract). The wrapper reads the
-/// same `&str` for the keyword-window walk; a recognizer that
-/// emitted entities relative to a different coordinate space
-/// would surface stale or panic-on-slice offsets.
-pub struct ContextEnhanced<R> {
-    inner: R,
-    enhancer: Enhancer,
-}
-
-impl<R> ContextEnhanced<R> {
-    /// Wrap `inner` with `enhancer`. After `recognize` produces
-    /// entities, `enhancer` runs over them in place.
-    pub fn new(inner: R, enhancer: Enhancer) -> Self {
-        Self { inner, enhancer }
-    }
-
-    /// Borrow the wrapped recognizer.
-    pub fn inner(&self) -> &R {
-        &self.inner
-    }
-
-    /// Borrow the enhancer applied to the inner recognizer's
-    /// output.
-    pub fn enhancer(&self) -> &Enhancer {
-        &self.enhancer
-    }
-}
-
-#[async_trait::async_trait]
-impl<R> EntityRecognizer<Text> for ContextEnhanced<R>
-where
-    R: EntityRecognizer<Text> + 'static,
-{
-    async fn recognize(&self, input: &RecognizerInput<Text>) -> Result<RecognizerOutput<Text>> {
-        let mut output = self.inner.recognize(input).await?;
-        if self.enhancer.is_empty() {
-            return Ok(output);
-        }
-        let mut ctx = Context::new(input.data.text.as_str()).with_hints(&input.context_hints);
-        if let Some(tokens) = input.artifacts.get::<Tokens>() {
-            ctx = ctx.with_tokens(tokens.as_slice());
-        }
-        if let Some(language) = input.language.as_ref() {
-            ctx = ctx.with_language(language);
-        }
-        self.enhancer.enhance(&mut output.entities, &ctx);
-        Ok(output)
-    }
-}
diff --git a/crates/nvisy-context/src/lib.rs b/crates/nvisy-context/src/lib.rs
deleted file mode 100644
index 7c470d062..000000000
--- a/crates/nvisy-context/src/lib.rs
+++ /dev/null
@@ -1,13 +0,0 @@
-#![forbid(unsafe_code)]
-#![cfg_attr(docsrs, feature(doc_cfg))]
-#![doc = include_str!("../README.md")]
-
-mod enhancer;
-mod io;
-mod matching;
-mod rule;
-
-pub use self::enhancer::{Context, Enhancer};
-pub use self::io::{ContextEnhanced, Token, Tokens};
-pub use self::matching::{KeywordMatcher, LemmaMatcher, SubstringMatcher};
-pub use self::rule::{BoostRule, DEFAULT_BOOST, DEFAULT_PREFIX_WORDS, DEFAULT_SUFFIX_WORDS};
diff --git a/crates/nvisy-context/src/matching/lemma.rs b/crates/nvisy-context/src/matching/lemma.rs
deleted file mode 100644
index 0e1d5b5e9..000000000
--- a/crates/nvisy-context/src/matching/lemma.rs
+++ /dev/null
@@ -1,65 +0,0 @@
-//! Lemma-aware [`KeywordMatcher`] implementation.
-
-use hipstr::HipStr;
-
-use super::matcher::{KeywordMatcher, SubstringMatcher};
-use crate::io::Token;
-
-/// Lemma-aware matcher. Compares each lemma in `tokens` against
-/// the keyword list with ASCII case-insensitive equality.
-///
-/// Falls back to [`SubstringMatcher`] semantics when `tokens` is
-/// empty (no shared NLP artifact was produced) so the enhancer
-/// runs uniformly regardless of whether the upstream pass emitted
-/// tokens.
-///
-/// Recognizes morphological variants the substring matcher cannot:
-/// `"running" → "run"`, `"dogs" → "dog"`, `"SSNs" → "ssn"`. Cost
-/// is one lowercase per keyword + one lowercase per lemma per
-/// match attempt.
-#[derive(Debug, Clone, Copy, Default)]
-pub struct LemmaMatcher;
-
-impl KeywordMatcher for LemmaMatcher {
-    fn any_match(&self, window: &str, tokens: &[Token], keywords: &[HipStr<'static>]) -> bool {
-        if tokens.is_empty() {
-            return SubstringMatcher.any_match(window, tokens, keywords);
-        }
-        let lowered_keywords: Vec<String> = keywords
-            .iter()
-            .map(|k| k.as_str().to_ascii_lowercase())
-            .collect();
-        tokens.iter().any(|tok| {
-            let lemma = tok.lemma.as_str().to_ascii_lowercase();
-            lowered_keywords.contains(&lemma)
-        })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn kws(items: &[&'static str]) -> Vec<HipStr<'static>> {
-        items.iter().copied().map(HipStr::from).collect()
-    }
-
-    #[test]
-    fn matches_morph_variants() {
-        let tokens = vec![
-            Token::from_text("the", 0..3),
-            Token::from_text("running", 4..11).with_lemma("run"),
-            Token::from_text("dogs", 12..16).with_lemma("dog"),
-        ];
-        let m = LemmaMatcher;
-        assert!(m.any_match("", &tokens, &kws(&["run"])));
-        assert!(m.any_match("", &tokens, &kws(&["dog"])));
-        assert!(!m.any_match("", &tokens, &kws(&["cat"])));
-    }
-
-    #[test]
-    fn falls_back_to_substring_without_tokens() {
-        let m = LemmaMatcher;
-        assert!(m.any_match("Your SSN: 123", &[], &kws(&["ssn"])));
-    }
-}
diff --git a/crates/nvisy-context/src/matching/matcher.rs b/crates/nvisy-context/src/matching/matcher.rs
deleted file mode 100644
index 5e7ac5605..000000000
--- a/crates/nvisy-context/src/matching/matcher.rs
+++ /dev/null
@@ -1,70 +0,0 @@
-//! [`KeywordMatcher`] trait + the default [`SubstringMatcher`].
-
-use hipstr::HipStr;
-
-use crate::io::Token;
-
-/// Decide whether any keyword from `keywords` fires within the
-/// candidate region around an entity match.
-///
-/// The strategy slot that lets the enhancer swap raw substring
-/// matching for lemma-aware matching (or a third-party
-/// fuzzy/word-boundary implementation) without changing its core
-/// pipeline.
-///
-/// Implementations receive both a raw `window` slice of the source
-/// text (for substring strategies) and the `tokens` covering that
-/// same range (for token/lemma strategies). Either or both may be
-/// ignored; `tokens` is empty when no NLP engine produced a token
-/// artifact.
-pub trait KeywordMatcher: Send + Sync {
-    /// `true` if at least one keyword from `keywords` appears in
-    /// the input.
-    fn any_match(&self, window: &str, tokens: &[Token], keywords: &[HipStr<'static>]) -> bool;
-}
-
-/// ASCII case-insensitive substring matcher. The default —
-/// runs whenever no token artifact was stamped on
-/// `RecognizerInput.artifacts`, or whenever the caller explicitly
-/// picks raw matching.
-///
-/// Fast, allocation-light, permissive: the keyword `"email"` fires
-/// inside `"MyEmailAddress"`. Ignores the `tokens` argument.
-#[derive(Debug, Clone, Copy, Default)]
-pub struct SubstringMatcher;
-
-impl KeywordMatcher for SubstringMatcher {
-    fn any_match(&self, window: &str, _tokens: &[Token], keywords: &[HipStr<'static>]) -> bool {
-        let lowered = window.to_ascii_lowercase();
-        keywords
-            .iter()
-            .any(|kw| lowered.contains(kw.as_str().to_ascii_lowercase().as_str()))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn kws(items: &[&'static str]) -> Vec<HipStr<'static>> {
-        items.iter().copied().map(HipStr::from).collect()
-    }
-
-    #[test]
-    fn substring_matches_case_insensitively() {
-        let m = SubstringMatcher;
-        assert!(m.any_match("Your SSN: 123", &[], &kws(&["ssn"])));
-        assert!(m.any_match(
-            "the SOCIAL SECURITY number",
-            &[],
-            &kws(&["social security"])
-        ));
-        assert!(!m.any_match("nothing here", &[], &kws(&["ssn"])));
-    }
-
-    #[test]
-    fn substring_is_permissive() {
-        let m = SubstringMatcher;
-        assert!(m.any_match("MyEmailAddress", &[], &kws(&["email"])));
-    }
-}
diff --git a/crates/nvisy-context/src/matching/mod.rs b/crates/nvisy-context/src/matching/mod.rs
deleted file mode 100644
index 3ffe59cf1..000000000
--- a/crates/nvisy-context/src/matching/mod.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-//! Keyword-matching strategies plugged into the [`Enhancer`].
-//!
-//! - [`KeywordMatcher`] is the trait the enhancer talks to.
-//! - [`SubstringMatcher`] is the default: ASCII case-insensitive
-//!   substring search over the raw text window. Runs whenever no
-//!   token artifact is present on `RecognizerInput.artifacts`.
-//! - [`LemmaMatcher`] reads lemmatized tokens an upstream NLP
-//!   engine stamped on `RecognizerInput.artifacts`. Recognizes
-//!   morphological variants substring matching misses.
-//!
-//! All three are re-exported at the crate root.
-//!
-//! [`Enhancer`]: crate::Enhancer
-
-mod lemma;
-mod matcher;
-
-pub use self::lemma::LemmaMatcher;
-pub use self::matcher::{KeywordMatcher, SubstringMatcher};
diff --git a/crates/nvisy-context/src/rule.rs b/crates/nvisy-context/src/rule.rs
deleted file mode 100644
index da3c809db..000000000
--- a/crates/nvisy-context/src/rule.rs
+++ /dev/null
@@ -1,181 +0,0 @@
-//! [`BoostRule`]: per-label keyword-boost rule.
-//!
-//! One rule per [`EntityLabelRef`] declares the keyword set that
-//! lifts confidence when one of those keywords appears within
-//! `prefix_words` words before or `suffix_words` words after an
-//! entity carrying that label. The window radii and the additive
-//! `boost` are resolved at rule construction time — there are no
-//! per-source overrides at apply time.
-//!
-//! Producers (the pattern crate today, future NER/LLM/custom
-//! recognizer authors) hand the engine a `Vec<BoostRule>` keyed by
-//! label. When several rules contribute to the same label (e.g.
-//! two different SSN detectors both contributing to
-//! `GOVERNMENT_ID`), the engine merges them by union of keywords —
-//! see [`BoostRule::merge`].
-//!
-//! [`EntityLabelRef`]: nvisy_core::entity::EntityLabelRef
-
-use std::collections::HashSet;
-
-use hipstr::HipStr;
-use nvisy_core::entity::EntityLabelRef;
-use nvisy_core::primitive::{Confidence, LanguageTag};
-
-/// Default window radius in words *before* an entity match.
-pub const DEFAULT_PREFIX_WORDS: usize = 5;
-
-/// Default window radius in words *after* an entity match. Set
-/// equal to [`DEFAULT_PREFIX_WORDS`] so trailing context like
-/// "123-45-6789 (social security)" boosts the same as leading
-/// context. Asymmetric windows surprise operators who rarely
-/// realize the asymmetry exists, so we pick symmetric defaults.
-pub const DEFAULT_SUFFIX_WORDS: usize = 5;
-
-/// Default additive boost applied when a keyword fires.
-pub const DEFAULT_BOOST: f64 = 0.35;
-
-/// Per-label boost rule the [`Enhancer`] applies at runtime.
-///
-/// [`Enhancer`]: super::Enhancer
-#[derive(Debug, Clone, PartialEq)]
-pub struct BoostRule {
-    /// Entity label this rule applies to. Each emitted
-    /// `Entity<Text>` whose [`label`] matches is checked against
-    /// this rule's keywords.
-    ///
-    /// [`label`]: nvisy_core::entity::Entity::label
-    pub label: EntityLabelRef,
-    /// Language scope. `None` means the rule applies regardless
-    /// of the per-call language hint; `Some(lang)` means the rule
-    /// only fires when the caller's language matches, or when no
-    /// hint is set (permissive fallback).
-    pub language: Option<LanguageTag>,
-    /// Keywords whose presence near a match lifts the entity's
-    /// confidence. Stored as [`HipStr`] for cheap clones across
-    /// per-pass rule sets.
-    pub keywords: Vec<HipStr<'static>>,
-    /// Window radius in words *before* the entity's match.
-    /// Counted against the token artifact on
-    /// `RecognizerInput.artifacts` when present, or via Unicode
-    /// word segmentation of the source text otherwise.
-    pub prefix_words: usize,
-    /// Window radius in words *after* the entity's match. Same
-    /// source as [`prefix_words`].
-    ///
-    /// [`prefix_words`]: Self::prefix_words
-    pub suffix_words: usize,
-    /// Additive boost applied to the entity's confidence when a
-    /// keyword fires. Clamped at the [`Confidence`] ceiling on
-    /// apply.
-    pub boost: Confidence,
-}
-
-impl BoostRule {
-    /// Construct a rule for `label` with explicit window radii
-    /// and `boost`. The rule is language-agnostic; use
-    /// [`with_language`] to scope it. Most callers want
-    /// [`BoostRule::for_label`] instead — it bakes in the default
-    /// window / boost values.
-    ///
-    /// [`with_language`]: Self::with_language
-    #[must_use]
-    pub fn new(
-        label: EntityLabelRef,
-        keywords: impl IntoIterator<Item = impl Into<HipStr<'static>>>,
-        prefix_words: usize,
-        suffix_words: usize,
-        boost: Confidence,
-    ) -> Self {
-        Self {
-            label,
-            language: None,
-            keywords: keywords.into_iter().map(Into::into).collect(),
-            prefix_words,
-            suffix_words,
-            boost,
-        }
-    }
-
-    /// Construct a rule for `label` using the crate's default
-    /// [`prefix_words`], [`suffix_words`], and [`boost`]
-    /// constants. The common case — recognizers building their
-    /// own boost rules from declared keywords don't need to
-    /// think about tuning knobs.
-    ///
-    /// [`prefix_words`]: DEFAULT_PREFIX_WORDS
-    /// [`suffix_words`]: DEFAULT_SUFFIX_WORDS
-    /// [`boost`]: DEFAULT_BOOST
-    #[must_use]
-    pub fn for_label(
-        label: EntityLabelRef,
-        keywords: impl IntoIterator<Item = impl Into<HipStr<'static>>>,
-    ) -> Self {
-        Self::new(
-            label,
-            keywords,
-            DEFAULT_PREFIX_WORDS,
-            DEFAULT_SUFFIX_WORDS,
-            Confidence::clamped(DEFAULT_BOOST),
-        )
-    }
-
-    /// Scope this rule to a single language.
-    ///
-    /// At apply time the rule fires only when the caller's
-    /// language hint matches `language`, or when no hint is set
-    /// (permissive fallback).
-    #[must_use]
-    pub fn with_language(mut self, language: LanguageTag) -> Self {
-        self.language = Some(language);
-        self
-    }
-
-    /// Return `true` when this rule applies under the per-call
-    /// language hint.
-    ///
-    /// - Language-agnostic rules (`self.language == None`)
-    ///   always apply.
-    /// - Language-scoped rules apply when the hint shares a
-    ///   primary subtag with the scope (so a rule scoped to
-    ///   `"en"` fires for `"en-US"` and `"en-GB"` hints), or
-    ///   when no hint is set (permissive fallback so callers
-    ///   who don't pass a language still get boosts).
-    #[must_use]
-    pub fn applies_to_language(&self, hint: Option<&LanguageTag>) -> bool {
-        match (&self.language, hint) {
-            (None, _) => true,
-            (Some(_), None) => true,
-            (Some(scope), Some(hint)) => scope.matches(hint),
-        }
-    }
-
-    /// Merge `other` into this rule by extending the keyword set
-    /// with any keywords not already present. Window radii and
-    /// `boost` are kept from `self` — callers that need different
-    /// values per source should construct independent rules and
-    /// keep them separate.
-    ///
-    /// # Panics
-    ///
-    /// Debug-asserts when the labels or languages differ. Merging
-    /// across keys is a caller bug — rules are keyed by
-    /// `(label, language)` and the engine looks them up by both.
-    pub fn merge(&mut self, other: BoostRule) {
-        debug_assert_eq!(
-            self.label, other.label,
-            "BoostRule::merge requires matching labels",
-        );
-        debug_assert_eq!(
-            self.language, other.language,
-            "BoostRule::merge requires matching languages",
-        );
-        let existing: HashSet<&str> = self.keywords.iter().map(HipStr::as_str).collect();
-        let additions: Vec<HipStr<'static>> = other
-            .keywords
-            .into_iter()
-            .filter(|kw| !existing.contains(kw.as_str()))
-            .collect();
-        self.keywords.extend(additions);
-    }
-}
diff --git a/crates/nvisy-core/Cargo.toml b/crates/nvisy-core/Cargo.toml
index 3d9e4c03b..9de0e6092 100644
--- a/crates/nvisy-core/Cargo.toml
+++ b/crates/nvisy-core/Cargo.toml
@@ -29,6 +29,10 @@ default = []
 test-utils = []
 
 [dependencies]
+# Framework
+elide = { workspace = true, features = ["image", "audio", "tabular", "serde", "sha2"] }
+elide-core = { workspace = true, features = ["serde", "image", "audio", "tabular"] }
+
 # Serialization
 serde = { workspace = true, features = [] }
 serde_json = { workspace = true, features = [] }
@@ -41,12 +45,10 @@ derive_more = { workspace = true, features = ["as_ref", "deref", "deref_mut", "d
 strum = { workspace = true, features = [] }
 
 # Primitive datatypes
-uuid = { workspace = true, features = [] }
-bytes = { workspace = true, features = [] }
+uuid = { workspace = true, features = ["v7"] }
 hipstr = { workspace = true, features = [] }
-oxilangtag = { workspace = true, features = [] }
-celes = { workspace = true, features = [] }
-type-map = { workspace = true, features = [] }
+jiff = { workspace = true, features = ["serde"] }
+semver = { workspace = true, features = ["serde"] }
 
 # Async runtime and parallelism
 async-trait = { workspace = true, features = [] }
diff --git a/crates/nvisy-core/src/context/analytic/embedding.rs b/crates/nvisy-core/src/context/analytic/embedding.rs
new file mode 100644
index 000000000..16e3066bb
--- /dev/null
+++ b/crates/nvisy-core/src/context/analytic/embedding.rs
@@ -0,0 +1,41 @@
+//! Embedding reference data for similarity search.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Distance metric for vector similarity comparison.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum DistanceMetric {
+    /// Cosine similarity (1 − cos θ).
+    Cosine,
+    /// Euclidean (L2) distance.
+    Euclidean,
+    /// Dot product (inner product).
+    DotProduct,
+    /// Manhattan (L1) distance.
+    Manhattan,
+}
+
+/// Pre-computed embedding vector for similarity comparison.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct EmbeddingData {
+    /// The embedding vector values.
+    pub vector: Vec<f32>,
+    /// Identifier of the model/algorithm that produced this embedding.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub algorithm: Option<String>,
+    /// Distance metric to use for comparison.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub distance_metric: Option<DistanceMetric>,
+}
+
+impl EmbeddingData {
+    /// Dimensionality of the embedding vector.
+    pub fn dimensions(&self) -> usize {
+        self.vector.len()
+    }
+}
diff --git a/crates/nvisy-core/src/context/analytic/mod.rs b/crates/nvisy-core/src/context/analytic/mod.rs
new file mode 100644
index 000000000..2487456a0
--- /dev/null
+++ b/crates/nvisy-core/src/context/analytic/mod.rs
@@ -0,0 +1,21 @@
+//! Computed representations for similarity search and pattern matching.
+
+mod embedding;
+mod pattern;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+pub use self::embedding::{DistanceMetric, EmbeddingData};
+pub use self::pattern::{GlobPattern, PatternData, PatternExpression, RegexPattern};
+
+/// Analytic computation variants.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum AnalyticVariant {
+    /// Pre-computed embedding vector for similarity search.
+    Embedding(EmbeddingData),
+    /// Regex or glob pattern for matching.
+    Pattern(PatternData),
+}
diff --git a/crates/nvisy-core/src/context/analytic/pattern.rs b/crates/nvisy-core/src/context/analytic/pattern.rs
new file mode 100644
index 000000000..d5248bc06
--- /dev/null
+++ b/crates/nvisy-core/src/context/analytic/pattern.rs
@@ -0,0 +1,46 @@
+//! Pattern reference data for regex/glob matching.
+
+use derive_more::From;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// A regular expression pattern.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct RegexPattern {
+    /// The regex expression string.
+    pub expression: String,
+}
+
+/// A shell-style glob pattern.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct GlobPattern {
+    /// The glob expression string.
+    pub expression: String,
+}
+
+/// A pattern expression with its syntax type.
+#[derive(Debug, Clone, PartialEq, Eq, From, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "syntax", rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum PatternExpression {
+    /// Regular expression pattern.
+    Regex(RegexPattern),
+    /// Shell-style glob pattern.
+    Glob(GlobPattern),
+}
+
+/// A named pattern for detection matching.
+///
+/// `label` describes the intent (for humans/LLMs); `pattern` carries
+/// the expression and its type.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct PatternData {
+    /// Human-readable label describing what this pattern matches.
+    pub label: String,
+    /// The pattern expression and its type.
+    #[serde(flatten)]
+    pub pattern: PatternExpression,
+}
diff --git a/crates/nvisy-core/src/context/biometric/face.rs b/crates/nvisy-core/src/context/biometric/face.rs
new file mode 100644
index 000000000..e10e8cb07
--- /dev/null
+++ b/crates/nvisy-core/src/context/biometric/face.rs
@@ -0,0 +1,54 @@
+//! Face biometric reference data.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::ContentSource;
+use crate::schema::BoundingBoxSchema;
+
+/// Reference face data for identity matching.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct FaceData {
+    /// Source pointer to the reference face image.
+    pub image_source: ContentSource,
+    /// Bounding box of the face within the image.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub region: Option<BoundingBoxSchema>,
+    /// Base64-encoded face embedding / template.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub template: Option<String>,
+    /// Algorithm that produced the template (e.g. `"arcface"`, `"facenet"`).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub algorithm: Option<String>,
+}
+
+impl FaceData {
+    /// Create face data pointing at a source image.
+    pub fn new(image_source: ContentSource) -> Self {
+        Self {
+            image_source,
+            region: None,
+            template: None,
+            algorithm: None,
+        }
+    }
+
+    /// Set the face bounding box.
+    pub fn with_region(mut self, region: BoundingBoxSchema) -> Self {
+        self.region = Some(region);
+        self
+    }
+
+    /// Set the encoded face template.
+    pub fn with_template(mut self, template: impl Into<String>) -> Self {
+        self.template = Some(template.into());
+        self
+    }
+
+    /// Set the extraction algorithm.
+    pub fn with_algorithm(mut self, algorithm: impl Into<String>) -> Self {
+        self.algorithm = Some(algorithm.into());
+        self
+    }
+}
diff --git a/crates/nvisy-core/src/context/biometric/mod.rs b/crates/nvisy-core/src/context/biometric/mod.rs
new file mode 100644
index 000000000..461a58865
--- /dev/null
+++ b/crates/nvisy-core/src/context/biometric/mod.rs
@@ -0,0 +1,21 @@
+//! Biometric reference data for identity verification.
+
+mod face;
+mod voice;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+pub use self::face::FaceData;
+pub use self::voice::VoiceData;
+
+/// Biometric identity verification variants.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum BiometricVariant {
+    /// Reference face for identity matching.
+    Face(FaceData),
+    /// Reference voiceprint for speaker identification.
+    Voice(VoiceData),
+}
diff --git a/crates/nvisy-core/src/context/biometric/voice.rs b/crates/nvisy-core/src/context/biometric/voice.rs
new file mode 100644
index 000000000..b3a21213d
--- /dev/null
+++ b/crates/nvisy-core/src/context/biometric/voice.rs
@@ -0,0 +1,54 @@
+//! Voice biometric reference data.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::ContentSource;
+use crate::schema::TimeSpanSchema;
+
+/// Reference voice data for speaker identification.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct VoiceData {
+    /// Source pointer to the reference audio.
+    pub audio_source: ContentSource,
+    /// Segment within the audio used for enrollment.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub segment: Option<TimeSpanSchema>,
+    /// Base64-encoded speaker embedding / voiceprint.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub template: Option<String>,
+    /// Algorithm that produced the voiceprint (e.g. `"ecapa-tdnn"`, `"x-vector"`).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub algorithm: Option<String>,
+}
+
+impl VoiceData {
+    /// Create voice data pointing at a source audio.
+    pub fn new(audio_source: ContentSource) -> Self {
+        Self {
+            audio_source,
+            segment: None,
+            template: None,
+            algorithm: None,
+        }
+    }
+
+    /// Set the enrollment segment.
+    pub fn with_segment(mut self, segment: TimeSpanSchema) -> Self {
+        self.segment = Some(segment);
+        self
+    }
+
+    /// Set the encoded voiceprint.
+    pub fn with_template(mut self, template: impl Into<String>) -> Self {
+        self.template = Some(template.into());
+        self
+    }
+
+    /// Set the extraction algorithm.
+    pub fn with_algorithm(mut self, algorithm: impl Into<String>) -> Self {
+        self.algorithm = Some(algorithm.into());
+        self
+    }
+}
diff --git a/crates/nvisy-core/src/context/document/mod.rs b/crates/nvisy-core/src/context/document/mod.rs
new file mode 100644
index 000000000..9d878b11e
--- /dev/null
+++ b/crates/nvisy-core/src/context/document/mod.rs
@@ -0,0 +1,21 @@
+//! Document templates and handwritten signatures.
+
+mod signature;
+mod template;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+pub use self::signature::SignatureData;
+pub use self::template::TemplateData;
+
+/// Document-related reference variants.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum DocumentVariant {
+    /// Reference document template for layout classification.
+    Template(TemplateData),
+    /// Handwritten signature for verification.
+    Signature(SignatureData),
+}
diff --git a/crates/nvisy-core/src/context/document/signature.rs b/crates/nvisy-core/src/context/document/signature.rs
new file mode 100644
index 000000000..1001437a9
--- /dev/null
+++ b/crates/nvisy-core/src/context/document/signature.rs
@@ -0,0 +1,48 @@
+//! Signature reference data.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::ContentSource;
+use crate::schema::BoundingBoxSchema;
+
+/// Reference handwritten signature for verification.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct SignatureData {
+    /// Source pointer to the reference signature image.
+    pub image_source: ContentSource,
+    /// Bounding box of the signature within the image.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub region: Option<BoundingBoxSchema>,
+    /// Identity of the signer this signature belongs to.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub signer_id: Option<String>,
+    /// Algorithm used for signature verification.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub algorithm: Option<String>,
+}
+
+impl SignatureData {
+    /// Create signature data pointing at a source image.
+    pub fn new(image_source: ContentSource) -> Self {
+        Self {
+            image_source,
+            region: None,
+            signer_id: None,
+            algorithm: None,
+        }
+    }
+
+    /// Set the signer identity.
+    pub fn with_signer_id(mut self, signer_id: impl Into<String>) -> Self {
+        self.signer_id = Some(signer_id.into());
+        self
+    }
+
+    /// Set the verification algorithm.
+    pub fn with_algorithm(mut self, algorithm: impl Into<String>) -> Self {
+        self.algorithm = Some(algorithm.into());
+        self
+    }
+}
diff --git a/crates/nvisy-core/src/context/document/template.rs b/crates/nvisy-core/src/context/document/template.rs
new file mode 100644
index 000000000..4b090a17c
--- /dev/null
+++ b/crates/nvisy-core/src/context/document/template.rs
@@ -0,0 +1,41 @@
+//! Document template reference data.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::ContentSource;
+use crate::schema::BoundingBoxSchema;
+
+/// Reference document template for layout/type classification.
+///
+/// Used to detect documents of a known type (ID cards, passports, forms,
+/// invoices) by comparing their visual layout against this reference.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TemplateData {
+    /// Source pointer to the reference template image.
+    pub image_source: ContentSource,
+    /// Optional sub-region of interest within the template.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub region: Option<BoundingBoxSchema>,
+    /// Document type label (e.g. `"passport"`, `"drivers_license"`, `"invoice"`).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub document_type: Option<String>,
+}
+
+impl TemplateData {
+    /// Create template data pointing at a source image.
+    pub fn new(image_source: ContentSource) -> Self {
+        Self {
+            image_source,
+            region: None,
+            document_type: None,
+        }
+    }
+
+    /// Set the document type label.
+    pub fn with_document_type(mut self, document_type: impl Into<String>) -> Self {
+        self.document_type = Some(document_type.into());
+        self
+    }
+}
diff --git a/crates/nvisy-core/src/context/entry.rs b/crates/nvisy-core/src/context/entry.rs
new file mode 100644
index 000000000..bc0fc9fa7
--- /dev/null
+++ b/crates/nvisy-core/src/context/entry.rs
@@ -0,0 +1,84 @@
+//! Context entry types for reference data.
+
+use derive_more::From;
+use jiff::Timestamp;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use uuid::Uuid;
+
+use super::analytic::AnalyticVariant;
+use super::biometric::BiometricVariant;
+use super::document::DocumentVariant;
+use super::geospatial::GeospatialVariant;
+use super::reference::ReferenceVariant;
+use super::temporal::TemporalVariant;
+
+/// Top-level domain classification for context reference data.
+///
+/// Each domain contains a nested enum of specific variants,
+/// keeping modality and semantic purpose cleanly separated.
+#[derive(Debug, Clone, From, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "domain", content = "data", rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum ContextEntryData {
+    /// Identity verification via biological traits.
+    Biometric(BiometricVariant),
+    /// Geographic regions and addresses.
+    Geospatial(GeospatialVariant),
+    /// Computed representations for similarity search and pattern matching.
+    Analytic(AnalyticVariant),
+    /// Raw data for direct comparison against input.
+    Reference(ReferenceVariant),
+    /// Date and time-based matching.
+    Temporal(TemporalVariant),
+    /// Document templates and handwritten signatures.
+    Document(DocumentVariant),
+}
+
+/// A single reference-data entry within a [`Context`].
+///
+/// [`Context`]: super::Context
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct ContextEntry {
+    /// Unique identifier for this entry.
+    pub id: Uuid,
+    /// Human-readable label.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub label: Option<String>,
+    /// When this entry was created.
+    #[schemars(with = "String")]
+    pub created_at: Timestamp,
+    /// When this entry should stop being used.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[schemars(with = "Option<String>")]
+    pub expires_at: Option<Timestamp>,
+    /// Semantically typed payload.
+    #[serde(flatten)]
+    pub data: ContextEntryData,
+}
+
+impl ContextEntry {
+    /// Create a new context entry with a generated UUID and current timestamp.
+    pub fn new(data: ContextEntryData) -> Self {
+        Self {
+            id: Uuid::now_v7(),
+            label: None,
+            created_at: Timestamp::now(),
+            expires_at: None,
+            data,
+        }
+    }
+
+    /// Set a human-readable label on this entry.
+    pub fn with_label(mut self, label: impl Into<String>) -> Self {
+        self.label = Some(label.into());
+        self
+    }
+
+    /// Set an expiration timestamp on this entry.
+    pub fn with_expires_at(mut self, expires_at: Timestamp) -> Self {
+        self.expires_at = Some(expires_at);
+        self
+    }
+}
diff --git a/crates/nvisy-core/src/context/geospatial/address.rs b/crates/nvisy-core/src/context/geospatial/address.rs
new file mode 100644
index 000000000..5b92413a8
--- /dev/null
+++ b/crates/nvisy-core/src/context/geospatial/address.rs
@@ -0,0 +1,30 @@
+//! Structured address reference data.
+
+use std::collections::BTreeMap;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Structured address reference.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct AddressData {
+    /// Street line (e.g. `"123 Main St"`).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub street: Option<String>,
+    /// City name.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub city: Option<String>,
+    /// State, province, or administrative area.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub state: Option<String>,
+    /// Postal / ZIP code.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub postal_code: Option<String>,
+    /// ISO 3166-1 alpha-2 country code.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub country: Option<String>,
+    /// Additional free-form address fields.
+    #[serde(default, skip_serializing_if = "BTreeMap::is_empty")]
+    pub extra: BTreeMap<String, String>,
+}
diff --git a/crates/nvisy-core/src/context/geospatial/coordinates.rs b/crates/nvisy-core/src/context/geospatial/coordinates.rs
new file mode 100644
index 000000000..4e6715546
--- /dev/null
+++ b/crates/nvisy-core/src/context/geospatial/coordinates.rs
@@ -0,0 +1,31 @@
+//! Geographic coordinate primitives.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::schema::PointSchema;
+
+/// A geographic coordinate (latitude/longitude).
+#[derive(Debug, Clone, Copy, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct GeoCoordinate {
+    /// Latitude in decimal degrees (−90 to 90).
+    pub lat: f64,
+    /// Longitude in decimal degrees (−180 to 180).
+    pub lng: f64,
+}
+
+impl GeoCoordinate {
+    /// Create a new coordinate.
+    pub fn new(lat: f64, lng: f64) -> Self {
+        Self { lat, lng }
+    }
+
+    /// Convert to a [`PointSchema`] for polygon operations.
+    pub fn to_point(self) -> PointSchema {
+        PointSchema {
+            x: self.lng,
+            y: self.lat,
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/context/geospatial/mod.rs b/crates/nvisy-core/src/context/geospatial/mod.rs
new file mode 100644
index 000000000..7fa7f131d
--- /dev/null
+++ b/crates/nvisy-core/src/context/geospatial/mod.rs
@@ -0,0 +1,23 @@
+//! Geospatial reference data for location-based matching.
+
+mod address;
+mod coordinates;
+mod region;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+pub use self::address::AddressData;
+pub use self::coordinates::GeoCoordinate;
+pub use self::region::{GeoBounds, GeoShape, RegionData};
+
+/// Geospatial location variants.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum GeospatialVariant {
+    /// Geographic region (bounds, circle, or polygon).
+    Region(RegionData),
+    /// Structured postal address.
+    Address(AddressData),
+}
diff --git a/crates/nvisy-core/src/context/geospatial/region.rs b/crates/nvisy-core/src/context/geospatial/region.rs
new file mode 100644
index 000000000..6e85838d2
--- /dev/null
+++ b/crates/nvisy-core/src/context/geospatial/region.rs
@@ -0,0 +1,100 @@
+//! Geospatial region data for location-based detection.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use super::coordinates::GeoCoordinate;
+use crate::schema::PolygonSchema;
+
+/// A geographic bounding box defined by its south-west and north-east corners.
+#[derive(Debug, Clone, Copy, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct GeoBounds {
+    /// South-west corner.
+    pub south_west: GeoCoordinate,
+    /// North-east corner.
+    pub north_east: GeoCoordinate,
+}
+
+impl GeoBounds {
+    /// Create bounds from corner coordinates.
+    pub fn new(south_west: GeoCoordinate, north_east: GeoCoordinate) -> Self {
+        Self {
+            south_west,
+            north_east,
+        }
+    }
+
+    /// Check whether a coordinate falls within this bounding box.
+    pub fn contains(&self, point: &GeoCoordinate) -> bool {
+        point.lat >= self.south_west.lat
+            && point.lat <= self.north_east.lat
+            && point.lng >= self.south_west.lng
+            && point.lng <= self.north_east.lng
+    }
+}
+
+/// Shape of a geospatial region.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "shape", rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum GeoShape {
+    /// Axis-aligned bounding rectangle.
+    Bounds(GeoBounds),
+    /// Circular region defined by center and radius.
+    Circle {
+        /// Center of the circle.
+        center: GeoCoordinate,
+        /// Radius in meters.
+        radius_m: f64,
+    },
+    /// Arbitrary polygon defined by vertices.
+    Polygon {
+        /// Boundary vertices in order (x = lng, y = lat).
+        boundary: PolygonSchema,
+    },
+}
+
+/// Geospatial region reference data for location-based detection.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct RegionData {
+    /// The geographic region.
+    #[serde(flatten)]
+    pub region: GeoShape,
+    /// Optional human-readable name for this region.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub name: Option<String>,
+}
+
+impl RegionData {
+    /// Create region data from bounds.
+    pub fn from_bounds(bounds: GeoBounds) -> Self {
+        Self {
+            region: GeoShape::Bounds(bounds),
+            name: None,
+        }
+    }
+
+    /// Create a circular region.
+    pub fn from_circle(center: GeoCoordinate, radius_m: f64) -> Self {
+        Self {
+            region: GeoShape::Circle { center, radius_m },
+            name: None,
+        }
+    }
+
+    /// Create a polygon region.
+    pub fn from_polygon(boundary: PolygonSchema) -> Self {
+        Self {
+            region: GeoShape::Polygon { boundary },
+            name: None,
+        }
+    }
+
+    /// Set a name on this region.
+    pub fn with_name(mut self, name: impl Into<String>) -> Self {
+        self.name = Some(name.into());
+        self
+    }
+}
diff --git a/crates/nvisy-core/src/context/mod.rs b/crates/nvisy-core/src/context/mod.rs
new file mode 100644
index 000000000..396a9ba8a
--- /dev/null
+++ b/crates/nvisy-core/src/context/mod.rs
@@ -0,0 +1,66 @@
+//! Persistent reference-data collections for detection.
+//!
+//! A [`Context`] holds reusable reference data — names, faces, voices,
+//! patterns, embeddings — that tells detection *what to look for*.  It is
+//! separate from policy (which controls *what to do* when something is found).
+//!
+//! ## Current status
+//!
+//! Today this subtree is **schema-only**. The ontology models the shape
+//! [`Context`]s would have (analytic / biometric / document / geospatial
+//! / reference / temporal entry kinds), but no detection backend in the
+//! workspace currently consumes a `Context` — recognizers receive their
+//! reference data through backend-specific configuration. The types
+//! exist so persisted contexts round-trip cleanly through the engine
+//! and so future backends have a stable target to build against.
+//! When a recognizer starts matching on a specific variant, that's the
+//! cue to evaluate whether the shape still fits and prune what doesn't.
+
+pub mod analytic;
+pub mod biometric;
+pub mod document;
+mod entry;
+pub mod geospatial;
+pub mod reference;
+pub mod temporal;
+
+use derive_builder::Builder;
+use schemars::JsonSchema;
+use semver::Version;
+use serde::{Deserialize, Serialize};
+use uuid::Uuid;
+
+pub use self::entry::{ContextEntry, ContextEntryData};
+
+/// A persistent, reusable collection of reference data for detection.
+#[derive(Debug, Clone, Builder, Serialize, Deserialize, JsonSchema)]
+#[builder(
+    name = "ContextBuilder",
+    pattern = "owned",
+    setter(into, strip_option, prefix = "with")
+)]
+#[serde(rename_all = "camelCase")]
+pub struct Context {
+    /// Unique identifier for this context.
+    #[builder(default = "Uuid::now_v7()")]
+    pub id: Uuid,
+    /// Human-readable label for this context.
+    pub name: String,
+    /// Context version.
+    #[schemars(with = "String")]
+    pub version: Version,
+    /// Optional longer description.
+    #[builder(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub description: Option<String>,
+    /// Reference-data entries.
+    #[builder(default)]
+    pub entries: Vec<ContextEntry>,
+}
+
+impl Context {
+    /// Start building a new context.
+    pub fn builder() -> ContextBuilder {
+        ContextBuilder::default()
+    }
+}
diff --git a/crates/nvisy-core/src/context/reference/credential.rs b/crates/nvisy-core/src/context/reference/credential.rs
new file mode 100644
index 000000000..eb4c63b95
--- /dev/null
+++ b/crates/nvisy-core/src/context/reference/credential.rs
@@ -0,0 +1,77 @@
+//! Credential reference data for secret detection.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Classification of credential secrets.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum CredentialKind {
+    /// API key.
+    ApiKey,
+    /// OAuth access or refresh token.
+    OauthToken,
+    /// Password or passphrase.
+    Password,
+    /// Private key (SSH, TLS, etc.).
+    PrivateKey,
+    /// Other credential type.
+    Other,
+}
+
+/// A reference credential for detecting leaked secrets.
+///
+/// The `value` field is intentionally excluded from serialization to
+/// prevent plaintext secrets from appearing in logs or API responses.
+/// It is only accepted during deserialization (ingest).
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct CredentialData {
+    /// The credential value (API key, token, etc.).
+    ///
+    /// Excluded from serialization output to prevent leaking secrets.
+    /// Round-trip deserialization yields an empty string.
+    #[serde(default, skip_serializing)]
+    pub value: String,
+    /// Classification of this credential.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub credential_kind: Option<CredentialKind>,
+    /// Service or provider this credential belongs to.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub provider: Option<String>,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn value_is_not_serialized() {
+        let cred = CredentialData {
+            value: "sk-secret-token".to_owned(),
+            credential_kind: Some(CredentialKind::ApiKey),
+            provider: Some("openai".to_owned()),
+        };
+        let json = serde_json::to_string(&cred).unwrap();
+        assert!(
+            !json.contains("sk-secret-token"),
+            "serialized JSON must not contain plaintext value, got: {json}"
+        );
+        assert!(json.contains(r#""credentialKind":"api_key""#));
+    }
+
+    #[test]
+    fn round_trip_defaults_value_to_empty() {
+        // Serialize without value, then deserialize: value must default to empty string.
+        let cred = CredentialData {
+            value: "secret".to_owned(),
+            credential_kind: None,
+            provider: None,
+        };
+        let json = serde_json::to_string(&cred).unwrap();
+        let back: CredentialData = serde_json::from_str(&json).unwrap();
+        assert_eq!(back.value, "");
+    }
+}
diff --git a/crates/nvisy-core/src/context/reference/image.rs b/crates/nvisy-core/src/context/reference/image.rs
new file mode 100644
index 000000000..d54bf4224
--- /dev/null
+++ b/crates/nvisy-core/src/context/reference/image.rs
@@ -0,0 +1,18 @@
+//! Image reference data for object matching.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::ContentSource;
+use crate::schema::BoundingBoxSchema;
+
+/// Reference image for object/scene matching.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct ImageData {
+    /// Source pointer to the reference image.
+    pub image_source: ContentSource,
+    /// Optional sub-region within the image.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub region: Option<BoundingBoxSchema>,
+}
diff --git a/crates/nvisy-core/src/context/reference/mod.rs b/crates/nvisy-core/src/context/reference/mod.rs
new file mode 100644
index 000000000..cf2656001
--- /dev/null
+++ b/crates/nvisy-core/src/context/reference/mod.rs
@@ -0,0 +1,29 @@
+//! Raw reference data for direct comparison against input.
+
+mod credential;
+mod image;
+mod tag;
+mod text;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+pub use self::credential::{CredentialData, CredentialKind};
+pub use self::image::ImageData;
+pub use self::tag::TagData;
+pub use self::text::{TextData, TextEntry};
+
+/// Direct comparison reference variants.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum ReferenceVariant {
+    /// Names, identifiers, or phrases to match.
+    Text(TextData),
+    /// Keyword tag for classification and routing.
+    Tag(TagData),
+    /// API keys, tokens, or known secret patterns.
+    Credential(CredentialData),
+    /// Reference image for object/scene matching.
+    Image(ImageData),
+}
diff --git a/crates/nvisy-core/src/context/reference/tag.rs b/crates/nvisy-core/src/context/reference/tag.rs
new file mode 100644
index 000000000..4753f4c4c
--- /dev/null
+++ b/crates/nvisy-core/src/context/reference/tag.rs
@@ -0,0 +1,15 @@
+//! Tag reference data for keyword-based routing.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// A keyword tag for classification and routing.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TagData {
+    /// The tag value (e.g. `"pii"`, `"financial"`, `"hipaa"`).
+    pub value: String,
+    /// Optional category grouping related tags.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub category: Option<String>,
+}
diff --git a/crates/nvisy-core/src/context/reference/text.rs b/crates/nvisy-core/src/context/reference/text.rs
new file mode 100644
index 000000000..f9bb46ce1
--- /dev/null
+++ b/crates/nvisy-core/src/context/reference/text.rs
@@ -0,0 +1,61 @@
+//! Text reference data for direct comparison.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::schema::LanguageTagSchema;
+
+/// A labeled text value for reference matching.
+///
+/// The `key` is a human/LLM-readable label describing what this value
+/// represents (e.g. `"full_name"`, `"account_number"`).  The `value` is
+/// the literal string used for pattern matching.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TextEntry {
+    /// Human/LLM-readable label.
+    pub key: String,
+    /// Literal value for pattern matching.
+    pub value: String,
+}
+
+/// Textual reference data as key-value pairs.
+///
+/// Keys describe *what* a value represents (for humans and LLMs);
+/// values are the literal strings used for matching.
+#[derive(Debug, Clone, Default, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TextData {
+    /// Key-value pairs for matching.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub entries: Vec<TextEntry>,
+    /// BCP-47 language tag for locale-sensitive matching.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub language: Option<LanguageTagSchema>,
+}
+
+impl TextData {
+    /// Create text data from a list of entries.
+    pub fn new(entries: Vec<TextEntry>) -> Self {
+        Self {
+            entries,
+            language: None,
+        }
+    }
+
+    /// Set the language for locale-sensitive matching.
+    pub fn with_language(mut self, language: LanguageTagSchema) -> Self {
+        self.language = Some(language);
+        self
+    }
+}
+
+impl TextEntry {
+    /// Create a labeled text entry.
+    pub fn new(key: impl Into<String>, value: impl Into<String>) -> Self {
+        Self {
+            key: key.into(),
+            value: value.into(),
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/context/temporal/date.rs b/crates/nvisy-core/src/context/temporal/date.rs
new file mode 100644
index 000000000..0a60b87ca
--- /dev/null
+++ b/crates/nvisy-core/src/context/temporal/date.rs
@@ -0,0 +1,36 @@
+//! Date and date-range reference data.
+
+use jiff::civil::Date;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// A date or date-range reference for temporal matching.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct DateData {
+    /// Start (or only) date.
+    #[schemars(with = "String")]
+    pub start: Date,
+    /// End date for a range. When `None` this represents a single date.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[schemars(with = "Option<String>")]
+    pub end: Option<Date>,
+}
+
+impl DateData {
+    /// Create a single-date reference.
+    pub fn single(date: Date) -> Self {
+        Self {
+            start: date,
+            end: None,
+        }
+    }
+
+    /// Create a date-range reference.
+    pub fn range(start: Date, end: Date) -> Self {
+        Self {
+            start,
+            end: Some(end),
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/context/temporal/datetime.rs b/crates/nvisy-core/src/context/temporal/datetime.rs
new file mode 100644
index 000000000..1e9e35f54
--- /dev/null
+++ b/crates/nvisy-core/src/context/temporal/datetime.rs
@@ -0,0 +1,43 @@
+//! Date-time reference data.
+
+use jiff::civil::DateTime;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// A date-time or date-time range reference for temporal matching.
+///
+/// Uses naive (timezone-unaware) date-times from [`DateTime`].
+/// For timezone-aware timestamps, use the entry-level `created_at` /
+/// `expires_at` fields on [`ContextEntry`].
+///
+/// [`ContextEntry`]: crate::context::ContextEntry
+/// [`DateTime`]: jiff::civil::DateTime
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct DateTimeData {
+    /// Start (or only) date-time.
+    #[schemars(with = "String")]
+    pub start: DateTime,
+    /// End date-time for a range. When `None` this represents a single instant.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[schemars(with = "Option<String>")]
+    pub end: Option<DateTime>,
+}
+
+impl DateTimeData {
+    /// Create a single date-time reference.
+    pub fn single(dt: DateTime) -> Self {
+        Self {
+            start: dt,
+            end: None,
+        }
+    }
+
+    /// Create a date-time range reference.
+    pub fn range(start: DateTime, end: DateTime) -> Self {
+        Self {
+            start,
+            end: Some(end),
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/context/temporal/mod.rs b/crates/nvisy-core/src/context/temporal/mod.rs
new file mode 100644
index 000000000..274551bba
--- /dev/null
+++ b/crates/nvisy-core/src/context/temporal/mod.rs
@@ -0,0 +1,29 @@
+//! Date, time, and datetime reference data.
+
+mod date;
+mod datetime;
+mod time;
+mod timespan;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+pub use self::date::DateData;
+pub use self::datetime::DateTimeData;
+pub use self::time::TimeData;
+pub use self::timespan::TimeSpanData;
+
+/// Temporal matching variants.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+#[non_exhaustive]
+pub enum TemporalVariant {
+    /// Date or date-range to match (e.g. birthdays, expiry dates).
+    Date(DateData),
+    /// Time-of-day or time range (e.g. business hours).
+    Time(TimeData),
+    /// Date-time or date-time range (e.g. appointment timestamps).
+    DateTime(DateTimeData),
+    /// Time span reference for matching audio/video segments.
+    TimeSpan(TimeSpanData),
+}
diff --git a/crates/nvisy-core/src/context/temporal/time.rs b/crates/nvisy-core/src/context/temporal/time.rs
new file mode 100644
index 000000000..c7734b500
--- /dev/null
+++ b/crates/nvisy-core/src/context/temporal/time.rs
@@ -0,0 +1,41 @@
+//! Time-of-day reference data.
+
+use jiff::civil::Time;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// A time-of-day or time range reference for temporal matching.
+///
+/// Uses naive (timezone-unaware) times from [`Time`].
+/// Useful for matching recurring time patterns (e.g. business hours).
+///
+/// [`Time`]: jiff::civil::Time
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TimeData {
+    /// Start (or only) time.
+    #[schemars(with = "String")]
+    pub start: Time,
+    /// End time for a range. When `None` this represents a single time.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[schemars(with = "Option<String>")]
+    pub end: Option<Time>,
+}
+
+impl TimeData {
+    /// Create a single time-of-day reference.
+    pub fn single(time: Time) -> Self {
+        Self {
+            start: time,
+            end: None,
+        }
+    }
+
+    /// Create a time range reference.
+    pub fn range(start: Time, end: Time) -> Self {
+        Self {
+            start,
+            end: Some(end),
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/context/temporal/timespan.rs b/crates/nvisy-core/src/context/temporal/timespan.rs
new file mode 100644
index 000000000..dc36633d8
--- /dev/null
+++ b/crates/nvisy-core/src/context/temporal/timespan.rs
@@ -0,0 +1,30 @@
+//! Time span reference data.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::schema::TimeSpanSchema;
+
+/// A time span reference for matching audio/video segments.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TimeSpanData {
+    /// The time interval to match.
+    pub span: TimeSpanSchema,
+    /// Optional human-readable label (e.g. `"intro"`, `"closing remarks"`).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub label: Option<String>,
+}
+
+impl TimeSpanData {
+    /// Create a time span reference.
+    pub fn new(span: TimeSpanSchema) -> Self {
+        Self { span, label: None }
+    }
+
+    /// Set a label.
+    pub fn with_label(mut self, label: impl Into<String>) -> Self {
+        self.label = Some(label.into());
+        self
+    }
+}
diff --git a/crates/nvisy-core/src/entity/annotation.rs b/crates/nvisy-core/src/entity/annotation.rs
deleted file mode 100644
index 5f9b419b2..000000000
--- a/crates/nvisy-core/src/entity/annotation.rs
+++ /dev/null
@@ -1,311 +0,0 @@
-//! User-supplied annotations for pre-identified regions and
-//! classification labels.
-//!
-//! Two kinds of annotation exist with different shapes because they
-//! describe different things:
-//!
-//! - [`Annotation<M>`] (modality-typed): per-region [`Inclusion`] and
-//!   [`Exclusion`] markers that target a concrete `M` location
-//!   (text span, image region, audio segment, tabular cell). Both
-//!   kinds carry an [`AnnotationStrength`] so the engine knows
-//!   whether the user asked for advisory bias or a hard constraint.
-//! - [`LabelAnnotation`] (modality-agnostic): document-wide
-//!   classification labels that don't reference a per-modality
-//!   location. Apply identically to every modality envelope spawned
-//!   from the same source.
-//!
-//! Value-style "treat this exact string as sensitive everywhere it
-//! appears" annotations are intentionally not part of this surface;
-//! that role belongs to the dictionary / deny-list pattern path,
-//! which is configured per run rather than per upload.
-//!
-//! [`Inclusion`]: AnnotationKind::Inclusion
-//! [`Exclusion`]: AnnotationKind::Exclusion
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::{AnnotationProvenance, Entity, EntityLabelRef, TrailProvenance, TrailStep, builtins};
-use crate::modality::{Modality, Overlap};
-use crate::primitive::Confidence;
-
-/// How firmly a region annotation should bind detector behavior.
-///
-/// Inclusion and exclusion both carry a strength so callers can
-/// distinguish "please look out for / try to skip this" from
-/// "this *is* / *is not* sensitive, no matter what the detector
-/// says."
-///
-/// Confidence lives on [`Hint`] only because asserting
-/// a thing is, by definition, certain — synthesised entities from
-/// asserted inclusions always materialise at full confidence
-/// (`1.0`).
-///
-/// [`Hint`]: Self::Hint
-#[derive(Debug, Clone, Copy, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-#[non_exhaustive]
-pub enum AnnotationStrength {
-    /// Advisory bias for LLM/VLM detectors. Rendered into the
-    /// detector's prompt. Detectors without a prompt surface
-    /// (regex, dictionary) ignore the hint; LLM detectors may still
-    /// reject it.
-    ///
-    /// `confidence` is forwarded to the prompt and recorded on
-    /// detected entities if the LLM honours the hint. `None`
-    /// defaults to `1.0` when materialised.
-    Hint {
-        /// Confidence to forward to the detector when the hint is
-        /// honoured. `None` defaults to `1.0` at materialisation time.
-        #[serde(skip_serializing_if = "Option::is_none")]
-        confidence: Option<Confidence>,
-    },
-    /// Hard constraint enforced by the engine regardless of which
-    /// detectors ran. An asserted [`Inclusion`] is materialised as
-    /// a synthetic entity at import time at full confidence (`1.0`);
-    /// an asserted [`Exclusion`] drops matching detections
-    /// post-filter.
-    ///
-    /// [`Inclusion`]: AnnotationKind::Inclusion
-    /// [`Exclusion`]: AnnotationKind::Exclusion
-    Assert,
-}
-
-/// The kind of region annotation with variant-specific data.
-///
-/// Document-level labels are not part of this enum — they live on
-/// [`LabelAnnotation`] because they don't carry a modality
-/// parameter.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum AnnotationKind<M: Modality> {
-    /// Pre-identified region the user wants treated as sensitive.
-    Inclusion {
-        /// Label to attach to the synthesised entity. `None`
-        /// when the user wants the region treated as sensitive
-        /// without committing to a kind — synthesised entities
-        /// then fall back to [`builtins::UNRESOLVED`].
-        ///
-        /// [`builtins::UNRESOLVED`]: super::builtins::UNRESOLVED
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        label: Option<EntityLabelRef>,
-        /// Modality-specific location this inclusion targets.
-        target: M::Location,
-        /// Whether this is an advisory [`Hint`] (LLM may reject) or
-        /// a hard [`Assert`] (engine enforces regardless of
-        /// detectors). Lives on `Inclusion` only because exclusions
-        /// are always assertions — there's no meaningful "maybe
-        /// safe" mode.
-        ///
-        /// [`Hint`]: AnnotationStrength::Hint
-        /// [`Assert`]: AnnotationStrength::Assert
-        strength: AnnotationStrength,
-    },
-    /// Known-safe region the user wants the engine to skip.
-    /// Exclusions are always treated as assertions — there is no
-    /// "hint" exclusion variant because letting the LLM second-guess
-    /// a user's safety claim would defeat the point.
-    Exclusion {
-        /// Modality-specific location this exclusion targets.
-        target: M::Location,
-    },
-}
-
-/// A user-provided region annotation on a content region.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct Annotation<M: Modality> {
-    /// Optional human-readable name for this annotation.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub name: Option<String>,
-    /// What kind of annotation and its variant-specific data.
-    #[serde(flatten)]
-    pub kind: AnnotationKind<M>,
-}
-
-impl<M: Modality> Annotation<M> {
-    /// Convert an [`Assert`]-strength [`Inclusion`] annotation into
-    /// an [`Entity<M>`]. Returns `None` for non-inclusion kinds and
-    /// for [`Hint`]-strength inclusions (which the caller should
-    /// render into a detector prompt instead).
-    ///
-    /// [`Inclusion`]: AnnotationKind::Inclusion
-    /// [`Assert`]: AnnotationStrength::Assert
-    /// [`Hint`]: AnnotationStrength::Hint
-    pub fn to_inclusion_entity(&self) -> Option<Entity<M>> {
-        let AnnotationKind::Inclusion {
-            label,
-            target,
-            strength: AnnotationStrength::Assert,
-        } = &self.kind
-        else {
-            return None;
-        };
-
-        let label_ref = label
-            .clone()
-            .unwrap_or_else(|| EntityLabelRef::from(builtins::UNRESOLVED.name.clone()));
-
-        let entity = Entity::builder()
-            .with_label(label_ref)
-            .with_trail(vec![TrailStep::recognition(
-                "annotation",
-                Confidence::MAX,
-                TrailProvenance::Annotation(AnnotationProvenance {
-                    name: self.name.clone(),
-                }),
-                "synthesised from asserted inclusion annotation",
-            )])
-            .with_confidence(Confidence::MAX)
-            .with_location(target.clone())
-            .build()
-            .expect("required fields provided");
-        Some(entity)
-    }
-}
-
-/// A document-wide classification label.
-///
-/// Labels are modality-agnostic: a `"medical"` or `"gdpr-request"`
-/// tag applies to the source document and propagates to every
-/// per-modality [`Document<M>`] spawned from it. Policy rules can
-/// gate on labels via conditions.
-///
-/// [`Document<M>`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/document/struct.Document.html
-#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct LabelAnnotation {
-    /// Optional human-readable name (provenance / source identifier).
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub name: Option<String>,
-    /// Label value (e.g. `"medical"`, `"gdpr-request"`).
-    pub label: String,
-}
-
-impl LabelAnnotation {
-    /// Build a [`LabelAnnotation`] from a label string.
-    pub fn new(label: impl Into<String>) -> Self {
-        Self {
-            name: None,
-            label: label.into(),
-        }
-    }
-}
-
-/// Check whether any [`Exclusion`] overlaps the given `entity`.
-/// Exclusions are always assertions — every exclusion participates
-/// in the post-detection filter.
-///
-/// [`Exclusion`]: AnnotationKind::Exclusion
-pub fn is_excluded<M>(annotations: &[Annotation<M>], entity: &Entity<M>) -> bool
-where
-    M: Modality,
-    M::Location: Overlap,
-{
-    annotations.iter().any(|ann| {
-        let AnnotationKind::Exclusion { target } = &ann.kind else {
-            return false;
-        };
-        entity.location.overlaps(target)
-    })
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use crate::entity::builtins;
-    use crate::modality::{Image, ImageLocation, Text, TextLocation};
-    use crate::primitive::BoundingBox;
-
-    fn inclusion(start: usize, end: usize, strength: AnnotationStrength) -> Annotation<Text> {
-        Annotation {
-            name: None,
-            kind: AnnotationKind::Inclusion {
-                label: Some(EntityLabelRef::from(builtins::PERSON_NAME.name.clone())),
-                target: TextLocation::new(start, end),
-                strength,
-            },
-        }
-    }
-
-    fn exclusion(start: usize, end: usize) -> Annotation<Text> {
-        Annotation {
-            name: None,
-            kind: AnnotationKind::Exclusion {
-                target: TextLocation::new(start, end),
-            },
-        }
-    }
-
-    fn hint() -> AnnotationStrength {
-        AnnotationStrength::Hint { confidence: None }
-    }
-
-    fn test_entity(start: usize, end: usize) -> Entity<Text> {
-        Entity::test_builder(start, end).test_build()
-    }
-
-    #[test]
-    fn assert_inclusion_produces_entity_at_location() {
-        let ann = inclusion(0, 10, AnnotationStrength::Assert);
-        let entity = ann.to_inclusion_entity().unwrap();
-        assert_eq!(entity.location, TextLocation::new(0, 10));
-        assert!((entity.confidence.get() - 1.0).abs() < f64::EPSILON);
-    }
-
-    #[test]
-    fn hint_inclusion_does_not_produce_entity() {
-        let ann = inclusion(0, 10, hint());
-        assert!(ann.to_inclusion_entity().is_none());
-    }
-
-    #[test]
-    fn unclassified_assert_inclusion_falls_back_to_unresolved() {
-        let ann: Annotation<Text> = Annotation {
-            name: None,
-            kind: AnnotationKind::Inclusion {
-                label: None,
-                target: TextLocation::new(0, 10),
-                strength: AnnotationStrength::Assert,
-            },
-        };
-        let entity = ann.to_inclusion_entity().unwrap();
-        assert_eq!(entity.label.as_str(), builtins::UNRESOLVED.name.as_str());
-    }
-
-    #[test]
-    fn to_inclusion_entity_returns_none_for_exclusion() {
-        let ann = exclusion(0, 4);
-        assert!(ann.to_inclusion_entity().is_none());
-    }
-
-    #[test]
-    fn assert_inclusion_image() {
-        let bbox = BoundingBox::new(0.0, 0.0, 10.0, 10.0);
-        let ann: Annotation<Image> = Annotation {
-            name: Some("face".into()),
-            kind: AnnotationKind::Inclusion {
-                label: Some(EntityLabelRef::from(builtins::FACE.name.clone())),
-                target: ImageLocation::new(bbox),
-                strength: AnnotationStrength::Assert,
-            },
-        };
-        let entity = ann.to_inclusion_entity().unwrap();
-        assert_eq!(entity.location.bounding_box, bbox);
-    }
-
-    #[test]
-    fn exclusion_by_location_overlap() {
-        let anns = vec![exclusion(5, 15)];
-        let entity = test_entity(10, 20);
-        assert!(is_excluded(&anns, &entity));
-    }
-
-    #[test]
-    fn exclusion_by_location_no_overlap() {
-        let anns = vec![exclusion(0, 5)];
-        let entity = test_entity(10, 20);
-        assert!(!is_excluded(&anns, &entity));
-    }
-}
diff --git a/crates/nvisy-core/src/entity/label/builtins.rs b/crates/nvisy-core/src/entity/label/builtins.rs
deleted file mode 100644
index 3d217fdea..000000000
--- a/crates/nvisy-core/src/entity/label/builtins.rs
+++ /dev/null
@@ -1,180 +0,0 @@
-//! Built-in [`EntityLabel`] constants.
-//!
-//! Each constant carries a category tag (`personal_identity`,
-//! `financial`, …) plus cross-cutting tags where applicable
-//! (`pii`, `phi`, `pci`). Selectors can match by label name *or*
-//! by tag without the workspace modelling categories as a
-//! separate enum.
-//!
-//! The `BUILT_INS` slice indexes every constant for the
-//! [`EntityLabelCatalog::with_builtins`] constructor; the
-//! constants themselves are public and reachable by name (e.g.
-//! `builtins::PERSON_NAME`).
-//!
-//! [`EntityLabelCatalog::with_builtins`]: super::EntityLabelCatalog::with_builtins
-
-use std::sync::LazyLock;
-
-use super::entity_label::EntityLabel;
-
-macro_rules! label {
-    ($vis:vis $ident:ident, $name:literal, $desc:literal, [ $($tag:literal),* $(,)? ]) => {
-        $vis static $ident: LazyLock<EntityLabel> = LazyLock::new(|| {
-            EntityLabel::from_static($name, Some($desc), &[$($tag),*])
-        });
-    };
-}
-
-label!(pub PERSON_NAME, "person_name","Person name (full, first, or last).", ["personal_identity", "pii"]);
-label!(pub DATE_OF_BIRTH, "date_of_birth","Date of birth.", ["personal_identity", "pii"]);
-label!(pub GOVERNMENT_ID, "government_id","Government-issued identification number (SSN, SIN, Aadhaar, national ID, etc.).", ["personal_identity", "pii"]);
-label!(pub TAX_ID, "tax_id","Tax identification number (ITIN, EIN, TIN, etc.).", ["personal_identity", "pii"]);
-label!(pub DRIVERS_LICENSE, "drivers_license","Driver's license number.", ["personal_identity", "pii"]);
-label!(pub PASSPORT_NUMBER, "passport_number","Passport number.", ["personal_identity", "pii"]);
-label!(pub NATIONAL_INSURANCE_NUMBER, "national_insurance_number","National insurance or social-security equivalent (NI, BSN, AHVN, etc.).", ["personal_identity", "pii"]);
-label!(pub VEHICLE_ID, "vehicle_id","Vehicle identification number (VIN).", ["personal_identity"]);
-label!(pub LICENSE_PLATE, "license_plate","License plate number.", ["personal_identity"]);
-label!(pub EMAIL_ADDRESS, "email_address","Email address.", ["contact_info", "pii"]);
-label!(pub PHONE_NUMBER, "phone_number","Phone number.", ["contact_info", "pii"]);
-label!(pub ADDRESS, "address","Physical or mailing address.", ["contact_info", "pii"]);
-label!(pub POSTAL_CODE, "postal_code","Postal or ZIP code.", ["contact_info"]);
-label!(pub URL, "url","URL or hyperlink.", ["contact_info"]);
-label!(pub AGE, "age","Age value.", ["demographic", "pii"]);
-label!(pub GENDER, "gender","Gender identity.", ["demographic", "pii"]);
-label!(pub ETHNICITY, "ethnicity","Racial or ethnic background.", ["demographic", "pii"]);
-label!(pub RELIGION, "religion","Religious affiliation.", ["demographic", "pii"]);
-label!(pub NATIONALITY, "nationality","Nationality.", ["demographic", "pii"]);
-label!(pub CITIZENSHIP, "citizenship","Citizenship status.", ["demographic", "pii"]);
-label!(pub LANGUAGE, "language","Language or dialect spoken.", ["demographic"]);
-label!(pub PAYMENT_CARD, "payment_card","Payment card number (credit or debit).", ["financial", "pci", "pii"]);
-label!(pub CARD_SECURITY_CODE, "card_security_code","Payment card security code (CVV/CVC).", ["financial", "pci"]);
-label!(pub CARD_EXPIRY, "card_expiry","Payment card expiration date.", ["financial", "pci"]);
-label!(pub BANK_ACCOUNT, "bank_account","Bank account number.", ["financial", "pii"]);
-label!(pub BANK_ROUTING, "bank_routing","Bank routing or transit number.", ["financial"]);
-label!(pub IBAN, "iban","International Bank Account Number (IBAN).", ["financial", "pii"]);
-label!(pub SWIFT_CODE, "swift_code","SWIFT/BIC code.", ["financial"]);
-label!(pub CRYPTO_ADDRESS, "crypto_address","Cryptocurrency wallet address.", ["financial", "pii"]);
-label!(pub CURRENCY, "currency","Currency code or symbol.", ["financial"]);
-label!(pub AMOUNT, "amount","Monetary amount.", ["financial"]);
-label!(pub MEDICAL_ID, "medical_id","Medical record number.", ["health", "phi", "pii"]);
-label!(pub INSURANCE_ID, "insurance_id","Health insurance identifier.", ["health", "phi", "pii"]);
-label!(pub PRESCRIPTION_ID, "prescription_id","Prescription identifier or medication regimen.", ["health", "phi"]);
-label!(pub DIAGNOSIS, "diagnosis","Medical diagnosis or condition.", ["health", "phi"]);
-label!(pub MEDICATION, "medication","Medication name.", ["health", "phi"]);
-label!(pub FINGERPRINT, "fingerprint","Fingerprint biometric data.", ["biometric", "pii"]);
-label!(pub VOICEPRINT, "voiceprint","Voiceprint biometric data.", ["biometric", "pii"]);
-label!(pub RETINA_SCAN, "retina_scan","Retina scan biometric data.", ["biometric", "pii"]);
-label!(pub FACIAL_GEOMETRY, "facial_geometry","Facial geometry biometric data.", ["biometric", "pii"]);
-label!(pub PASSWORD, "password","Password.", ["credentials", "secret"]);
-label!(pub API_KEY, "api_key","API key.", ["credentials", "secret"]);
-label!(pub AUTH_TOKEN, "auth_token","Authentication token (OAuth, JWT, session token).", ["credentials", "secret"]);
-label!(pub PRIVATE_KEY, "private_key","Private cryptographic key.", ["credentials", "secret"]);
-label!(pub IP_ADDRESS, "ip_address","IP address (v4 or v6).", ["network_identifier", "pii"]);
-label!(pub MAC_ADDRESS, "mac_address","MAC address.", ["network_identifier", "pii"]);
-label!(pub DEVICE_ID, "device_id","Device identifier (IMEI, UDID, etc.).", ["network_identifier", "pii"]);
-label!(pub USERNAME, "username","Username or handle.", ["network_identifier", "pii"]);
-label!(pub COORDINATES, "coordinates","GPS coordinates (latitude/longitude).", ["location", "pii"]);
-label!(pub GEOLOCATION_METADATA, "geolocation_metadata","Geolocation metadata.", ["location", "pii"]);
-label!(pub FACE, "face","Human face detected in an image or video frame.", ["visual", "pii"]);
-label!(pub HANDWRITING, "handwriting","Handwritten text.", ["visual"]);
-label!(pub SIGNATURE, "signature","Handwritten signature.", ["visual", "pii"]);
-label!(pub LOGO, "logo","Brand or organisation logo.", ["visual"]);
-label!(pub BARCODE, "barcode","Barcode or QR code.", ["visual"]);
-label!(pub ORGANIZATION_NAME, "organization_name","Organization or company name.", ["organization"]);
-label!(pub COMPANY_ID, "company_id","Public company-registry identifier (Handelsregisternummer, Companies House number, etc.).", ["organization"]);
-label!(pub DEPARTMENT_NAME, "department_name","Department or business-unit name.", ["organization"]);
-label!(pub FACILITY_NAME, "facility_name","Physical facility or location name.", ["organization"]);
-label!(pub CASE_NUMBER, "case_number","Case, matter, or docket number.", ["organization"]);
-label!(pub INTERNAL_ID, "internal_id","Operator-defined internal identifier.", ["organization"]);
-label!(pub DATE_TIME, "date_time","Date or time value.", ["temporal"]);
-label!(pub EVENT, "event","Named event reference.", ["temporal"]);
-label!(pub OCCUPATION, "occupation","Occupation or job title.", ["organization"]);
-label!(pub PRODUCT, "product","Product name.", ["organization"]);
-label!(pub QUANTITY, "quantity","Numerical quantity.", ["quantity"]);
-label!(pub UNRESOLVED, "unresolved","Entity kind not yet identified.", ["unresolved"]);
-
-/// Every built-in label constant, indexed for catalog construction.
-pub(super) static BUILT_INS: &[&LazyLock<EntityLabel>] = &[
-    &PERSON_NAME,
-    &DATE_OF_BIRTH,
-    &GOVERNMENT_ID,
-    &TAX_ID,
-    &DRIVERS_LICENSE,
-    &PASSPORT_NUMBER,
-    &NATIONAL_INSURANCE_NUMBER,
-    &VEHICLE_ID,
-    &LICENSE_PLATE,
-    &EMAIL_ADDRESS,
-    &PHONE_NUMBER,
-    &ADDRESS,
-    &POSTAL_CODE,
-    &URL,
-    &AGE,
-    &GENDER,
-    &ETHNICITY,
-    &RELIGION,
-    &NATIONALITY,
-    &CITIZENSHIP,
-    &LANGUAGE,
-    &PAYMENT_CARD,
-    &CARD_SECURITY_CODE,
-    &CARD_EXPIRY,
-    &BANK_ACCOUNT,
-    &BANK_ROUTING,
-    &IBAN,
-    &SWIFT_CODE,
-    &CRYPTO_ADDRESS,
-    &CURRENCY,
-    &AMOUNT,
-    &MEDICAL_ID,
-    &INSURANCE_ID,
-    &PRESCRIPTION_ID,
-    &DIAGNOSIS,
-    &MEDICATION,
-    &FINGERPRINT,
-    &VOICEPRINT,
-    &RETINA_SCAN,
-    &FACIAL_GEOMETRY,
-    &PASSWORD,
-    &API_KEY,
-    &AUTH_TOKEN,
-    &PRIVATE_KEY,
-    &IP_ADDRESS,
-    &MAC_ADDRESS,
-    &DEVICE_ID,
-    &USERNAME,
-    &COORDINATES,
-    &GEOLOCATION_METADATA,
-    &FACE,
-    &HANDWRITING,
-    &SIGNATURE,
-    &LOGO,
-    &BARCODE,
-    &ORGANIZATION_NAME,
-    &COMPANY_ID,
-    &DEPARTMENT_NAME,
-    &FACILITY_NAME,
-    &CASE_NUMBER,
-    &INTERNAL_ID,
-    &DATE_TIME,
-    &EVENT,
-    &OCCUPATION,
-    &PRODUCT,
-    &QUANTITY,
-    &UNRESOLVED,
-];
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn well_known_built_ins_have_expected_tags() {
-        assert_eq!(PAYMENT_CARD.name, "payment_card");
-        assert!(PAYMENT_CARD.has_tag("financial"));
-        assert!(PAYMENT_CARD.has_tag("pci"));
-        assert!(PAYMENT_CARD.has_tag("pii"));
-        assert_eq!(PERSON_NAME.name, "person_name");
-        assert!(PERSON_NAME.has_tag("personal_identity"));
-    }
-}
diff --git a/crates/nvisy-core/src/entity/label/catalog.rs b/crates/nvisy-core/src/entity/label/catalog.rs
deleted file mode 100644
index 89f02d124..000000000
--- a/crates/nvisy-core/src/entity/label/catalog.rs
+++ /dev/null
@@ -1,115 +0,0 @@
-//! [`EntityLabelCatalog`] — name-indexed lookup over a set of
-//! [`EntityLabel`]s.
-//!
-//! Constructed at runtime configuration time. Recognizers'
-//! supported labels and selectors' tag-matching path both walk a
-//! `EntityLabelCatalog`. The workspace ships a built-in catalog through
-//! [`EntityLabelCatalog::with_builtins`]; deployments can register their own
-//! labels alongside or instead of the built-ins via
-//! [`EntityLabelCatalog::with_label`] / [`EntityLabelCatalog::with_labels`].
-
-use std::collections::HashMap;
-use std::sync::LazyLock;
-
-use hipstr::HipStr;
-
-use super::builtins::BUILT_INS;
-use super::entity_label::EntityLabel;
-
-/// Name-indexed catalog of [`EntityLabel`]s.
-///
-/// Built from a list of labels (mixing workspace-shipped built-ins
-/// with deployment-defined custom labels). Construction copies each
-/// label into a [`HashMap`] keyed by `HipStr` clone of the label's
-/// name; subsequent lookups are O(1).
-#[derive(Debug, Clone, Default)]
-pub struct EntityLabelCatalog {
-    by_name: HashMap<HipStr<'static>, EntityLabel>,
-}
-
-impl EntityLabelCatalog {
-    /// Empty catalog. Built-ins must be registered explicitly via
-    /// [`Self::with_label`] / [`Self::with_labels`]; use
-    /// [`Self::with_builtins`] for the workspace-shipped set.
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// EntityLabelCatalog pre-populated with every workspace-shipped built-in
-    /// label.
-    pub fn with_builtins() -> Self {
-        let mut cat = Self::new();
-        for lazy in BUILT_INS {
-            cat.insert(LazyLock::force(lazy).clone());
-        }
-        cat
-    }
-
-    /// Register a single label. Replaces any prior entry sharing
-    /// the same [`EntityLabel::name`].
-    pub fn insert(&mut self, label: EntityLabel) {
-        self.by_name.insert(label.name.clone(), label);
-    }
-
-    /// Builder-style sibling of [`Self::insert`] returning `Self`.
-    #[must_use]
-    pub fn with_label(mut self, label: EntityLabel) -> Self {
-        self.insert(label);
-        self
-    }
-
-    /// Bulk-register a sequence of labels.
-    #[must_use]
-    pub fn with_labels<I>(mut self, labels: I) -> Self
-    where
-        I: IntoIterator<Item = EntityLabel>,
-    {
-        for l in labels {
-            self.insert(l);
-        }
-        self
-    }
-
-    /// Look up a label by name. Returns `None` for names not
-    /// registered in this catalog.
-    pub fn lookup(&self, name: &str) -> Option<&EntityLabel> {
-        self.by_name.get(name)
-    }
-
-    /// Iterator over every registered label, in no particular order.
-    pub fn iter(&self) -> impl Iterator<Item = &EntityLabel> + '_ {
-        self.by_name.values()
-    }
-
-    /// Number of labels in the catalog.
-    pub fn len(&self) -> usize {
-        self.by_name.len()
-    }
-
-    /// `true` when the catalog is empty.
-    pub fn is_empty(&self) -> bool {
-        self.by_name.is_empty()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn builtin_catalog_resolves_known_names() {
-        let cat = EntityLabelCatalog::with_builtins();
-        let l = cat.lookup("payment_card").expect("built-in");
-        assert!(l.has_tag("financial"));
-        assert!(cat.lookup("acme_internal_id").is_none());
-    }
-
-    #[test]
-    fn catalog_accepts_custom_labels_alongside_builtins() {
-        let custom = EntityLabel::new("acme_internal_id").with_tags(["custom"]);
-        let cat = EntityLabelCatalog::with_builtins().with_label(custom);
-        assert!(cat.lookup("payment_card").is_some());
-        let acme = cat.lookup("acme_internal_id").expect("custom registered");
-        assert!(acme.has_tag("custom"));
-    }
-}
diff --git a/crates/nvisy-core/src/entity/label/entity_label.rs b/crates/nvisy-core/src/entity/label/entity_label.rs
deleted file mode 100644
index f282a0fa9..000000000
--- a/crates/nvisy-core/src/entity/label/entity_label.rs
+++ /dev/null
@@ -1,178 +0,0 @@
-//! [`EntityLabel`] — open vocabulary tag for detected entities.
-//!
-//! Any recognizer can mint a label by name, ship it through the
-//! pipeline, and have the audit reference it verbatim. The workspace
-//! ships a catalog of built-in labels in [`super::builtins`];
-//! recognizers and policy authors are free to invent new ones
-//! (`acme-internal-id`, `medical-record-no`) without touching
-//! workspace code.
-//!
-//! ## Identity
-//!
-//! Labels are identified by [`name`]; two labels with the same name
-//! are considered the same entity kind regardless of differences in
-//! [`description`] or [`tags`]. Selectors match by name.
-//!
-//! ## Tags
-//!
-//! [`tags`] is a free-form list of short identifiers that policy
-//! selectors can match against. Built-in labels carry category
-//! tags (`personal_identity`, `contact_info`, `financial`, etc.)
-//! plus cross-cutting tags (`pii`, `phi`, `pci`). Custom labels
-//! can ship with zero tags; selectors targeting tags only match
-//! labels that carry them.
-//!
-//! [`name`]: EntityLabel::name
-//! [`description`]: EntityLabel::description
-//! [`tags`]: EntityLabel::tags
-
-use std::fmt;
-
-use hipstr::HipStr;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::EntityLabelRef;
-
-/// Open-vocabulary entity label: identity, optional description,
-/// and zero or more tags.
-#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct EntityLabel {
-    /// Canonical name of the label (e.g. `"person_name"`,
-    /// `"acme_internal_id"`). Selectors match by this value.
-    #[schemars(with = "String")]
-    pub name: HipStr<'static>,
-    /// Optional human-readable description of what the label
-    /// represents. Surfaced in audits and policy author tooling.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub description: Option<HipStr<'static>>,
-    /// Free-form tags grouping this label with related ones.
-    /// Built-in labels carry category tags
-    /// (`personal_identity`, `financial`, …) plus cross-cutting
-    /// tags where applicable (`pii`, `phi`, `pci`). Empty for
-    /// untagged custom labels.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    #[schemars(with = "Vec<String>")]
-    pub tags: Vec<HipStr<'static>>,
-}
-
-impl EntityLabel {
-    /// Construct a label from a runtime name. Description and tags
-    /// default to empty; use [`Self::with_description`] and
-    /// [`Self::with_tags`] to add them.
-    pub fn new(name: impl Into<HipStr<'static>>) -> Self {
-        Self {
-            name: name.into(),
-            description: None,
-            tags: Vec::new(),
-        }
-    }
-
-    /// Construct a label entirely from `&'static str` literals.
-    /// Used by the built-in catalog in [`super::builtins`] so the
-    /// strings live in static storage and runtime construction is
-    /// just one `Vec::from` per built-in.
-    pub fn from_static(
-        name: &'static str,
-        description: Option<&'static str>,
-        tags: &'static [&'static str],
-    ) -> Self {
-        Self {
-            name: HipStr::from_static(name),
-            description: description.map(HipStr::from_static),
-            tags: tags.iter().copied().map(HipStr::from_static).collect(),
-        }
-    }
-
-    /// Attach a description.
-    #[must_use]
-    pub fn with_description(mut self, description: impl Into<HipStr<'static>>) -> Self {
-        self.description = Some(description.into());
-        self
-    }
-
-    /// Attach tags. Replaces any previously set tags.
-    #[must_use]
-    pub fn with_tags<I, S>(mut self, tags: I) -> Self
-    where
-        I: IntoIterator<Item = S>,
-        S: Into<HipStr<'static>>,
-    {
-        self.tags = tags.into_iter().map(Into::into).collect();
-        self
-    }
-
-    /// Returns `true` when this label carries `tag` in its tag
-    /// list. Comparison is byte-for-byte.
-    #[must_use]
-    pub fn has_tag(&self, tag: &str) -> bool {
-        self.tags.iter().any(|t| t == tag)
-    }
-
-    /// Construct a name-only [`EntityLabelRef`] handle to this
-    /// label. Clones the underlying [`HipStr`] (a refcount bump
-    /// for `from_static` labels — no allocation).
-    ///
-    /// [`EntityLabelRef`]: super::EntityLabelRef
-    #[must_use]
-    pub fn label_ref(&self) -> EntityLabelRef {
-        EntityLabelRef::from(self.name.clone())
-    }
-}
-
-impl AsRef<str> for EntityLabel {
-    fn as_ref(&self) -> &str {
-        &self.name
-    }
-}
-
-impl fmt::Display for EntityLabel {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        fmt::Display::fmt(&self.name, f)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn from_static_round_trips() {
-        let l = EntityLabel::from_static(
-            "email_address",
-            Some("Email address."),
-            &["contact_info", "pii"],
-        );
-        assert_eq!(l.name, "email_address");
-        assert_eq!(l.description.as_deref(), Some("Email address."));
-        assert!(l.has_tag("contact_info"));
-        assert!(l.has_tag("pii"));
-        assert!(!l.has_tag("financial"));
-    }
-
-    #[test]
-    fn builder_setters_chain() {
-        let l = EntityLabel::new("acme_internal_id")
-            .with_description("ACME corp internal record id")
-            .with_tags(["custom", "acme"]);
-        assert_eq!(l.name, "acme_internal_id");
-        assert_eq!(
-            l.description.as_deref(),
-            Some("ACME corp internal record id"),
-        );
-        assert!(l.has_tag("acme"));
-    }
-
-    #[test]
-    fn equality_ignores_metadata() {
-        // NOTE: deliberately *not* the behaviour today — `derive(PartialEq)`
-        // makes equality structural. If selectors need name-only equality
-        // they should compare `.name` explicitly. This test documents the
-        // current contract so a future change is intentional.
-        let a = EntityLabel::new("person_name").with_tags(["pii"]);
-        let b = EntityLabel::new("person_name");
-        assert_ne!(a, b);
-    }
-}
diff --git a/crates/nvisy-core/src/entity/label/mod.rs b/crates/nvisy-core/src/entity/label/mod.rs
deleted file mode 100644
index 8fc9f7439..000000000
--- a/crates/nvisy-core/src/entity/label/mod.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! Entity label types and catalog.
-//!
-//! Four concerns split across this folder:
-//!
-//! - [`EntityLabel`] — full catalog entry (name + description +
-//!   tags). Authored once per label; consumed by selectors and
-//!   audit-rendering tooling that need the metadata.
-//! - [`EntityLabelRef`] — name-only handle stored on every
-//!   detected [`Entity`]. Cheap-clone wrapper around
-//!   [`HipStr<'static>`].
-//! - [`EntityLabelCatalog`] — name-indexed lookup over a
-//!   collection of `EntityLabel`s. The workspace ships a built-in
-//!   catalog constructed from [`EntityLabelCatalog::with_builtins`];
-//!   consumers can register custom labels alongside or instead of
-//!   the built-ins.
-//! - [`builtins`] — every built-in `EntityLabel` constant
-//!   (`builtins::PERSON_NAME`, `builtins::EMAIL_ADDRESS`, …) plus
-//!   the internal `BUILT_INS` slice the catalog walks at
-//!   construction time.
-//!
-//! [`Entity`]: crate::entity::Entity
-//! [`HipStr<'static>`]: hipstr::HipStr
-
-pub mod builtins;
-mod catalog;
-mod entity_label;
-mod reference;
-
-pub use self::catalog::EntityLabelCatalog;
-pub use self::entity_label::EntityLabel;
-pub use self::reference::EntityLabelRef;
diff --git a/crates/nvisy-core/src/entity/label/reference.rs b/crates/nvisy-core/src/entity/label/reference.rs
deleted file mode 100644
index 43cfea6db..000000000
--- a/crates/nvisy-core/src/entity/label/reference.rs
+++ /dev/null
@@ -1,148 +0,0 @@
-//! [`EntityLabelRef`] — name-only handle to an [`EntityLabel`].
-//!
-//! Per-entity hot paths (`Entity::label`, audit refs, selector
-//! matching) carry only the label's identifying name, not its full
-//! catalog metadata. [`EntityLabelRef`] wraps a [`HipStr<'static>`]
-//! so the surface is a single newtype rather than a bare string,
-//! giving us a typed receiver for ergonomics like
-//! `entity.label.matches("payment_card")`.
-//!
-//! Catalog-side metadata (description, tags) lives on
-//! [`EntityLabel`] and is dereferenced through
-//! [`EntityLabelCatalog::lookup`] when a consumer needs it.
-//!
-//! [`EntityLabel`]: super::EntityLabel
-//! [`EntityLabelCatalog::lookup`]: super::EntityLabelCatalog::lookup
-
-use std::fmt;
-
-use hipstr::HipStr;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Name-only handle to an [`EntityLabel`]. Cheap-clone wrapper
-/// around [`HipStr<'static>`].
-///
-/// Carried on every [`Entity`] in place of the full catalog
-/// metadata. Two refs are equal when their names are equal
-/// byte-for-byte.
-///
-/// [`EntityLabel`]: super::EntityLabel
-/// [`Entity`]: crate::entity::Entity
-#[derive(Debug, Clone, PartialEq, Eq, Hash, Serialize, Deserialize, JsonSchema)]
-#[serde(transparent)]
-#[schemars(with = "String")]
-pub struct EntityLabelRef(HipStr<'static>);
-
-impl EntityLabelRef {
-    /// Wrap a name.
-    pub fn new(name: impl Into<HipStr<'static>>) -> Self {
-        Self(name.into())
-    }
-
-    /// Wrap a `&'static str` without allocating.
-    #[must_use]
-    pub const fn from_static(name: &'static str) -> Self {
-        Self(HipStr::from_static(name))
-    }
-
-    /// Borrow the underlying name.
-    #[must_use]
-    pub fn as_str(&self) -> &str {
-        self.0.as_str()
-    }
-
-    /// Borrow the inner [`HipStr`].
-    #[must_use]
-    pub fn as_hipstr(&self) -> &HipStr<'static> {
-        &self.0
-    }
-
-    /// Consume the ref and return the inner [`HipStr`].
-    #[must_use]
-    pub fn into_hipstr(self) -> HipStr<'static> {
-        self.0
-    }
-
-    /// `true` when this ref names `name` byte-for-byte.
-    #[must_use]
-    pub fn matches(&self, name: &str) -> bool {
-        self.0 == name
-    }
-}
-
-impl AsRef<str> for EntityLabelRef {
-    fn as_ref(&self) -> &str {
-        self.as_str()
-    }
-}
-
-impl fmt::Display for EntityLabelRef {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        fmt::Display::fmt(&self.0, f)
-    }
-}
-
-impl From<HipStr<'static>> for EntityLabelRef {
-    fn from(value: HipStr<'static>) -> Self {
-        Self(value)
-    }
-}
-
-impl From<EntityLabelRef> for HipStr<'static> {
-    fn from(value: EntityLabelRef) -> Self {
-        value.0
-    }
-}
-
-impl From<&'static str> for EntityLabelRef {
-    fn from(value: &'static str) -> Self {
-        Self::from_static(value)
-    }
-}
-
-impl From<String> for EntityLabelRef {
-    fn from(value: String) -> Self {
-        Self(HipStr::from(value))
-    }
-}
-
-impl PartialEq<str> for EntityLabelRef {
-    fn eq(&self, other: &str) -> bool {
-        self.0 == other
-    }
-}
-
-impl PartialEq<&str> for EntityLabelRef {
-    fn eq(&self, other: &&str) -> bool {
-        self.0 == *other
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn from_static_no_alloc() {
-        let r = EntityLabelRef::from_static("payment_card");
-        assert_eq!(r.as_str(), "payment_card");
-        assert!(r.matches("payment_card"));
-        assert!(!r.matches("person_name"));
-    }
-
-    #[test]
-    fn equality_with_str() {
-        let r = EntityLabelRef::from_static("email_address");
-        assert_eq!(r, "email_address");
-    }
-
-    #[test]
-    fn serde_transparent() {
-        let r = EntityLabelRef::from_static("ssn");
-        let s = serde_json::to_string(&r).unwrap();
-        assert_eq!(s, "\"ssn\"");
-        let back: EntityLabelRef = serde_json::from_str(&s).unwrap();
-        assert_eq!(back, r);
-    }
-}
diff --git a/crates/nvisy-core/src/entity/method/mod.rs b/crates/nvisy-core/src/entity/method/mod.rs
deleted file mode 100644
index 615347341..000000000
--- a/crates/nvisy-core/src/entity/method/mod.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-//! Per-entity provenance: who detected, who refined, and the
-//! score-adjustment trail.
-//!
-//! Document-level "how was the content produced" (text-layer parse,
-//! OCR, transcription) lives separately on the modality-specific
-//! extraction tag carried on [`Document<M>::meta`]
-//! ([`TextExtraction`], [`ImageExtraction`], [`AudioExtraction`],
-//! [`TabularExtraction`]) — that axis is a property of the
-//! [`Document<M>`], not of the entities inside it.
-//!
-//! [`Document<M>`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/document/struct.Document.html
-//! [`Document<M>::meta`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/document/struct.Document.html#structfield.meta
-//! [`TextExtraction`]: crate::modality::TextExtraction
-//! [`ImageExtraction`]: crate::modality::ImageExtraction
-//! [`AudioExtraction`]: crate::modality::AudioExtraction
-//! [`TabularExtraction`]: crate::modality::TabularExtraction
-
-mod provenance;
-mod trail;
-
-pub use self::provenance::{AnnotationProvenance, ModelProvenance, PatternProvenance};
-pub use self::trail::{TrailProvenance, TrailStep, TrailStepKind};
diff --git a/crates/nvisy-core/src/entity/method/provenance.rs b/crates/nvisy-core/src/entity/method/provenance.rs
deleted file mode 100644
index 712f154dd..000000000
--- a/crates/nvisy-core/src/entity/method/provenance.rs
+++ /dev/null
@@ -1,124 +0,0 @@
-//! Provenance metadata attached to recognition methods.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Provenance for a pattern-based detection (regex, dictionary,
-/// deny-list). Each variant carries only the fields meaningful for
-/// that matcher.
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-#[non_exhaustive]
-pub enum PatternProvenance {
-    /// Regular expression matched against the full text.
-    Regex {
-        /// Name of the pattern that matched (e.g. "ssn", "email").
-        name: String,
-        /// The actual regex string that matched, when the pattern
-        /// engine surfaces it. Audit/compliance consumers store this
-        /// to prove which regex triggered a redaction; recognizers
-        /// that don't expose the pattern source leave it `None`.
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        regex: Option<String>,
-        /// Name of the validator that confirmed the match (e.g.
-        /// "luhn", "iban"), when one ran.
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        validator: Option<String>,
-        /// Whether contextual analysis (keyword co-occurrence)
-        /// adjusted the confidence score for this match.
-        #[serde(default, skip_serializing_if = "std::ops::Not::not")]
-        contextual: bool,
-    },
-    /// Exact-match lookup in a curated dictionary.
-    Dictionary {
-        /// Name of the dictionary pattern that matched.
-        name: String,
-        /// Whether contextual analysis (keyword co-occurrence)
-        /// adjusted the confidence score for this match.
-        #[serde(default, skip_serializing_if = "std::ops::Not::not")]
-        contextual: bool,
-    },
-    /// Caller-supplied deny-list value forced into results. Carries
-    /// no per-match identity — the matched value is the same one the
-    /// caller supplied.
-    DenyList,
-}
-
-impl PatternProvenance {
-    /// Mark this provenance as contextually adjusted. No-op for
-    /// `DenyList`, which doesn't track contextual adjustment.
-    pub fn mark_contextual(&mut self) {
-        match self {
-            Self::Regex { contextual, .. } | Self::Dictionary { contextual, .. } => {
-                *contextual = true;
-            }
-            Self::DenyList => {}
-        }
-    }
-
-    /// Pattern name that produced this match (regex or dictionary),
-    /// or `None` for deny-list matches.
-    pub fn name(&self) -> Option<&str> {
-        match self {
-            Self::Regex { name, .. } | Self::Dictionary { name, .. } => Some(name),
-            Self::DenyList => None,
-        }
-    }
-}
-
-/// Provenance for a model-based detection.
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct ModelProvenance {
-    /// Model name (e.g. `"spacy-en-core-web-lg"`, `"gpt-4"`).
-    pub name: String,
-    /// Model version string.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub version: Option<String>,
-    /// Whether contextual analysis (keyword co-occurrence) adjusted
-    /// the confidence score for this match. Mirrors the
-    /// `contextual` flag on [`PatternProvenance`] so post-recognition
-    /// enhancers can record their decision uniformly across
-    /// pattern- and model-based detections.
-    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
-    pub contextual: bool,
-}
-
-impl ModelProvenance {
-    /// Create a new model provenance with the given name.
-    pub fn new(name: impl Into<String>) -> Self {
-        Self {
-            name: name.into(),
-            version: None,
-            contextual: false,
-        }
-    }
-
-    /// Set the model version.
-    pub fn with_version(mut self, version: impl Into<String>) -> Self {
-        self.version = Some(version.into());
-        self
-    }
-
-    /// Mark this provenance as contextually adjusted. Set by the
-    /// `ContextEnhancer` when keyword co-occurrence boosted the
-    /// match's confidence.
-    pub fn mark_contextual(&mut self) {
-        self.contextual = true;
-    }
-}
-
-/// Provenance for an annotation (pre-identified region from upload).
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct AnnotationProvenance {
-    /// Name of the annotation as supplied by the uploader. Comes
-    /// from [`Annotation::name`] at conversion time; downstream
-    /// consumers use it as a human-readable label / source tag.
-    ///
-    /// [`Annotation::name`]: crate::entity::Annotation::name
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub name: Option<String>,
-}
diff --git a/crates/nvisy-core/src/entity/method/trail.rs b/crates/nvisy-core/src/entity/method/trail.rs
deleted file mode 100644
index 9243d78df..000000000
--- a/crates/nvisy-core/src/entity/method/trail.rs
+++ /dev/null
@@ -1,242 +0,0 @@
-//! Per-entity score trail.
-//!
-//! [`Entity<M>::trail`] carries the chronological list of
-//! [`TrailStep`]s explaining how the entity reached its final
-//! confidence: which recognizer fired, which post-detection steps
-//! adjusted the score, and the score before and after each step.
-//!
-//! This single trail replaces the prior parallel pair of
-//! `recognition_methods` + `refinement_methods` and the standalone
-//! "analysis explanation": every score-affecting event recorded
-//! on the entity lives here. Consumers asking "who recognized this"
-//! filter by [`TrailStepKind::Recognition`]; "what refinements ran"
-//! by [`TrailStepKind::Refinement`] / [`TrailStepKind::Verification`]
-//! / [`TrailStepKind::Fusion`] / [`TrailStepKind::Calibration`].
-//!
-//! `source` is the recognizer's self-reported identifier for
-//! recognizer-produced steps; well-known constants (`"dedup"`,
-//! `"calibration"`) cover post-recognition steps. `reason` is free
-//! text so custom recognizers can describe novel adjustments without
-//! extending an enum.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use strum::{Display, EnumString};
-
-use super::provenance::{AnnotationProvenance, ModelProvenance, PatternProvenance};
-use crate::primitive::Confidence;
-
-/// One score-adjustment step in an [`Entity`]'s trail.
-///
-/// [`Entity`]: crate::entity::Entity
-#[derive(Debug, Clone, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct TrailStep {
-    /// Identifier of the source that produced this step. For
-    /// recognizer-produced steps this is the recognizer's
-    /// self-reported name (`"pattern"`, `"ner"`, `"llm"`, `"vlm"`,
-    /// or a custom recognizer's name). For post-recognition steps
-    /// it is a well-known constant (`"dedup"`, `"calibration"`).
-    pub source: String,
-    /// Which category of step this is — recognition, refinement,
-    /// verification, fusion, or calibration.
-    pub kind: TrailStepKind,
-    /// Score before this step ran. `None` for the first step (no
-    /// prior score to record).
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub original: Option<Confidence>,
-    /// Score after this step ran.
-    pub adjusted: Confidence,
-    /// Typed provenance details for this step. [`TrailProvenance::None`]
-    /// when the step has no structured metadata to attach (e.g.
-    /// dedup fusion, calibration adjustment).
-    #[serde(default, skip_serializing_if = "TrailProvenance::is_none")]
-    pub provenance: TrailProvenance,
-    /// Free-text reason explaining what this step did and why.
-    pub reason: String,
-}
-
-impl TrailStep {
-    /// Construct a base [`Recognition`]
-    /// step — a recognizer firing for the first time. `original` is
-    /// `None`.
-    ///
-    /// [`Recognition`]: TrailStepKind::Recognition
-    pub fn recognition(
-        source: impl Into<String>,
-        adjusted: Confidence,
-        provenance: TrailProvenance,
-        reason: impl Into<String>,
-    ) -> Self {
-        Self {
-            source: source.into(),
-            kind: TrailStepKind::Recognition,
-            original: None,
-            adjusted,
-            provenance,
-            reason: reason.into(),
-        }
-    }
-
-    /// Construct a [`Refinement`] step —
-    /// a post-recognition score tweak (context boost or penalty,
-    /// validator pass, etc.).
-    ///
-    /// [`Refinement`]: TrailStepKind::Refinement
-    pub fn refinement(
-        source: impl Into<String>,
-        original: Confidence,
-        adjusted: Confidence,
-        reason: impl Into<String>,
-    ) -> Self {
-        Self {
-            source: source.into(),
-            kind: TrailStepKind::Refinement,
-            original: Some(original),
-            adjusted,
-            provenance: TrailProvenance::None,
-            reason: reason.into(),
-        }
-    }
-
-    /// Construct a [`Verification`]
-    /// step — an LLM/VLM verify pass confirmed or rejected the
-    /// detection.
-    ///
-    /// [`Verification`]: TrailStepKind::Verification
-    pub fn verification(
-        source: impl Into<String>,
-        original: Confidence,
-        adjusted: Confidence,
-        provenance: TrailProvenance,
-        reason: impl Into<String>,
-    ) -> Self {
-        Self {
-            source: source.into(),
-            kind: TrailStepKind::Verification,
-            original: Some(original),
-            adjusted,
-            provenance,
-            reason: reason.into(),
-        }
-    }
-
-    /// Construct a [`Fusion`] step — dedup
-    /// merged overlapping matches into this entity.
-    ///
-    /// [`Fusion`]: TrailStepKind::Fusion
-    pub fn fusion(original: Confidence, adjusted: Confidence, reason: impl Into<String>) -> Self {
-        Self {
-            source: "dedup".to_owned(),
-            kind: TrailStepKind::Fusion,
-            original: Some(original),
-            adjusted,
-            provenance: TrailProvenance::None,
-            reason: reason.into(),
-        }
-    }
-
-    /// Construct a [`Calibration`] step
-    /// — the per-method calibration multiplier was applied.
-    ///
-    /// [`Calibration`]: TrailStepKind::Calibration
-    pub fn calibration(
-        original: Confidence,
-        adjusted: Confidence,
-        reason: impl Into<String>,
-    ) -> Self {
-        Self {
-            source: "calibration".to_owned(),
-            kind: TrailStepKind::Calibration,
-            original: Some(original),
-            adjusted,
-            provenance: TrailProvenance::None,
-            reason: reason.into(),
-        }
-    }
-}
-
-/// Discriminant for [`TrailStep`].
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-#[derive(Display, EnumString, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-#[strum(serialize_all = "snake_case")]
-#[non_exhaustive]
-pub enum TrailStepKind {
-    /// Base step: a recognizer fired and produced this entity for
-    /// the first time.
-    Recognition,
-    /// Post-recognition tweak (validator pass, context-rule boost
-    /// or penalty, contextual adjustment).
-    Refinement,
-    /// LLM/VLM verify pass confirmed or rejected the detection.
-    Verification,
-    /// Deduplication merged overlapping matches into this entity.
-    Fusion,
-    /// Per-method calibration multiplier was applied.
-    Calibration,
-}
-
-/// Typed provenance metadata attached to a [`TrailStep`].
-///
-/// Carries the structured details for built-in step kinds (pattern
-/// name, model name, annotation name). Custom recognizers register
-/// under their own name and rely on the step's `reason` for
-/// human-readable detail, leaving this field as
-/// [`TrailProvenance::None`].
-#[derive(Debug, Clone, PartialEq, Default)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-#[non_exhaustive]
-pub enum TrailProvenance {
-    /// Pattern-based detection — regex, dictionary, or deny-list.
-    Pattern(PatternProvenance),
-    /// Model-based detection — NER, LLM, or VLM. The
-    /// [`ModelProvenance`] carries name + kind + optional version.
-    Model(ModelProvenance),
-    /// Annotation-derived detection — a hint or assertion supplied
-    /// alongside the uploaded file materialised into an entity.
-    Annotation(AnnotationProvenance),
-    /// No structured provenance for this step. Used by custom
-    /// recognizers and by post-recognition steps (dedup, calibration)
-    /// whose details fit in the `reason` text.
-    #[default]
-    None,
-}
-
-impl TrailProvenance {
-    /// True when this provenance carries no structured data.
-    #[must_use]
-    pub fn is_none(&self) -> bool {
-        matches!(self, Self::None)
-    }
-
-    /// The source recognizer / model / annotation name that
-    /// produced this step, when one is available. Used by the
-    /// post-recognition enhancer to look up the source's declared
-    /// context keywords in its registry.
-    ///
-    /// Returns `None` for `DenyList` (no per-match identity) and
-    /// for [`TrailProvenance::None`] (no structured provenance).
-    #[must_use]
-    pub fn name(&self) -> Option<&str> {
-        match self {
-            Self::Pattern(p) => p.name(),
-            Self::Model(m) => Some(m.name.as_str()),
-            Self::Annotation(a) => a.name.as_deref(),
-            Self::None => None,
-        }
-    }
-
-    /// Mark this provenance as contextually adjusted. Dispatches to
-    /// the per-variant flag; no-op for variants that don't track
-    /// contextual adjustment (`Annotation`, `None`).
-    pub fn mark_contextual(&mut self) {
-        match self {
-            Self::Pattern(p) => p.mark_contextual(),
-            Self::Model(m) => m.mark_contextual(),
-            Self::Annotation(_) | Self::None => {}
-        }
-    }
-}
diff --git a/crates/nvisy-core/src/entity/mod.rs b/crates/nvisy-core/src/entity/mod.rs
deleted file mode 100644
index efef6d5f4..000000000
--- a/crates/nvisy-core/src/entity/mod.rs
+++ /dev/null
@@ -1,161 +0,0 @@
-//! Sensitive-data entity types and detection metadata.
-//!
-//! [`Entity<M>`] represents a single occurrence of sensitive data
-//! detected within a document. Detected entities are accumulated on
-//! the run-scoped [`Audit<M>`] as [`EntityRecord<M>`]s — each record
-//! bundles the entity with the optional [`AuditEntry<M>`] produced
-//! for it during redaction.
-//!
-//! Each entity carries a chronological [`TrailStep`] list explaining
-//! how it reached its final confidence: the base recognizer firing,
-//! any refinement / verification / fusion / calibration steps, and
-//! the score before and after each. This single trail replaces the
-//! prior parallel pair of recognition + refinement method lists.
-//!
-//! [`Audit<M>`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/provenance/struct.Audit.html
-//! [`EntityRecord<M>`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/provenance/struct.EntityRecord.html
-//! [`AuditEntry<M>`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/provenance/struct.AuditEntry.html
-
-mod annotation;
-pub mod label;
-mod method;
-mod source;
-
-use derive_builder::Builder;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-pub use self::annotation::{
-    Annotation, AnnotationKind, AnnotationStrength, LabelAnnotation, is_excluded,
-};
-pub use self::label::{EntityLabel, EntityLabelCatalog, EntityLabelRef, builtins};
-pub use self::method::{
-    AnnotationProvenance, ModelProvenance, PatternProvenance, TrailProvenance, TrailStep,
-    TrailStepKind,
-};
-pub use self::source::ContentSource;
-use crate::modality::Modality;
-#[cfg(any(test, feature = "test-utils"))]
-use crate::modality::{Text, TextLocation};
-use crate::primitive::{Confidence, LanguageTag};
-
-/// A detected sensitive data occurrence within a document.
-///
-/// The trail of score-affecting steps lives on [`trail`].
-///
-/// [`trail`]: Self::trail
-#[derive(Debug, Clone, PartialEq, Builder)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[builder(
-    name = "EntityBuilder",
-    pattern = "owned",
-    setter(into, strip_option, prefix = "with")
-)]
-#[serde(rename_all = "camelCase")]
-pub struct Entity<M: Modality> {
-    /// Unique identifier for this entity (UUIDv7).
-    #[builder(default = "Uuid::now_v7()")]
-    pub id: Uuid,
-    /// NER/LLM-assigned coreference identifier, stable across
-    /// coreferent mentions within one detection call. Two entities
-    /// sharing `entity_id` refer to the same real-world entity.
-    #[builder(default, setter(into = false))]
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub entity_id: Option<String>,
-    /// Open-vocabulary classification of the entity. Wraps the
-    /// label's identifying name; full catalog metadata
-    /// (description, tags) is dereferenced through an
-    /// [`EntityLabelCatalog`].
-    pub label: EntityLabelRef,
-    /// Modality-specific location of the entity within the document.
-    pub location: M::Location,
-    /// Detection confidence score in the range `[0.0, 1.0]`. Equals
-    /// the `adjusted` score on the final step in `trail`.
-    pub confidence: Confidence,
-    /// Chronological trail of score-affecting steps: the base
-    /// recognition, any refinement / verification / fusion /
-    /// calibration adjustments, each tagged with `original` /
-    /// `adjusted` score and a free-text reason.
-    #[builder(default)]
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub trail: Vec<TrailStep>,
-    /// BCP-47 language tag for the span the entity covers, when the
-    /// recognizer has one to assign. Locale-aware anonymizers (the
-    /// fake-data generator in `nvisy-fake`, format-preserving
-    /// templates) read this to pick a locale. `None` falls back to
-    /// the document-level language metadata.
-    #[builder(default)]
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub language: Option<LanguageTag>,
-}
-
-impl<M: Modality> Entity<M> {
-    /// Create a new [`EntityBuilder`].
-    pub fn builder() -> EntityBuilder<M> {
-        EntityBuilder::default()
-    }
-
-    /// Original recognition score, before any post-recognition
-    /// adjustments. Reads from the first step's `original` (or
-    /// `adjusted` if it had none), returning `None` only if the
-    /// trail is empty.
-    #[must_use]
-    pub fn original_score(&self) -> Option<Confidence> {
-        self.trail.first().map(|s| s.original.unwrap_or(s.adjusted))
-    }
-
-    /// Final confidence — same as [`Self::confidence`], exposed as a
-    /// method for symmetry with [`Self::original_score`].
-    #[must_use]
-    pub fn final_score(&self) -> Confidence {
-        self.confidence
-    }
-
-    /// Names of the recognizers that produced this entity (one or
-    /// more, in dispatch order). Reads the [`TrailStepKind::Recognition`]
-    /// steps' `source` field.
-    pub fn recognizers(&self) -> impl Iterator<Item = &str> {
-        self.trail
-            .iter()
-            .filter(|s| s.kind == TrailStepKind::Recognition)
-            .map(|s| s.source.as_str())
-    }
-}
-
-#[cfg(any(test, feature = "test-utils"))]
-impl Entity<Text> {
-    /// Create a pre-filled [`EntityBuilder`] for tests.
-    ///
-    /// Defaults: `PersonName` / a synthetic `pattern` recognition
-    /// step / confidence `0.9` / text location at `start..end`.
-    pub fn test_builder(start: usize, end: usize) -> EntityBuilder<Text> {
-        let conf = Confidence::clamped(0.9);
-        Entity::builder()
-            .with_label(EntityLabelRef::from(
-                self::builtins::PERSON_NAME.name.clone(),
-            ))
-            .with_trail(vec![TrailStep::recognition(
-                "pattern",
-                conf,
-                TrailProvenance::Pattern(PatternProvenance::Regex {
-                    name: "test".to_owned(),
-                    regex: None,
-                    validator: None,
-                    contextual: false,
-                }),
-                "test fixture",
-            )])
-            .with_confidence(conf)
-            .with_location(TextLocation::new(start, end))
-    }
-}
-
-#[cfg(any(test, feature = "test-utils"))]
-impl<M: Modality> EntityBuilder<M> {
-    /// Build the entity, panicking on missing fields.
-    pub fn test_build(self) -> Entity<M> {
-        self.build().expect("test entity builder: missing fields")
-    }
-}
diff --git a/crates/nvisy-core/src/extraction/artifacts.rs b/crates/nvisy-core/src/extraction/artifacts.rs
deleted file mode 100644
index f9bda88d7..000000000
--- a/crates/nvisy-core/src/extraction/artifacts.rs
+++ /dev/null
@@ -1,63 +0,0 @@
-//! [`Artifacts`]: heterogeneous typed-map newtype attached to a
-//! [`Span<M>`] so extractors (and any future stage) can read or
-//! stamp out-of-band enrichments alongside the payload without
-//! widening the call surface for new entry types.
-//!
-//! Keyed by Rust type, so each entry is a distinct typed slot.
-//! Backed by [`TypeMap`] today; the newtype exists so consumer code
-//! never names the underlying crate and we can swap representations
-//! later without churn.
-//!
-//! [`Span<M>`]: crate::extraction::Span
-//! [`TypeMap`]: type_map::concurrent::TypeMap
-
-use std::fmt;
-
-use type_map::concurrent::TypeMap;
-
-/// Heterogeneous typed-map of per-span artifacts.
-///
-/// Entries are keyed by their concrete Rust type; only one value of
-/// any given type can be present at a time. Consumers do
-/// `artifacts.get::<T>()` / `artifacts.insert(value)`; when the entry
-/// is absent the consumer silently degrades.
-///
-/// `Send + Sync` so spans cross thread boundaries; cloning is *not*
-/// supported (entries may not themselves be `Clone`), matching
-/// `TypeMap`'s own surface.
-#[derive(Default)]
-pub struct Artifacts(TypeMap);
-
-impl Artifacts {
-    /// Build an empty bundle.
-    pub fn new() -> Self {
-        Self(TypeMap::new())
-    }
-
-    /// Insert one typed entry. Returns the previous value of that
-    /// type if one was already present.
-    pub fn insert<T: Send + Sync + 'static>(&mut self, value: T) -> Option<T> {
-        self.0.insert(value)
-    }
-
-    /// Borrow the entry of type `T`, if present.
-    pub fn get<T: Send + Sync + 'static>(&self) -> Option<&T> {
-        self.0.get::<T>()
-    }
-
-    /// Remove and return the entry of type `T`, if present.
-    pub fn take<T: Send + Sync + 'static>(&mut self) -> Option<T> {
-        self.0.remove::<T>()
-    }
-
-    /// Whether an entry of type `T` is present.
-    pub fn contains<T: Send + Sync + 'static>(&self) -> bool {
-        self.0.contains::<T>()
-    }
-}
-
-impl fmt::Debug for Artifacts {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("Artifacts").finish_non_exhaustive()
-    }
-}
diff --git a/crates/nvisy-core/src/extraction/data_at.rs b/crates/nvisy-core/src/extraction/data_at.rs
deleted file mode 100644
index 06c69416a..000000000
--- a/crates/nvisy-core/src/extraction/data_at.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-//! [`DataAt`]: resolve a per-modality location to its full
-//! [`M::Data`] payload — the same shape an [`Anonymizer<M>`] needs.
-//!
-//! Sibling to [`TextAt`]: where `TextAt` answers "what display text
-//! represents this location" (the read-only string a dedup layer or
-//! validation check inspects), `DataAt` answers "what is the full
-//! typed payload at this location" so a redaction operator can run
-//! against it. The two intentionally have different return shapes —
-//! `TextAt` lets consumers stay text-agnostic across modalities;
-//! `DataAt` preserves the modality-specific [`M::Data`] envelope
-//! the operator was written against.
-//!
-//! Concrete implementations live where the underlying resolver
-//! lives: `nvisy-engine` ships impls on `DocumentTree<M>` that
-//! consult the codec handle's `read` method.
-//!
-//! [`TextAt`]: super::TextAt
-//! [`Anonymizer<M>`]: # "lives in nvisy-toolkit"
-//! [`M::Data`]: crate::modality::Modality::Data
-
-use crate::modality::Modality;
-
-/// Resolve a modality-typed location to its full [`M::Data`]
-/// payload. Generic per-phase code bounds over `&impl DataAt<M>`
-/// and dispatches uniformly across modalities.
-///
-/// [`M::Data`]: Modality::Data
-#[async_trait::async_trait]
-pub trait DataAt<M: Modality>: Sync {
-    /// Resolve a location to its typed source payload, or `None`
-    /// when no payload exists at the location (out-of-bounds, the
-    /// handle doesn't expose that modality, …).
-    async fn data_at(&self, location: &M::Location) -> Option<M::Data>;
-}
diff --git a/crates/nvisy-core/src/extraction/mod.rs b/crates/nvisy-core/src/extraction/mod.rs
deleted file mode 100644
index ab37c0507..000000000
--- a/crates/nvisy-core/src/extraction/mod.rs
+++ /dev/null
@@ -1,62 +0,0 @@
-//! Extraction-side primitives shared across the runtime.
-//!
-//! - [`Extractor<M>`] — the per-modality extraction trait every
-//!   backend implements. Symmetric to [`EntityRecognizer`] on the
-//!   producer side but uses its own input / output shapes.
-//! - [`Span<M>`] — per-call input: payload + per-modality location +
-//!   optional language / correlation id + typed [`Artifacts`].
-//! - [`ExtractorOutput<M, T>`] — paired return shape: the
-//!   backend-shaped `value` plus the per-modality provenance value
-//!   the document stamps.
-//! - [`Artifacts`] — heterogeneous typed-map newtype attached to a
-//!   [`Span`] so extractors carry out-of-band enrichments alongside
-//!   the payload.
-//! - [`Modality`] — extension trait naming
-//!   `M::Extraction`.
-//! - [`TextAt`] — trait every extraction-aware consumer (dedup
-//!   layer, validation check) bounds on to read source *text* at a
-//!   per-modality location.
-//! - [`DataAt`] — sibling trait returning the full per-modality
-//!   [`M::Data`] payload an `Anonymizer<M>` operates on; bounded
-//!   by the redaction phase.
-//!
-//! [`M::Data`]: crate::modality::Modality::Data
-//! [`EntityRecognizer`]: crate::recognition::EntityRecognizer
-
-mod artifacts;
-mod data_at;
-mod output;
-mod span;
-mod text_at;
-
-pub use self::artifacts::Artifacts;
-pub use self::data_at::DataAt;
-pub use self::output::ExtractorOutput;
-pub use self::span::Span;
-pub use self::text_at::TextAt;
-use crate::Result;
-use crate::modality::Modality;
-
-/// Per-modality extractor: convert a per-call [`Span<M>`] into a
-/// backend-shaped `value` plus the modality-keyed provenance the
-/// document stamps at extraction time.
-///
-/// Object-safe so heterogeneous extractors live behind
-/// `Arc<dyn Extractor<M, Output = O>>` in consumer-side registries.
-#[async_trait::async_trait]
-pub trait Extractor<M>: Send + Sync
-where
-    M: Modality,
-{
-    /// The extractor's modality-specific return shape. Pick whatever
-    /// the underlying backend naturally produces; consumer glue
-    /// translates it into per-document [`Block<M>`] values.
-    ///
-    /// [`Block<M>`]: # "carrier owned by nvisy-engine"
-    type Output: Send;
-
-    /// Extract from `span`, returning the modality-specific output
-    /// alongside the provenance value the consumer stamps into the
-    /// document's metadata.
-    async fn extract(&self, span: &Span<M>) -> Result<ExtractorOutput<M, Self::Output>>;
-}
diff --git a/crates/nvisy-core/src/extraction/output.rs b/crates/nvisy-core/src/extraction/output.rs
deleted file mode 100644
index f67e602e0..000000000
--- a/crates/nvisy-core/src/extraction/output.rs
+++ /dev/null
@@ -1,33 +0,0 @@
-//! [`ExtractorOutput`]: paired return shape for an
-//! [`Extractor<M>`].
-//!
-//! Bundles the extractor's backend-shaped `value` with the modality-
-//! keyed `extraction` provenance the document stamps at extraction
-//! time. Lifting provenance into the return type means consumers
-//! capture both pieces from a single call rather than reaching for a
-//! separate `extraction()` accessor on the extractor.
-//!
-//! [`Extractor<M>`]: crate::extraction::Extractor
-
-use crate::modality::Modality;
-
-/// Combined extractor return shape: a per-call `value` and the
-/// modality-keyed `extraction` provenance produced alongside it.
-#[derive(Debug, Clone)]
-pub struct ExtractorOutput<M: Modality, T> {
-    /// Backend-shaped extractor output (e.g. `Vec<OcrOutput>` for an
-    /// image extractor, `Transcription` for an audio extractor).
-    pub value: T,
-    /// Per-modality provenance value the document's metadata records
-    /// at extraction time. Same value for every call from a given
-    /// extractor instance — folded into the output so consumers
-    /// don't keep a separate accessor.
-    pub extraction: M::Extraction,
-}
-
-impl<M: Modality, T> ExtractorOutput<M, T> {
-    /// Construct an output from its two parts.
-    pub fn new(value: T, extraction: M::Extraction) -> Self {
-        Self { value, extraction }
-    }
-}
diff --git a/crates/nvisy-core/src/extraction/span.rs b/crates/nvisy-core/src/extraction/span.rs
deleted file mode 100644
index bad572ce9..000000000
--- a/crates/nvisy-core/src/extraction/span.rs
+++ /dev/null
@@ -1,79 +0,0 @@
-//! [`Span<M>`]: per-call input handed to an [`Extractor<M>`].
-//!
-//! Pairs the modality payload with where in the source it lives
-//! ([`M::Location`]) and a typed [`Artifacts`] bundle for
-//! out-of-band enrichments. Sibling to [`RecognizerInput<M>`] on the
-//! recognizer side — the two stay separate while their concerns
-//! diverge: recognizers carry hints/labels/candidate languages,
-//! extractors carry a location and a typed artifacts bundle.
-//!
-//! [`Extractor<M>`]: crate::extraction::Extractor
-//! [`M::Location`]: crate::modality::Modality::Location
-//! [`RecognizerInput<M>`]: crate::recognition::RecognizerInput
-
-use uuid::Uuid;
-
-use super::Artifacts;
-use crate::modality::Modality;
-use crate::primitive::LanguageTag;
-
-/// Per-call extraction input: the payload, where it lives in the
-/// source, an optional language assertion / correlation id, and a
-/// typed bundle of [`Artifacts`].
-///
-/// Modalities that recognizers care about (today: every recognizer
-/// modality) implement [`Modality`]; extractors reuse the same
-/// `M::Data` shape rather than defining a parallel payload type.
-#[derive(Debug)]
-pub struct Span<M: Modality> {
-    /// Modality-specific payload the extractor will process.
-    pub data: M::Data,
-    /// Where this payload lives in the source (whole-image for OCR,
-    /// full-stream time span for STT, …). Modality-specific via
-    /// [`Modality::Location`].
-    pub location: M::Location,
-    /// Caller-asserted language. Backends that support per-call
-    /// language hinting use this; the rest ignore it.
-    pub language: Option<LanguageTag>,
-    /// Correlation UUID propagated through the tracing span for this
-    /// call.
-    pub correlation_id: Option<Uuid>,
-    /// Heterogeneous typed bundle of per-span enrichments. Empty by
-    /// default.
-    pub artifacts: Artifacts,
-}
-
-impl<M: Modality> Span<M> {
-    /// Construct a span with the payload and location set;
-    /// language, correlation id, and artifacts default to empty.
-    pub fn new(data: M::Data, location: M::Location) -> Self {
-        Self {
-            data,
-            location,
-            language: None,
-            correlation_id: None,
-            artifacts: Artifacts::new(),
-        }
-    }
-
-    /// Set the asserted language.
-    #[must_use]
-    pub fn with_language(mut self, language: LanguageTag) -> Self {
-        self.language = Some(language);
-        self
-    }
-
-    /// Set the correlation id propagated through the tracing span.
-    #[must_use]
-    pub fn with_correlation_id(mut self, id: Uuid) -> Self {
-        self.correlation_id = Some(id);
-        self
-    }
-
-    /// Replace the artifacts bundle with `artifacts`.
-    #[must_use]
-    pub fn with_artifacts(mut self, artifacts: Artifacts) -> Self {
-        self.artifacts = artifacts;
-        self
-    }
-}
diff --git a/crates/nvisy-core/src/extraction/text_at.rs b/crates/nvisy-core/src/extraction/text_at.rs
deleted file mode 100644
index 129b8e3a8..000000000
--- a/crates/nvisy-core/src/extraction/text_at.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-//! [`TextAt`]: resolve a per-modality location to its source text.
-//!
-//! The toolkit-side phase code (deduplication layers, validation
-//! checks, redaction strategy bindings) all need to read back the
-//! original text at a modality-typed location. The trait lives here
-//! in core because every layer of the stack — toolkit components,
-//! document phase drivers, custom user pipelines — bounds on it.
-//!
-//! Sibling to [`DataAt`]: where `DataAt` answers "what raw modality
-//! payload lives at this location" (the typed `M::Data` envelope an
-//! anonymizer rewrites), `TextAt` answers "what display text
-//! represents this location" — the string a dedup layer or
-//! validation check substring-scans. For text/tabular the two are
-//! trivially related (the data *is* text). For image/audio they
-//! diverge: `DataAt` reads raw pixels / samples through the codec,
-//! while `TextAt` reads OCR'd / transcribed text from the document's
-//! blocks.
-//!
-//! Concrete implementations live where the underlying resolver lives:
-//! `nvisy-engine` ships impls on `DocumentTree<M>` that consult the
-//! codec handle for text/tabular and the document blocks for
-//! image/audio.
-//!
-//! [`DataAt`]: super::DataAt
-
-use crate::modality::Modality;
-
-/// Resolve a modality-typed location to the corresponding source
-/// text. Generic per-phase code bounds over `&impl TextAt<M>` and
-/// dispatches uniformly across modalities.
-#[async_trait::async_trait]
-pub trait TextAt<M: Modality>: Sync {
-    /// Resolve a location to its source-text representation, or
-    /// `None` when no readable text exists at the location.
-    async fn text_at(&self, location: &M::Location) -> Option<String>;
-}
diff --git a/crates/nvisy-core/src/file.rs b/crates/nvisy-core/src/file.rs
new file mode 100644
index 000000000..a8b67d0a4
--- /dev/null
+++ b/crates/nvisy-core/src/file.rs
@@ -0,0 +1,102 @@
+//! [`FileMetadata`]: persisted descriptor for an uploaded file.
+//!
+//! A file in the engine is a `(metadata, bytes)` pair: the bytes
+//! live in a blob-separated keyspace, the metadata in a small
+//! JSON keyspace. The split lets `list_files` enumerate every
+//! file for an actor without paying the cost of loading the
+//! bytes. Both keyspaces key by `(actor_id, file_id)`.
+//!
+//! The descriptor mirrors what [`DocumentInput`] carries on a
+//! [`StartBatch`] — the same `descriptor_labels` /
+//! `descriptor_metadata` gate policies via
+//! [`DocumentPredicate`]. When a run references a stored file,
+//! the run inherits these gates.
+//!
+//! [`DocumentInput`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/runs/struct.DocumentInput.html
+//! [`StartBatch`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/runs/struct.StartBatch.html
+//! [`DocumentPredicate`]: crate::policy::DocumentPredicate
+
+use std::collections::HashMap;
+
+use hipstr::HipStr;
+use jiff::Timestamp;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use uuid::Uuid;
+
+/// Descriptor for one stored file. Persisted as a small JSON
+/// blob next to the raw bytes; clients consume this through
+/// `GET /files` (list) and `GET /files/{id}` (one).
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct FileMetadata {
+    /// Engine-minted UUIDv7. The `(actor_id, id)` pair keys both
+    /// the metadata blob and the raw bytes.
+    pub id: Uuid,
+    /// Original filename, when the upload carried one (e.g. via
+    /// `Content-Disposition: attachment; filename="scan.pdf"`).
+    /// Display-only — the codec resolves on `extension`.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[schemars(with = "Option<String>")]
+    pub filename: Option<HipStr<'static>>,
+    /// Caller-supplied MIME hint (e.g. `application/pdf`). The
+    /// codec uses `extension`, not this; recorded for audit and
+    /// for clients that round-trip metadata.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[schemars(with = "Option<String>")]
+    pub content_type: Option<HipStr<'static>>,
+    /// File extension the codec registry resolves on (e.g.
+    /// `"txt"`, `"pdf"`, `"png"`). Case-insensitive, no leading
+    /// dot. Required at upload time — the engine derives it from
+    /// `filename` when the upload omits it explicitly.
+    #[schemars(with = "String")]
+    pub extension: HipStr<'static>,
+    /// Length of the raw bytes in bytes.
+    pub size: u64,
+    /// Hex-encoded SHA-256 of the raw bytes. Stable per upload —
+    /// re-uploading the same bytes produces the same digest.
+    #[schemars(with = "String")]
+    pub digest: HipStr<'static>,
+    /// Engine timestamp at upload time. UUIDv7's encoded
+    /// timestamp also orders files, but this is the source of
+    /// truth for display and filtering.
+    #[schemars(with = "String")]
+    pub uploaded_at: Timestamp,
+    /// Where this file came from. `None` for uploaded files;
+    /// [`FileLineage::RedactedFrom`] for files produced by a
+    /// redaction apply.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub lineage: Option<FileLineage>,
+    /// Doc-level labels that gate
+    /// [`DocumentPredicate::HasLabel`] policies when a run
+    /// references this file.
+    ///
+    /// [`DocumentPredicate::HasLabel`]: crate::policy::DocumentPredicate::HasLabel
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub descriptor_labels: Vec<String>,
+    /// Doc-level metadata that gates
+    /// [`DocumentPredicate::HasMetadata`] policies, merged with
+    /// any per-request metadata at run-time.
+    ///
+    /// [`DocumentPredicate::HasMetadata`]: crate::policy::DocumentPredicate::HasMetadata
+    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
+    pub descriptor_metadata: HashMap<String, String>,
+}
+
+/// Provenance for a [`FileMetadata`]. Uploaded files carry
+/// `None`; engine-produced files (today: redaction apply
+/// outputs) carry one of these variants so audits and clients
+/// can trace any file back to what produced it.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "camelCase")]
+pub enum FileLineage {
+    /// Output of a redaction apply. `runId` is the run that
+    /// produced this file; `sourceFileId` is the input file the
+    /// run read.
+    RedactedFrom {
+        /// Run that produced this file (`/redactions/{runId}`).
+        run_id: Uuid,
+        /// Original input file the run read.
+        source_file_id: Uuid,
+    },
+}
diff --git a/crates/nvisy-core/src/lib.rs b/crates/nvisy-core/src/lib.rs
index ea75f4cca..471ec0a1e 100644
--- a/crates/nvisy-core/src/lib.rs
+++ b/crates/nvisy-core/src/lib.rs
@@ -2,13 +2,14 @@
 #![cfg_attr(docsrs, feature(doc_cfg))]
 #![doc = include_str!("../README.md")]
 
-pub mod entity;
-pub mod extraction;
-pub mod health;
-pub mod modality;
-pub mod primitive;
-pub mod recognition;
-pub mod redaction;
+pub mod context;
+pub mod file;
+pub mod plan;
+pub mod policy;
+pub mod schema;
+pub mod service;
+pub mod source;
 
-mod error;
-pub use self::error::{Error, ErrorKind, Result};
+pub use self::file::{FileLineage, FileMetadata};
+pub use self::service::{Error, ErrorKind, Result};
+pub use self::source::ContentSource;
diff --git a/crates/nvisy-core/src/modality/audio.rs b/crates/nvisy-core/src/modality/audio.rs
deleted file mode 100644
index e9b9fa6be..000000000
--- a/crates/nvisy-core/src/modality/audio.rs
+++ /dev/null
@@ -1,163 +0,0 @@
-//! [`Audio`] modality marker, [`AudioLocation`] coordinate type,
-//! [`AudioData`] per-call payload, and [`AudioExtraction`] provenance
-//! enum.
-
-use std::cmp::Ordering;
-
-use bytes::Bytes;
-use hipstr::HipStr;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-use super::{Modality, Overlap};
-use crate::entity::ModelProvenance;
-use crate::primitive::TimeSpan;
-use crate::redaction::AudioReplacement;
-
-/// Audio modality marker (zero-sized).
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Audio;
-
-impl Modality for Audio {
-    type Data = AudioData;
-    type Extraction = AudioExtraction;
-    type Location = AudioLocation;
-    type Replacement = AudioReplacement;
-
-    const KIND: super::ModalityKind = super::ModalityKind::Audio;
-    const NAME: &'static str = "audio";
-}
-
-/// A time interval within audio content.
-#[derive(Debug, Clone, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct AudioLocation {
-    /// Time interval.
-    pub time_span: TimeSpan,
-    /// Speaker identifier from diarization.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub speaker_id: Option<String>,
-    /// Links this interval to a specific audio document.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub audio_id: Option<Uuid>,
-}
-
-impl AudioLocation {
-    /// Create an [`AudioLocation`] covering `time_span`, with no
-    /// speaker attribution or audio-document link.
-    pub fn new(time_span: TimeSpan) -> Self {
-        Self {
-            time_span,
-            speaker_id: None,
-            audio_id: None,
-        }
-    }
-}
-
-impl Eq for AudioLocation {}
-
-impl Ord for AudioLocation {
-    /// Lex order over `(time_span.start_us, time_span.end_us)`.
-    /// `speaker_id` and `audio_id` are ignored.
-    fn cmp(&self, other: &Self) -> Ordering {
-        (self.time_span.start_us, self.time_span.end_us)
-            .cmp(&(other.time_span.start_us, other.time_span.end_us))
-    }
-}
-
-impl PartialOrd for AudioLocation {
-    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
-        Some(self.cmp(other))
-    }
-}
-
-/// Per-call payload for [`Audio`] extractors.
-///
-/// Audio backends (STT, diarization) take encoded bytes plus an
-/// optional original filename. No dimensions or sample-rate metadata
-/// is carried at this layer — providers parse the container
-/// themselves.
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-pub struct AudioData {
-    /// Encoded audio bytes.
-    pub bytes: Bytes,
-    /// Original filename, when known. Threaded through to providers
-    /// like OpenAI Whisper that key on the source filename's
-    /// extension to pick a decoder.
-    pub filename: Option<HipStr<'static>>,
-}
-
-impl AudioData {
-    /// Construct with the encoded bytes; filename is initially unset.
-    pub fn new(bytes: impl Into<Bytes>) -> Self {
-        Self {
-            bytes: bytes.into(),
-            filename: None,
-        }
-    }
-
-    /// Attach an original filename hint.
-    pub fn with_filename(mut self, filename: impl Into<HipStr<'static>>) -> Self {
-        self.filename = Some(filename.into());
-        self
-    }
-
-    /// Extension derived from [`filename`], or `"mp3"` when no
-    /// filename is set or the filename has no extension.
-    ///
-    /// [`filename`]: Self::filename
-    pub fn extension(&self) -> &str {
-        self.filename
-            .as_deref()
-            .and_then(|name| name.rsplit_once('.'))
-            .map(|(_, ext)| ext)
-            .unwrap_or("mp3")
-    }
-
-    /// View the encoded bytes.
-    pub fn as_bytes(&self) -> &Bytes {
-        &self.bytes
-    }
-}
-
-impl Overlap for AudioLocation {
-    /// Two audio intervals overlap only when they target the same
-    /// stream (matching `audio_id`) on the same `speaker_id` and
-    /// their time spans intersect. Two voices captured on the same
-    /// time interval are physically distinct redaction targets
-    /// (source separation can suppress one speaker while keeping
-    /// another), so different speaker IDs are not treated as
-    /// overlapping even when the time spans intersect.
-    fn overlaps(&self, other: &Self) -> bool {
-        self.audio_id == other.audio_id
-            && self.speaker_id == other.speaker_id
-            && self.time_span.overlaps(&other.time_span)
-    }
-}
-
-/// How a [`Document<Audio>`]'s content was produced.
-///
-/// [`Pending`] is the importer-time placeholder before any extractor
-/// has run; the extractor stage replaces it with the concrete variant
-/// carrying the backend's [`ModelProvenance`].
-///
-/// [`Document<Audio>`]: # "carrier owned by nvisy-engine"
-/// [`Pending`]: Self::Pending
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-#[non_exhaustive]
-pub enum AudioExtraction {
-    /// No extractor has run yet. Importer stamps this; the STT or
-    /// diarization backend replaces it once samples are processed.
-    Pending,
-    /// Speech-to-text transcription: audio samples converted into
-    /// text segments.
-    Transcription(ModelProvenance),
-    /// Speaker diarization: audio segmented by speaker identity
-    /// before recognition attributes utterances.
-    Diarization(ModelProvenance),
-}
diff --git a/crates/nvisy-core/src/modality/image.rs b/crates/nvisy-core/src/modality/image.rs
deleted file mode 100644
index 69e4b8573..000000000
--- a/crates/nvisy-core/src/modality/image.rs
+++ /dev/null
@@ -1,183 +0,0 @@
-//! [`Image`] modality marker, [`ImageLocation`] coordinate type,
-//! [`ImageData`] per-call payload, and [`ImageExtraction`] provenance
-//! enum.
-
-use std::cmp::Ordering;
-
-use bytes::Bytes;
-use hipstr::HipStr;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-use super::{Modality, Overlap};
-use crate::entity::ModelProvenance;
-use crate::primitive::{BoundingBox, Dimensions, Polygon};
-use crate::redaction::ImageReplacement;
-
-/// Image modality marker (zero-sized).
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Image;
-
-impl Modality for Image {
-    type Data = ImageData;
-    type Extraction = ImageExtraction;
-    type Location = ImageLocation;
-    type Replacement = ImageReplacement;
-
-    const KIND: super::ModalityKind = super::ModalityKind::Image;
-    const NAME: &'static str = "image";
-}
-
-/// A region within image content.
-#[derive(Debug, Clone, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct ImageLocation {
-    /// Axis-aligned bounding box of the region.
-    pub bounding_box: BoundingBox,
-    /// Polygon vertices for the region when the source produced a
-    /// rotated or quadrilateral shape (OCR engines that emit 4-point
-    /// polygons populate this; axis-aligned-only sources leave it
-    /// unset).
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub polygon: Option<Polygon>,
-    /// Links this region to a specific image document.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub image_id: Option<Uuid>,
-    /// 1-based page number (for multi-page documents like PDFs).
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub page_number: Option<u32>,
-}
-
-impl ImageLocation {
-    /// Create an [`ImageLocation`] from the bounding box alone, with
-    /// every optional field unset.
-    pub fn new(bounding_box: BoundingBox) -> Self {
-        Self {
-            bounding_box,
-            polygon: None,
-            image_id: None,
-            page_number: None,
-        }
-    }
-
-    /// Area of the bounding box in pixels (`width * height`).
-    pub fn area(&self) -> f64 {
-        self.bounding_box.width * self.bounding_box.height
-    }
-}
-
-impl PartialOrd for ImageLocation {
-    /// Reading order: top-to-bottom by `bounding_box.y`, then
-    /// left-to-right by `bounding_box.x`, then `height`, then
-    /// `width`. Float fields compare via [`f64::total_cmp`]. `Ord`
-    /// is not implemented — `f64` is not `Eq`, so a total order
-    /// can't be promised at the trait level.
-    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
-        let bb = &self.bounding_box;
-        let ob = &other.bounding_box;
-        Some(
-            bb.y.total_cmp(&ob.y)
-                .then_with(|| bb.x.total_cmp(&ob.x))
-                .then_with(|| bb.height.total_cmp(&ob.height))
-                .then_with(|| bb.width.total_cmp(&ob.width)),
-        )
-    }
-}
-
-/// Per-call payload for [`Image`] recognizers and extractors.
-///
-/// The pixel dimensions are needed alongside the encoded bytes
-/// because recognizers that emit normalised bounding boxes scale
-/// them to pixel coordinates using `dims`.
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-pub struct ImageData {
-    /// Encoded image bytes.
-    pub bytes: Bytes,
-    /// Pixel dimensions of the encoded image.
-    pub dims: Dimensions,
-    /// Original filename, when known. Useful for diagnostics and for
-    /// downstream consumers that want to infer the encoding from the
-    /// extension.
-    pub filename: Option<HipStr<'static>>,
-}
-
-impl ImageData {
-    /// Construct with the encoded bytes and pixel dimensions; filename
-    /// is initially unset.
-    pub fn new(bytes: impl Into<Bytes>, dims: Dimensions) -> Self {
-        Self {
-            bytes: bytes.into(),
-            dims,
-            filename: None,
-        }
-    }
-
-    /// Attach an original filename hint.
-    pub fn with_filename(mut self, filename: impl Into<HipStr<'static>>) -> Self {
-        self.filename = Some(filename.into());
-        self
-    }
-
-    /// Extension derived from [`filename`], or `"png"` when no
-    /// filename is set or the filename has no extension.
-    ///
-    /// [`filename`]: Self::filename
-    pub fn extension(&self) -> &str {
-        self.filename
-            .as_deref()
-            .and_then(|name| name.rsplit_once('.'))
-            .map(|(_, ext)| ext)
-            .unwrap_or("png")
-    }
-}
-
-impl Overlap for ImageLocation {
-    /// Two image regions overlap only when they target the same
-    /// image (matching `image_id`) on the same page (matching
-    /// `page_number`) and their bounding boxes intersect. Without
-    /// the image/page gates, two regions with the same bbox
-    /// coordinates on different pages or different uploads would
-    /// false-positive as overlapping.
-    fn overlaps(&self, other: &Self) -> bool {
-        self.image_id == other.image_id
-            && self.page_number == other.page_number
-            && self.bounding_box.overlaps(&other.bounding_box)
-    }
-}
-
-/// How a [`Document<Image>`]'s content was produced.
-///
-/// Every image-modality document is the output of *some* recognition
-/// pass over pixels; the variant names which pass ran. [`Pending`] is
-/// the importer-time placeholder before any extractor has run; the
-/// extractor stage replaces it with the concrete variant carrying the
-/// backend's [`ModelProvenance`].
-///
-/// [`Document<Image>`]: # "carrier owned by nvisy-engine"
-/// [`Pending`]: Self::Pending
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-#[non_exhaustive]
-pub enum ImageExtraction {
-    /// No extractor has run yet. Importer stamps this; the extractor
-    /// stage replaces it once an OCR / scene-text / object-detection /
-    /// layout backend produces blocks.
-    Pending,
-    /// Optical character recognition: raster text (printed or
-    /// handwritten) converted into machine-readable characters.
-    Ocr(ModelProvenance),
-    /// Scene text detection: text embedded in natural images (signs,
-    /// screens, whiteboards) localised prior to OCR.
-    SceneText(ModelProvenance),
-    /// Object detection: regions of interest located and labelled
-    /// within an image or video frame.
-    ObjectDetection(ModelProvenance),
-    /// Document layout analysis: structural regions (headers, footers,
-    /// signature blocks, form fields) identified by spatial
-    /// arrangement rather than content.
-    LayoutAnalysis(ModelProvenance),
-}
diff --git a/crates/nvisy-core/src/modality/mod.rs b/crates/nvisy-core/src/modality/mod.rs
deleted file mode 100644
index f9fab8229..000000000
--- a/crates/nvisy-core/src/modality/mod.rs
+++ /dev/null
@@ -1,190 +0,0 @@
-//! Per-modality marker types and the [`Modality`] trait.
-//!
-//! Generic containers ([`Entity<M>`]) are parameterised over
-//! `M: Modality`, where `M` is one of the zero-sized markers
-//! [`Text`], [`Image`], [`Audio`], or [`Tabular`]. Each marker's
-//! [`Modality::Location`] associated type names the actual
-//! coordinate value those containers store ([`TextLocation`],
-//! [`ImageLocation`], [`AudioLocation`], [`TabularLocation`]).
-//!
-//! Splitting the marker from the data lets the type system carry
-//! "which modality" cleanly while the data can grow shape (e.g. an
-//! image location adding polygon variants) without touching the
-//! marker.
-//!
-//! [`Modality`] bundles the four per-modality associated types every
-//! consumer reaches for: [`Location`] (coordinate), [`Data`] (per-call
-//! payload the recognizer/extractor scans), [`Replacement`] (redaction
-//! record), and [`Extraction`] (per-modality provenance enum stamped
-//! onto a document at extractor time), plus the [`KIND`] runtime tag
-//! ([`ModalityKind`]) used wherever `M` has been erased. The
-//! document-shape side (`Block`, `Metadata`) lives in `nvisy-engine`
-//! via a `DocumentModality` extension trait — toolkit and document
-//! don't pollute core.
-//!
-//! [`KIND`]: Modality::KIND
-//!
-//! [`Location`]: Modality::Location
-//! [`Data`]: Modality::Data
-//! [`Replacement`]: Modality::Replacement
-//! [`Extraction`]: Modality::Extraction
-//!
-//! [`Entity<M>`]: crate::entity::Entity
-
-mod audio;
-mod image;
-mod tabular;
-mod text;
-
-use std::any::TypeId;
-use std::fmt::Debug;
-use std::hash::Hash;
-
-use schemars::JsonSchema;
-use serde::Serialize;
-use serde::de::DeserializeOwned;
-
-pub use self::audio::{Audio, AudioData, AudioExtraction, AudioLocation};
-pub use self::image::{Image, ImageData, ImageExtraction, ImageLocation};
-pub use self::tabular::{Tabular, TabularExtraction, TabularLocation};
-pub use self::text::{ContextWindow, Text, TextData, TextExtraction, TextLocation};
-
-/// Runtime tag identifying which [`Modality`] a generic container
-/// carries. Used wherever the typed `M` is erased — the codec
-/// registry returning `UntypedDocumentHandle`, redaction
-/// applicator switching on the audit modality, etc.
-///
-/// Closed set today (Text/Tabular/Image/Audio). Adding a fifth
-/// modality is a workspace-wide change — it also requires a new
-/// variant on every `Untyped*` enum and a [`Modality`] impl.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-#[derive(serde::Serialize, serde::Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum ModalityKind {
-    /// [`Text`] modality.
-    Text,
-    /// [`Tabular`] modality.
-    Tabular,
-    /// [`Image`] modality.
-    Image,
-    /// [`Audio`] modality.
-    Audio,
-}
-
-impl ModalityKind {
-    /// Return the [`ModalityKind`] for a typed `M: Modality` at the
-    /// call site.
-    ///
-    /// Resolved at runtime via [`TypeId`]; the match is exhaustive
-    /// over the four built-in marker types. An unknown `M` panics
-    /// — the modality set is closed.
-    #[must_use]
-    pub fn of<M: Modality>() -> Self {
-        let id = TypeId::of::<M>();
-        if id == TypeId::of::<Text>() {
-            Self::Text
-        } else if id == TypeId::of::<Tabular>() {
-            Self::Tabular
-        } else if id == TypeId::of::<Image>() {
-            Self::Image
-        } else if id == TypeId::of::<Audio>() {
-            Self::Audio
-        } else {
-            unreachable!("Modality must be one of Text/Tabular/Image/Audio");
-        }
-    }
-}
-
-/// Marker trait implemented by every per-modality marker type
-/// ([`Text`], [`Image`], [`Audio`], [`Tabular`]).
-///
-/// Bundles the per-modality associated types every consumer reaches
-/// for:
-///
-/// - [`Location`] — coordinate value carried by generic containers
-///   parameterised on `M` (`Entity<M>::location` is `M::Location`,
-///   etc.). Always present.
-/// - [`Data`] — per-call payload (the bytes / text / dimensions a
-///   recognizer or extractor actually scans). Tabular shares
-///   [`TextData`] with Text since tabular cells are text-shaped at
-///   the per-call level.
-/// - [`Replacement`] — redaction record an anonymizer emits and any
-///   write-back path lands at the entity's location.
-/// - [`Extraction`] — per-modality provenance enum stamped onto a
-///   `Document<M>`'s metadata at extractor time
-///   (`TextExtraction`, `ImageExtraction`, etc.).
-///
-/// The marker type itself is zero-sized; everything else lives
-/// behind the associated types.
-///
-/// [`Location`]: Self::Location
-/// [`Data`]: Self::Data
-/// [`Replacement`]: Self::Replacement
-/// [`Extraction`]: Self::Extraction
-pub trait Modality: Copy + Default + Debug + PartialEq + Eq + Send + Sync + 'static {
-    /// Stable lowercase identifier for this modality (e.g.
-    /// `"text"`, `"image"`). Used as the value of structured
-    /// telemetry fields and as a wire-friendly modality tag in
-    /// logs. Each marker advertises its own; downstream code never
-    /// pattern-matches on the closed set.
-    const NAME: &'static str;
-
-    /// Runtime tag for this modality. Used wherever the typed `M`
-    /// is erased — the codec registry, untyped document handles,
-    /// the redaction applicator's audit dispatch, etc.
-    const KIND: ModalityKind;
-
-    /// Coordinate value carried by generic containers parameterised
-    /// on this modality.
-    type Location: Clone
-        + Debug
-        + PartialEq
-        + Send
-        + Sync
-        + Serialize
-        + DeserializeOwned
-        + JsonSchema
-        + 'static;
-
-    /// Per-call modality-specific payload: the bytes / text /
-    /// dimensions a recognizer or extractor actually scans.
-    type Data: Clone + Hash + Debug + Send + Sync;
-
-    /// What an anonymizer writes at the entity's location. Embedded
-    /// in audit records, so it carries the full serde + schemars
-    /// bundle to derive its codecs transparently.
-    type Replacement: Clone
-        + Debug
-        + PartialEq
-        + Send
-        + Sync
-        + Serialize
-        + DeserializeOwned
-        + JsonSchema
-        + 'static;
-
-    /// Per-modality provenance enum recording how a [`Document<M>`]'s
-    /// primary content was produced (e.g. native text-layer parse vs
-    /// OCR'd image-backed page).
-    ///
-    /// [`Document<M>`]: # "carrier owned by nvisy-engine"
-    type Extraction: Clone + Debug + PartialEq + Send + Sync + 'static;
-}
-
-/// Check whether two coordinates of the same modality overlap.
-///
-/// Implemented on the per-modality [`Location`] types, not on the
-/// markers — the markers carry no data and have nothing to compare.
-/// Semantics vary by modality:
-/// - **Text**: byte-range interval overlap (`start < other.end && other.start < end`).
-/// - **Image**: bounding box intersection.
-/// - **Audio**: time span overlap.
-/// - **Tabular**: same cell (row + column), with optional intra-cell
-///   byte-range check when offsets are present.
-///
-/// [`Location`]: Modality::Location
-pub trait Overlap {
-    /// Whether this coordinate overlaps `other` in its modality's
-    /// natural sense.
-    fn overlaps(&self, other: &Self) -> bool;
-}
diff --git a/crates/nvisy-core/src/modality/tabular.rs b/crates/nvisy-core/src/modality/tabular.rs
deleted file mode 100644
index fc9927635..000000000
--- a/crates/nvisy-core/src/modality/tabular.rs
+++ /dev/null
@@ -1,193 +0,0 @@
-//! [`Tabular`] modality marker, [`TabularLocation`] coordinate type,
-//! and the [`TabularExtraction`] provenance enum.
-
-use std::cmp::Ordering;
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::{Modality, Overlap, TextData};
-use crate::entity::ModelProvenance;
-use crate::redaction::TabularReplacement;
-
-/// Tabular modality marker (zero-sized).
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Tabular;
-
-impl Modality for Tabular {
-    type Data = TextData;
-    type Extraction = TabularExtraction;
-    type Location = TabularLocation;
-    type Replacement = TabularReplacement;
-
-    const KIND: super::ModalityKind = super::ModalityKind::Tabular;
-    const NAME: &'static str = "tabular";
-}
-
-/// A cell (or sub-cell range) within tabular content.
-#[derive(Debug, Clone, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct TabularLocation {
-    /// Row index (0-based).
-    pub row_index: u32,
-    /// Column index (0-based).
-    pub column_index: u32,
-    /// Byte offset within the cell where the range starts, if applicable.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub start_offset: Option<usize>,
-    /// Byte offset within the cell where the range ends, if applicable.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub end_offset: Option<usize>,
-    /// Column name or header label.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub column_name: Option<String>,
-    /// Sheet or table name (for multi-sheet documents).
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub sheet_name: Option<String>,
-}
-
-impl TabularLocation {
-    /// Create a [`TabularLocation`] for the given cell coordinates,
-    /// with every optional field (intra-cell offsets, column name,
-    /// sheet name) unset.
-    pub fn new(row_index: u32, column_index: u32) -> Self {
-        Self {
-            row_index,
-            column_index,
-            start_offset: None,
-            end_offset: None,
-            column_name: None,
-            sheet_name: None,
-        }
-    }
-
-    /// Inclusive byte range within the cell, treating absent offsets
-    /// as "whole cell". Returns `(start, end)` with `start == 0`
-    /// when no `start_offset` is set and `end == usize::MAX` when
-    /// no `end_offset` is set — a sentinel range that overlaps any
-    /// concrete subrange and merges into anything.
-    fn cell_range(&self) -> (usize, usize) {
-        (
-            self.start_offset.unwrap_or(0),
-            self.end_offset.unwrap_or(usize::MAX),
-        )
-    }
-}
-
-impl Ord for TabularLocation {
-    /// Lex order over `(sheet_name, row_index, column_index,
-    /// start_offset, end_offset)`. Absent intra-cell offsets sort
-    /// as `0` / `usize::MAX` respectively so a whole-cell location
-    /// brackets any sub-cell range. `column_name` is ignored.
-    fn cmp(&self, other: &Self) -> Ordering {
-        let (s1, e1) = self.cell_range();
-        let (s2, e2) = other.cell_range();
-        (&self.sheet_name, self.row_index, self.column_index, s1, e1).cmp(&(
-            &other.sheet_name,
-            other.row_index,
-            other.column_index,
-            s2,
-            e2,
-        ))
-    }
-}
-
-impl PartialOrd for TabularLocation {
-    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
-        Some(self.cmp(other))
-    }
-}
-
-impl Overlap for TabularLocation {
-    /// Two tabular ranges overlap only when they target the same
-    /// cell — matching `row_index`, `column_index`, **and**
-    /// `sheet_name` — and their intra-cell byte ranges intersect.
-    /// Without the sheet gate, two cells at the same row/col across
-    /// sheets of a workbook would false-positive as overlapping.
-    fn overlaps(&self, other: &Self) -> bool {
-        if self.row_index != other.row_index
-            || self.column_index != other.column_index
-            || self.sheet_name != other.sheet_name
-        {
-            return false;
-        }
-        let (s1, e1) = self.cell_range();
-        let (s2, e2) = other.cell_range();
-        s1 < e2 && s2 < e1
-    }
-}
-
-/// How a [`Document<Tabular>`]'s structure was produced.
-///
-/// [`Document<Tabular>`]: # "carrier owned by nvisy-engine"
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-#[non_exhaustive]
-pub enum TabularExtraction {
-    /// Header names and column types known from the source format
-    /// (Parquet, Avro, CSV with a header row).
-    SchemaTyped,
-    /// Column semantics inferred from column data (header-less CSV,
-    /// positional conventions).
-    SchemaInferred,
-    /// Tabular structure recovered from an image (row/column
-    /// reconstruction from a scanned table); preserves cell
-    /// relationships that plain OCR loses.
-    Recovered(ModelProvenance),
-}
-
-impl TabularExtraction {
-    /// Map the codec's `has_header()` signal to the matching
-    /// extraction variant.
-    ///
-    /// `None` means the codec couldn't decide (non-tabular handle
-    /// somehow reached this site) — fall back to `SchemaInferred`
-    /// rather than panicking, since over-reporting "inferred" is safe.
-    pub fn from_header_signal(has_header: Option<bool>) -> Self {
-        match has_header {
-            Some(true) => Self::SchemaTyped,
-            Some(false) | None => Self::SchemaInferred,
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn cell_with_offsets(row: u32, col: u32, start: usize, end: usize) -> TabularLocation {
-        TabularLocation {
-            start_offset: Some(start),
-            end_offset: Some(end),
-            ..TabularLocation::new(row, col)
-        }
-    }
-
-    #[test]
-    fn overlap_same_cell_no_offsets() {
-        assert!(TabularLocation::new(0, 0).overlaps(&TabularLocation::new(0, 0)));
-    }
-
-    #[test]
-    fn no_overlap_different_row() {
-        assert!(!TabularLocation::new(0, 0).overlaps(&TabularLocation::new(1, 0)));
-    }
-
-    #[test]
-    fn overlap_same_cell_intersecting_offsets() {
-        assert!(cell_with_offsets(0, 0, 0, 10).overlaps(&cell_with_offsets(0, 0, 5, 15)));
-    }
-
-    #[test]
-    fn no_overlap_same_cell_disjoint_offsets() {
-        assert!(!cell_with_offsets(0, 0, 0, 5).overlaps(&cell_with_offsets(0, 0, 5, 10)));
-    }
-
-    #[test]
-    fn overlap_same_cell_one_has_offsets() {
-        assert!(TabularLocation::new(0, 0).overlaps(&cell_with_offsets(0, 0, 5, 10)));
-    }
-}
diff --git a/crates/nvisy-core/src/modality/text.rs b/crates/nvisy-core/src/modality/text.rs
deleted file mode 100644
index 9a86a0c7a..000000000
--- a/crates/nvisy-core/src/modality/text.rs
+++ /dev/null
@@ -1,259 +0,0 @@
-//! [`Text`] modality marker, [`TextLocation`] coordinate type,
-//! [`TextData`] per-call payload, and [`TextExtraction`] provenance
-//! enum.
-
-use std::cmp::Ordering;
-use std::ops::Range;
-
-use derive_more::{AsRef, Deref, Display, From};
-use hipstr::HipStr;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::{Modality, Overlap};
-use crate::entity::ModelProvenance;
-use crate::redaction::TextReplacement;
-
-/// Text modality marker (zero-sized).
-///
-/// Used as the type parameter on generic carriers (`Entity<Text>`,
-/// `Hint<Text>`, `RecognizerInput<Text>`, …). The per-call payload
-/// (a [`TextLocation`]) is stored as `M::Location` on those carriers.
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Text;
-
-impl Modality for Text {
-    type Data = TextData;
-    type Extraction = TextExtraction;
-    type Location = TextLocation;
-    type Replacement = TextReplacement;
-
-    const KIND: super::ModalityKind = super::ModalityKind::Text;
-    const NAME: &'static str = "text";
-}
-
-/// Half-open `[start, end)` byte range around a [`TextLocation`],
-/// used for the optional surrounding context window. The newtype
-/// makes the "both endpoints or none" invariant unrepresentable —
-/// the previous twin-`Option` fields allowed a `(Some, None)`
-/// half-state with no meaningful semantics.
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct ContextWindow {
-    /// Byte offset where the context window starts.
-    pub start: usize,
-    /// Byte offset where the context window ends (exclusive).
-    pub end: usize,
-}
-
-impl ContextWindow {
-    /// Construct a window covering `start..end`.
-    pub fn new(start: usize, end: usize) -> Self {
-        Self { start, end }
-    }
-}
-
-/// A range within text content.
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct TextLocation {
-    /// Byte or character offset where the range starts.
-    pub start: usize,
-    /// Byte or character offset where the range ends.
-    pub end: usize,
-    /// Surrounding context window for redaction, when known.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub context: Option<ContextWindow>,
-    /// 1-based page number.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub page_number: Option<u32>,
-}
-
-impl TextLocation {
-    /// Create a [`TextLocation`] covering `start..end` with all
-    /// optional fields unset.
-    pub fn new(start: usize, end: usize) -> Self {
-        Self {
-            start,
-            end,
-            context: None,
-            page_number: None,
-        }
-    }
-
-    /// Byte length of the range (`end - start`).
-    pub fn len(&self) -> usize {
-        self.end.saturating_sub(self.start)
-    }
-
-    /// Whether the range is empty (zero length).
-    pub fn is_empty(&self) -> bool {
-        self.len() == 0
-    }
-
-    /// Translate a value-local byte range to a parent-local
-    /// [`TextLocation`], assuming the value is a verbatim slice of
-    /// the source covered by `self` (no escapes, no decoding).
-    ///
-    /// Returns `None` when `value_range.end > self.len()` or the
-    /// range is malformed (`start > end`).
-    ///
-    /// Used by [`Handler::lift_chunk`] implementations whose chunk
-    /// data and source bytes coincide — TXT lines, HTML text
-    /// nodes, PDF page text, DOCX text runs, etc.
-    ///
-    /// [`Handler::lift_chunk`]: # "see nvisy-codec"
-    #[must_use]
-    pub fn subslice(&self, value_range: Range<usize>) -> Option<TextLocation> {
-        if value_range.start > value_range.end || value_range.end > self.len() {
-            return None;
-        }
-        Some(TextLocation {
-            start: self.start + value_range.start,
-            end: self.start + value_range.end,
-            context: self.context,
-            page_number: self.page_number,
-        })
-    }
-}
-
-impl Ord for TextLocation {
-    /// Lex order over `(start, end)`. `context` and `page_number`
-    /// are ignored.
-    fn cmp(&self, other: &Self) -> Ordering {
-        (self.start, self.end).cmp(&(other.start, other.end))
-    }
-}
-
-impl PartialOrd for TextLocation {
-    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
-        Some(self.cmp(other))
-    }
-}
-
-/// How a [`Document<Text>`]'s text content was produced.
-///
-/// [`Document<Text>`]: # "carrier owned by nvisy-engine"
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-#[non_exhaustive]
-pub enum TextExtraction {
-    /// Structural parse of a text-bearing format: PDF text layer,
-    /// DOCX XML runs, HTML, plain UTF-8.
-    Native,
-    /// Text obtained by OCR'ing an image-backed page (image-only PDF,
-    /// scanned document).
-    Recognized(ModelProvenance),
-}
-
-/// Per-call payload for [`Text`] recognizers, extractors, and codec
-/// chunk reads.
-///
-/// Held as a [`HipStr<'static>`] so cheap clones (atomic refcount
-/// for non-inline text, inline copy for short strings) let the
-/// caller share one payload across multiple recognizers without
-/// duplicating the source bytes.
-///
-/// Shared per-call enrichment (lemmatized tokens, language
-/// detections, stopword sets) lives on the surrounding
-/// [`Span<Text>`]'s [`Artifacts`] bundle, not on [`TextData`] —
-/// the same typed-map is reused for every modality and every
-/// recognizer/extractor stage.
-///
-/// [`Span<Text>`]: crate::extraction::Span
-/// [`Artifacts`]: crate::extraction::Artifacts
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-#[derive(Display, From, AsRef, Deref)]
-#[as_ref(forward)]
-#[display("{text}")]
-pub struct TextData {
-    /// The text the recognizer should scan. Byte offsets in emitted
-    /// entities refer back into this string.
-    #[deref]
-    pub text: HipStr<'static>,
-}
-
-impl TextData {
-    /// Construct from anything convertible to [`HipStr<'static>`] —
-    /// owned `String`, borrowed `&'static str`, an existing
-    /// `HipStr`, …
-    pub fn new(text: impl Into<HipStr<'static>>) -> Self {
-        Self { text: text.into() }
-    }
-
-    /// View the inner string slice.
-    pub fn as_str(&self) -> &str {
-        self.text.as_str()
-    }
-
-    /// Consume the wrapper and return the content as a [`String`],
-    /// allocating only when the underlying [`HipStr`] is borrowed.
-    pub fn into_string(self) -> String {
-        self.text.into()
-    }
-}
-
-impl From<&str> for TextData {
-    fn from(s: &str) -> Self {
-        Self::new(HipStr::from(s))
-    }
-}
-
-impl From<String> for TextData {
-    fn from(s: String) -> Self {
-        Self::new(HipStr::from(s))
-    }
-}
-
-impl PartialEq<&str> for TextData {
-    fn eq(&self, other: &&str) -> bool {
-        self.text.as_str() == *other
-    }
-}
-
-impl Overlap for TextLocation {
-    /// Two text ranges overlap only when they share a page (or both
-    /// have `page_number: None`) and their byte ranges intersect.
-    /// Without the page gate, two ranges on different pages of the
-    /// same document that happen to share byte offsets would
-    /// false-positive as overlapping.
-    fn overlaps(&self, other: &Self) -> bool {
-        self.page_number == other.page_number && self.start < other.end && other.start < self.end
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn len_and_is_empty() {
-        assert_eq!(TextLocation::new(0, 10).len(), 10);
-        assert!(!TextLocation::new(0, 10).is_empty());
-        assert!(TextLocation::new(5, 5).is_empty());
-    }
-
-    #[test]
-    fn overlap_intersecting() {
-        assert!(TextLocation::new(0, 10).overlaps(&TextLocation::new(5, 15)));
-    }
-
-    #[test]
-    fn overlap_contained() {
-        assert!(TextLocation::new(0, 10).overlaps(&TextLocation::new(2, 5)));
-    }
-
-    #[test]
-    fn no_overlap_adjacent() {
-        assert!(!TextLocation::new(0, 5).overlaps(&TextLocation::new(5, 10)));
-    }
-
-    #[test]
-    fn no_overlap_disjoint() {
-        assert!(!TextLocation::new(0, 5).overlaps(&TextLocation::new(10, 15)));
-    }
-}
diff --git a/crates/nvisy-core/src/plan/analyzer.rs b/crates/nvisy-core/src/plan/analyzer.rs
new file mode 100644
index 000000000..d6df38d8f
--- /dev/null
+++ b/crates/nvisy-core/src/plan/analyzer.rs
@@ -0,0 +1,45 @@
+//! Top-level analyzer params.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use super::deduplication::DeduplicationParams;
+use super::enricher::EnricherParams;
+use super::recognizer::RecognizerParams;
+use super::scope::ScopeParams;
+use crate::schema::LabelSchema;
+
+/// Full description of how to build an analyzer for one
+/// request.
+///
+/// Engine compiles this into [`elide::detection::Analyzer<M>`] per
+/// modality the request targets. The compile step
+/// monomorphises each per-modality analyzer from the same
+/// params — text picks up the text-applicable recognizers,
+/// image picks up the image ones, etc.
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct AnalyzerParams {
+    /// Recognizer slots: pattern (at-most-one), ner, llm
+    /// (each a list, identified by name).
+    #[serde(default)]
+    pub recognizers: RecognizerParams,
+    /// Enricher slots: language, ocr, stt (each at-most-one).
+    /// Enrichers run sequentially before recognition; the
+    /// engine picks a canonical order (language → ocr → stt).
+    #[serde(default)]
+    pub enrichers: EnricherParams,
+    /// Deduplication pipeline applied after recognition:
+    /// calibrate, fuse, resolve, filter. Defaults assemble the
+    /// canonical pipeline at compile time.
+    #[serde(default)]
+    pub deduplication: DeduplicationParams,
+    /// Caller-asserted scope (languages, jurisdictions).
+    /// Threaded into every recognizer's context.
+    #[serde(default)]
+    pub scope: ScopeParams,
+    /// Per-request entity-label catalog. Drives tag-based
+    /// selector matching downstream in the anonymizer compile.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub label_catalog: Vec<LabelSchema>,
+}
diff --git a/crates/nvisy-core/src/plan/deduplication.rs b/crates/nvisy-core/src/plan/deduplication.rs
new file mode 100644
index 000000000..0d799362e
--- /dev/null
+++ b/crates/nvisy-core/src/plan/deduplication.rs
@@ -0,0 +1,99 @@
+//! Dedup-pipeline specs.
+//!
+//! Mirrors elide's [`Calibrate → Fuse → Resolve → Filter`] layers.
+//! Each component is parameterised by a serialisable strategy enum;
+//! engine maps each spec variant to the matching elide strategy at
+//! compile time.
+
+use std::collections::HashMap;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Dedup pipeline applied after recognition. Layers run in the
+/// canonical order: calibrate → fuse → resolve → filter.
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct DeduplicationParams {
+    /// Per-recognizer confidence weights. An empty map skips the
+    /// calibrate layer.
+    #[serde(default, skip_serializing_if = "CalibrationMap::is_empty")]
+    pub calibration: CalibrationMap,
+    /// How overlapping entities are fused into one. Defaults to
+    /// [`FusionStrategyParams::MaxConfidence`].
+    #[serde(default)]
+    pub fusion: FusionStrategyParams,
+    /// How an entity's identity is resolved when several remain
+    /// after fusion. Defaults to
+    /// [`ResolutionStrategyParams::HighestConfidence`].
+    #[serde(default)]
+    pub resolution: ResolutionStrategyParams,
+    /// Minimum confidence the filter layer admits. `None` falls
+    /// back to elide's `ConfidenceThreshold::BASELINE` at compile
+    /// time.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub min_confidence: Option<f32>,
+}
+
+/// Per-recognizer calibration weights (recognizer-name → multiplier).
+///
+/// Wire shape: `{ "pattern": 1.0, "ner": 0.85 }`. Engine builds the
+/// elide [`CalibrationMap`] from this; absent recognizers default to
+/// `1.0` (identity).
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(transparent)]
+pub struct CalibrationMap(pub HashMap<String, f64>);
+
+impl CalibrationMap {
+    /// `true` when no weights are set; engine skips the calibrate
+    /// layer in that case.
+    pub fn is_empty(&self) -> bool {
+        self.0.is_empty()
+    }
+}
+
+/// Strategy the fuse layer uses to combine overlapping entities.
+#[derive(
+    Debug,
+    Clone,
+    Copy,
+    Default,
+    PartialEq,
+    Eq,
+    Serialize,
+    Deserialize,
+    JsonSchema
+)]
+#[serde(rename_all = "snake_case")]
+pub enum FusionStrategyParams {
+    /// Keep the entity with the highest confidence.
+    #[default]
+    MaxConfidence,
+    /// Arithmetic mean of overlapping confidences.
+    Mean,
+    /// Probabilistic noisy-or combination — treats recognizers as
+    /// independent evidence sources.
+    NoisyOr,
+}
+
+/// Strategy the resolve layer uses to break ties when several
+/// entities cover the same span.
+#[derive(
+    Debug,
+    Clone,
+    Copy,
+    Default,
+    PartialEq,
+    Eq,
+    Serialize,
+    Deserialize,
+    JsonSchema
+)]
+#[serde(rename_all = "snake_case")]
+pub enum ResolutionStrategyParams {
+    /// Keep the highest-confidence entity.
+    #[default]
+    HighestConfidence,
+    /// Keep the entity covering the longest span.
+    LongestSpan,
+}
diff --git a/crates/nvisy-core/src/plan/enricher.rs b/crates/nvisy-core/src/plan/enricher.rs
new file mode 100644
index 000000000..b33355d7c
--- /dev/null
+++ b/crates/nvisy-core/src/plan/enricher.rs
@@ -0,0 +1,107 @@
+//! Enricher params: per-kind slots inside an
+//! [`AnalyzerParams`].
+//!
+//! Three enricher kinds — language detection, OCR, STT — each
+//! at-most-one per analyzer. Enrichers run sequentially before
+//! recognition; each writes to the per-call working context so
+//! recognizers downstream see what it wrote.
+//!
+//! [`AnalyzerParams`]: super::AnalyzerParams
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Enricher slots an analyzer can fill. Each slot is
+/// at-most-one; the slot name is the kind.
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct EnricherParams {
+    /// Detect the document's primary language(s) and write them
+    /// into the recognizer context. Drives jurisdiction-aware
+    /// recognizer dispatch downstream.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub language: Option<LanguageEnricherParams>,
+    /// OCR the image (or PDF page raster) and stamp the
+    /// recognised [`Layout`] onto the recognizer context, so
+    /// downstream text recognizers can match on the OCR'd text.
+    ///
+    /// Image modality only.
+    ///
+    /// [`Layout`]: elide_core::modality::image::Layout
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub ocr: Option<OcrEnricherParams>,
+    /// Speech-to-text: transcribe the audio stream and stamp
+    /// the resulting [`TranscriptSegment`]s onto the recognizer
+    /// context, so downstream text recognizers can match
+    /// against the transcript.
+    ///
+    /// Audio modality only.
+    ///
+    /// [`TranscriptSegment`]: elide_core::modality::audio::TranscriptSegment
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub stt: Option<SttEnricherParams>,
+}
+
+/// Params for the language-detection enricher.
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct LanguageEnricherParams {
+    /// Minimum confidence in `[0.0, 1.0]` the detector must
+    /// report before a language is asserted into the context.
+    /// Lower values write more languages; higher values are
+    /// stricter. `None` lets the engine pick the default.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub min_confidence: Option<f32>,
+}
+
+/// Params for the OCR enricher (image modality).
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct OcrEnricherParams {
+    /// OCR backend choice.
+    pub backend: OcrBackendParams,
+}
+
+/// How to instantiate the OCR backend.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+pub enum OcrBackendParams {
+    /// No-op backend; recognises no blocks. For tests, offline
+    /// wiring, or skeleton runs.
+    Mock,
+    /// BentoML-hosted OCR service. Engine wires the shared
+    /// `elide-bento` client; per-request URL + model come from
+    /// this variant.
+    Bento {
+        /// Base URL of the BentoML service.
+        base_url: String,
+        /// Model identifier the backend should target.
+        model: String,
+    },
+}
+
+/// Params for the STT enricher (audio modality).
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct SttEnricherParams {
+    /// STT backend choice.
+    pub backend: SttBackendParams,
+}
+
+/// How to instantiate the STT backend.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+pub enum SttBackendParams {
+    /// No-op backend; emits no transcript segments. For tests
+    /// and skeleton runs.
+    Mock,
+    /// BentoML-hosted STT service. Per-request URL + model come
+    /// from this variant. Engine wiring lands when
+    /// `elide-bento` ships a `BentoStt` client.
+    Bento {
+        /// Base URL of the BentoML service.
+        base_url: String,
+        /// Model identifier the backend should target.
+        model: String,
+    },
+}
diff --git a/crates/nvisy-core/src/plan/mod.rs b/crates/nvisy-core/src/plan/mod.rs
new file mode 100644
index 000000000..c55332087
--- /dev/null
+++ b/crates/nvisy-core/src/plan/mod.rs
@@ -0,0 +1,41 @@
+//! Analyzer plan: serialisable description of how to build an
+//! [`elide::detection::Analyzer`] for a request.
+//!
+//! Symmetric with [`crate::policy`]: where Policy describes
+//! redaction governance (which entities to hide and how), the plan
+//! describes recognition (which entities to find and how). Both are
+//! pure data — engine compiles them into elide runtime values at
+//! request time.
+//!
+//! ## Layout
+//!
+//! - [`AnalyzerParams`] — top-level: recognizers, enrichers, dedup
+//!   pipeline, scope, label catalog.
+//! - [`recognizer`] — per-recognizer specs (Pattern, NER, LLM).
+//! - [`enricher`] — per-enricher specs (language detection).
+//! - [`deduplication`] — fuse / resolve / filter / calibration
+//!   strategies.
+//! - [`scope`] — language + jurisdiction assertions the caller
+//!   passes into recognition.
+//!
+//! [`elide::detection::Analyzer`]: https://docs.rs/elide/latest/elide/struct.Analyzer.html
+
+mod analyzer;
+pub mod deduplication;
+pub mod enricher;
+pub mod recognizer;
+pub mod scope;
+
+pub use self::analyzer::AnalyzerParams;
+pub use self::deduplication::{
+    CalibrationMap, DeduplicationParams, FusionStrategyParams, ResolutionStrategyParams,
+};
+pub use self::enricher::{
+    EnricherParams, LanguageEnricherParams, OcrBackendParams, OcrEnricherParams, SttBackendParams,
+    SttEnricherParams,
+};
+pub use self::recognizer::{
+    LlmBackendParams, LlmRecognizerParams, NerBackendParams, NerRecognizerParams,
+    PatternRecognizerParams, RecognizerParams,
+};
+pub use self::scope::ScopeParams;
diff --git a/crates/nvisy-core/src/plan/recognizer.rs b/crates/nvisy-core/src/plan/recognizer.rs
new file mode 100644
index 000000000..27ca941b0
--- /dev/null
+++ b/crates/nvisy-core/src/plan/recognizer.rs
@@ -0,0 +1,142 @@
+//! Recognizer params: per-kind slots inside an
+//! [`AnalyzerParams`].
+//!
+//! Three recognizer kinds — pattern, NER, LLM — each with its
+//! own slot on [`RecognizerParams`]:
+//!
+//! - **Pattern** is at-most-one. Single regex/dictionary engine
+//!   per analyzer; multi-pattern means accumulating into one
+//!   instance's pattern list, not running two engines.
+//! - **NER** is a list. Multiple NER backends (different model
+//!   versions, language-specialised models, ensemble lanes) run
+//!   in parallel; each is identified by [`name`] for
+//!   provenance.
+//! - **LLM** is a list. Multiple providers (OpenAI, Anthropic,
+//!   …) or per-label-class specialist models, identified by
+//!   [`name`].
+//!
+//! [`AnalyzerParams`]: super::AnalyzerParams
+//! [`name`]: NerRecognizerParams::name
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Recognizer slots an analyzer can fill. The slot name is the
+/// kind; pattern is at-most-one, NER and LLM accept many
+/// (identified by `name`).
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct RecognizerParams {
+    /// Built-in pattern + dictionary recognizer (`elide-pattern`).
+    /// At most one per analyzer.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub pattern: Option<PatternRecognizerParams>,
+    /// External-backend NER recognizers (`elide-ner`). Each
+    /// must have a unique `name`; runs in parallel with the
+    /// rest.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub ner: Vec<NerRecognizerParams>,
+    /// External-backend LLM recognizers (`elide-llm`). Each
+    /// must have a unique `name`.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub llm: Vec<LlmRecognizerParams>,
+}
+
+/// Params for the `elide-pattern` recognizer.
+#[derive(
+    Debug,
+    Clone,
+    Default,
+    PartialEq,
+    Eq,
+    Serialize,
+    Deserialize,
+    JsonSchema
+)]
+#[serde(rename_all = "camelCase")]
+pub struct PatternRecognizerParams {
+    /// Load every pattern + dictionary shipped with
+    /// `elide-pattern`. Implies the country-scoped
+    /// jurisdictional pattern packs are active for the scope's
+    /// jurisdictions.
+    #[serde(default)]
+    pub builtins: bool,
+    /// Wrap the bare pattern recognizer in elide's
+    /// `Enhanced<PatternRecognizer>` layer so per-label context
+    /// keywords boost low-confidence matches before they leave
+    /// the recognizer.
+    #[serde(default)]
+    pub context_enhanced: bool,
+}
+
+/// Params for one `elide-ner` recognizer instance.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct NerRecognizerParams {
+    /// Recognizer name — surfaces on the per-entity provenance
+    /// trail so audits can tell which NER instance fired. Must
+    /// be unique within the parent
+    /// [`RecognizerParams::ner`](super::RecognizerParams::ner)
+    /// list.
+    pub name: String,
+    /// Backend instantiation choice.
+    pub backend: NerBackendParams,
+}
+
+/// How to instantiate the NER backend.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+pub enum NerBackendParams {
+    /// No-op backend; emits no entities. For tests, offline
+    /// wiring, or skeleton runs.
+    Mock,
+    /// BentoML-hosted NER service. Engine wires the shared
+    /// `elide-bento` client; per-request URL + model come from
+    /// this variant.
+    Bento {
+        /// Base URL of the BentoML service.
+        base_url: String,
+        /// Model identifier the backend should target.
+        model: String,
+    },
+}
+
+/// Params for one `elide-llm` recognizer instance.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct LlmRecognizerParams {
+    /// Recognizer name — surfaces on the per-entity provenance
+    /// trail. Must be unique within the parent
+    /// [`RecognizerParams::llm`](super::RecognizerParams::llm)
+    /// list.
+    pub name: String,
+    /// Backend (provider) choice.
+    pub backend: LlmBackendParams,
+    /// Optional custom prompt template. `None` uses elide's
+    /// default recognition prompt.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub prompt: Option<String>,
+}
+
+/// How to instantiate the LLM backend.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+pub enum LlmBackendParams {
+    /// No-op backend; emits no entities.
+    Mock,
+    /// OpenAI GPT provider.
+    Openai {
+        /// Model identifier (e.g. `"gpt-4o-mini"`).
+        model: String,
+    },
+    /// Anthropic Claude provider.
+    Anthropic {
+        /// Model identifier (e.g. `"claude-3-5-sonnet-20241022"`).
+        model: String,
+    },
+    /// Google Gemini provider.
+    Google {
+        /// Model identifier (e.g. `"gemini-1.5-flash"`).
+        model: String,
+    },
+}
diff --git a/crates/nvisy-core/src/plan/scope.rs b/crates/nvisy-core/src/plan/scope.rs
new file mode 100644
index 000000000..72665ef79
--- /dev/null
+++ b/crates/nvisy-core/src/plan/scope.rs
@@ -0,0 +1,27 @@
+//! Caller-asserted scope: languages + jurisdictions the recognizer
+//! context carries.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::schema::LanguageTagSchema;
+
+/// Per-request scope assertions. Engine threads these into every
+/// recognizer's [`elide::recognition::Scope`].
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct ScopeParams {
+    /// BCP 47 language tags the caller asserts (or has detected
+    /// upstream). Each becomes an asserted [`Language`] in the
+    /// recognizer context.
+    ///
+    /// [`Language`]: elide_core::primitive::Language
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub languages: Vec<LanguageTagSchema>,
+    /// ISO 3166-1 alpha-2 country codes the caller asserts as
+    /// applicable jurisdictions. Drives jurisdiction-scoped pattern
+    /// packs in elide-pattern.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub jurisdictions: Vec<String>,
+}
diff --git a/crates/nvisy-engine/src/policy/audit.rs b/crates/nvisy-core/src/policy/audit.rs
similarity index 64%
rename from crates/nvisy-engine/src/policy/audit.rs
rename to crates/nvisy-core/src/policy/audit.rs
index 85264fa09..2357cfeb2 100644
--- a/crates/nvisy-engine/src/policy/audit.rs
+++ b/crates/nvisy-core/src/policy/audit.rs
@@ -1,13 +1,5 @@
 //! [`AuditAction`]: payload of an [`Action::Audit`] rule.
 //!
-//! Flags entities that need extra human attention without
-//! transforming them. The detection pass already produces an
-//! audit entry per matched entity; this action tags those entries
-//! with a severity hint so downstream review tooling can prioritise
-//! them. Designed as a struct (not a unit type) so future fields
-//! (`category`, `notify`, `reviewer`, …) can land without a
-//! wire-break.
-//!
 //! [`Action::Audit`]: super::rule::Action::Audit
 
 use hipstr::HipStr;
diff --git a/crates/nvisy-core/src/policy/document.rs b/crates/nvisy-core/src/policy/document.rs
new file mode 100644
index 000000000..f5d1cda5b
--- /dev/null
+++ b/crates/nvisy-core/src/policy/document.rs
@@ -0,0 +1,49 @@
+//! Document-level predicate that gates whether a whole policy
+//! applies to a given document.
+//!
+//! Evaluated once per document (cf. [`Predicate`], which is
+//! evaluated per entity). Engine merges the document's content
+//! descriptor with the caller's per-request metadata, then checks
+//! the predicate against that union.
+//!
+//! [`Predicate`]: super::Predicate
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Predicate over document-level facts.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "camelCase")]
+pub enum DocumentPredicate {
+    /// Document carries `label` (e.g. `"contract"`, `"medical"`).
+    /// Labels come from the content descriptor; the importer
+    /// populates them at ingest time.
+    HasLabel {
+        /// Required document label.
+        label: String,
+    },
+    /// Document metadata has `key` set, optionally to `value`. When
+    /// `value` is `None` the key just has to be present.
+    HasMetadata {
+        /// Metadata key to read.
+        key: String,
+        /// Required value, or `None` to match on presence alone.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        value: Option<String>,
+    },
+    /// All sub-predicates must hold (AND).
+    All {
+        /// Conjunction members.
+        all: Vec<DocumentPredicate>,
+    },
+    /// At least one sub-predicate must hold (OR).
+    Any {
+        /// Disjunction members.
+        any: Vec<DocumentPredicate>,
+    },
+    /// Sub-predicate must not hold (NOT).
+    Not {
+        /// Negated predicate.
+        not: Box<DocumentPredicate>,
+    },
+}
diff --git a/crates/nvisy-core/src/policy/mod.rs b/crates/nvisy-core/src/policy/mod.rs
new file mode 100644
index 000000000..d2b97b2a4
--- /dev/null
+++ b/crates/nvisy-core/src/policy/mod.rs
@@ -0,0 +1,93 @@
+//! Policy types: authored vocabulary for redaction governance.
+//!
+//! A request submits `Vec<Policy>` in precedence order. Engine
+//! walks them; for each policy whose [`Policy::applies_when`]
+//! holds against the document, it walks [`Policy::rules`] in order
+//! and runs the first matching rule's [`RuleAction`]. If no rule
+//! in a policy matches, the policy's [`Policy::fallback`] runs
+//! (and the chain halts) if set; otherwise the engine moves to
+//! the next policy. If no policy matches and no policy carries a
+//! fallback, the entity is skipped.
+//!
+//! Identity is UUID-keyed: every [`Policy`] and every [`Rule`]
+//! carries a stable [`Uuid`]. Engine stamps `policy.id` and
+//! `rule.id` into the redaction event's [`Attribution`] so
+//! reviewers can trace any redaction back to the exact rule that
+//! fired.
+//!
+//! [`Attribution`]: elide_core::entity::provenance::Attribution
+
+mod audit;
+mod document;
+mod predicate;
+pub mod redaction;
+mod retention;
+mod rule;
+mod suppress;
+
+use hipstr::HipStr;
+use schemars::JsonSchema;
+use semver::Version;
+use serde::{Deserialize, Serialize};
+use uuid::Uuid;
+
+pub use self::audit::AuditAction;
+pub use self::document::DocumentPredicate;
+pub use self::predicate::Predicate;
+pub use self::redaction::AnyRedaction;
+pub use self::retention::{Retention, RetentionPolicy, RetentionScope};
+pub use self::rule::{Rule, RuleAction};
+pub use self::suppress::SuppressAction;
+use crate::schema::LabelSchema;
+
+/// A named, versioned governance policy.
+///
+/// Identity is the UUID; `name` is display-only. `version` is the
+/// policy body's semver — two submissions of the same
+/// `(id, version)` pair should produce identical decisions.
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct Policy {
+    /// Stable identifier. UUIDv7 recommended (time-ordered);
+    /// customer-supplied so re-submissions carry the same id.
+    /// Engine stamps this into the redaction event's
+    /// [`Attribution::policy_id`] so reviewers can find this
+    /// policy from any redaction it drove.
+    ///
+    /// [`Attribution::policy_id`]: elide_core::entity::provenance::Attribution::policy_id
+    pub id: Uuid,
+    /// Human-readable name. Display-only — does not key anything.
+    #[schemars(with = "String")]
+    pub name: HipStr<'static>,
+    /// Semver of the policy body.
+    #[schemars(with = "String")]
+    pub version: Version,
+    /// Optional description for reviewers.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub description: Option<String>,
+    /// Document-level gate. The whole policy (rules + fallback)
+    /// is skipped when this is `Some(...)` and the predicate is
+    /// false for the document. Evaluated once per document.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub applies_when: Option<DocumentPredicate>,
+    /// Vocabulary the policy operates over. Engine unions every
+    /// submitted policy's `labels` into a per-request
+    /// [`LabelCatalog`] used to drive recognizer dispatch and
+    /// tag-based [`Predicate::TagOneOf`] matching.
+    ///
+    /// [`LabelCatalog`]: elide_core::entity::LabelCatalog
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub labels: Vec<LabelSchema>,
+    /// Ordered rules. First match wins within this policy.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub rules: Vec<Rule>,
+    /// Per-policy catch-all. Fires when no rule in this policy
+    /// matched. Presence halts the chain; absence falls through
+    /// to the next policy. [`Option`] enforces "at most one
+    /// fallback per policy" at the type level.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub fallback: Option<RuleAction>,
+    /// Lifecycle rules for content under this policy.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub retention: Vec<RetentionPolicy>,
+}
diff --git a/crates/nvisy-core/src/policy/predicate.rs b/crates/nvisy-core/src/policy/predicate.rs
new file mode 100644
index 000000000..e8f359f38
--- /dev/null
+++ b/crates/nvisy-core/src/policy/predicate.rs
@@ -0,0 +1,60 @@
+//! Composable, entity-level predicates for [`Rule::predicate`].
+//!
+//! Each variant is serialisable. Engine compiles a `Predicate` into
+//! a closure passed to [`elide::redaction::Anonymizer::with_catalog_predicate`]
+//! (or routed to [`with_label`] / [`with_tag`] fast paths for the
+//! degenerate single-label / single-tag shapes). Leaf variants
+//! inspect entity facts (label, tag, confidence, coref); the
+//! composing variants ([`All`], [`Any`], [`Not`]) wire boolean
+//! algebra over them.
+//!
+//! [`Rule::predicate`]: super::Rule::predicate
+//! [`with_label`]: elide::redaction::Anonymizer::with_label
+//! [`with_tag`]: elide::redaction::Anonymizer::with_tag
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Predicate over a recognised entity. The wire format uses an
+/// internally tagged enum so authors write
+/// `{ "kind": "confidence", "min": 0.7 }` etc.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "camelCase")]
+pub enum Predicate {
+    /// Entity confidence is at least `min`.
+    Confidence {
+        /// Minimum confidence in `[0.0, 1.0]`.
+        min: f32,
+    },
+    /// Entity label is one of `labels`.
+    LabelOneOf {
+        /// Allowed labels.
+        labels: Vec<String>,
+    },
+    /// Entity label carries one of `tags`, per the per-request
+    /// label catalog.
+    TagOneOf {
+        /// Allowed tags.
+        tags: Vec<String>,
+    },
+    /// Entity carries the given coreference cluster id.
+    CoRef {
+        /// Cluster id to match.
+        coref: String,
+    },
+    /// All sub-predicates must hold (AND).
+    All {
+        /// Conjunction members.
+        all: Vec<Predicate>,
+    },
+    /// At least one sub-predicate must hold (OR).
+    Any {
+        /// Disjunction members.
+        any: Vec<Predicate>,
+    },
+    /// Sub-predicate must not hold (NOT).
+    Not {
+        /// Negated predicate.
+        not: Box<Predicate>,
+    },
+}
diff --git a/crates/nvisy-core/src/policy/redaction/any.rs b/crates/nvisy-core/src/policy/redaction/any.rs
new file mode 100644
index 000000000..b279da558
--- /dev/null
+++ b/crates/nvisy-core/src/policy/redaction/any.rs
@@ -0,0 +1,69 @@
+//! [`AnyRedaction`]: type-erased redaction operator spec.
+//!
+//! Carries the modality tag and the typed per-modality `*Redaction`
+//! enum that a policy author would pick. Used at API boundaries
+//! where the caller has not pinned an `M: Modality` type yet — most
+//! notably the redaction-override surface (`Replace`, `Add`).
+//!
+//! Wire shape:
+//!
+//! ```json
+//! { "modality": "text", "kind": "mask", ... }
+//! ```
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use strum::EnumTryAs;
+
+use super::audio::AudioRedaction;
+use super::image::ImageRedaction;
+use super::tabular::TabularRedaction;
+use super::text::TextRedaction;
+
+/// Which modality an [`AnyRedaction`] applies to.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum RedactionModality {
+    /// Text-modality entities (txt, json, html, xml, …).
+    Text,
+    /// Tabular-modality entities (csv, xlsx, …).
+    Tabular,
+    /// Image-modality entities (png, jpeg, tiff, …).
+    Image,
+    /// Audio-modality entities (wav, mp3, …).
+    Audio,
+}
+
+/// Type-erased redaction operator spec.
+///
+/// Use the typed [`TextRedaction`] / [`ImageRedaction`] /
+/// [`AudioRedaction`] / [`TabularRedaction`] enums directly inside
+/// policy rules where the modality is known at compile time; reach
+/// for [`AnyRedaction`] only at the dynamic API boundary.
+#[derive(Debug, Clone, PartialEq, EnumTryAs)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "modality", rename_all = "snake_case")]
+pub enum AnyRedaction {
+    /// Text-modality redaction operator.
+    Text(TextRedaction),
+    /// Tabular-modality redaction operator.
+    Tabular(TabularRedaction),
+    /// Image-modality redaction operator.
+    Image(ImageRedaction),
+    /// Audio-modality redaction operator.
+    Audio(AudioRedaction),
+}
+
+impl AnyRedaction {
+    /// The modality this redaction operator belongs to.
+    #[must_use]
+    pub fn modality(&self) -> RedactionModality {
+        match self {
+            Self::Text(_) => RedactionModality::Text,
+            Self::Tabular(_) => RedactionModality::Tabular,
+            Self::Image(_) => RedactionModality::Image,
+            Self::Audio(_) => RedactionModality::Audio,
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/policy/redaction/audio.rs b/crates/nvisy-core/src/policy/redaction/audio.rs
new file mode 100644
index 000000000..a192da067
--- /dev/null
+++ b/crates/nvisy-core/src/policy/redaction/audio.rs
@@ -0,0 +1,58 @@
+//! [`AudioRedaction`]: the operator spec an audio-modality policy
+//! rule carries.
+//!
+//! Each variant mirrors an elide built-in operator the engine
+//! constructs at apply time:
+//!
+//! - [`AudioRedaction::Erase`] → [`elide::redaction::operators::Erase`]
+//!   cuts the interval out, shortening the clip.
+//! - [`AudioRedaction::Keep`] → [`elide::redaction::operators::Keep`]
+//!   passes the interval through unchanged.
+//! - [`AudioRedaction::Silence`] →
+//!   [`elide::redaction::operators::Silence`] zeroes the interval in
+//!   place, preserving the clip's length and downstream timing.
+//! - [`AudioRedaction::Beep`] → [`elide::redaction::operators::Beep`]
+//!   overlays a tone over the interval — the broadcast-bleep
+//!   treatment.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::schema::WaveformSchema;
+
+/// Operator spec a `redact` audio rule carries.
+#[derive(Debug, Clone, Copy, PartialEq)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+pub enum AudioRedaction {
+    /// Cut the matched interval out, shortening the clip.
+    Erase,
+    /// Pass the interval through unchanged.
+    Keep,
+    /// Zero the matched interval in place, preserving the clip
+    /// length and the timing of everything after it.
+    Silence,
+    /// Overlay a tone over the matched interval (the broadcast
+    /// "bleep"); preserves duration.
+    Beep {
+        /// Tone frequency in Hertz. Default 1 kHz, the broadcast
+        /// convention.
+        #[serde(default = "default_beep_hz")]
+        hz: f32,
+        /// Peak amplitude in `0.0..=1.0` of full scale. Default 0.5
+        /// — audible but not painful, and never clips.
+        #[serde(default = "default_beep_amplitude")]
+        amplitude: f32,
+        /// Tone shape. Default sine.
+        #[serde(default)]
+        waveform: WaveformSchema,
+    },
+}
+
+fn default_beep_hz() -> f32 {
+    1000.0
+}
+
+fn default_beep_amplitude() -> f32 {
+    0.5
+}
diff --git a/crates/nvisy-core/src/policy/redaction/image.rs b/crates/nvisy-core/src/policy/redaction/image.rs
new file mode 100644
index 000000000..f83b4c03d
--- /dev/null
+++ b/crates/nvisy-core/src/policy/redaction/image.rs
@@ -0,0 +1,61 @@
+//! [`ImageRedaction`]: the operator spec an image-modality policy
+//! rule carries.
+//!
+//! Each variant mirrors an elide built-in operator the engine
+//! constructs at apply time:
+//!
+//! - [`ImageRedaction::Erase`] → [`elide::redaction::operators::Erase`]
+//! - [`ImageRedaction::Keep`] → [`elide::redaction::operators::Keep`]
+//! - [`ImageRedaction::Blur`] → [`elide::redaction::operators::Blur`]
+//! - [`ImageRedaction::Pixelate`] →
+//!   [`elide::redaction::operators::Pixelate`]
+//! - [`ImageRedaction::Blackbox`] →
+//!   [`elide::redaction::operators::Blackbox`]
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use crate::schema::ColorSchema;
+
+/// Operator spec a `redact` image rule carries.
+#[derive(Debug, Clone, PartialEq)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+pub enum ImageRedaction {
+    /// Clear the matched region.
+    Erase,
+    /// Pass the region through unchanged.
+    Keep,
+    /// Gaussian-blur the region.
+    Blur {
+        /// Standard deviation of the Gaussian kernel, in pixels.
+        /// Larger is blurrier (and harder to reverse).
+        #[serde(default = "default_blur_sigma")]
+        sigma: f32,
+    },
+    /// Mosaic-pixelate the region.
+    Pixelate {
+        /// Side length of each mosaic block, in pixels. Larger
+        /// blocks are coarser (and harder to reverse).
+        #[serde(default = "default_pixelate_block")]
+        block_size: u32,
+    },
+    /// Cover the region with a solid color fill (black by default).
+    Blackbox {
+        /// Fill color the codec rasterises over the region.
+        #[serde(default = "default_blackbox_color")]
+        color: ColorSchema,
+    },
+}
+
+fn default_blur_sigma() -> f32 {
+    16.0
+}
+
+fn default_pixelate_block() -> u32 {
+    16
+}
+
+fn default_blackbox_color() -> ColorSchema {
+    ColorSchema { r: 0, g: 0, b: 0 }
+}
diff --git a/crates/nvisy-core/src/policy/redaction/mod.rs b/crates/nvisy-core/src/policy/redaction/mod.rs
new file mode 100644
index 000000000..00c397bdf
--- /dev/null
+++ b/crates/nvisy-core/src/policy/redaction/mod.rs
@@ -0,0 +1,99 @@
+//! Per-modality redaction operator specs — the closed wire vocabulary
+//! a policy author can choose from inside a `redact` rule — plus the
+//! [`ModalityRedactions`] map a rule carries to wire one operator
+//! per modality, and the [`AnyRedaction`] erasure used at the
+//! override boundary.
+//!
+//! Each modality has its own enum because the operator catalogue
+//! differs by modality. Text gets the full elide built-in set
+//! ([`Replace`], [`Mask`], [`Hash`], [`Erase`], [`Keep`]) plus a
+//! `Custom` escape hatch. Image / Audio / Tabular currently expose
+//! only `Custom` — elide ships no built-in operators wired into the
+//! policy wire format for those modalities.
+//!
+//! The split between these enums (the spec) and elide's
+//! [`Operator<M>`] trait is intentional: the spec is the
+//! serialisable, author-facing wire shape; the engine compiles each
+//! variant into the matching runtime operator instance at apply
+//! time.
+//!
+//! [`Replace`]: elide::redaction::operators::Replace
+//! [`Mask`]: elide::redaction::operators::Mask
+//! [`Hash`]: elide::redaction::operators::Sha2Hash
+//! [`Erase`]: elide::redaction::operators::Erase
+//! [`Keep`]: elide::redaction::operators::Keep
+//! [`Operator<M>`]: elide_core::operator::Operator
+
+mod any;
+mod audio;
+mod image;
+mod tabular;
+mod text;
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+pub use self::any::{AnyRedaction, RedactionModality};
+pub use self::audio::AudioRedaction;
+pub use self::image::ImageRedaction;
+pub use self::tabular::TabularRedaction;
+pub use self::text::{HashAlgorithm, TextRedaction};
+
+/// Per-modality operator specs carried by a `redact` rule.
+///
+/// A single rule can name an operator for every modality the
+/// workspace supports. At apply time the redaction phase picks the
+/// operator matching the entity's modality (engine reads the right
+/// field directly — `redactions.text.as_ref()` etc.); modalities
+/// the rule didn't cover fall through to the deployment-wide
+/// default, and entities with no operator from either source are
+/// skipped.
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub struct ModalityRedactions {
+    /// Operator for text-modality entities.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub text: Option<TextRedaction>,
+    /// Operator for tabular-modality entities.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tabular: Option<TabularRedaction>,
+    /// Operator for image-modality entities.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub image: Option<ImageRedaction>,
+    /// Operator for audio-modality entities.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub audio: Option<AudioRedaction>,
+}
+
+impl ModalityRedactions {
+    /// `true` when no operator is set for any modality. A rule whose
+    /// `redact` field is empty after deserialisation is an author
+    /// error — the request validator rejects it.
+    #[must_use]
+    pub fn is_empty(&self) -> bool {
+        self.text.is_none()
+            && self.tabular.is_none()
+            && self.image.is_none()
+            && self.audio.is_none()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn empty_when_no_modality_set() {
+        let r = ModalityRedactions::default();
+        assert!(r.is_empty());
+    }
+
+    #[test]
+    fn non_empty_when_any_modality_set() {
+        let r = ModalityRedactions {
+            text: Some(TextRedaction::Erase),
+            ..Default::default()
+        };
+        assert!(!r.is_empty());
+    }
+}
diff --git a/crates/nvisy-core/src/policy/redaction/tabular.rs b/crates/nvisy-core/src/policy/redaction/tabular.rs
new file mode 100644
index 000000000..f6ac7df66
--- /dev/null
+++ b/crates/nvisy-core/src/policy/redaction/tabular.rs
@@ -0,0 +1,77 @@
+//! [`TabularRedaction`]: the operator spec a tabular-modality policy
+//! rule carries.
+//!
+//! Tabular is `TextBacked` in elide — cells reuse the text payload
+//! and replacement types — so the operator catalogue is identical
+//! to [`TextRedaction`]. The spec is structurally the same; we
+//! declare it as a separate enum so the wire format can diverge if
+//! tabular ever grows table-shape-aware operators (per-column
+//! masking, header-preserving transforms, …) that wouldn't make
+//! sense on text.
+
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+use super::text::HashAlgorithm;
+
+/// Operator spec a `redact` tabular rule carries.
+#[derive(Debug, Clone, PartialEq, Eq)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+pub enum TabularRedaction {
+    /// Delete the matched cell entirely.
+    Erase,
+    /// Pass the value through unchanged.
+    Keep,
+    /// Character-replacement masking.
+    Mask {
+        /// The character that replaces masked positions.
+        #[serde(default = "default_mask_char")]
+        mask_char: char,
+        /// Characters to leave unmasked at the start of the value.
+        #[serde(default, skip_serializing_if = "is_zero")]
+        keep_prefix: usize,
+        /// Characters to leave unmasked at the end of the value.
+        #[serde(default, skip_serializing_if = "is_zero")]
+        keep_suffix: usize,
+    },
+    /// Substitute the cell with a fixed template (`{label}`,
+    /// `{value}`, `{coref}` placeholders).
+    Replace {
+        /// Template string. Default `[{label}]`.
+        #[serde(default = "default_replace_template")]
+        template: String,
+    },
+    /// One-way SHA-2 hash with optional salt.
+    Hash {
+        /// SHA-256 (default) or SHA-512.
+        #[serde(default)]
+        algorithm: HashAlgorithm,
+        /// Salt prepended to the value before hashing.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        salt: Option<String>,
+    },
+    /// Vault-backed pseudonym: every mention of the same entity
+    /// reads the same surrogate.
+    Pseudonymize,
+    /// Reversible AES-256-GCM ciphertext. The engine wires the
+    /// per-tenant key provider.
+    Encrypt,
+    /// Drop the entire row a matched cell sits in.
+    DropRow,
+    /// Drop the entire column a matched cell sits in (header
+    /// included).
+    DropColumn,
+}
+
+fn default_replace_template() -> String {
+    "[{label}]".to_string()
+}
+
+fn default_mask_char() -> char {
+    '*'
+}
+
+fn is_zero(n: &usize) -> bool {
+    *n == 0
+}
diff --git a/crates/nvisy-core/src/policy/redaction/text.rs b/crates/nvisy-core/src/policy/redaction/text.rs
new file mode 100644
index 000000000..ac3b8494d
--- /dev/null
+++ b/crates/nvisy-core/src/policy/redaction/text.rs
@@ -0,0 +1,115 @@
+//! [`TextRedaction`]: the operator spec a text-modality policy rule
+//! carries.
+//!
+//! Each variant mirrors an elide built-in operator the engine
+//! constructs at apply time:
+//!
+//! - [`TextRedaction::Erase`] → [`elide::redaction::operators::Erase`]
+//! - [`TextRedaction::Keep`] → [`elide::redaction::operators::Keep`]
+//! - [`TextRedaction::Mask`] → [`elide::redaction::operators::Mask`]
+//! - [`TextRedaction::Replace`] → [`elide::redaction::operators::Replace`]
+//! - [`TextRedaction::Hash`] → [`elide::redaction::operators::Sha2Hash`]
+//! - [`TextRedaction::Pseudonymize`] →
+//!   [`elide::redaction::operators::Pseudonymize`]
+//! - [`TextRedaction::Encrypt`] →
+//!   [`elide::redaction::operators::AesEncrypt`] (engine wires the
+//!   per-tenant key provider)
+//!
+//! No `Custom` escape hatch — every operator the wire format admits
+//! is predefined. New built-ins land in elide first, then surface
+//! here as new variants.
+
+use elide::redaction::operators::Sha2Algorithm;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// SHA-2 variant for the [`TextRedaction::Hash`] operator.
+///
+/// Spec mirror of elide's [`Sha2Algorithm`]; the [`From`] impl below
+/// keeps the wire vocabulary independent of upstream's serde gating.
+///
+/// [`Sha2Algorithm`]: elide::redaction::operators::Sha2Algorithm
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum HashAlgorithm {
+    /// SHA-256 — 32-byte digest, 64-char hex.
+    #[default]
+    Sha256,
+    /// SHA-512 — 64-byte digest, 128-char hex.
+    Sha512,
+}
+
+impl From<HashAlgorithm> for Sha2Algorithm {
+    fn from(value: HashAlgorithm) -> Self {
+        match value {
+            HashAlgorithm::Sha256 => Sha2Algorithm::Sha256,
+            HashAlgorithm::Sha512 => Sha2Algorithm::Sha512,
+        }
+    }
+}
+
+/// Operator spec a `redact` text rule carries.
+#[derive(Debug, Clone, PartialEq, Eq)]
+#[derive(Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+pub enum TextRedaction {
+    /// Delete the matched span entirely.
+    Erase,
+    /// Pass the value through unchanged.
+    Keep,
+    /// Character-replacement masking. Leaves `keep_prefix` leading
+    /// and `keep_suffix` trailing characters visible; masks the
+    /// rest with `mask_char`.
+    Mask {
+        /// The character that replaces masked positions.
+        #[serde(default = "default_mask_char")]
+        mask_char: char,
+        /// Characters to leave unmasked at the start of the value.
+        /// `0` (the default) masks from the start.
+        #[serde(default, skip_serializing_if = "is_zero")]
+        keep_prefix: usize,
+        /// Characters to leave unmasked at the end of the value.
+        /// `0` (the default) masks through to the end.
+        #[serde(default, skip_serializing_if = "is_zero")]
+        keep_suffix: usize,
+    },
+    /// Substitute the span with a fixed template. Supports
+    /// `{label}` / `{value}` / `{coref}` placeholders.
+    Replace {
+        /// Template string. Default `[{label}]`.
+        #[serde(default = "default_replace_template")]
+        template: String,
+    },
+    /// One-way SHA-2 hash with optional salt.
+    Hash {
+        /// SHA-256 (default) or SHA-512.
+        #[serde(default)]
+        algorithm: HashAlgorithm,
+        /// Salt prepended to the value before hashing.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        salt: Option<String>,
+    },
+    /// Vault-backed pseudonym: every mention of the same entity
+    /// reads the same surrogate. The engine wires a per-request
+    /// vault + the default [`RandomToken`] generator.
+    ///
+    /// [`RandomToken`]: elide::redaction::generator::RandomToken
+    Pseudonymize,
+    /// Reversible AES-256-GCM ciphertext. The engine supplies the
+    /// per-tenant AES key at construction so raw key material never
+    /// lives in serialised policy.
+    Encrypt,
+}
+
+fn default_replace_template() -> String {
+    "[{label}]".to_string()
+}
+
+fn default_mask_char() -> char {
+    '*'
+}
+
+fn is_zero(n: &usize) -> bool {
+    *n == 0
+}
diff --git a/crates/nvisy-engine/src/policy/retention/duration.rs b/crates/nvisy-core/src/policy/retention/duration.rs
similarity index 100%
rename from crates/nvisy-engine/src/policy/retention/duration.rs
rename to crates/nvisy-core/src/policy/retention/duration.rs
diff --git a/crates/nvisy-engine/src/policy/retention/mod.rs b/crates/nvisy-core/src/policy/retention/mod.rs
similarity index 100%
rename from crates/nvisy-engine/src/policy/retention/mod.rs
rename to crates/nvisy-core/src/policy/retention/mod.rs
diff --git a/crates/nvisy-core/src/policy/rule.rs b/crates/nvisy-core/src/policy/rule.rs
new file mode 100644
index 000000000..7894b1719
--- /dev/null
+++ b/crates/nvisy-core/src/policy/rule.rs
@@ -0,0 +1,88 @@
+//! One rule inside a [`Policy`]: shared identity/description
+//! fields, the match predicate, and the action ([`RuleAction`]).
+//!
+//! The engine compiles a rule's [`predicate`](Rule::predicate)
+//! into an elide anonymizer rule at request time. Three shapes
+//! are recognised as fast paths and route to the matching
+//! [`Anonymizer`] builder method; everything else compiles to a
+//! catalog-aware closure:
+//!
+//! - [`Predicate::LabelOneOf`] with a single label → [`Anonymizer::with_label`]
+//! - [`Predicate::TagOneOf`] with a single tag → [`Anonymizer::with_tag`]
+//! - any composite (or multi-label / multi-tag) → [`Anonymizer::with_catalog_predicate`]
+//!
+//! Author-facing wire format: rules carry one composable
+//! `predicate` field. No need for separate label/tag/predicate
+//! kinds — `Predicate::LabelOneOf { labels: ["email"] }` is the
+//! same rule as the old `RuleKind::Label { label: "email" }` and
+//! compiles down to the same fast path.
+//!
+//! [`Policy`]: super::Policy
+//! [`Anonymizer`]: elide::redaction::Anonymizer
+//! [`Anonymizer::with_label`]: elide::redaction::Anonymizer::with_label
+//! [`Anonymizer::with_tag`]: elide::redaction::Anonymizer::with_tag
+//! [`Anonymizer::with_catalog_predicate`]: elide::redaction::Anonymizer::with_catalog_predicate
+
+use hipstr::HipStr;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+use uuid::Uuid;
+
+use super::audit::AuditAction;
+use super::predicate::Predicate;
+use super::redaction::ModalityRedactions;
+use super::suppress::SuppressAction;
+
+/// One rule inside a [`Policy`]. Identity is the UUID; `name` /
+/// `description` are display-only.
+///
+/// [`Policy`]: super::Policy
+#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct Rule {
+    /// Stable identifier. UUIDv7 recommended (time-ordered);
+    /// customer-supplied so re-submissions carry the same id.
+    /// Engine stamps this into the redaction event's
+    /// [`Attribution::reason`] so reviewers can trace back which
+    /// rule fired.
+    ///
+    /// [`Attribution::reason`]: elide_core::entity::provenance::Attribution::reason
+    pub id: Uuid,
+    /// Human-readable name. Display-only — does not key anything.
+    #[schemars(with = "String")]
+    pub name: HipStr<'static>,
+    /// Optional description for reviewers.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub description: Option<String>,
+    /// Entity-level predicate that decides whether the rule fires
+    /// on a given recognised entity. Composable — see
+    /// [`Predicate`] for the full grammar.
+    pub predicate: Predicate,
+    /// What to do on match.
+    pub action: RuleAction,
+}
+
+/// What a rule does when its [`predicate`](Rule::predicate)
+/// matches.
+///
+/// Three verbs: [`Redact`](RuleAction::Redact) transforms the
+/// entity with one operator per modality;
+/// [`Suppress`](RuleAction::Suppress) drops the entity entirely
+/// (false-positive marker) and stamps a reason onto the audit;
+/// [`Audit`](RuleAction::Audit) flags it for human review without
+/// transforming.
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(tag = "kind", rename_all = "camelCase")]
+pub enum RuleAction {
+    /// Redact matching entities. The carried map names operators
+    /// per modality; modalities the rule didn't cover fall through
+    /// to the policy fallback (or the next policy in the chain).
+    Redact(ModalityRedactions),
+    /// Suppress matching entities — treat as a false positive and
+    /// skip redaction. The audit records the suppression.
+    Suppress(SuppressAction),
+    /// Flag matching entities for human review. The entity is
+    /// left untouched; the audit entry carries the severity hint
+    /// for downstream tooling.
+    Audit(AuditAction),
+}
diff --git a/crates/nvisy-engine/src/policy/suppress.rs b/crates/nvisy-core/src/policy/suppress.rs
similarity index 79%
rename from crates/nvisy-engine/src/policy/suppress.rs
rename to crates/nvisy-core/src/policy/suppress.rs
index 99fa62096..55c937486 100644
--- a/crates/nvisy-engine/src/policy/suppress.rs
+++ b/crates/nvisy-core/src/policy/suppress.rs
@@ -1,9 +1,5 @@
 //! [`SuppressAction`]: payload of an [`Action::Suppress`] rule.
 //!
-//! Today carries an optional `reason` string that surfaces in the
-//! audit entry. Designed as a struct (not a unit type) so future
-//! fields (`expires`, `severity`, …) can land without a wire-break.
-//!
 //! [`Action::Suppress`]: super::rule::Action::Suppress
 
 use hipstr::HipStr;
diff --git a/crates/nvisy-core/src/primitive/confidence/mod.rs b/crates/nvisy-core/src/primitive/confidence/mod.rs
deleted file mode 100644
index f2b89e4ee..000000000
--- a/crates/nvisy-core/src/primitive/confidence/mod.rs
+++ /dev/null
@@ -1,14 +0,0 @@
-//! Detection confidence as a validated numeric primitive.
-//!
-//! [`Confidence`] is a value in `[0.0, 1.0]` (NaN / ±∞ rejected).
-//! [`ConfidenceThreshold`] is a lower bound on `Confidence`, with a
-//! single named operation [`admits`] so threshold semantics don't
-//! drift across consumer sites.
-//!
-//! [`admits`]: ConfidenceThreshold::admits
-
-mod threshold;
-mod value;
-
-pub use self::threshold::ConfidenceThreshold;
-pub use self::value::Confidence;
diff --git a/crates/nvisy-core/src/primitive/confidence/threshold.rs b/crates/nvisy-core/src/primitive/confidence/threshold.rs
deleted file mode 100644
index 6617fde8a..000000000
--- a/crates/nvisy-core/src/primitive/confidence/threshold.rs
+++ /dev/null
@@ -1,159 +0,0 @@
-//! [`ConfidenceThreshold`] — a validated lower bound for
-//! [`Confidence`] values.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Deserializer, Serialize};
-
-use super::Confidence;
-
-/// A lower-bound threshold on [`Confidence`]. Wraps a [`Confidence`]
-/// so it inherits the `[0.0, 1.0]` + finite-float invariant, and
-/// exposes [`admits`] as the single named operation for the "is this
-/// entity above the cutoff?" question.
-///
-/// The shape exists so producers can't reach for a bare `f64` that
-/// silently accepts `1.5` or `NaN`, and so consumers don't open-code
-/// `entity.confidence.get() >= threshold.get()` everywhere with the
-/// inequality direction free to drift. The contract is fixed: the
-/// threshold *admits* a confidence at or above its value (inclusive
-/// on the boundary).
-///
-/// Wire shape: `ConfidenceThreshold` serialises as a bare number,
-/// same as [`Confidence`]. Existing `"confidenceThreshold": 0.85`
-/// payloads round-trip without change.
-///
-/// # Examples
-///
-/// ```
-/// use nvisy_core::primitive::{Confidence, ConfidenceThreshold};
-///
-/// let cutoff = ConfidenceThreshold::new(0.85).unwrap();
-/// assert!(cutoff.admits(Confidence::new(0.90).unwrap()));
-/// assert!(cutoff.admits(Confidence::new(0.85).unwrap())); // inclusive
-/// assert!(!cutoff.admits(Confidence::new(0.84).unwrap()));
-/// ```
-///
-/// [`admits`]: Self::admits
-#[derive(Debug, Clone, Copy, PartialEq, PartialOrd, Serialize, JsonSchema)]
-#[serde(transparent)]
-pub struct ConfidenceThreshold(Confidence);
-
-impl<'de> Deserialize<'de> for ConfidenceThreshold {
-    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
-        Confidence::deserialize(deserializer).map(Self)
-    }
-}
-
-impl ConfidenceThreshold {
-    /// Upper bound. Admits only [`Confidence::MAX`]. Equivalent to
-    /// `ConfidenceThreshold::from_confidence(Confidence::MAX)`, but
-    /// `const`. Mirrors [`Confidence::MAX`].
-    pub const MAX: Self = Self::from_confidence(Confidence::MAX);
-    /// Lower bound. Admits every valid [`Confidence`] (no filtering).
-    /// Equivalent to `ConfidenceThreshold::from_confidence(Confidence::MIN)`,
-    /// but `const`. Mirrors [`Confidence::MIN`].
-    pub const MIN: Self = Self::from_confidence(Confidence::MIN);
-
-    /// Construct a [`ConfidenceThreshold`] from a raw score,
-    /// returning [`None`] when the value is outside `[0.0, 1.0]` or
-    /// non-finite. Mirrors [`Confidence::new`] — the threshold's
-    /// validity domain is exactly the same as a confidence score.
-    pub fn new(value: f64) -> Option<Self> {
-        Confidence::new(value).map(Self)
-    }
-
-    /// Construct a [`ConfidenceThreshold`] by clamping a raw score
-    /// into `[0.0, 1.0]`, panicking on `NaN` / `±∞`. Mirrors
-    /// [`Confidence::clamped`] — use when the value is a literal /
-    /// internal constant known to be in range modulo float rounding.
-    /// For floats from external sources, prefer
-    /// [`Self::new`] and decide what to do with the rejection at
-    /// the boundary.
-    pub fn clamped(value: f64) -> Self {
-        Self(Confidence::clamped(value))
-    }
-
-    /// Wrap an already-validated [`Confidence`] as a threshold.
-    pub const fn from_confidence(value: Confidence) -> Self {
-        Self(value)
-    }
-
-    /// Returns the inner threshold as a [`Confidence`].
-    pub fn confidence(&self) -> Confidence {
-        self.0
-    }
-
-    /// Returns the inner score as a raw `f64`. Prefer [`Self::admits`]
-    /// at comparison sites — this accessor exists for serialisation,
-    /// telemetry, and prompt construction where the raw number is
-    /// what's needed.
-    pub fn get(&self) -> f64 {
-        self.0.get()
-    }
-
-    /// Returns `true` when `confidence` is **at or above** the
-    /// threshold. The boundary is inclusive: a threshold of `0.85`
-    /// admits a confidence of exactly `0.85`.
-    pub fn admits(self, confidence: Confidence) -> bool {
-        confidence.get() >= self.0.get()
-    }
-}
-
-impl From<Confidence> for ConfidenceThreshold {
-    fn from(c: Confidence) -> Self {
-        Self::from_confidence(c)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn admits_above_below_and_exact() {
-        let cutoff = ConfidenceThreshold::new(0.5).unwrap();
-        assert!(cutoff.admits(Confidence::new(0.6).unwrap()));
-        assert!(cutoff.admits(Confidence::new(0.5).unwrap()));
-        assert!(!cutoff.admits(Confidence::new(0.49).unwrap()));
-    }
-
-    #[test]
-    fn min_max_constants_equal_boundary_values() {
-        assert_eq!(ConfidenceThreshold::MIN.get(), 0.0);
-        assert_eq!(ConfidenceThreshold::MAX.get(), 1.0);
-    }
-
-    #[test]
-    fn boundaries() {
-        assert!(ConfidenceThreshold::MIN.admits(Confidence::MIN));
-        assert!(ConfidenceThreshold::MAX.admits(Confidence::MAX));
-        assert!(!ConfidenceThreshold::MAX.admits(Confidence::new(0.99).unwrap()));
-    }
-
-    #[test]
-    fn rejects_out_of_range() {
-        assert!(ConfidenceThreshold::new(-0.1).is_none());
-        assert!(ConfidenceThreshold::new(1.1).is_none());
-        assert!(ConfidenceThreshold::new(f64::NAN).is_none());
-        assert!(ConfidenceThreshold::new(f64::INFINITY).is_none());
-    }
-
-    #[test]
-    fn clamped_clamps_into_range() {
-        assert_eq!(ConfidenceThreshold::clamped(0.5).get(), 0.5);
-        assert_eq!(ConfidenceThreshold::clamped(1.5).get(), 1.0);
-        assert_eq!(ConfidenceThreshold::clamped(-0.5).get(), 0.0);
-    }
-
-    #[test]
-    #[should_panic(expected = "non-finite confidence")]
-    fn clamped_panics_on_nan() {
-        let _ = ConfidenceThreshold::clamped(f64::NAN);
-    }
-
-    #[test]
-    fn deserialize_rejects_out_of_range() {
-        assert!(serde_json::from_str::<ConfidenceThreshold>("1.5").is_err());
-        assert!(serde_json::from_str::<ConfidenceThreshold>("-0.1").is_err());
-    }
-}
diff --git a/crates/nvisy-core/src/primitive/confidence/value.rs b/crates/nvisy-core/src/primitive/confidence/value.rs
deleted file mode 100644
index 928e02b68..000000000
--- a/crates/nvisy-core/src/primitive/confidence/value.rs
+++ /dev/null
@@ -1,204 +0,0 @@
-//! [`Confidence`] — a validated `[0.0, 1.0]` confidence score.
-
-use schemars::JsonSchema;
-use serde::de::Error as DeError;
-use serde::{Deserialize, Deserializer, Serialize};
-
-/// A confidence score in the closed range `[0.0, 1.0]`.
-///
-/// Construct with [`Confidence::new`]; the constructor returns
-/// [`None`] for values outside the valid range or non-finite floats
-/// (`NaN`, `±∞`). Read the inner score with [`Confidence::get`].
-/// The API mirrors [`NonZero::new`] / [`NonZero::get`].
-///
-/// `Confidence` is `Copy` and cheap to pass by value. Operations
-/// that combine confidences (averaging, multiplying) are not
-/// provided directly — callers should compute on `f64` via `.get()`
-/// and re-construct.
-///
-/// Deserialization is strict: an input outside `[0.0, 1.0]` or any
-/// non-finite float fails with an error. The newtype is a real
-/// invariant, not a hint.
-///
-/// # Examples
-///
-/// ```
-/// use nvisy_core::primitive::Confidence;
-///
-/// let c = Confidence::new(0.85).unwrap();
-/// assert_eq!(c.get(), 0.85);
-///
-/// assert!(Confidence::new(1.5).is_none());
-/// assert!(Confidence::new(f64::NAN).is_none());
-/// ```
-///
-/// [`Confidence::new`]: Self::new
-/// [`Confidence::get`]: Self::get
-/// [`NonZero::new`]: std::num::NonZero::new
-/// [`NonZero::get`]: std::num::NonZero::get
-#[derive(Debug, Clone, Copy, PartialEq, PartialOrd, Serialize, JsonSchema)]
-#[serde(transparent)]
-pub struct Confidence(f64);
-
-impl<'de> Deserialize<'de> for Confidence {
-    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
-        let value = f64::deserialize(deserializer)?;
-        Self::new(value).ok_or_else(|| {
-            DeError::custom(format!(
-                "confidence {value} out of range [0.0, 1.0] or non-finite"
-            ))
-        })
-    }
-}
-
-impl Confidence {
-    /// Upper bound of the valid range. Equivalent to
-    /// `Confidence::new(1.0).unwrap()`, but `const` so it works in
-    /// constant contexts. Mirrors the `MIN` / `MAX` constants on
-    /// primitive integer types (`u32::MAX`, `i64::MAX`, …).
-    pub const MAX: Self = Self(1.0);
-    /// Lower bound of the valid range. Equivalent to
-    /// `Confidence::new(0.0).unwrap()`, but `const` so it works in
-    /// constant contexts. Mirrors the `MIN` / `MAX` constants on
-    /// primitive integer types (`u32::MIN`, `i64::MIN`, …).
-    pub const MIN: Self = Self(0.0);
-
-    /// Construct a [`Confidence`] from a raw score, returning
-    /// [`None`] when `value` is `NaN`, `±∞`, or outside the closed
-    /// range `[0.0, 1.0]`.
-    pub fn new(value: f64) -> Option<Self> {
-        if value.is_finite() && (0.0..=1.0).contains(&value) {
-            Some(Self(value))
-        } else {
-            None
-        }
-    }
-
-    /// Construct a [`Confidence`] by clamping a raw score into
-    /// `[0.0, 1.0]`. Use when the input is *known to be in range
-    /// modulo float rounding* — a softmax mean that might come out
-    /// as `1.0000000000000002`, a calibration offset that nudges a
-    /// 0.99 slightly above 1.0, etc.
-    ///
-    /// **Panics** on `NaN` or `±∞`. Non-finite values cannot have
-    /// come from a well-behaved producer; loud failure beats the
-    /// previous silent-degrade-to-`0.0` behaviour, which masked
-    /// upstream bugs by quietly dropping detections.
-    ///
-    /// For floats from external / untrusted sources (network
-    /// payloads, model server responses, user input), use
-    /// [`Self::try_clamped`] and decide what to do with the rejection
-    /// at the boundary.
-    pub fn clamped(value: f64) -> Self {
-        assert!(
-            value.is_finite(),
-            "non-finite confidence: {value} (NaN/±∞ indicates an upstream bug)"
-        );
-        Self(value.clamp(0.0, 1.0))
-    }
-
-    /// Fallible counterpart to [`Self::clamped`] for floats whose
-    /// finiteness isn't pre-validated. Returns [`None`] on `NaN` /
-    /// `±∞`; otherwise clamps into `[0.0, 1.0]`. Use at the boundary
-    /// where untrusted floats enter the pipeline (model server
-    /// responses, deserialised user input) and the caller knows
-    /// what fallback to apply on rejection.
-    pub fn try_clamped(value: f64) -> Option<Self> {
-        value.is_finite().then(|| Self(value.clamp(0.0, 1.0)))
-    }
-
-    /// Returns the inner score.
-    pub fn get(&self) -> f64 {
-        self.0
-    }
-
-    /// Add `delta` to the score, saturating at `1.0`. Negative deltas
-    /// saturate at `0.0`. Non-finite deltas (`NaN`, `±∞`) are treated
-    /// as `0.0` — same posture as [`Self::clamped`].
-    ///
-    /// Distinct from combining two confidences (which the type
-    /// deliberately doesn't expose): this is a scalar adjustment to
-    /// a single score, used by context-aware boosting / penalty.
-    pub fn saturating_add(self, delta: f64) -> Self {
-        debug_assert!(delta.is_finite(), "non-finite confidence delta: {delta}");
-        let delta = if delta.is_finite() { delta } else { 0.0 };
-        Self((self.0 + delta).clamp(0.0, 1.0))
-    }
-
-    /// Subtract `delta` from the score, saturating at `0.0`. Negative
-    /// deltas saturate at `1.0`. Symmetric to [`Self::saturating_add`].
-    pub fn saturating_sub(self, delta: f64) -> Self {
-        self.saturating_add(-delta)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_in_range_values() {
-        assert_eq!(Confidence::new(0.0).unwrap().get(), 0.0);
-        assert_eq!(Confidence::new(1.0).unwrap().get(), 1.0);
-        assert_eq!(Confidence::new(0.5).unwrap().get(), 0.5);
-    }
-
-    #[test]
-    fn min_max_constants_equal_boundary_values() {
-        assert_eq!(Confidence::MIN.get(), 0.0);
-        assert_eq!(Confidence::MAX.get(), 1.0);
-        assert_eq!(Confidence::MIN, Confidence::new(0.0).unwrap());
-        assert_eq!(Confidence::MAX, Confidence::new(1.0).unwrap());
-    }
-
-    #[test]
-    fn rejects_out_of_range_values() {
-        assert!(Confidence::new(-0.1).is_none());
-        assert!(Confidence::new(1.1).is_none());
-    }
-
-    #[test]
-    fn rejects_non_finite_values() {
-        assert!(Confidence::new(f64::NAN).is_none());
-        assert!(Confidence::new(f64::INFINITY).is_none());
-        assert!(Confidence::new(f64::NEG_INFINITY).is_none());
-    }
-
-    #[test]
-    fn clamped_handles_rounding_above_one() {
-        assert_eq!(Confidence::clamped(1.0000000000000002).get(), 1.0);
-        assert_eq!(Confidence::clamped(-0.0).get(), 0.0);
-    }
-
-    #[test]
-    #[should_panic(expected = "non-finite confidence")]
-    fn clamped_panics_on_nan() {
-        let _ = Confidence::clamped(f64::NAN);
-    }
-
-    #[test]
-    #[should_panic(expected = "non-finite confidence")]
-    fn clamped_panics_on_infinity() {
-        let _ = Confidence::clamped(f64::INFINITY);
-    }
-
-    #[test]
-    fn try_clamped_accepts_finite() {
-        assert_eq!(Confidence::try_clamped(0.5).unwrap().get(), 0.5);
-        assert_eq!(Confidence::try_clamped(2.0).unwrap().get(), 1.0);
-        assert_eq!(Confidence::try_clamped(-1.0).unwrap().get(), 0.0);
-    }
-
-    #[test]
-    fn try_clamped_rejects_non_finite() {
-        assert!(Confidence::try_clamped(f64::NAN).is_none());
-        assert!(Confidence::try_clamped(f64::INFINITY).is_none());
-        assert!(Confidence::try_clamped(f64::NEG_INFINITY).is_none());
-    }
-
-    #[test]
-    fn deserialize_rejects_out_of_range() {
-        assert!(serde_json::from_str::<Confidence>("1.4").is_err());
-        assert!(serde_json::from_str::<Confidence>("-0.2").is_err());
-    }
-}
diff --git a/crates/nvisy-core/src/primitive/country.rs b/crates/nvisy-core/src/primitive/country.rs
deleted file mode 100644
index 6fc9fca07..000000000
--- a/crates/nvisy-core/src/primitive/country.rs
+++ /dev/null
@@ -1,179 +0,0 @@
-//! [ISO 3166-1 alpha-2] country code type.
-//!
-//! Thin wrapper around [`celes::Country`] that exposes only the
-//! alpha-2 surface — alpha-3, numeric, and long-name forms are
-//! reachable via [`CountryCode::into_inner`] for the rare consumer
-//! that needs them.
-//!
-//! [ISO 3166-1 alpha-2]: https://en.wikipedia.org/wiki/ISO_3166-1_alpha-2
-
-use std::fmt;
-use std::str::FromStr;
-
-use serde::{Deserialize, Serialize};
-
-use crate::Error;
-
-/// A validated [ISO 3166-1 alpha-2] country code.
-///
-/// Two-letter uppercase code identifying a country or region.
-/// Construction accepts any case (`"us"`, `"US"`, `"uS"`) and
-/// rejects anything that isn't a known ISO 3166-1 alpha-2 code.
-///
-/// # Examples
-///
-/// ```
-/// use nvisy_core::primitive::CountryCode;
-///
-/// let us = CountryCode::new("us").unwrap();
-/// assert_eq!(us.as_str(), "US");
-///
-/// assert!(CountryCode::new("USA").is_err());
-/// assert!(CountryCode::new("XZ").is_err());
-/// ```
-///
-/// [ISO 3166-1 alpha-2]: https://en.wikipedia.org/wiki/ISO_3166-1_alpha-2
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize)]
-#[serde(try_from = "String", into = "String")]
-pub struct CountryCode(celes::Country);
-
-impl CountryCode {
-    /// Parse and validate a country code.
-    ///
-    /// Input is case-insensitive; the canonical form returned by
-    /// [`as_str`] is uppercase.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when `code` is not a known
-    /// ISO 3166-1 alpha-2 code.
-    ///
-    /// [`as_str`]: Self::as_str
-    pub fn new(code: &str) -> Result<Self, Error> {
-        celes::Country::from_alpha2(code).map(Self).map_err(|_| {
-            Error::validation(
-                format!("country code `{code}` is not a known ISO 3166-1 alpha-2 code"),
-                "nvisy-core",
-            )
-        })
-    }
-
-    /// Return the canonical (uppercase) alpha-2 representation.
-    #[must_use]
-    pub fn as_str(&self) -> &'static str {
-        self.0.alpha2
-    }
-
-    /// Borrow the underlying [`celes::Country`] for callers that
-    /// need alpha-3, numeric, or the country's long name.
-    #[must_use]
-    pub fn into_inner(self) -> celes::Country {
-        self.0
-    }
-}
-
-impl fmt::Display for CountryCode {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.write_str(self.as_str())
-    }
-}
-
-impl FromStr for CountryCode {
-    type Err = Error;
-
-    fn from_str(s: &str) -> Result<Self, Self::Err> {
-        Self::new(s)
-    }
-}
-
-impl TryFrom<String> for CountryCode {
-    type Error = Error;
-
-    fn try_from(value: String) -> Result<Self, Self::Error> {
-        Self::new(&value)
-    }
-}
-
-impl From<CountryCode> for String {
-    fn from(code: CountryCode) -> Self {
-        code.as_str().to_owned()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_uppercase_alpha2() {
-        let us = CountryCode::new("US").unwrap();
-        assert_eq!(us.as_str(), "US");
-    }
-
-    #[test]
-    fn accepts_lowercase_alpha2_and_normalises() {
-        let us = CountryCode::new("us").unwrap();
-        assert_eq!(us.as_str(), "US");
-    }
-
-    #[test]
-    fn accepts_mixed_case() {
-        let gb = CountryCode::new("Gb").unwrap();
-        assert_eq!(gb.as_str(), "GB");
-    }
-
-    #[test]
-    fn rejects_alpha3_code() {
-        assert!(CountryCode::new("USA").is_err());
-        assert!(CountryCode::new("GBR").is_err());
-    }
-
-    #[test]
-    fn rejects_unassigned_two_letter_code() {
-        assert!(CountryCode::new("XZ").is_err());
-        assert!(CountryCode::new("ZZ").is_err());
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(CountryCode::new("U").is_err());
-        assert!(CountryCode::new("").is_err());
-        assert!(CountryCode::new("USAB").is_err());
-    }
-
-    #[test]
-    fn rejects_non_alpha() {
-        assert!(CountryCode::new("U1").is_err());
-        assert!(CountryCode::new("00").is_err());
-    }
-
-    #[test]
-    fn equality_is_canonical() {
-        // `us` and `US` should compare equal once parsed.
-        let lower = CountryCode::new("us").unwrap();
-        let upper = CountryCode::new("US").unwrap();
-        assert_eq!(lower, upper);
-    }
-
-    #[test]
-    fn serde_roundtrip_uppercase() {
-        let us = CountryCode::new("us").unwrap();
-        let json = serde_json::to_string(&us).unwrap();
-        assert_eq!(json, "\"US\"");
-        let back: CountryCode = serde_json::from_str(&json).unwrap();
-        assert_eq!(back, us);
-    }
-
-    #[test]
-    fn from_str_parses_alpha2() {
-        let de: CountryCode = "DE".parse().unwrap();
-        assert_eq!(de.as_str(), "DE");
-    }
-
-    #[test]
-    fn display_writes_alpha2() {
-        let fr = CountryCode::new("FR").unwrap();
-        assert_eq!(format!("{fr}"), "FR");
-    }
-}
diff --git a/crates/nvisy-core/src/primitive/geometry/bounding_box.rs b/crates/nvisy-core/src/primitive/geometry/bounding_box.rs
deleted file mode 100644
index f6888c179..000000000
--- a/crates/nvisy-core/src/primitive/geometry/bounding_box.rs
+++ /dev/null
@@ -1,205 +0,0 @@
-//! Axis-aligned bounding boxes — floating-point ([`BoundingBox`])
-//! and integer pixel ([`IBoundingBox`]).
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Integer pixel-coordinate bounding box for rendering operations.
-///
-/// Converted from [`BoundingBox`] by rounding each field to the nearest
-/// integer. Use this at the rendering boundary where pixel-exact
-/// coordinates are required.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-pub struct IBoundingBox {
-    /// Horizontal offset of the top-left corner in pixels.
-    pub x: u32,
-    /// Vertical offset of the top-left corner in pixels.
-    pub y: u32,
-    /// Width in pixels.
-    pub width: u32,
-    /// Height in pixels.
-    pub height: u32,
-}
-
-impl From<&BoundingBox> for IBoundingBox {
-    fn from(bb: &BoundingBox) -> Self {
-        bb.to_pixel()
-    }
-}
-
-impl From<BoundingBox> for IBoundingBox {
-    fn from(bb: BoundingBox) -> Self {
-        Self::from(&bb)
-    }
-}
-
-/// Axis-aligned bounding box for image-based entity locations.
-///
-/// Coordinates are `f64` to support both pixel and normalized (0.0–1.0)
-/// values from detection models.
-#[derive(Debug, Clone, Copy, Default, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct BoundingBox {
-    /// Horizontal offset of the top-left corner (pixels or normalized).
-    pub x: f64,
-    /// Vertical offset of the top-left corner (pixels or normalized).
-    pub y: f64,
-    /// Width of the bounding box.
-    pub width: f64,
-    /// Height of the bounding box.
-    pub height: f64,
-}
-
-impl BoundingBox {
-    /// Create a new bounding box.
-    pub fn new(x: f64, y: f64, width: f64, height: f64) -> Self {
-        Self {
-            x,
-            y,
-            width,
-            height,
-        }
-    }
-
-    /// Area of the bounding box.
-    pub fn area(&self) -> f64 {
-        self.width * self.height
-    }
-
-    /// Right edge (`x + width`).
-    pub fn right(&self) -> f64 {
-        self.x + self.width
-    }
-
-    /// Bottom edge (`y + height`).
-    pub fn bottom(&self) -> f64 {
-        self.y + self.height
-    }
-
-    /// Center point `(cx, cy)`.
-    pub fn center(&self) -> (f64, f64) {
-        (self.x + self.width / 2.0, self.y + self.height / 2.0)
-    }
-
-    /// Returns `true` if this box overlaps with `other`.
-    pub fn overlaps(&self, other: &BoundingBox) -> bool {
-        self.x < other.right()
-            && other.x < self.right()
-            && self.y < other.bottom()
-            && other.y < self.bottom()
-    }
-
-    /// Returns the smallest box that encloses both `self` and `other`.
-    pub fn union(&self, other: &BoundingBox) -> BoundingBox {
-        let x = self.x.min(other.x);
-        let y = self.y.min(other.y);
-        let right = self.right().max(other.right());
-        let bottom = self.bottom().max(other.bottom());
-        BoundingBox::new(x, y, right - x, bottom - y)
-    }
-
-    /// Convert to integer pixel coordinates by rounding each field.
-    ///
-    /// Saturating: negative coordinates clamp to `0`, values above
-    /// `u32::MAX` clamp to `u32::MAX`, and `NaN` produces `0`. Use
-    /// for renderers / overlay drawing where a best-effort pixel
-    /// box is preferable to a hard failure. For producers that want
-    /// to detect malformed input (negative width, NaN coords) use
-    /// [`Self::try_to_pixel`] instead.
-    pub fn to_pixel(&self) -> IBoundingBox {
-        IBoundingBox {
-            x: self.x.round() as u32,
-            y: self.y.round() as u32,
-            width: self.width.round() as u32,
-            height: self.height.round() as u32,
-        }
-    }
-
-    /// Fallible counterpart to [`Self::to_pixel`]. Returns [`None`]
-    /// when any field is non-finite (`NaN` / `±∞`) or negative —
-    /// values that have no meaningful pixel representation and
-    /// almost always indicate an upstream bug in the producer.
-    pub fn try_to_pixel(&self) -> Option<IBoundingBox> {
-        let all_finite = self.x.is_finite()
-            && self.y.is_finite()
-            && self.width.is_finite()
-            && self.height.is_finite();
-        let non_negative =
-            self.x >= 0.0 && self.y >= 0.0 && self.width >= 0.0 && self.height >= 0.0;
-        (all_finite && non_negative).then(|| self.to_pixel())
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn edges_and_center() {
-        let bb = BoundingBox::new(10.0, 20.0, 30.0, 40.0);
-        assert!((bb.right() - 40.0).abs() < f64::EPSILON);
-        assert!((bb.bottom() - 60.0).abs() < f64::EPSILON);
-        let (cx, cy) = bb.center();
-        assert!((cx - 25.0).abs() < f64::EPSILON);
-        assert!((cy - 40.0).abs() < f64::EPSILON);
-    }
-
-    #[test]
-    fn overlaps() {
-        let a = BoundingBox::new(0.0, 0.0, 10.0, 10.0);
-        let b = BoundingBox::new(5.0, 5.0, 10.0, 10.0);
-        let c = BoundingBox::new(10.0, 0.0, 10.0, 10.0);
-        assert!(a.overlaps(&b));
-        assert!(!a.overlaps(&c)); // touching at edge = no overlap
-    }
-
-    #[test]
-    fn try_to_pixel_accepts_finite_non_negative() {
-        let bb = BoundingBox::new(10.4, 20.6, 30.0, 40.0);
-        let px = bb.try_to_pixel().expect("finite, non-negative");
-        assert_eq!(px.x, 10);
-        assert_eq!(px.y, 21);
-        assert_eq!(px.width, 30);
-        assert_eq!(px.height, 40);
-    }
-
-    #[test]
-    fn try_to_pixel_rejects_negative_or_non_finite() {
-        assert!(
-            BoundingBox::new(-1.0, 0.0, 10.0, 10.0)
-                .try_to_pixel()
-                .is_none()
-        );
-        assert!(
-            BoundingBox::new(0.0, 0.0, f64::NAN, 10.0)
-                .try_to_pixel()
-                .is_none()
-        );
-        assert!(
-            BoundingBox::new(0.0, 0.0, 10.0, f64::INFINITY)
-                .try_to_pixel()
-                .is_none()
-        );
-    }
-
-    #[test]
-    fn union_boxes() {
-        let a = BoundingBox::new(0.0, 0.0, 5.0, 5.0);
-        let b = BoundingBox::new(3.0, 3.0, 7.0, 7.0);
-        let u = a.union(&b);
-        assert!((u.x).abs() < f64::EPSILON);
-        assert!((u.y).abs() < f64::EPSILON);
-        assert!((u.width - 10.0).abs() < f64::EPSILON);
-        assert!((u.height - 10.0).abs() < f64::EPSILON);
-    }
-
-    #[test]
-    fn from_bounding_box_rounds() {
-        let bb = BoundingBox::new(1.4, 2.6, 3.5, 4.4);
-        let px = IBoundingBox::from(bb);
-        assert_eq!(px.x, 1);
-        assert_eq!(px.y, 3);
-        assert_eq!(px.width, 4);
-        assert_eq!(px.height, 4);
-    }
-}
diff --git a/crates/nvisy-core/src/primitive/geometry/dimensions.rs b/crates/nvisy-core/src/primitive/geometry/dimensions.rs
deleted file mode 100644
index b23a7e0de..000000000
--- a/crates/nvisy-core/src/primitive/geometry/dimensions.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-//! Image / canvas dimensions in integer pixels.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Pixel dimensions of an image or any 2D canvas.
-///
-/// Used to convert between normalized `[0, 1]` coordinates (what
-/// vision models typically emit) and absolute pixel coordinates
-/// (what renderers and the rest of our pipeline consume). See
-/// [`NormalizedBoundingBox::to_pixel`] for the conversion.
-///
-/// [`NormalizedBoundingBox::to_pixel`]: super::NormalizedBoundingBox::to_pixel
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Dimensions {
-    /// Width in pixels.
-    pub width: u32,
-    /// Height in pixels.
-    pub height: u32,
-}
-
-impl Dimensions {
-    /// Create a `Dimensions` from explicit width and height.
-    pub fn new(width: u32, height: u32) -> Self {
-        Self { width, height }
-    }
-}
-
-impl From<(u32, u32)> for Dimensions {
-    fn from((width, height): (u32, u32)) -> Self {
-        Self { width, height }
-    }
-}
diff --git a/crates/nvisy-core/src/primitive/geometry/mod.rs b/crates/nvisy-core/src/primitive/geometry/mod.rs
deleted file mode 100644
index 71f8ba351..000000000
--- a/crates/nvisy-core/src/primitive/geometry/mod.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-//! 2D coordinate geometry: axis-aligned bounding boxes (pixel and
-//! normalised forms), pixel dimensions, and free-form polygons.
-//!
-//! - [`BoundingBox`] — floating-point pixel form used everywhere
-//!   pixel coordinates may be sub-pixel (OCR output, intermediate
-//!   computation).
-//! - [`IBoundingBox`] — integer pixel form used at the rendering
-//!   boundary, where pixel-exact coordinates are required.
-//! - [`NormalizedBoundingBox`] — `[0, 1]` floating-point form
-//!   used at API boundaries where pixel dimensions are unknown
-//!   (vision-model outputs). Convert to [`BoundingBox`] via
-//!   [`NormalizedBoundingBox::to_pixel`] once [`Dimensions`] are
-//!   available.
-//! - [`Dimensions`] — pixel width + height; the conversion
-//!   reference between normalised and pixel coordinates.
-//! - [`Polygon`] (with [`Vertex`]) — rotated or non-rectangular
-//!   regions, typically skewed text reported by OCR.
-
-mod bounding_box;
-mod dimensions;
-mod normalized_bounding_box;
-mod polygon;
-
-pub use self::bounding_box::{BoundingBox, IBoundingBox};
-pub use self::dimensions::Dimensions;
-pub use self::normalized_bounding_box::NormalizedBoundingBox;
-pub use self::polygon::{Polygon, Vertex};
diff --git a/crates/nvisy-core/src/primitive/geometry/normalized_bounding_box.rs b/crates/nvisy-core/src/primitive/geometry/normalized_bounding_box.rs
deleted file mode 100644
index b518f992d..000000000
--- a/crates/nvisy-core/src/primitive/geometry/normalized_bounding_box.rs
+++ /dev/null
@@ -1,59 +0,0 @@
-//! Normalised `[0, 1]` axis-aligned bounding box.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::{BoundingBox, Dimensions};
-
-/// Axis-aligned bounding box in normalised `[0, 1]` coordinates.
-///
-/// Used at API boundaries where pixel dimensions are unknown to
-/// the producer — most commonly the output of vision-language
-/// models that don't see the original image's pixel size.
-/// `(0, 0)` is the top-left corner of the image, `(1, 1)` is the
-/// bottom-right.
-///
-/// Field values are **not** clamped to `[0, 1]`; the type carries
-/// the *intent* of normalised coordinates, not a hard invariant.
-/// Producers should keep values in range; conversion to pixel
-/// space ([`NormalizedBoundingBox::to_pixel`]) is mechanical
-/// multiplication regardless.
-#[derive(Debug, Clone, Copy, Default, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct NormalizedBoundingBox {
-    /// Top-left x in `[0, 1]` (fraction of image width).
-    pub x: f64,
-    /// Top-left y in `[0, 1]` (fraction of image height).
-    pub y: f64,
-    /// Width in `[0, 1]` (fraction of image width).
-    pub width: f64,
-    /// Height in `[0, 1]` (fraction of image height).
-    pub height: f64,
-}
-
-impl NormalizedBoundingBox {
-    /// Create a normalised bounding box from explicit fields.
-    pub fn new(x: f64, y: f64, width: f64, height: f64) -> Self {
-        Self {
-            x,
-            y,
-            width,
-            height,
-        }
-    }
-
-    /// Convert to pixel-space [`BoundingBox`] by multiplying each
-    /// dimension by the corresponding axis of `dims`.
-    ///
-    /// Out-of-range input ([0, 1] violations) propagates as
-    /// out-of-bounds pixel values — caller's responsibility to
-    /// keep producers honest.
-    pub fn to_pixel(&self, dims: Dimensions) -> BoundingBox {
-        BoundingBox::new(
-            self.x * f64::from(dims.width),
-            self.y * f64::from(dims.height),
-            self.width * f64::from(dims.width),
-            self.height * f64::from(dims.height),
-        )
-    }
-}
diff --git a/crates/nvisy-core/src/primitive/geometry/polygon.rs b/crates/nvisy-core/src/primitive/geometry/polygon.rs
deleted file mode 100644
index ed9ce1313..000000000
--- a/crates/nvisy-core/src/primitive/geometry/polygon.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-//! Polygon type for non-rectangular regions.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// A single vertex in a [`Polygon`].
-#[derive(Debug, Clone, Copy, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Vertex {
-    /// Horizontal coordinate (pixels or normalized).
-    pub x: f64,
-    /// Vertical coordinate (pixels or normalized).
-    pub y: f64,
-}
-
-/// A closed polygon defined by its vertices.
-///
-/// Used for rotated or non-rectangular regions such as skewed text
-/// detected by OCR. Vertices are ordered (typically clockwise from
-/// top-left) and coordinates are `f64` to support both pixel and
-/// normalized values.
-#[derive(Debug, Clone, Default, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Polygon {
-    /// Ordered vertices defining the polygon outline.
-    pub vertices: Vec<Vertex>,
-}
diff --git a/crates/nvisy-core/src/primitive/language/detection.rs b/crates/nvisy-core/src/primitive/language/detection.rs
deleted file mode 100644
index dd9a4ee51..000000000
--- a/crates/nvisy-core/src/primitive/language/detection.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-//! [`LanguageDetection`] — single language-detection result with
-//! [`LanguageProvenance`] distinguishing detected from
-//! caller-asserted answers, plus the [`LanguageSpan`] byte-offset
-//! range it applies to when the detector reports per-region results.
-
-use serde::{Deserialize, Serialize};
-
-use super::LanguageTag;
-use crate::primitive::Confidence;
-
-/// Provenance of a [`LanguageDetection`].
-///
-/// Lets consumers distinguish "the engine ran a detector and got
-/// this answer" from "the caller asserted this language and bypassed
-/// detection" — without overloading `confidence: None`.
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-#[derive(Serialize, Deserialize)]
-#[serde(rename_all = "camelCase")]
-pub enum LanguageProvenance {
-    /// Produced by a language-detection backend.
-    Detected,
-    /// Asserted by the caller, bypassing detection.
-    Asserted,
-}
-
-/// A byte-offset range within the analyzed text.
-///
-/// Attached to a [`LanguageDetection`] when the detector knows the
-/// span its answer covers (mixed-language input produces multiple
-/// detections, each with a distinct span). Single-language
-/// detections from non-segmenting backends, and caller-asserted
-/// answers, typically leave the span as `None`.
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-#[derive(Serialize, Deserialize)]
-pub struct LanguageSpan {
-    /// Byte offset of the span start in the original text.
-    pub start: usize,
-    /// Byte offset of the span end in the original text.
-    pub end: usize,
-}
-
-/// A single language detection result.
-///
-/// Carries the detected language plus an optional confidence and an
-/// optional byte-offset [`LanguageSpan`]. Backends that don't expose
-/// confidence (or where confidence isn't meaningful) leave it as
-/// `None`; single-language detectors that don't track per-region
-/// information leave `span` as `None`.
-///
-/// The `provenance` field records whether this answer came from a
-/// real detector run or was asserted by the caller; backends only
-/// ever produce [`LanguageProvenance::Detected`], with `Asserted`
-/// reserved for callers that bypass detection.
-#[derive(Debug, Clone, PartialEq)]
-#[derive(Serialize, Deserialize)]
-#[serde(rename_all = "camelCase")]
-pub struct LanguageDetection {
-    /// The detected language.
-    pub language: LanguageTag,
-    /// Optional confidence score. `None` when the backend doesn't
-    /// expose one.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub confidence: Option<Confidence>,
-    /// How this language was obtained: detected or caller-asserted.
-    pub provenance: LanguageProvenance,
-    /// Byte-offset range this detection applies to, when the backend
-    /// reports per-region detections. Single-language detectors that
-    /// answer "the whole text is X" leave this `None`.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub span: Option<LanguageSpan>,
-}
diff --git a/crates/nvisy-core/src/primitive/language/detections.rs b/crates/nvisy-core/src/primitive/language/detections.rs
deleted file mode 100644
index 08508a21c..000000000
--- a/crates/nvisy-core/src/primitive/language/detections.rs
+++ /dev/null
@@ -1,72 +0,0 @@
-//! [`LanguageDetections`]: typed wrapper around a list of
-//! [`LanguageDetection`] suitable for storage in a typed-map
-//! artifact bundle.
-//!
-//! Producers (language-detection backends in `nvisy-ner` and other
-//! NLP crates) construct one of these for the text they scan and
-//! insert it on an [`Artifacts`] map; consumers that care about
-//! language (recognizers gating on a language hint, per-language
-//! stopword resolution) fetch by type.
-//!
-//! Whole-document detections store one entry with `span = None`;
-//! multi-language documents store one entry per language with the
-//! byte range covered.
-//!
-//! [`Artifacts`]: crate::extraction::Artifacts
-
-use super::LanguageDetection;
-
-/// Languages a language-detection backend resolved for one text scan.
-///
-/// Newtype around `Vec<LanguageDetection>` so the type-map sees a
-/// distinct typed entry.
-#[derive(Debug, Clone, Default, PartialEq)]
-pub struct LanguageDetections(pub Vec<LanguageDetection>);
-
-impl LanguageDetections {
-    /// Construct from a list of detections.
-    #[must_use]
-    pub fn new(detections: Vec<LanguageDetection>) -> Self {
-        Self(detections)
-    }
-
-    /// Borrow the underlying detections.
-    #[must_use]
-    pub fn as_slice(&self) -> &[LanguageDetection] {
-        &self.0
-    }
-
-    /// The language covering the most bytes of the source text,
-    /// breaking ties on detector confidence.
-    ///
-    /// Monolingual docs return the single detection; mixed-language
-    /// docs return the largest-coverage span; caller-asserted
-    /// languages (no `span`) are treated as covering the whole
-    /// document and therefore win against any one region.
-    ///
-    /// Returns `None` iff the list is empty.
-    pub fn dominant(&self) -> Option<&LanguageDetection> {
-        self.0.iter().max_by(|a, b| {
-            span_bytes(a)
-                .cmp(&span_bytes(b))
-                .then_with(|| confidence_key(a).total_cmp(&confidence_key(b)))
-        })
-    }
-}
-
-impl From<Vec<LanguageDetection>> for LanguageDetections {
-    fn from(detections: Vec<LanguageDetection>) -> Self {
-        Self::new(detections)
-    }
-}
-
-fn span_bytes(d: &LanguageDetection) -> usize {
-    match d.span {
-        Some(s) => s.end.saturating_sub(s.start),
-        None => usize::MAX,
-    }
-}
-
-fn confidence_key(d: &LanguageDetection) -> f64 {
-    d.confidence.map(|c| c.get()).unwrap_or(f64::NEG_INFINITY)
-}
diff --git a/crates/nvisy-core/src/primitive/language/mod.rs b/crates/nvisy-core/src/primitive/language/mod.rs
deleted file mode 100644
index 589052d1f..000000000
--- a/crates/nvisy-core/src/primitive/language/mod.rs
+++ /dev/null
@@ -1,15 +0,0 @@
-//! Language-related primitives.
-//!
-//! BCP-47 [`LanguageTag`] plus [`LanguageDetection`] — the result
-//! shape every language-detection backend returns, with provenance
-//! and an optional per-region byte-offset span — plus
-//! [`LanguageDetections`], the typed-map artifact key that bundles
-//! the per-document detection set.
-
-mod detection;
-mod detections;
-mod tag;
-
-pub use self::detection::{LanguageDetection, LanguageProvenance, LanguageSpan};
-pub use self::detections::LanguageDetections;
-pub use self::tag::LanguageTag;
diff --git a/crates/nvisy-core/src/primitive/language/tag.rs b/crates/nvisy-core/src/primitive/language/tag.rs
deleted file mode 100644
index 412dc65b4..000000000
--- a/crates/nvisy-core/src/primitive/language/tag.rs
+++ /dev/null
@@ -1,113 +0,0 @@
-//! BCP-47 language tag type.
-
-use derive_more::{Display, FromStr};
-use serde::{Deserialize, Serialize};
-
-use crate::Error;
-
-/// A validated [BCP-47] language tag.
-///
-/// Wraps [`LanguageTag`] with serde support. Use `#[schemars(with =
-/// "String")]` on fields of this type for JSON Schema generation.
-///
-/// # Examples
-///
-/// ```
-/// use nvisy_core::primitive::LanguageTag;
-///
-/// let tag: LanguageTag = "en-US".parse().unwrap();
-/// assert_eq!(tag.as_str(), "en-US");
-/// assert_eq!(tag.primary_language(), "en");
-/// ```
-///
-/// [`LanguageTag`]: oxilangtag::LanguageTag
-/// [BCP-47]: https://www.rfc-editor.org/info/bcp47
-#[derive(Debug, Clone, PartialEq, Eq, Hash, Display, FromStr)]
-#[derive(Serialize, Deserialize)]
-#[serde(transparent)]
-pub struct LanguageTag(oxilangtag::LanguageTag<String>);
-
-impl LanguageTag {
-    /// Parse a BCP-47 language tag from a string.
-    ///
-    /// Convenience over the `FromStr` impl when the input is
-    /// already a `&str` literal.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when `tag` is not a valid
-    /// BCP-47 tag.
-    pub fn new(tag: &str) -> Result<Self, Error> {
-        tag.parse().map_err(|e| {
-            Error::validation(
-                format!("invalid BCP-47 language tag `{tag}`: {e}"),
-                "nvisy-core",
-            )
-        })
-    }
-
-    /// Returns the tag as a string slice.
-    pub fn as_str(&self) -> &str {
-        self.0.as_str()
-    }
-
-    /// Returns the primary language subtag (e.g. `"en"` from `"en-US"`).
-    pub fn primary_language(&self) -> &str {
-        self.0.primary_language()
-    }
-
-    /// Return `true` when `self` and `other` share the same
-    /// primary language subtag.
-    ///
-    /// Compares only the primary subtag, so `"en"`, `"en-US"`, and
-    /// `"en-GB"` all match each other; `"en"` does not match
-    /// `"de"`. ASCII case-insensitive (BCP-47 tags are
-    /// case-insensitive by spec).
-    ///
-    /// # Examples
-    ///
-    /// ```
-    /// use nvisy_core::primitive::LanguageTag;
-    ///
-    /// let en = LanguageTag::new("en").unwrap();
-    /// let en_us = LanguageTag::new("en-US").unwrap();
-    /// let de = LanguageTag::new("de").unwrap();
-    ///
-    /// assert!(en.matches(&en_us));
-    /// assert!(en_us.matches(&en));
-    /// assert!(!en.matches(&de));
-    /// ```
-    #[must_use]
-    pub fn matches(&self, other: &Self) -> bool {
-        self.primary_language()
-            .eq_ignore_ascii_case(other.primary_language())
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn tag(s: &str) -> LanguageTag {
-        LanguageTag::new(s).expect("valid BCP-47 tag")
-    }
-
-    #[test]
-    fn matches_same_primary_subtag() {
-        assert!(tag("en").matches(&tag("en-US")));
-        assert!(tag("en-US").matches(&tag("en")));
-        assert!(tag("en-US").matches(&tag("en-GB")));
-        assert!(tag("en").matches(&tag("en")));
-    }
-
-    #[test]
-    fn matches_rejects_distinct_primary_subtags() {
-        assert!(!tag("en").matches(&tag("de")));
-        assert!(!tag("en-US").matches(&tag("de-DE")));
-    }
-
-    #[test]
-    fn matches_is_case_insensitive() {
-        assert!(tag("EN").matches(&tag("en-us")));
-    }
-}
diff --git a/crates/nvisy-core/src/primitive/mod.rs b/crates/nvisy-core/src/primitive/mod.rs
deleted file mode 100644
index 1744a0b79..000000000
--- a/crates/nvisy-core/src/primitive/mod.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-//! Primitive types used across the ontology.
-//!
-//! Submodules group related primitives:
-//! - `confidence` — [`Confidence`] scores + [`ConfidenceThreshold`]
-//!   cutoffs.
-//! - `geometry` — 2D coordinate types ([`BoundingBox`], [`Polygon`]).
-//! - `language` — language tags + per-region detection results
-//!   ([`LanguageTag`], [`LanguageDetection`]).
-//! - `rendering` — output-side knobs ([`Color`], [`Dpi`]).
-//!
-//! Root-level primitives: [`CountryCode`] (ISO 3166-1 alpha-2),
-//! [`TimeSpan`] (temporal intervals).
-
-mod confidence;
-mod country;
-mod geometry;
-mod language;
-mod rendering;
-mod time_span;
-
-pub use self::confidence::{Confidence, ConfidenceThreshold};
-pub use self::country::CountryCode;
-pub use self::geometry::{
-    BoundingBox, Dimensions, IBoundingBox, NormalizedBoundingBox, Polygon, Vertex,
-};
-pub use self::language::{
-    LanguageDetection, LanguageDetections, LanguageProvenance, LanguageSpan, LanguageTag,
-};
-pub use self::rendering::{Color, Dpi};
-pub use self::time_span::TimeSpan;
diff --git a/crates/nvisy-core/src/primitive/rendering/color.rs b/crates/nvisy-core/src/primitive/rendering/color.rs
deleted file mode 100644
index 0813017bf..000000000
--- a/crates/nvisy-core/src/primitive/rendering/color.rs
+++ /dev/null
@@ -1,28 +0,0 @@
-//! RGB color type for image redaction.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// An RGB color value.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Color {
-    /// Red channel (`0..=255`).
-    pub r: u8,
-    /// Green channel (`0..=255`).
-    pub g: u8,
-    /// Blue channel (`0..=255`).
-    pub b: u8,
-}
-
-impl Color {
-    /// Opaque black (`#000000`). The redaction default — image
-    /// handlers fall back to this when an operator omits a fill.
-    pub const BLACK: Self = Self { r: 0, g: 0, b: 0 };
-}
-
-impl Default for Color {
-    fn default() -> Self {
-        Self::BLACK
-    }
-}
diff --git a/crates/nvisy-core/src/primitive/rendering/dpi.rs b/crates/nvisy-core/src/primitive/rendering/dpi.rs
deleted file mode 100644
index 26ecc576c..000000000
--- a/crates/nvisy-core/src/primitive/rendering/dpi.rs
+++ /dev/null
@@ -1,44 +0,0 @@
-//! Dots-per-inch resolution type.
-
-use derive_more::{Display, From, Into};
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Dots-per-inch resolution for rasterisation and rendering.
-///
-/// Wraps a [`u16`] to prevent accidental misuse of raw integers as DPI
-/// values. Common presets are available as associated constants.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
-#[derive(From, Into, Display, Serialize, Deserialize, JsonSchema)]
-#[display("{_0} dpi")]
-pub struct Dpi(u16);
-
-impl Dpi {
-    /// High-quality OCR resolution: 300 DPI.
-    pub const OCR: Self = Self(300);
-    /// Standard print resolution: 150 DPI.
-    pub const PRINT: Self = Self(150);
-    /// Screen resolution (72 DPI): matches PDF point units.
-    pub const SCREEN: Self = Self(72);
-
-    /// Create a DPI value from a raw `u16`.
-    pub const fn new(value: u16) -> Self {
-        Self(value)
-    }
-
-    /// Return the raw numeric value.
-    pub const fn value(self) -> u16 {
-        self.0
-    }
-
-    /// Compute the scale factor relative to PDF points (1 pt = 1/72 in).
-    pub fn scale_factor(self) -> f32 {
-        self.0 as f32 / Self::SCREEN.0 as f32
-    }
-}
-
-impl Default for Dpi {
-    fn default() -> Self {
-        Self::OCR
-    }
-}
diff --git a/crates/nvisy-core/src/primitive/rendering/mod.rs b/crates/nvisy-core/src/primitive/rendering/mod.rs
deleted file mode 100644
index 2b5c3c0b9..000000000
--- a/crates/nvisy-core/src/primitive/rendering/mod.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-//! Rendering-side knobs: color, resolution.
-//!
-//! These types are consumed at the rasterisation / output boundary
-//! (image redaction overlay color, PDF / image rendering resolution).
-//! They live separately from the geometry types because they're
-//! producer-facing presentation choices rather than coordinates.
-
-mod color;
-mod dpi;
-
-pub use self::color::Color;
-pub use self::dpi::Dpi;
diff --git a/crates/nvisy-core/src/primitive/time_span.rs b/crates/nvisy-core/src/primitive/time_span.rs
deleted file mode 100644
index 294d21bc3..000000000
--- a/crates/nvisy-core/src/primitive/time_span.rs
+++ /dev/null
@@ -1,204 +0,0 @@
-//! Temporal interval type with microsecond precision.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Microseconds per second.
-const US_PER_SEC: i64 = 1_000_000;
-
-/// A time interval within an audio stream, in microseconds.
-///
-/// Uses `i64` microseconds for sample-level precision without floating
-/// point rounding. At 48kHz sample rate, one sample is ~20.8μs — well
-/// within the 1μs resolution.
-///
-/// Use [`from_secs`] and [`start_secs`]
-/// for ergonomic conversion to/from seconds.
-///
-/// [`from_secs`]: Self::from_secs
-/// [`start_secs`]: Self::start_secs
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct TimeSpan {
-    /// Start time in microseconds from the beginning of the stream.
-    pub start_us: i64,
-    /// End time in microseconds from the beginning of the stream.
-    pub end_us: i64,
-}
-
-impl TimeSpan {
-    /// Create a time span from microsecond offsets.
-    pub fn new(start_us: i64, end_us: i64) -> Self {
-        Self { start_us, end_us }
-    }
-
-    /// Create a time span from seconds (converted to microseconds).
-    pub fn from_secs(start: f64, end: f64) -> Self {
-        Self {
-            start_us: (start * US_PER_SEC as f64) as i64,
-            end_us: (end * US_PER_SEC as f64) as i64,
-        }
-    }
-
-    /// Start time in seconds.
-    pub fn start_secs(&self) -> f64 {
-        self.start_us as f64 / US_PER_SEC as f64
-    }
-
-    /// End time in seconds.
-    pub fn end_secs(&self) -> f64 {
-        self.end_us as f64 / US_PER_SEC as f64
-    }
-
-    /// Duration in microseconds.
-    pub fn duration_us(&self) -> i64 {
-        self.end_us - self.start_us
-    }
-
-    /// Duration in seconds.
-    pub fn duration_secs(&self) -> f64 {
-        self.duration_us() as f64 / US_PER_SEC as f64
-    }
-
-    /// Midpoint in microseconds.
-    pub fn midpoint_us(&self) -> i64 {
-        (self.start_us + self.end_us) / 2
-    }
-
-    /// Returns `true` if `t` (microseconds) falls within `[start, end)`.
-    pub fn contains_us(&self, t: i64) -> bool {
-        t >= self.start_us && t < self.end_us
-    }
-
-    /// Returns `true` if this span overlaps with `other`.
-    pub fn overlaps(&self, other: &TimeSpan) -> bool {
-        self.start_us < other.end_us && other.start_us < self.end_us
-    }
-
-    /// Returns the intersection of two spans, or `None` if they don't overlap.
-    pub fn intersection(&self, other: &TimeSpan) -> Option<TimeSpan> {
-        let start = self.start_us.max(other.start_us);
-        let end = self.end_us.min(other.end_us);
-        if start < end {
-            Some(TimeSpan::new(start, end))
-        } else {
-            None
-        }
-    }
-
-    /// Returns the smallest span that covers both `self` and `other`.
-    pub fn union(&self, other: &TimeSpan) -> TimeSpan {
-        TimeSpan::new(
-            self.start_us.min(other.start_us),
-            self.end_us.max(other.end_us),
-        )
-    }
-
-    /// Convert this span to a `[start_sample, end_sample)` index range
-    /// into a channel-interleaved sample buffer.
-    ///
-    /// Rounds half-up at the frame boundary, then multiplies by
-    /// `channels` so the returned indices land on frame boundaries
-    /// (no stereo channel swap on partial-frame edits).
-    pub fn sample_range(&self, sample_rate: u32, channels: u16) -> (usize, usize) {
-        let start_frame = us_to_frame(self.start_us, sample_rate);
-        let end_frame = us_to_frame(self.end_us, sample_rate);
-        let channels = channels as usize;
-        (
-            start_frame.saturating_mul(channels),
-            end_frame.saturating_mul(channels),
-        )
-    }
-}
-
-/// Convert a microsecond offset to a frame index at `sample_rate`.
-/// Half-up rounding keeps samples on either side of the boundary
-/// consistently assigned.
-fn us_to_frame(us: i64, sample_rate: u32) -> usize {
-    if us <= 0 {
-        return 0;
-    }
-    let num = (us as u128) * (sample_rate as u128) + 500_000;
-    (num / 1_000_000) as usize
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn from_secs_roundtrip() {
-        let span = TimeSpan::from_secs(1.5, 3.25);
-        assert_eq!(span.start_us, 1_500_000);
-        assert_eq!(span.end_us, 3_250_000);
-        assert!((span.start_secs() - 1.5).abs() < 1e-9);
-        assert!((span.end_secs() - 3.25).abs() < 1e-9);
-    }
-
-    #[test]
-    fn duration() {
-        let span = TimeSpan::new(1_000_000, 3_500_000);
-        assert_eq!(span.duration_us(), 2_500_000);
-        assert!((span.duration_secs() - 2.5).abs() < 1e-9);
-    }
-
-    #[test]
-    fn contains() {
-        let span = TimeSpan::new(1_000_000, 5_000_000);
-        assert!(span.contains_us(1_000_000));
-        assert!(span.contains_us(3_000_000));
-        assert!(!span.contains_us(5_000_000)); // exclusive end
-        assert!(!span.contains_us(500_000));
-    }
-
-    #[test]
-    fn overlaps() {
-        let a = TimeSpan::new(1_000_000, 3_000_000);
-        let b = TimeSpan::new(2_000_000, 4_000_000);
-        let c = TimeSpan::new(3_000_000, 5_000_000);
-        assert!(a.overlaps(&b));
-        assert!(!a.overlaps(&c)); // touching at boundary = no overlap
-    }
-
-    #[test]
-    fn intersection() {
-        let a = TimeSpan::new(1_000_000, 4_000_000);
-        let b = TimeSpan::new(2_000_000, 5_000_000);
-        let i = a.intersection(&b).unwrap();
-        assert_eq!(i.start_us, 2_000_000);
-        assert_eq!(i.end_us, 4_000_000);
-
-        let c = TimeSpan::new(4_000_000, 6_000_000);
-        assert!(a.intersection(&c).is_none());
-    }
-
-    #[test]
-    fn union() {
-        let a = TimeSpan::new(1_000_000, 3_000_000);
-        let b = TimeSpan::new(2_000_000, 5_000_000);
-        let u = a.union(&b);
-        assert_eq!(u.start_us, 1_000_000);
-        assert_eq!(u.end_us, 5_000_000);
-    }
-
-    #[test]
-    fn sample_range_mono_aligns_to_frames() {
-        let span = TimeSpan::new(3_000, 6_000);
-        let (start, end) = span.sample_range(1000, 1);
-        assert_eq!((start, end), (3, 6));
-    }
-
-    #[test]
-    fn sample_range_stereo_aligns_to_frames() {
-        let span = TimeSpan::new(3_000, 6_000);
-        let (start, end) = span.sample_range(1000, 2);
-        assert_eq!((start, end), (6, 12));
-    }
-
-    #[test]
-    fn sample_range_clamps_negative_to_zero() {
-        let span = TimeSpan::new(-1_000, 2_000);
-        let (start, end) = span.sample_range(1000, 1);
-        assert_eq!((start, end), (0, 2));
-    }
-}
diff --git a/crates/nvisy-core/src/recognition/hint.rs b/crates/nvisy-core/src/recognition/hint.rs
deleted file mode 100644
index 1f9fa9c7d..000000000
--- a/crates/nvisy-core/src/recognition/hint.rs
+++ /dev/null
@@ -1,65 +0,0 @@
-//! [`Hint<M>`]: uploader-supplied annotation region in modality-native
-//! coordinates.
-//!
-//! A hint marks a span the uploader believes might contain a
-//! sensitive entity (and optionally claims an entity kind and a
-//! display name for it). Recognizers that support hint adjudication
-//! — typically LLM-based ones — fold hints into their detection pass
-//! so the model can confirm, relocate, or implicitly reject each one
-//! alongside open-ended discovery. Recognizers that don't (pattern,
-//! dictionary, generic NER backends) ignore [`RecognizerInput::hints`]
-//! entirely.
-//!
-//! `Hint<M>` mirrors [`Entity<M>`] structurally: both carry a
-//! modality-native coordinate. The difference is provenance — an
-//! entity is a recognizer's *output* with a confidence and a trail;
-//! a hint is a recognizer's *input* with no trail yet.
-//!
-//! [`Entity<M>`]: crate::entity::Entity
-//! [`RecognizerInput::hints`]: super::RecognizerInput::hints
-
-use crate::entity::EntityLabelRef;
-use crate::modality::Modality;
-
-/// Uploader-supplied annotation region in modality-native
-/// coordinates (`Text` byte range, `Image` bounding box, etc).
-#[derive(Debug, Clone, PartialEq)]
-pub struct Hint<M: Modality> {
-    /// Uploader-supplied display name (optional). Recognizers that
-    /// confirm or relocate this hint forward the name into the
-    /// emitted entity's recognition trail step.
-    pub name: Option<String>,
-    /// Uploader-claimed label (optional). When set, recognizers
-    /// that confirm the hint stamp this on the emitted entity's
-    /// `label` field.
-    pub label: Option<EntityLabelRef>,
-    /// Region in modality-native coordinates.
-    pub location: M::Location,
-}
-
-impl<M: Modality> Hint<M> {
-    /// Construct a hint with only the location set; name and label
-    /// default to `None`.
-    #[must_use]
-    pub fn new(location: M::Location) -> Self {
-        Self {
-            name: None,
-            label: None,
-            location,
-        }
-    }
-
-    /// Attach an uploader-supplied display name.
-    #[must_use]
-    pub fn with_name(mut self, name: impl Into<String>) -> Self {
-        self.name = Some(name.into());
-        self
-    }
-
-    /// Attach an uploader-claimed label.
-    #[must_use]
-    pub fn with_label(mut self, label: impl Into<EntityLabelRef>) -> Self {
-        self.label = Some(label.into());
-        self
-    }
-}
diff --git a/crates/nvisy-core/src/recognition/input.rs b/crates/nvisy-core/src/recognition/input.rs
deleted file mode 100644
index 6da30fd78..000000000
--- a/crates/nvisy-core/src/recognition/input.rs
+++ /dev/null
@@ -1,202 +0,0 @@
-//! [`RecognizerInput<M>`]: per-call input for an
-//! [`EntityRecognizer<M>`].
-//!
-//! Flat per-call surface for recognizers: the modality payload plus
-//! the per-call concerns recognizers actually use (language hints,
-//! candidate-language whitelist, uploader-supplied [`Hint<M>`]
-//! regions, document-level labels, correlation id, and the typed
-//! [`Artifacts`] bundle for shared NLP enrichment).
-//!
-//! Location is intentionally absent — recognizers emit entities in
-//! modality-local coordinates and don't read a per-call location.
-//! Extractors use [`Span<M>`] instead, which carries the location.
-//!
-//! [`Artifacts`]: crate::extraction::Artifacts
-//! [`EntityRecognizer<M>`]: super::EntityRecognizer
-//! [`Span<M>`]: crate::extraction::Span
-
-use uuid::Uuid;
-
-use super::Hint;
-use crate::extraction::Artifacts;
-use crate::modality::Modality;
-use crate::primitive::{CountryCode, LanguageTag};
-
-/// Per-call input for an [`EntityRecognizer<M>`].
-///
-/// Bundles the modality payload with the per-call concerns
-/// recognizers actually use; location lives on the
-/// extractor-side [`Span<M>`] instead since recognizers don't need
-/// it.
-///
-/// [`EntityRecognizer<M>`]: super::EntityRecognizer
-/// [`Span<M>`]: crate::extraction::Span
-#[derive(Debug)]
-pub struct RecognizerInput<M: Modality> {
-    /// Modality-specific payload (text bytes, image bytes + dims,
-    /// …).
-    pub data: M::Data,
-    /// Shared per-call NLP enrichment, keyed by Rust type.
-    /// Recognizers that don't care leave it empty.
-    pub artifacts: Artifacts,
-    /// Caller-asserted language. When `Some`, recognizers that
-    /// support per-call language hinting (typically NER / LLM
-    /// backends) skip their own detection.
-    pub language: Option<LanguageTag>,
-    /// Restrict language auto-detection to this subset when
-    /// [`language`] is `None`. Empty means "any".
-    ///
-    /// [`language`]: Self::language
-    pub candidate_languages: Vec<LanguageTag>,
-    /// Caller-asserted jurisdiction. When `Some`, recognizers
-    /// that carry per-rule `countries` scopes skip rules that
-    /// don't match. `None` means "any" — rules that declare
-    /// countries still run as a permissive fallback so callers
-    /// who don't pass a hint don't lose detections.
-    pub country: Option<CountryCode>,
-    /// Uploader-supplied hint regions in modality-native coordinates.
-    /// Recognizers that support hint adjudication (LLM-based NER, VLM)
-    /// read this; recognizers that don't (pattern, dictionary) ignore
-    /// it.
-    pub hints: Vec<Hint<M>>,
-    /// Document-level classification labels (e.g. `"medical"`,
-    /// `"gdpr-request"`). Recognizers may use these to bias their
-    /// behavior for domain-specific terms; those that don't ignore the
-    /// field.
-    pub labels: Vec<String>,
-    /// Out-of-band context strings the caller wants treated as
-    /// in-context for confidence boosting (e.g. the column header
-    /// of a CSV cell, the JSON object key of a string value, the
-    /// log field name a value sits under). Recognizers that run a
-    /// context enhancer feed these to the enhancer alongside the
-    /// in-text word window; recognizers without an enhancer ignore
-    /// the field.
-    pub context_hints: Vec<String>,
-    /// Correlation UUID propagated through the tracing span for this
-    /// call. Recognizer bodies do not read this directly; it's set
-    /// on the span by the caller.
-    pub correlation_id: Option<Uuid>,
-}
-
-impl<M: Modality> RecognizerInput<M> {
-    /// Construct an input with only the modality payload set;
-    /// every other field defaults to empty.
-    pub fn new(data: M::Data) -> Self {
-        Self {
-            data,
-            artifacts: Artifacts::new(),
-            language: None,
-            candidate_languages: Vec::new(),
-            country: None,
-            hints: Vec::new(),
-            labels: Vec::new(),
-            context_hints: Vec::new(),
-            correlation_id: None,
-        }
-    }
-
-    /// Replace the artifacts bundle.
-    #[must_use]
-    pub fn with_artifacts(mut self, artifacts: Artifacts) -> Self {
-        self.artifacts = artifacts;
-        self
-    }
-
-    /// Set the asserted language.
-    #[must_use]
-    pub fn with_language(mut self, language: LanguageTag) -> Self {
-        self.language = Some(language);
-        self
-    }
-
-    /// Set the candidate languages for auto-detection.
-    #[must_use]
-    pub fn with_candidate_languages(mut self, languages: Vec<LanguageTag>) -> Self {
-        self.candidate_languages = languages;
-        self
-    }
-
-    /// Set the asserted jurisdiction.
-    #[must_use]
-    pub fn with_country(mut self, country: CountryCode) -> Self {
-        self.country = Some(country);
-        self
-    }
-
-    /// Attach uploader-supplied hint regions.
-    #[must_use]
-    pub fn with_hints(mut self, hints: Vec<Hint<M>>) -> Self {
-        self.hints = hints;
-        self
-    }
-
-    /// Attach document-level classification labels.
-    #[must_use]
-    pub fn with_labels(mut self, labels: Vec<String>) -> Self {
-        self.labels = labels;
-        self
-    }
-
-    /// Attach out-of-band context hint strings (column headers,
-    /// JSON keys, …) the enhancer should treat as in-context.
-    #[must_use]
-    pub fn with_context_hints(mut self, hints: Vec<String>) -> Self {
-        self.context_hints = hints;
-        self
-    }
-
-    /// Set the correlation id propagated through the tracing span.
-    #[must_use]
-    pub fn with_correlation_id(mut self, id: Uuid) -> Self {
-        self.correlation_id = Some(id);
-        self
-    }
-
-    /// Whether a recognizer rule scoped to `allowed` languages
-    /// should run for this call.
-    ///
-    /// - An empty `allowed` list means the rule is language-agnostic
-    ///   and always runs.
-    /// - When `allowed` is non-empty and [`language`] is `Some(_)`,
-    ///   the rule runs when the hint shares a primary subtag with
-    ///   any entry in `allowed` (so an `["en"]` rule fires for
-    ///   `"en-US"` and `"en-GB"` hints).
-    /// - When [`language`] is `None`, the rule still runs — we can't
-    ///   disprove applicability without a hint.
-    ///
-    /// [`language`]: Self::language
-    #[must_use]
-    pub fn applies_to_language(&self, allowed: &[LanguageTag]) -> bool {
-        if allowed.is_empty() {
-            return true;
-        }
-        match self.language.as_ref() {
-            Some(hint) => allowed.iter().any(|a| a.matches(hint)),
-            None => true,
-        }
-    }
-
-    /// Whether a recognizer rule scoped to `allowed` countries
-    /// should run for this call.
-    ///
-    /// - An empty `allowed` list means the rule is jurisdiction-
-    ///   agnostic and always runs.
-    /// - When `allowed` is non-empty and [`country`] is `Some(_)`,
-    ///   the rule runs only when the hint is in `allowed`.
-    /// - When [`country`] is `None`, the rule still runs — we
-    ///   can't disprove applicability without a hint, and
-    ///   silently dropping detections would surprise callers
-    ///   who simply forgot to set the field.
-    ///
-    /// [`country`]: Self::country
-    #[must_use]
-    pub fn applies_to_country(&self, allowed: &[CountryCode]) -> bool {
-        if allowed.is_empty() {
-            return true;
-        }
-        match self.country.as_ref() {
-            Some(hint) => allowed.iter().any(|a| a == hint),
-            None => true,
-        }
-    }
-}
diff --git a/crates/nvisy-core/src/recognition/label_map.rs b/crates/nvisy-core/src/recognition/label_map.rs
deleted file mode 100644
index b19fcd34d..000000000
--- a/crates/nvisy-core/src/recognition/label_map.rs
+++ /dev/null
@@ -1,156 +0,0 @@
-//! [`LabelMap`]: backend label → canonical label-name translation.
-//!
-//! Shared translation table used by every model-driven recognizer
-//! (NER backends, LLM recognizers, …). Lets a recognizer consume
-//! raw model labels uniformly regardless of which backend produced
-//! them — swap backends without re-implementing translation.
-//!
-//! The map is bidirectional in spirit (look up an entity label
-//! name to find the canonical model label a backend should be asked
-//! for) but the primary path is model-label → entity-label-name.
-//! The reverse lookup is a linear scan.
-
-use std::borrow::Cow;
-use std::collections::HashMap;
-
-use serde::{Deserialize, Serialize};
-
-use crate::entity::{EntityLabelCatalog, EntityLabelRef};
-
-/// Translation table from raw model labels to entity label names.
-///
-/// The default ([`LabelMap::canonical`]) maps every name in a
-/// [`EntityLabelCatalog`] to itself, so backends that already return canonical
-/// names pass through unchanged. Custom backends register their
-/// own model-specific labels via [`with_entry`].
-///
-/// [`with_entry`]: Self::with_entry
-#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
-#[serde(transparent)]
-pub struct LabelMap {
-    entries: HashMap<String, EntityLabelRef>,
-}
-
-impl LabelMap {
-    /// Empty map. Backends with no recognizable labels see every
-    /// span dropped — typically you want [`Self::canonical`] /
-    /// [`Self::canonical_from`] instead.
-    #[must_use]
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Identity map from every label name in the workspace
-    /// built-in [`EntityLabelCatalog`] to itself. Convenience wrapper around
-    /// [`Self::canonical_from`] for callers that don't need custom
-    /// labels.
-    #[must_use]
-    pub fn canonical() -> Self {
-        Self::canonical_from(&EntityLabelCatalog::with_builtins())
-    }
-
-    /// Identity map over every name in the supplied catalog.
-    /// Backends that already return canonical names — or that
-    /// have been pre-registered with the catalog — pass through
-    /// unchanged.
-    #[must_use]
-    pub fn canonical_from(catalog: &EntityLabelCatalog) -> Self {
-        let entries = catalog
-            .iter()
-            .map(|label| {
-                (
-                    label.name.to_string(),
-                    EntityLabelRef::from(label.name.clone()),
-                )
-            })
-            .collect();
-        Self { entries }
-    }
-
-    /// Register one model-label → entity-label-ref entry. Last
-    /// write wins on duplicates.
-    #[must_use]
-    pub fn with_entry(
-        mut self,
-        model_label: impl Into<Cow<'static, str>>,
-        entity_label: impl Into<EntityLabelRef>,
-    ) -> Self {
-        self.entries
-            .insert(model_label.into().into_owned(), entity_label.into());
-        self
-    }
-
-    /// Register many entries.
-    #[must_use]
-    pub fn with_entries<I, K, V>(mut self, entries: I) -> Self
-    where
-        I: IntoIterator<Item = (K, V)>,
-        K: Into<String>,
-        V: Into<EntityLabelRef>,
-    {
-        for (model_label, entity_label) in entries {
-            self.entries.insert(model_label.into(), entity_label.into());
-        }
-        self
-    }
-
-    /// Look up a raw model label. `None` when not registered.
-    #[must_use]
-    pub fn lookup(&self, model_label: &str) -> Option<&EntityLabelRef> {
-        self.entries.get(model_label)
-    }
-
-    /// Find a model label that maps to the given entity label
-    /// name. Linear scan; returns the first match. Used by
-    /// zero-shot backends that need to format requested-labels as
-    /// raw model labels for the service.
-    #[must_use]
-    pub fn model_label_for(&self, entity_label: &str) -> Option<&str> {
-        self.entries
-            .iter()
-            .find_map(|(m, e)| (e.as_str() == entity_label).then_some(m.as_str()))
-    }
-
-    /// Number of registered entries.
-    #[must_use]
-    pub fn len(&self) -> usize {
-        self.entries.len()
-    }
-
-    /// Whether the map is empty.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.entries.is_empty()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use crate::entity::builtins;
-
-    #[test]
-    fn canonical_map_resolves_known_labels() {
-        let map = LabelMap::canonical();
-        assert_eq!(
-            map.lookup("email_address").map(|r| r.as_str()),
-            Some("email_address")
-        );
-        assert!(map.lookup("ssn").is_none());
-    }
-
-    #[test]
-    fn custom_entries_override_canonical() {
-        let map = LabelMap::canonical().with_entry(
-            "PER",
-            EntityLabelRef::from(builtins::PERSON_NAME.name.clone()),
-        );
-        assert_eq!(map.lookup("PER").map(|r| r.as_str()), Some("person_name"));
-    }
-
-    #[test]
-    fn model_label_for_round_trips() {
-        let map = LabelMap::canonical();
-        assert_eq!(map.model_label_for("email_address"), Some("email_address"));
-    }
-}
diff --git a/crates/nvisy-core/src/recognition/mod.rs b/crates/nvisy-core/src/recognition/mod.rs
deleted file mode 100644
index 25786aafa..000000000
--- a/crates/nvisy-core/src/recognition/mod.rs
+++ /dev/null
@@ -1,63 +0,0 @@
-//! [`EntityRecognizer<M>`]: the entity-detection trait.
-//!
-//! Every detector that emits [`Entity<M>`] for some modality `M`
-//! implements this trait — pattern recognizers, NER bento clients,
-//! LLM agents, OCR pipelines, plus any third-party recognizer a
-//! consumer wires into their pipeline. Object-safe so heterogeneous
-//! recognizers live behind `Arc<dyn EntityRecognizer<M>>` in
-//! consumer-side registries.
-//!
-//! # Layering
-//!
-//! - [`Modality`] extends [`Modality`] with an associated `Data`
-//!   type — the modality-specific payload (text bytes, image bytes +
-//!   dims, …) recognizers actually scan.
-//! - [`RecognizerInput<M>`] carries the payload plus the per-call
-//!   concerns recognizers use ([`Artifacts`], language hints,
-//!   candidate-language whitelist, uploader-supplied [`Hint<M>`]
-//!   regions, document-level labels, correlation id). Location is
-//!   intentionally absent — recognizers emit modality-local
-//!   coordinates and don't read a per-call location; extractors use
-//!   [`Span<M>`] instead.
-//! - [`EntityRecognizer<M>`] takes `&RecognizerInput<M>` and emits a
-//!   [`RecognizerOutput<M>`].
-//!
-//! [`Artifacts`]: crate::extraction::Artifacts
-//! [`Entity<M>`]: crate::entity::Entity
-//! [`Modality`]: crate::modality::Modality
-//! [`Span<M>`]: crate::extraction::Span
-
-mod hint;
-mod input;
-mod label_map;
-mod output;
-
-pub use self::hint::Hint;
-pub use self::input::RecognizerInput;
-pub use self::label_map::LabelMap;
-pub use self::output::RecognizerOutput;
-use crate::Result;
-use crate::modality::Modality;
-
-/// Recognizer for a single [`Modality`] `M`.
-///
-/// Implementors emit a [`RecognizerOutput<M>`] for one document or
-/// one scan unit, reading whatever per-call configuration they need
-/// from [`RecognizerInput<M>`]. Each consumer composes their own
-/// list of recognizers; the trait does not assume a central
-/// registry.
-///
-/// Recognizers are expected to be stateless across calls. Any
-/// per-document state a long-lived implementation needs is its
-/// own responsibility to clear (e.g. an `Arc<Mutex<…>>` reset at
-/// the top of `recognize`).
-///
-/// [`Modality`]: crate::modality::Modality
-#[async_trait::async_trait]
-pub trait EntityRecognizer<M: Modality>: Send + Sync {
-    /// Detect entities in `input` and return them in modality-local
-    /// coordinates. Downstream callers rebase text offsets into
-    /// document coordinates when stitching results back into a
-    /// multi-block document; image entities pass through unchanged.
-    async fn recognize(&self, input: &RecognizerInput<M>) -> Result<RecognizerOutput<M>>;
-}
diff --git a/crates/nvisy-core/src/recognition/output.rs b/crates/nvisy-core/src/recognition/output.rs
deleted file mode 100644
index 02938b89c..000000000
--- a/crates/nvisy-core/src/recognition/output.rs
+++ /dev/null
@@ -1,53 +0,0 @@
-//! [`RecognizerOutput<M>`]: per-call output of an
-//! [`EntityRecognizer<M>`].
-//!
-//! Wraps the emitted entities in a named struct (rather than a bare
-//! `Vec<Entity<M>>`) so future per-call metadata — drop counters,
-//! telemetry, partial-failure flags — can land alongside without
-//! churning every recognizer signature.
-//!
-//! [`EntityRecognizer<M>`]: super::EntityRecognizer
-
-use crate::entity::Entity;
-use crate::modality::Modality;
-
-/// Per-call output of an [`EntityRecognizer`].
-///
-/// [`EntityRecognizer`]: super::EntityRecognizer
-#[derive(Debug, Clone)]
-pub struct RecognizerOutput<M: Modality> {
-    /// Entities the recognizer emitted in modality-local coordinates.
-    pub entities: Vec<Entity<M>>,
-}
-
-impl<M: Modality> RecognizerOutput<M> {
-    /// Construct from the underlying entity list.
-    #[must_use]
-    pub fn new(entities: Vec<Entity<M>>) -> Self {
-        Self { entities }
-    }
-
-    /// Empty output — no entities emitted.
-    #[must_use]
-    pub fn empty() -> Self {
-        Self::new(Vec::new())
-    }
-}
-
-impl<M: Modality> Default for RecognizerOutput<M> {
-    fn default() -> Self {
-        Self::empty()
-    }
-}
-
-impl<M: Modality> From<Vec<Entity<M>>> for RecognizerOutput<M> {
-    fn from(entities: Vec<Entity<M>>) -> Self {
-        Self::new(entities)
-    }
-}
-
-impl<M: Modality> From<RecognizerOutput<M>> for Vec<Entity<M>> {
-    fn from(output: RecognizerOutput<M>) -> Self {
-        output.entities
-    }
-}
diff --git a/crates/nvisy-core/src/redaction/anonymizer.rs b/crates/nvisy-core/src/redaction/anonymizer.rs
deleted file mode 100644
index fd4e90f26..000000000
--- a/crates/nvisy-core/src/redaction/anonymizer.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! [`Anonymizer<M>`]: the per-modality redaction operator trait.
-
-use super::LeakProfile;
-use crate::Result;
-use crate::entity::Entity;
-use crate::modality::Modality;
-
-/// Per-modality redaction operator.
-///
-/// `apply` rewrites one entity into a [`Modality::Replacement`]. The
-/// engine batches calls; per-document state (counters, caches) lives
-/// on the operator instance — operators are constructed once and
-/// shared across runs. Locale-aware operators (e.g. fake-data
-/// generators) read `entity.language` to pick a locale.
-///
-/// Object-safe so registries can hold heterogeneous operators behind
-/// `Arc<dyn Anonymizer<M>>`.
-///
-/// [`Modality::Replacement`]: crate::modality::Modality::Replacement
-#[async_trait::async_trait]
-pub trait Anonymizer<M: Modality>: Send + Sync {
-    /// What the operator's output leaks about the original. Used by
-    /// policy-authoring tools and audit reporting; not consulted in
-    /// the hot path.
-    fn leak_profile(&self) -> LeakProfile;
-
-    /// Rewrite `entity` into a replacement value. `source` is the
-    /// per-document payload the recognizer scanned — the operator
-    /// uses it to read the original bytes the entity points at.
-    async fn apply(&self, entity: &Entity<M>, source: &M::Data) -> Result<M::Replacement>;
-}
diff --git a/crates/nvisy-core/src/redaction/deanonymizer.rs b/crates/nvisy-core/src/redaction/deanonymizer.rs
deleted file mode 100644
index 5e198baee..000000000
--- a/crates/nvisy-core/src/redaction/deanonymizer.rs
+++ /dev/null
@@ -1,52 +0,0 @@
-//! [`Deanonymizer<M>`]: inverse of [`Anonymizer<M>`].
-//!
-//! Recovers the original payload that was captured when an
-//! [`Entity<M>`] was anonymized. The trait accepts both the entity
-//! and the replacement the anonymizer wrote, so a single
-//! deanonymizer registry can dispatch over implementations that
-//! use either (or both) as their recovery key:
-//!
-//! - **Audit-keyed**: the original was stashed at apply-time in a
-//!   side store keyed on `entity.id`. The impl ignores
-//!   `replacement` and looks up by entity.
-//! - **Self-contained**: the recovery material lives inside the
-//!   replacement itself (e.g. an AES-256-GCM ciphertext blob).
-//!   The impl ignores `entity` and decodes `replacement`.
-//! - **Hybrid**: future operators may need both — e.g. an
-//!   encrypted blob whose key is derived from the entity id.
-//!
-//! Operators whose redaction is mathematically one-way (`Hash`,
-//! `Redact`, `Replace` without audit, `Mask` without prefix-keep,
-//! `Fake` without audit) cannot implement this trait. That's a
-//! feature, not a gap — calling `revert` on an irreversible
-//! operator should be a type error, not a runtime `None`.
-//!
-//! [`Anonymizer<M>`]: super::Anonymizer
-//! [`Entity<M>`]: crate::entity::Entity
-
-use crate::Result;
-use crate::entity::Entity;
-use crate::modality::Modality;
-
-/// Inverse of [`Anonymizer<M>`]. Given the entity that was
-/// anonymized and the replacement the operator wrote, recover the
-/// original payload.
-///
-/// Returns `Ok(None)` when there's nothing to recover — the
-/// replacement is a "removed" / "column dropped" variant, or the
-/// side store has no record for the entity. Reserve `Err` for
-/// backend failures (decryption failed, storage unreachable, …).
-///
-/// [`Anonymizer<M>`]: super::Anonymizer
-#[async_trait::async_trait]
-pub trait Deanonymizer<M: Modality>: Send + Sync {
-    /// Recover the original payload from `(entity, replacement)`.
-    /// Audit-keyed impls read `entity.id`; self-contained impls
-    /// decode `replacement`. See the module docs for which mode
-    /// applies to a given operator.
-    async fn revert(
-        &self,
-        entity: &Entity<M>,
-        replacement: &M::Replacement,
-    ) -> Result<Option<M::Data>>;
-}
diff --git a/crates/nvisy-core/src/redaction/leak_profile.rs b/crates/nvisy-core/src/redaction/leak_profile.rs
deleted file mode 100644
index 99d1cba57..000000000
--- a/crates/nvisy-core/src/redaction/leak_profile.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-//! [`LeakProfile`]: how much an operator's output leaks about the
-//! original it replaced.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// What a redacted output leaks about the original it replaced.
-///
-/// Variants are ordered from most-leaky to least-leaky, so
-/// `Recoverable < Partial < Irrecoverable`. Surfaced through
-/// [`Anonymizer::leak_profile`] for policy authoring and audit
-/// reporting.
-///
-/// [`Anonymizer::leak_profile`]: super::Anonymizer::leak_profile
-#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum LeakProfile {
-    /// The original value is recoverable from the output given the
-    /// right metadata (encryption key, token vault, pseudonym map,
-    /// or the candidate entity list against a hash).
-    Recoverable,
-    /// The original value is gone, but observable shape leaks:
-    /// position, length, bounding box, cell coordinates, or a known
-    /// silence on the timeline.
-    Partial,
-    /// No trace of the original value or its shape remains in the
-    /// output.
-    Irrecoverable,
-}
diff --git a/crates/nvisy-core/src/redaction/memoized.rs b/crates/nvisy-core/src/redaction/memoized.rs
deleted file mode 100644
index 9f96d7f5d..000000000
--- a/crates/nvisy-core/src/redaction/memoized.rs
+++ /dev/null
@@ -1,86 +0,0 @@
-//! [`Memoized`]: caches an inner [`Anonymizer`]'s output so the
-//! same input payload always yields the same replacement.
-//!
-//! The cache is a [`Store<M::Replacement>`] keyed by a stable hash
-//! of the source payload. First call for a given payload runs
-//! `inner.apply`; subsequent calls short-circuit to the cached
-//! replacement. This buys document-level coherence — every mention
-//! of the same payload gets the same fake — without requiring the
-//! recogniser to tag coreference.
-//!
-//! The wrapped operator's [`LeakProfile`] is preserved as-is. There
-//! is no [`Deanonymizer`] impl: the cached entry is the *fake*, not
-//! the original.
-//!
-//! `Memoized` is wired for every modality. Whether memoization is
-//! worth its cost depends on the inner operator — for cheap
-//! descriptors like [`ImageReplacement::Blur`] the savings are
-//! negligible; for ML-based synthesis (deepfake faces, voice
-//! cloning) they're significant.
-//!
-//! [`Anonymizer`]: super::Anonymizer
-//! [`Deanonymizer`]: super::Deanonymizer
-//! [`ImageReplacement::Blur`]: super::ImageReplacement::Blur
-//! [`Store<M::Replacement>`]: super::Store
-
-use std::hash::{DefaultHasher, Hash, Hasher};
-use std::sync::Arc;
-
-use super::store::Store;
-use super::{Anonymizer, LeakProfile};
-use crate::Result;
-use crate::entity::Entity;
-use crate::modality::Modality;
-
-/// Memoizing wrapper around any [`Anonymizer<M>`].
-#[derive(Debug, Clone)]
-pub struct Memoized<A, S> {
-    inner: A,
-    store: Arc<S>,
-}
-
-impl<A, S> Memoized<A, S> {
-    /// Build a `Memoized` wrapper around `inner`, caching into
-    /// `store`.
-    pub fn new(inner: A, store: Arc<S>) -> Self {
-        Self { inner, store }
-    }
-
-    /// Borrow the inner anonymizer.
-    pub fn inner(&self) -> &A {
-        &self.inner
-    }
-
-    /// Borrow the shared cache store.
-    pub fn store(&self) -> &Arc<S> {
-        &self.store
-    }
-}
-
-#[async_trait::async_trait]
-impl<M, A, S> Anonymizer<M> for Memoized<A, S>
-where
-    M: Modality,
-    A: Anonymizer<M>,
-    S: Store<M::Replacement> + 'static,
-{
-    fn leak_profile(&self) -> LeakProfile {
-        self.inner.leak_profile()
-    }
-
-    async fn apply(&self, entity: &Entity<M>, source: &M::Data) -> Result<M::Replacement> {
-        let key = data_token(source);
-        if let Some(cached) = self.store.get(&key).await? {
-            return Ok(cached);
-        }
-        let replacement = self.inner.apply(entity, source).await?;
-        self.store.put(&key, replacement.clone()).await?;
-        Ok(replacement)
-    }
-}
-
-fn data_token<T: Hash>(payload: &T) -> String {
-    let mut hasher = DefaultHasher::new();
-    payload.hash(&mut hasher);
-    format!("{:016x}", hasher.finish())
-}
diff --git a/crates/nvisy-core/src/redaction/mod.rs b/crates/nvisy-core/src/redaction/mod.rs
deleted file mode 100644
index e14bd0922..000000000
--- a/crates/nvisy-core/src/redaction/mod.rs
+++ /dev/null
@@ -1,52 +0,0 @@
-//! Redaction-side primitives shared across the runtime.
-//!
-//! - [`Anonymizer<M>`] + [`LeakProfile`] — the per-modality
-//!   redaction operator trait, plus the leak-profile enum operators
-//!   self-report. Built-in implementations (mask, hash, encrypt,
-//!   keep, redact, replace) ship in `nvisy-toolkit`; the fake-data
-//!   generator ships in `nvisy-fake`.
-//! - [`TextReplacement`], [`ImageReplacement`], [`AudioReplacement`],
-//!   [`TabularReplacement`] — what an anonymizer emits at the
-//!   entity's location. Each modality binds one of these via
-//!   [`Modality::Replacement`].
-//! - [`Redactions<M>`] — typed batch of `(M::Location, M::Replacement)`
-//!   pairs handed from the producer side (anonymizer / registry) to
-//!   the write-back side.
-//! - [`RedactAt<M>`] — the write-back trait every sink implements
-//!   (codec's `DocumentHandle`, engine's `DocumentTree`).
-//! - [`Deanonymizer`] — audit-keyed inverse of `Anonymizer`.
-//!   Implemented by recoverable wrappers that persist the original
-//!   at apply-time keyed on `entity.id`.
-//! - [`Store`] — pluggable token vault keyed on opaque strings.
-//!   Backends (in-memory, key-value) ship in `nvisy-toolkit`.
-//! - [`Memoized`] — wraps any inner `Anonymizer` and caches its
-//!   output so the same payload always gets the same replacement.
-//!   No recovery; preserves the inner's `LeakProfile`.
-//!
-//! The read-side siblings [`TextAt`] / [`DataAt`] live in
-//! [`extraction`].
-//!
-//! [`DataAt`]: crate::extraction::DataAt
-//! [`extraction`]: crate::extraction
-//! [`Modality::Replacement`]: crate::modality::Modality::Replacement
-//! [`TextAt`]: crate::extraction::TextAt
-
-mod anonymizer;
-mod deanonymizer;
-mod leak_profile;
-mod memoized;
-mod redact_at;
-mod redactions;
-mod replacement;
-mod store;
-
-pub use self::anonymizer::Anonymizer;
-pub use self::deanonymizer::Deanonymizer;
-pub use self::leak_profile::LeakProfile;
-pub use self::memoized::Memoized;
-pub use self::redact_at::RedactAt;
-pub use self::redactions::Redactions;
-pub use self::replacement::{
-    AudioReplacement, ImageReplacement, TabularReplacement, TextReplacement,
-};
-pub use self::store::Store;
diff --git a/crates/nvisy-core/src/redaction/redact_at.rs b/crates/nvisy-core/src/redaction/redact_at.rs
deleted file mode 100644
index 9b379a752..000000000
--- a/crates/nvisy-core/src/redaction/redact_at.rs
+++ /dev/null
@@ -1,32 +0,0 @@
-//! [`RedactAt<M>`]: write-back sibling of [`DataAt`] / [`TextAt`].
-//!
-//! Where [`DataAt`] reads a typed payload at a per-modality location
-//! and [`TextAt`] reads the display text at it, `RedactAt` writes
-//! per-modality replacements back into the underlying source.
-//!
-//! Implementations include the codec layer's typed
-//! `DocumentHandle<M>` (writes into the format-specific bytes) and
-//! the engine's `DocumentTree<M>` (forwards through the codec
-//! handle).
-//!
-//! The trait takes a batched [`Redactions<M>`] because most
-//! implementations need ordering control (right-to-left for text /
-//! audio so earlier shifts don't invalidate later coordinates,
-//! batched per page for PDF, …).
-//!
-//! [`DataAt`]: crate::extraction::DataAt
-//! [`TextAt`]: crate::extraction::TextAt
-
-use crate::Result;
-use crate::modality::Modality;
-use crate::redaction::Redactions;
-
-/// Apply a batch of `(location, replacement)` pairs to a per-modality
-/// source. Producer guarantees non-overlapping locations; the
-/// implementation reorders the batch as needed.
-#[async_trait::async_trait]
-pub trait RedactAt<M: Modality>: Send + Sync {
-    /// Apply the batch in whatever order is correct for this
-    /// implementation. The first error aborts the batch.
-    async fn redact_at(&mut self, redactions: Redactions<M>) -> Result<()>;
-}
diff --git a/crates/nvisy-core/src/redaction/redactions.rs b/crates/nvisy-core/src/redaction/redactions.rs
deleted file mode 100644
index f311f1b52..000000000
--- a/crates/nvisy-core/src/redaction/redactions.rs
+++ /dev/null
@@ -1,110 +0,0 @@
-//! [`Redactions<M>`]: collection of `(location, replacement)` pairs
-//! handed from the operator side to the write-back side.
-//!
-//! The collection is a thin newtype over `Vec<(M::Location,
-//! M::Replacement)>`. Insertion is an unconditional push — no overlap
-//! detection, no merging.
-//!
-//! # Why no dedup
-//!
-//! Upstream phases (e.g. the document crate's deduplication step)
-//! enforce that no two surviving entities have overlapping locations.
-//! By the time a batch is built, every entry's location is disjoint
-//! from every other entry's.
-//!
-//! # Ordering
-//!
-//! The producer makes no ordering guarantee.
-//! [`RedactAt::redact_at`] implementations reorder the batch as needed
-//! (right-to-left for text or audio so earlier shifts don't
-//! invalidate later coordinates, batched per page for PDF, …).
-//! Callers that need a single replacement build a one-element batch
-//! with [`Redactions::single`].
-//!
-//! [`RedactAt::redact_at`]: super::RedactAt::redact_at
-
-use std::fmt;
-
-use crate::modality::Modality;
-
-/// A list of `(location, replacement)` pairs handed to a
-/// [`RedactAt<M>`] implementation. Producer guarantees non-overlapping
-/// locations; the implementation is free to reorder for efficiency.
-///
-/// [`RedactAt<M>`]: super::RedactAt
-pub struct Redactions<M: Modality> {
-    pub(crate) items: Vec<(M::Location, M::Replacement)>,
-}
-
-impl<M: Modality> Redactions<M> {
-    /// Create an empty collection.
-    pub fn new() -> Self {
-        Self { items: Vec::new() }
-    }
-
-    /// One-element batch — the common "single replacement" convenience.
-    pub fn single(location: M::Location, replacement: M::Replacement) -> Self {
-        Self {
-            items: vec![(location, replacement)],
-        }
-    }
-
-    /// Total number of replacements queued.
-    pub fn len(&self) -> usize {
-        self.items.len()
-    }
-
-    /// Returns `true` if the collection holds no replacements.
-    pub fn is_empty(&self) -> bool {
-        self.items.is_empty()
-    }
-
-    /// Append a `(location, replacement)` pair to the batch. No
-    /// overlap check, no merge. Caller is responsible for non-overlap.
-    pub fn push(&mut self, location: M::Location, replacement: M::Replacement) {
-        self.items.push((location, replacement));
-    }
-
-    /// Consume the collection and return the underlying
-    /// `(location, replacement)` pairs in insertion order.
-    pub fn into_items(self) -> Vec<(M::Location, M::Replacement)> {
-        self.items
-    }
-
-    /// Borrow the underlying pairs in insertion order.
-    pub fn items(&self) -> &[(M::Location, M::Replacement)] {
-        &self.items
-    }
-}
-
-impl<M: Modality> Redactions<M>
-where
-    M::Location: Ord,
-{
-    /// Sort the batch right-to-left by location (descending). Used by
-    /// [`RedactAt::redact_at`] implementations whose write step
-    /// shrinks or shifts the underlying buffer so an earlier shrink
-    /// doesn't invalidate later coordinates.
-    ///
-    /// Only implemented when `M::Location: Ord` — image redaction
-    /// writes pixels in place and doesn't need this method.
-    ///
-    /// [`RedactAt::redact_at`]: super::RedactAt::redact_at
-    pub fn sort_descending(&mut self) {
-        self.items.sort_by(|(a, _), (b, _)| b.cmp(a));
-    }
-}
-
-impl<M: Modality> Default for Redactions<M> {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
-impl<M: Modality> fmt::Debug for Redactions<M> {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("Redactions")
-            .field("len", &self.len())
-            .finish()
-    }
-}
diff --git a/crates/nvisy-core/src/redaction/replacement.rs b/crates/nvisy-core/src/redaction/replacement.rs
deleted file mode 100644
index 0b988851a..000000000
--- a/crates/nvisy-core/src/redaction/replacement.rs
+++ /dev/null
@@ -1,137 +0,0 @@
-//! Per-modality [`Modality::Replacement`] types — the rich,
-//! byte-level instruction an anonymizer produces and the codec
-//! consumes.
-//!
-//! These are not "audit tags" — they carry the full parameters the
-//! codec needs to actually rewrite the underlying bytes. The audit
-//! record stores the same value, so an audit replay reproduces the
-//! redaction exactly.
-//!
-//! [`Modality::Replacement`]: super::Modality::Replacement
-
-use bytes::Bytes;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use crate::primitive::Color;
-
-/// What a text-modality anonymizer writes at the entity's location.
-#[derive(Debug, Clone, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum TextReplacement {
-    /// Substitute the span with this string. Covers `Replace`,
-    /// `Mask`, `Hash`, `Encrypt`, and every other text operator
-    /// that emits a string.
-    Substituted {
-        /// The string the codec writes over the span.
-        value: String,
-    },
-    /// Delete the span entirely. Covers `Redact`.
-    Removed,
-}
-
-impl TextReplacement {
-    /// Build a `Substituted` replacement.
-    pub fn substituted(value: impl Into<String>) -> Self {
-        Self::Substituted {
-            value: value.into(),
-        }
-    }
-
-    /// Borrow the replacement text. `Removed` answers `None`; the
-    /// caller should treat that as the empty string (span deleted).
-    pub fn replacement_value(&self) -> Option<&str> {
-        match self {
-            Self::Substituted { value } => Some(value),
-            Self::Removed => None,
-        }
-    }
-}
-
-/// What an image-modality anonymizer writes at the entity's
-/// location. Each variant carries the full parameter the codec
-/// rasterizer needs.
-#[derive(Debug, Clone, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "method", rename_all = "snake_case")]
-pub enum ImageReplacement {
-    /// Gaussian blur applied to the region with the given sigma.
-    Blur {
-        /// Standard deviation of the Gaussian kernel in pixels.
-        sigma: f32,
-    },
-    /// Opaque solid-colour block overlay on the region.
-    Block {
-        /// Fill colour the codec rasterises over the region.
-        color: Color,
-    },
-    /// Mosaic pixelation with a fixed block size in pixels.
-    Pixelate {
-        /// Side length in pixels of each mosaic block.
-        block_size: u32,
-    },
-    /// Region replaced with the supplied encoded image bytes.
-    Replace {
-        /// Encoded replacement image (PNG / JPEG / TIFF bytes).
-        data: Bytes,
-    },
-}
-
-/// What a tabular-modality anonymizer writes at the entity's
-/// location.
-///
-/// Tabular operators write a per-cell string or drop a column
-/// outright. Per-column drops are recorded once per affected
-/// column; downstream callers fan that single record out across
-/// every row of the column.
-#[derive(Debug, Clone, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum TabularReplacement {
-    /// Substitute the cell with this string.
-    Substituted {
-        /// The string the codec writes into the cell.
-        value: String,
-    },
-    /// Drop the column schema-wide.
-    ColumnDropped,
-}
-
-impl TabularReplacement {
-    /// Build a `Substituted` replacement.
-    pub fn substituted(value: impl Into<String>) -> Self {
-        Self::Substituted {
-            value: value.into(),
-        }
-    }
-
-    /// Borrow the cell's replacement text. `ColumnDropped` answers
-    /// `None`; the caller should treat that as a schema-level drop
-    /// signal, not as an empty-string substitution.
-    pub fn replacement_value(&self) -> Option<&str> {
-        match self {
-            Self::Substituted { value } => Some(value),
-            Self::ColumnDropped => None,
-        }
-    }
-}
-
-/// What an audio-modality anonymizer writes at the entity's
-/// location. Each variant carries the parameter the codec needs to
-/// produce the replacement samples.
-#[derive(Debug, Clone, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "method", rename_all = "snake_case")]
-pub enum AudioReplacement {
-    /// The time-span is silenced.
-    Silence,
-    /// The time-span is removed (downstream samples shift left).
-    Remove,
-    /// The time-span is replaced with the supplied encoded audio
-    /// bytes.
-    Replace {
-        /// Encoded replacement audio (WAV / MP3 bytes).
-        data: Bytes,
-    },
-}
diff --git a/crates/nvisy-core/src/redaction/store.rs b/crates/nvisy-core/src/redaction/store.rs
deleted file mode 100644
index f8db4421d..000000000
--- a/crates/nvisy-core/src/redaction/store.rs
+++ /dev/null
@@ -1,32 +0,0 @@
-//! [`Store<V>`]: pluggable token vault keyed on opaque strings,
-//! holding cloneable values.
-//!
-//! [`Memoized`] uses `Store<M::Replacement>` as its cache backing —
-//! same input data hashes to the same key, so the inner anonymizer
-//! runs only once per distinct payload.
-//!
-//! Implementations pick their own backing (in-memory map, KV store,
-//! KMS-backed encrypted blob).
-//!
-//! [`Memoized`]: super::Memoized
-
-use crate::Result;
-
-/// Token vault keyed on opaque strings, parameterised on the stored
-/// value type `V`.
-///
-/// Implementations must be safe to share across tasks and serve
-/// concurrent reads/writes. Keys are opaque tokens emitted by the
-/// anonymizer; values are whatever payload the operator chose to
-/// persist.
-#[async_trait::async_trait]
-pub trait Store<V: Clone + Send + Sync>: Send + Sync {
-    /// Persist `value` under `token`. Overwriting an existing token
-    /// replaces the prior value.
-    async fn put(&self, token: &str, value: V) -> Result<()>;
-
-    /// Look up the value previously stored under `token`. Returns
-    /// `Ok(None)` for unknown tokens; reserve `Err` for backend
-    /// failures.
-    async fn get(&self, token: &str) -> Result<Option<V>>;
-}
diff --git a/crates/nvisy-core/src/schema/color.rs b/crates/nvisy-core/src/schema/color.rs
new file mode 100644
index 000000000..136f6d0f0
--- /dev/null
+++ b/crates/nvisy-core/src/schema/color.rs
@@ -0,0 +1,41 @@
+//! [`ColorSchema`]: wire shape for [`elide_core::primitive::Color`].
+
+use elide_core::primitive::Color;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Wire-shape proxy for [`elide_core::primitive::Color`] — an 8-bit
+/// RGB triple. Used by [`ImageRedaction::Blackbox`] to specify the
+/// fill color.
+///
+/// [`ImageRedaction::Blackbox`]: crate::policy::redaction::ImageRedaction::Blackbox
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[schemars(rename = "Color")]
+pub struct ColorSchema {
+    /// Red channel.
+    pub r: u8,
+    /// Green channel.
+    pub g: u8,
+    /// Blue channel.
+    pub b: u8,
+}
+
+impl From<ColorSchema> for Color {
+    fn from(s: ColorSchema) -> Self {
+        Color {
+            r: s.r,
+            g: s.g,
+            b: s.b,
+        }
+    }
+}
+
+impl From<Color> for ColorSchema {
+    fn from(c: Color) -> Self {
+        Self {
+            r: c.r,
+            g: c.g,
+            b: c.b,
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/schema/geometry.rs b/crates/nvisy-core/src/schema/geometry.rs
new file mode 100644
index 000000000..172e0f608
--- /dev/null
+++ b/crates/nvisy-core/src/schema/geometry.rs
@@ -0,0 +1,71 @@
+//! Wire shapes for `elide_core::primitive` geometry types:
+//! [`PointSchema`], [`BoundingBoxSchema`], [`PolygonSchema`].
+
+use elide_core::primitive::{BoundingBox, Point, Polygon};
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Wire-shape proxy for [`elide_core::primitive::Point`].
+#[derive(Debug, Clone, Copy, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[schemars(rename = "Point")]
+pub struct PointSchema {
+    /// Horizontal coordinate.
+    pub x: f64,
+    /// Vertical coordinate.
+    pub y: f64,
+}
+
+impl From<PointSchema> for Point {
+    fn from(s: PointSchema) -> Self {
+        Point::new(s.x, s.y)
+    }
+}
+
+impl From<Point> for PointSchema {
+    fn from(p: Point) -> Self {
+        Self { x: p.x, y: p.y }
+    }
+}
+
+/// Wire-shape proxy for [`elide_core::primitive::BoundingBox`].
+#[derive(Debug, Clone, Copy, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[schemars(rename = "BoundingBox")]
+pub struct BoundingBoxSchema {
+    /// Minimum corner (top-left).
+    pub min: PointSchema,
+    /// Maximum corner (bottom-right).
+    pub max: PointSchema,
+}
+
+impl From<BoundingBoxSchema> for BoundingBox {
+    fn from(s: BoundingBoxSchema) -> Self {
+        BoundingBox::new(s.min.into(), s.max.into())
+    }
+}
+
+impl From<BoundingBox> for BoundingBoxSchema {
+    fn from(b: BoundingBox) -> Self {
+        Self {
+            min: b.min.into(),
+            max: b.max.into(),
+        }
+    }
+}
+
+/// Wire-shape proxy for [`elide_core::primitive::Polygon`].
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
+#[serde(transparent)]
+#[schemars(rename = "Polygon", transparent)]
+pub struct PolygonSchema(pub Vec<PointSchema>);
+
+impl From<PolygonSchema> for Polygon {
+    fn from(s: PolygonSchema) -> Self {
+        Polygon::new(s.0.into_iter().map(Into::into).collect::<Vec<Point>>())
+    }
+}
+
+impl From<Polygon> for PolygonSchema {
+    fn from(p: Polygon) -> Self {
+        Self(p.vertices().iter().copied().map(Into::into).collect())
+    }
+}
diff --git a/crates/nvisy-core/src/schema/label.rs b/crates/nvisy-core/src/schema/label.rs
new file mode 100644
index 000000000..cbedf9480
--- /dev/null
+++ b/crates/nvisy-core/src/schema/label.rs
@@ -0,0 +1,44 @@
+//! [`LabelSchema`]: wire shape for [`elide_core::entity::Label`].
+
+use elide_core::entity::Label;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Wire-shape proxy for [`elide_core::entity::Label`].
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+#[schemars(rename = "Label")]
+pub struct LabelSchema {
+    /// Stable identifier, e.g. `"email_address"`.
+    pub name: String,
+    /// Optional human-readable description.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub description: Option<String>,
+    /// Free-form tags policy selectors can target.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub tags: Vec<String>,
+}
+
+impl From<LabelSchema> for Label {
+    fn from(s: LabelSchema) -> Self {
+        let label = match s.description {
+            Some(desc) => Label::described(s.name, desc),
+            None => Label::new(s.name),
+        };
+        if s.tags.is_empty() {
+            label
+        } else {
+            label.with_tags(s.tags)
+        }
+    }
+}
+
+impl From<Label> for LabelSchema {
+    fn from(l: Label) -> Self {
+        Self {
+            name: l.name().to_owned(),
+            description: l.description().map(str::to_owned),
+            tags: l.tags().iter().map(|t| t.as_str().to_owned()).collect(),
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/schema/language.rs b/crates/nvisy-core/src/schema/language.rs
new file mode 100644
index 000000000..346bac239
--- /dev/null
+++ b/crates/nvisy-core/src/schema/language.rs
@@ -0,0 +1,29 @@
+//! [`LanguageTagSchema`]: wire shape for
+//! [`elide_core::primitive::LanguageTag`].
+
+use elide_core::primitive::LanguageTag;
+use elide_core::{Error, ErrorKind};
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Wire-shape proxy for [`elide_core::primitive::LanguageTag`].
+///
+/// A BCP 47 tag string (e.g. `"en"`, `"de-CH"`).
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(transparent)]
+#[schemars(rename = "LanguageTag", transparent)]
+pub struct LanguageTagSchema(pub String);
+
+impl TryFrom<LanguageTagSchema> for LanguageTag {
+    type Error = Error;
+
+    fn try_from(s: LanguageTagSchema) -> Result<Self, Self::Error> {
+        LanguageTag::parse(s.0).map_err(|e| Error::new(ErrorKind::Validation, e))
+    }
+}
+
+impl From<LanguageTag> for LanguageTagSchema {
+    fn from(t: LanguageTag) -> Self {
+        Self(t.as_str().to_owned())
+    }
+}
diff --git a/crates/nvisy-core/src/schema/mod.rs b/crates/nvisy-core/src/schema/mod.rs
new file mode 100644
index 000000000..f3ce07398
--- /dev/null
+++ b/crates/nvisy-core/src/schema/mod.rs
@@ -0,0 +1,30 @@
+//! Wire-format proxies for elide-core types embedded in nvisy-core
+//! shapes, plus round-trip [`From`] / [`Into`] conversions to bridge
+//! them back to the elide vocabulary at request time.
+//!
+//! Elide-core stays free of `schemars` — schema generation is an
+//! HTTP-layer concern the toolkit doesn't model. These proxies *are*
+//! the wire shape nvisy-core types use; engine converts at the seam
+//! when it needs an elide-typed value.
+//!
+//! Each proxy is named `<ElideName>Schema` (e.g. [`LabelSchema`]
+//! mirrors [`elide_core::entity::Label`]). The proxies derive
+//! [`schemars::JsonSchema`] plus serde + clone basics; their
+//! conversions are pure [`From`] / [`Into`] (or [`TryFrom`] when the
+//! elide constructor is fallible).
+
+mod color;
+mod geometry;
+mod label;
+mod language;
+mod operator;
+mod time;
+mod waveform;
+
+pub use self::color::ColorSchema;
+pub use self::geometry::{BoundingBoxSchema, PointSchema, PolygonSchema};
+pub use self::label::LabelSchema;
+pub use self::language::LanguageTagSchema;
+pub use self::operator::OperatorIdSchema;
+pub use self::time::TimeSpanSchema;
+pub use self::waveform::WaveformSchema;
diff --git a/crates/nvisy-core/src/schema/operator.rs b/crates/nvisy-core/src/schema/operator.rs
new file mode 100644
index 000000000..f9e802a33
--- /dev/null
+++ b/crates/nvisy-core/src/schema/operator.rs
@@ -0,0 +1,32 @@
+//! [`OperatorIdSchema`]: wire shape for
+//! [`elide_core::operator::OperatorId`].
+
+use elide_core::operator::OperatorId;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Wire-shape proxy for [`elide_core::operator::OperatorId`].
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+#[schemars(rename = "OperatorId")]
+pub struct OperatorIdSchema {
+    /// Stable operator name (e.g. `"mask"`, `"aes-gcm-encrypt"`).
+    pub name: String,
+    /// Operator version at the time it was applied.
+    pub version: String,
+}
+
+impl From<OperatorIdSchema> for OperatorId {
+    fn from(s: OperatorIdSchema) -> Self {
+        OperatorId::new(s.name, s.version)
+    }
+}
+
+impl From<OperatorId> for OperatorIdSchema {
+    fn from(o: OperatorId) -> Self {
+        Self {
+            name: o.name.as_str().to_owned(),
+            version: o.version.as_str().to_owned(),
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/schema/time.rs b/crates/nvisy-core/src/schema/time.rs
new file mode 100644
index 000000000..82a3f865b
--- /dev/null
+++ b/crates/nvisy-core/src/schema/time.rs
@@ -0,0 +1,34 @@
+//! [`TimeSpanSchema`]: wire shape for [`elide_core::primitive::TimeSpan`].
+
+use elide_core::primitive::TimeSpan;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Wire-shape proxy for [`elide_core::primitive::TimeSpan`].
+///
+/// Microsecond half-open interval `[start, end)`.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+#[schemars(rename = "TimeSpan")]
+pub struct TimeSpanSchema {
+    /// Start of the interval, microseconds from the stream start.
+    pub start_us: u64,
+    /// End of the interval (exclusive), microseconds from the stream
+    /// start.
+    pub end_us: u64,
+}
+
+impl From<TimeSpanSchema> for TimeSpan {
+    fn from(s: TimeSpanSchema) -> Self {
+        TimeSpan::new(s.start_us, s.end_us)
+    }
+}
+
+impl From<TimeSpan> for TimeSpanSchema {
+    fn from(t: TimeSpan) -> Self {
+        Self {
+            start_us: t.start_micros(),
+            end_us: t.end_micros(),
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/schema/waveform.rs b/crates/nvisy-core/src/schema/waveform.rs
new file mode 100644
index 000000000..b1b3b57d6
--- /dev/null
+++ b/crates/nvisy-core/src/schema/waveform.rs
@@ -0,0 +1,51 @@
+//! [`WaveformSchema`]: wire shape for the audio
+//! [`elide_core::modality::audio::Waveform`] tone shape, used by
+//! [`AudioRedaction::Beep`].
+//!
+//! [`AudioRedaction::Beep`]: crate::policy::redaction::AudioRedaction::Beep
+
+use elide_core::modality::audio::Waveform;
+use schemars::JsonSchema;
+use serde::{Deserialize, Serialize};
+
+/// Tone shape for the [`AudioRedaction::Beep`] operator.
+///
+/// [`AudioRedaction::Beep`]: crate::policy::redaction::AudioRedaction::Beep
+#[derive(
+    Debug,
+    Clone,
+    Copy,
+    PartialEq,
+    Eq,
+    Default,
+    Serialize,
+    Deserialize,
+    JsonSchema
+)]
+#[serde(rename_all = "snake_case")]
+#[schemars(rename = "Waveform")]
+pub enum WaveformSchema {
+    /// Pure sine. The broadcast censor-beep convention.
+    #[default]
+    Sine,
+    /// Square wave. Harsher and richer in harmonics — the "retro" bleep.
+    Square,
+}
+
+impl From<WaveformSchema> for Waveform {
+    fn from(s: WaveformSchema) -> Self {
+        match s {
+            WaveformSchema::Sine => Waveform::Sine,
+            WaveformSchema::Square => Waveform::Square,
+        }
+    }
+}
+
+impl From<Waveform> for WaveformSchema {
+    fn from(w: Waveform) -> Self {
+        match w {
+            Waveform::Sine => Self::Sine,
+            Waveform::Square => Self::Square,
+        }
+    }
+}
diff --git a/crates/nvisy-core/src/error.rs b/crates/nvisy-core/src/service/error.rs
similarity index 100%
rename from crates/nvisy-core/src/error.rs
rename to crates/nvisy-core/src/service/error.rs
diff --git a/crates/nvisy-core/src/health.rs b/crates/nvisy-core/src/service/health.rs
similarity index 100%
rename from crates/nvisy-core/src/health.rs
rename to crates/nvisy-core/src/service/health.rs
diff --git a/crates/nvisy-core/src/service/mod.rs b/crates/nvisy-core/src/service/mod.rs
new file mode 100644
index 000000000..7855464a9
--- /dev/null
+++ b/crates/nvisy-core/src/service/mod.rs
@@ -0,0 +1,9 @@
+//! Service-level concerns: the runtime's error vocabulary and the
+//! healthcheck composition trait. Distinct from the elide toolkit's
+//! own error type — runtime adds request-scoped context and surface
+//! categories the toolkit doesn't model.
+
+pub mod error;
+pub mod health;
+
+pub use self::error::{Error, ErrorKind, Result};
diff --git a/crates/nvisy-core/src/entity/source.rs b/crates/nvisy-core/src/source.rs
similarity index 100%
rename from crates/nvisy-core/src/entity/source.rs
rename to crates/nvisy-core/src/source.rs
diff --git a/crates/nvisy-engine/Cargo.toml b/crates/nvisy-engine/Cargo.toml
index 97a99643c..58e3f342d 100644
--- a/crates/nvisy-engine/Cargo.toml
+++ b/crates/nvisy-engine/Cargo.toml
@@ -2,8 +2,8 @@
 
 [package]
 name = "nvisy-engine"
-description = "Whole-document carrier + pipeline orchestrator for Nvisy: Document<M>, Audit, Policy, ingestion, the document runner"
-keywords = ["nvisy", "document", "pipeline", "orchestrator"]
+description = "Runtime adapter over elide: Policy compile, multi-tenant registry, request orchestration"
+keywords = ["nvisy", "engine", "elide", "redaction"]
 categories = ["data-structures", "concurrency"]
 readme = "README.md"
 
@@ -11,7 +11,7 @@ version = { workspace = true }
 rust-version = { workspace = true }
 edition = { workspace = true }
 license = { workspace = true }
-publish = true
+publish = { workspace = true }
 
 authors = { workspace = true }
 repository = { workspace = true }
@@ -19,19 +19,11 @@ homepage = { workspace = true }
 documentation = { workspace = true }
 
 [features]
-default = ["tabular", "image", "audio", "rich"]
-test-utils = ["dep:tempfile"]
+default = ["tabular", "image", "audio"]
 
-tabular = ["nvisy-codec/tabular"]
-image = ["nvisy-codec/image"]
-audio = ["nvisy-codec/audio"]
-rich = ["image", "nvisy-codec/rich"]
-
-openai = ["nvisy-llm/openai-gpt"]
-anthropic = ["nvisy-llm/anthropic-claude"]
-google = ["nvisy-llm/google-gemini"]
-
-bento = ["nvisy-toolkit/bento", "nvisy-ner/bento", "nvisy-ocr/bento"]
+tabular = ["elide/codec-tabular"]
+image = ["elide/codec-image"]
+audio = ["elide/codec-audio"]
 
 [package.metadata.docs.rs]
 all-features = true
@@ -39,61 +31,53 @@ rustdoc-args = ["--cfg", "docsrs"]
 
 [dependencies]
 # Internal crates
-nvisy-codec = { workspace = true, features = ["text"] }
 nvisy-core = { workspace = true, features = [] }
-nvisy-llm = { workspace = true, features = [] }
-nvisy-ner = { workspace = true, features = [] }
-nvisy-ocr = { workspace = true, features = [] }
-nvisy-pattern = { workspace = true, features = [] }
-nvisy-stt = { workspace = true, features = [] }
-nvisy-toolkit = { workspace = true, features = [] }
-
-# Serialization
-serde = { workspace = true, features = [] }
-serde_json = { workspace = true, features = [] }
-schemars = { workspace = true, features = [] }
-toml = { workspace = true, features = [] }
-humantime-serde = { workspace = true, features = [] }
-
-# Derive macros and error handling
-anyhow = { workspace = true, features = [] }
-derive_builder = { workspace = true, features = [] }
-derive_more = { workspace = true, features = ["deref", "deref_mut", "display", "from", "from_str", "into", "into_iterator"] }
-strum = { workspace = true, features = [] }
 
-# Primitive datatypes
-uuid = { workspace = true, features = [] }
-bytes = { workspace = true, features = [] }
-hipstr = { workspace = true, features = [] }
-jiff = { workspace = true, features = [] }
-semver = { workspace = true, features = ["serde"] }
+# Elide toolkit (upstream). `mock` pulls the no-op NER/LLM backends
+# the analyzer compile uses for `NerBackendParams::Mock` /
+# `LlmBackendParams::Mock`; real LLM providers land when key-provider
+# wiring lands.
+elide = { workspace = true, features = ["codec", "codec-text", "pattern", "ner", "llm", "mock", "serde", "sha2"] }
+elide-core = { workspace = true, features = ["serde", "image", "audio", "tabular"] }
 
-# Encoding and hashing
-base64 = { workspace = true, features = [] }
-aes-gcm = { workspace = true, features = [] }
+# Recognition-side elide crates not re-exported through the elide
+# umbrella the way the recognizer crates are (`elide::recognition::*`
+# is the natural reach for those). `elide-ocr` is the OCR enricher
+# layer engine wires from the image-modality compile path.
+elide-ocr = { workspace = true, features = ["mock"] }
+elide-stt = { workspace = true, features = ["mock"] }
 
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-futures = { workspace = true, features = [] }
-tokio = { workspace = true, features = ["rt", "sync", "time", "macros"] }
-tokio-util = { workspace = true, features = [] }
+# Runtime-owned elide extensions
+elide-bento = { workspace = true, features = ["ner", "ocr"] }
 
-# Observability
-tracing = { workspace = true, features = [] }
+# Derive macros
+derive_more = { workspace = true, features = ["as_ref", "deref", "from"] }
 
-# Storage and file-type detection
+# Primitive datatypes
+uuid = { workspace = true, features = ["v7"] }
+semver = { workspace = true, features = ["serde"] }
+hipstr = { workspace = true, features = ["serde"] }
+jiff = { workspace = true, features = ["serde"] }
+bytes = { workspace = true, features = [] }
+serde = { workspace = true, features = ["derive"] }
+sha2 = { workspace = true, features = [] }
+
+# Async runtime — `time` powers the per-doc timeout wrapping the
+# analyze fan-out. We do **not** depend on `rt`/`spawn`-style
+# concurrency: the engine drives concurrent analyzes via
+# `futures::stream::buffer_unordered` so the surrounding borrows
+# of the codec registry + analyzer spec can stay non-`'static`.
+tokio = { workspace = true, features = ["macros", "time"] }
+
+# Concurrent stream combinators for the analyze fan-out.
+# `stream::iter(...).buffer_unordered(N)` is the engine's
+# non-spawn concurrency primitive — caps in-flight work at N
+# while keeping every per-doc future borrowing the codec
+# registry + spec for free.
+futures = { workspace = true, features = ["std"] }
+
+# Multi-tenant storage
 fjall = { workspace = true, features = [] }
 
-# Utilities
-validator = { workspace = true, features = [] }
-rand = { workspace = true, features = [] }
-tempfile = { workspace = true, features = [], optional = true }
-
-[dev-dependencies]
-nvisy-engine = { path = ".", features = ["test-utils"] }
-nvisy-core = { workspace = true, features = ["test-utils"] }
-nvisy-codec = { workspace = true, features = ["txt", "csv", "test-utils"] }
-tempfile = { workspace = true, features = [] }
-
-[package.metadata.cargo-machete]
-ignored = ["humantime-serde"]
+# Resource serialisation (policy / context / run blobs in fjall)
+serde_json = { workspace = true, features = [] }
diff --git a/crates/nvisy-engine/src/analyzer/audio.rs b/crates/nvisy-engine/src/analyzer/audio.rs
new file mode 100644
index 000000000..bc42e4d62
--- /dev/null
+++ b/crates/nvisy-engine/src/analyzer/audio.rs
@@ -0,0 +1,80 @@
+//! Compile the audio-applicable parts of [`AnalyzerParams`]
+//! into an [`elide::detection::Analyzer<Audio>`].
+//!
+//! Audio runs Pattern and NER over the transcript text. An STT
+//! enricher stamps `TranscriptSegment`s onto the recognizer
+//! artifacts before recognition; the per-modality compile path
+//! wires it via `elide-stt`. LLM is not implemented on Audio in
+//! elide today — `RecognizerParams::llm` returns a `Validation`
+//! error.
+//!
+//! [`AnalyzerParams`]: nvisy_core::plan::AnalyzerParams
+
+use elide::detection::Analyzer;
+use elide_core::modality::audio::Audio;
+use elide_core::recognition::Scope;
+use elide_core::{Error, ErrorKind};
+use elide_stt::{MockBackend as MockSttBackend, SttEnricher};
+use nvisy_core::plan::{AnalyzerParams, SttBackendParams, SttEnricherParams};
+
+use super::common::{
+    attach_dedup, attach_ner, attach_pattern, build_catalog, reject_language_enricher,
+};
+use super::scope::compile_scope;
+
+/// Compile `spec` into an audio-modality analyzer + its compiled
+/// [`Scope`].
+pub fn compile_audio(spec: &AnalyzerParams) -> Result<(Analyzer<Audio>, Scope<Audio>), Error> {
+    let scope = compile_scope::<Audio>(&spec.scope)?;
+    let catalog = build_catalog(spec);
+    let mut analyzer = Analyzer::<Audio>::new();
+
+    if spec.enrichers.language.is_some() {
+        analyzer = reject_language_enricher::<Audio>()?;
+    }
+    if spec.enrichers.ocr.is_some() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: OCR enricher is only valid on the image modality",
+        ));
+    }
+    if let Some(stt) = &spec.enrichers.stt {
+        analyzer = attach_stt(analyzer, stt)?;
+    }
+
+    if let Some(pattern) = &spec.recognizers.pattern {
+        analyzer = attach_pattern(analyzer, pattern)?;
+    }
+    for ner in &spec.recognizers.ner {
+        analyzer = attach_ner(analyzer, ner)?;
+    }
+    if !spec.recognizers.llm.is_empty() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: LLM recognizer is not available on the audio modality \
+             (elide-llm has no LlmModality impl for Audio today)",
+        ));
+    }
+
+    analyzer = attach_dedup(analyzer, &spec.deduplication);
+    let _ = catalog;
+    Ok((analyzer, scope))
+}
+
+/// Attach an [`SttEnricher`] for the audio modality. `Mock`
+/// uses elide-stt's in-process no-op backend; `Bento` returns a
+/// clean "not wired yet" error until `elide-bento` ships a
+/// `BentoStt` client.
+fn attach_stt(
+    analyzer: Analyzer<Audio>,
+    spec: &SttEnricherParams,
+) -> Result<Analyzer<Audio>, Error> {
+    match &spec.backend {
+        SttBackendParams::Mock => Ok(analyzer.with_enricher(SttEnricher::new(MockSttBackend))),
+        SttBackendParams::Bento { .. } => Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: BentoML STT backend needs an elide-bento `BentoStt` \
+             client; not wired into the compile surface yet",
+        )),
+    }
+}
diff --git a/crates/nvisy-engine/src/analyzer/common.rs b/crates/nvisy-engine/src/analyzer/common.rs
new file mode 100644
index 000000000..8463ce36a
--- /dev/null
+++ b/crates/nvisy-engine/src/analyzer/common.rs
@@ -0,0 +1,136 @@
+//! Modality-generic compile helpers shared by every per-modality
+//! analyzer compile function.
+//!
+//! `PatternRecognizer` and `NerRecognizer` are themselves
+//! modality-generic (both impl `Recognizer<M>` for any
+//! `M: TextRecognizable`), so this module hosts the pattern/NER
+//! attach + dedup + scope helpers once and lets the per-modality
+//! files specialise only LLM (which is generic over a different
+//! `LlmModality` trait) and the per-modality rejection of unwired
+//! spec variants.
+
+use elide::detection::Analyzer;
+use elide::detection::calibrate::{CalibrateLayer, CalibrationMap as ElideCalibrationMap};
+use elide::detection::filter::FilterLayer;
+use elide::detection::fuse::{FuseLayer, MaxConfidence, Mean, NoisyOr};
+use elide::detection::resolve::{HighestConfidence, LongestSpan, ResolveLayer};
+use elide::recognition::ner::NerRecognizer;
+use elide::recognition::pattern::PatternRecognizer;
+use elide_bento::BentoNer;
+use elide_core::entity::LabelCatalog;
+use elide_core::modality::TextRecognizable;
+use elide_core::primitive::ConfidenceThreshold;
+use elide_core::recognition::Recognizer;
+use elide_core::{Error, ErrorKind};
+use nvisy_core::plan::{
+    AnalyzerParams, DeduplicationParams, FusionStrategyParams, NerBackendParams,
+    NerRecognizerParams, PatternRecognizerParams, ResolutionStrategyParams,
+};
+
+/// Build the per-request label catalog from `spec`. Engine does not
+/// pre-seed builtins; the caller picks.
+pub(crate) fn build_catalog(spec: &AnalyzerParams) -> LabelCatalog {
+    let mut catalog = LabelCatalog::new();
+    for label in &spec.label_catalog {
+        catalog.insert(label.clone().into());
+    }
+    catalog
+}
+
+/// Reject the language enricher: `elide-ner/lingua` wiring isn't
+/// exposed through the compile surface yet. Per-modality compile
+/// fns call this when they see `params.enrichers.language` set
+/// (every modality supports language detection in principle, but
+/// no modality currently has the backend wired).
+pub(super) fn reject_language_enricher<M>() -> Result<Analyzer<M>, Error>
+where
+    M: elide_core::modality::Modality,
+{
+    Err(Error::new(
+        ErrorKind::Validation,
+        "analyzer compile: language enricher needs elide-ner/lingua wiring; \
+         not exposed through the compile surface yet",
+    ))
+}
+
+/// Attach a [`PatternRecognizer`] built from `spec`. The same
+/// recognizer instance — bare or wrapped in elide's `Enhanced`
+/// layer — serves any `M: TextRecognizable`, so the helper is
+/// uniform across modalities.
+pub(super) fn attach_pattern<M>(
+    analyzer: Analyzer<M>,
+    spec: &PatternRecognizerParams,
+) -> Result<Analyzer<M>, Error>
+where
+    M: TextRecognizable,
+    PatternRecognizer: Recognizer<M> + 'static,
+    elide::recognition::context::Enhanced<PatternRecognizer>: Recognizer<M> + 'static,
+{
+    let mut builder = PatternRecognizer::builder();
+    if spec.builtins {
+        builder = builder.with_builtin_patterns().with_builtin_dictionaries();
+    }
+    if spec.context_enhanced {
+        Ok(analyzer.with_recognizer(builder.build_context_enhanced()?))
+    } else {
+        Ok(analyzer.with_recognizer(builder.build()?))
+    }
+}
+
+/// Attach a [`NerRecognizer`] built from `spec`. Like pattern,
+/// modality-generic for any `M: TextRecognizable`.
+pub(super) fn attach_ner<M>(
+    analyzer: Analyzer<M>,
+    spec: &NerRecognizerParams,
+) -> Result<Analyzer<M>, Error>
+where
+    M: TextRecognizable,
+    NerRecognizer: Recognizer<M> + 'static,
+{
+    let mut builder = NerRecognizer::builder().with_name(spec.name.clone());
+    match &spec.backend {
+        NerBackendParams::Mock => {
+            builder = builder.with_mock_backend();
+        }
+        NerBackendParams::Bento { base_url, model } => {
+            builder = builder.with_backend(BentoNer::new(base_url.clone(), model.clone())?);
+        }
+    }
+    Ok(analyzer.with_recognizer(builder.build()?))
+}
+
+/// Append the deduplication layers: calibrate → fuse → resolve →
+/// filter. Calibrate is skipped when the calibration map is empty.
+pub(super) fn attach_dedup<M>(mut analyzer: Analyzer<M>, spec: &DeduplicationParams) -> Analyzer<M>
+where
+    M: elide_core::modality::Modality,
+{
+    if !spec.calibration.is_empty() {
+        let mut map = ElideCalibrationMap::new();
+        for (recognizer, weight) in &spec.calibration.0 {
+            map.insert(recognizer.clone(), *weight);
+        }
+        analyzer = analyzer.with_layer(CalibrateLayer::new(map));
+    }
+
+    analyzer = match spec.fusion {
+        FusionStrategyParams::MaxConfidence => analyzer.with_layer(FuseLayer::new(MaxConfidence)),
+        FusionStrategyParams::Mean => analyzer.with_layer(FuseLayer::new(Mean)),
+        FusionStrategyParams::NoisyOr => analyzer.with_layer(FuseLayer::new(NoisyOr)),
+    };
+
+    analyzer = match spec.resolution {
+        ResolutionStrategyParams::HighestConfidence => {
+            analyzer.with_layer(ResolveLayer::new(HighestConfidence))
+        }
+        ResolutionStrategyParams::LongestSpan => {
+            analyzer.with_layer(ResolveLayer::new(LongestSpan))
+        }
+    };
+
+    let threshold = match spec.min_confidence {
+        Some(v) => ConfidenceThreshold::clamped(v),
+        None => ConfidenceThreshold::BASELINE,
+    };
+    analyzer.with_layer(FilterLayer::new().with_threshold(threshold))
+}
diff --git a/crates/nvisy-engine/src/analyzer/image.rs b/crates/nvisy-engine/src/analyzer/image.rs
new file mode 100644
index 000000000..0682e4c53
--- /dev/null
+++ b/crates/nvisy-engine/src/analyzer/image.rs
@@ -0,0 +1,103 @@
+//! Compile the image-applicable parts of [`AnalyzerParams`] into
+//! an [`elide::detection::Analyzer<Image>`].
+//!
+//! Image is the fullest non-text modality: Pattern and NER run
+//! over the OCR'd text (the OCR enricher stamps a `Layout` onto
+//! the recognizer artifacts upstream), and LLM is available
+//! image-natively for vision-language models.
+//!
+//! [`AnalyzerParams`]: nvisy_core::plan::AnalyzerParams
+
+use elide::detection::Analyzer;
+use elide::recognition::llm::LlmRecognizer;
+use elide_bento::BentoOcr;
+use elide_core::modality::image::Image;
+use elide_core::recognition::Scope;
+use elide_core::{Error, ErrorKind};
+use elide_ocr::{MockBackend as MockOcrBackend, OcrEnricher};
+use nvisy_core::plan::{
+    AnalyzerParams, LlmBackendParams, LlmRecognizerParams, OcrBackendParams, OcrEnricherParams,
+};
+
+use super::common::{
+    attach_dedup, attach_ner, attach_pattern, build_catalog, reject_language_enricher,
+};
+use super::scope::compile_scope;
+
+/// Compile `spec` into an image-modality analyzer + its compiled
+/// [`Scope`].
+pub fn compile_image(spec: &AnalyzerParams) -> Result<(Analyzer<Image>, Scope<Image>), Error> {
+    let scope = compile_scope::<Image>(&spec.scope)?;
+    let catalog = build_catalog(spec);
+    let mut analyzer = Analyzer::<Image>::new();
+
+    if spec.enrichers.language.is_some() {
+        analyzer = reject_language_enricher::<Image>()?;
+    }
+    if let Some(ocr) = &spec.enrichers.ocr {
+        analyzer = attach_ocr(analyzer, ocr)?;
+    }
+    if spec.enrichers.stt.is_some() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: STT enricher is only valid on the audio modality",
+        ));
+    }
+
+    if let Some(pattern) = &spec.recognizers.pattern {
+        analyzer = attach_pattern(analyzer, pattern)?;
+    }
+    for ner in &spec.recognizers.ner {
+        analyzer = attach_ner(analyzer, ner)?;
+    }
+    for llm in &spec.recognizers.llm {
+        analyzer = attach_llm(analyzer, llm)?;
+    }
+
+    analyzer = attach_dedup(analyzer, &spec.deduplication);
+    let _ = catalog;
+    Ok((analyzer, scope))
+}
+
+fn attach_ocr(
+    analyzer: Analyzer<Image>,
+    spec: &OcrEnricherParams,
+) -> Result<Analyzer<Image>, Error> {
+    let enricher = match &spec.backend {
+        OcrBackendParams::Mock => OcrEnricher::new(MockOcrBackend),
+        OcrBackendParams::Bento { base_url, model } => {
+            OcrEnricher::new(BentoOcr::new(base_url.clone(), model.clone())?)
+        }
+    };
+    Ok(analyzer.with_enricher(enricher))
+}
+
+fn attach_llm(
+    analyzer: Analyzer<Image>,
+    spec: &LlmRecognizerParams,
+) -> Result<Analyzer<Image>, Error> {
+    let mut builder = LlmRecognizer::<Image>::builder().with_name(spec.name.clone());
+    match &spec.backend {
+        LlmBackendParams::Mock => {
+            builder = builder.with_mock_backend();
+        }
+        LlmBackendParams::Openai { .. }
+        | LlmBackendParams::Anthropic { .. }
+        | LlmBackendParams::Google { .. } => {
+            return Err(Error::new(
+                ErrorKind::Validation,
+                "analyzer compile: real LLM providers need engine-side credential + \
+                 rate-limit wiring; not exposed through the compile surface yet",
+            ));
+        }
+    }
+    if spec.prompt.is_some() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: custom LLM prompts need the elide Prompt trait wiring; \
+             pass `prompt: null` to use the default prompt for now",
+        ));
+    }
+    builder = builder.with_default_prompt();
+    Ok(analyzer.with_recognizer(builder.build()?))
+}
diff --git a/crates/nvisy-engine/src/analyzer/mod.rs b/crates/nvisy-engine/src/analyzer/mod.rs
new file mode 100644
index 000000000..598577f2d
--- /dev/null
+++ b/crates/nvisy-engine/src/analyzer/mod.rs
@@ -0,0 +1,36 @@
+//! Compile a [`nvisy_core::plan::AnalyzerParams`] into an
+//! [`elide::detection::Analyzer`] per modality at request time.
+//!
+//! Symmetric with [`crate::anonymizer`]: the spec is pure data
+//! (recognizer + enricher + dedup + scope choices); engine walks
+//! it and assembles the matching elide runtime values.
+//!
+//! ## Per-modality coverage
+//!
+//! | Modality | Pattern | NER | LLM |
+//! |----------|---------|-----|-----|
+//! | Text     | yes     | yes | yes |
+//! | Tabular  | yes     | yes | (no upstream `LlmModality` impl) |
+//! | Image    | yes     | yes | yes |
+//! | Audio    | yes     | yes | (no upstream `LlmModality` impl) |
+//!
+//! Pattern (bare + context-enhanced) and NER work on any
+//! `TextRecognizable` modality (cells, OCR'd text, transcripts).
+//! LLM is per-modality typed in elide and requires an `LlmModality`
+//! impl, which today exists only for Text + Image.
+//!
+//! [`Analyzer`]: elide::detection::Analyzer
+//! [`AnalyzerParams`]: nvisy_core::plan::AnalyzerParams
+
+mod audio;
+mod common;
+mod image;
+mod scope;
+mod tabular;
+mod text;
+
+pub use self::audio::compile_audio;
+pub(crate) use self::common::build_catalog;
+pub use self::image::compile_image;
+pub use self::tabular::compile_tabular;
+pub use self::text::compile_text;
diff --git a/crates/nvisy-engine/src/analyzer/scope.rs b/crates/nvisy-engine/src/analyzer/scope.rs
new file mode 100644
index 000000000..4af871e91
--- /dev/null
+++ b/crates/nvisy-engine/src/analyzer/scope.rs
@@ -0,0 +1,31 @@
+//! Compile a [`nvisy_core::plan::ScopeParams`] into an
+//! [`elide::recognition::Scope`].
+
+use elide_core::Error;
+use elide_core::modality::Modality;
+use elide_core::primitive::{CountryCode, Language, LanguageTag};
+use elide_core::recognition::Scope;
+use nvisy_core::plan::ScopeParams;
+
+/// Translate `spec` into an [`elide::recognition::Scope<M>`].
+///
+/// Languages are turned into asserted [`Language`] entries;
+/// jurisdictions parse into [`CountryCode`]s. Returns the first
+/// parse error encountered.
+pub(super) fn compile_scope<M: Modality>(spec: &ScopeParams) -> Result<Scope<M>, Error> {
+    let mut scope = Scope::<M>::new();
+    for lang in &spec.languages {
+        let tag = LanguageTag::try_from(lang.clone())?;
+        scope = scope.with_language(Language::asserted(tag));
+    }
+    for code in &spec.jurisdictions {
+        let country = CountryCode::from_alpha2(code).map_err(|e| {
+            Error::new(
+                elide_core::ErrorKind::Validation,
+                format!("scope: jurisdiction `{code}`: {e}"),
+            )
+        })?;
+        scope = scope.with_country(country);
+    }
+    Ok(scope)
+}
diff --git a/crates/nvisy-engine/src/analyzer/tabular.rs b/crates/nvisy-engine/src/analyzer/tabular.rs
new file mode 100644
index 000000000..16e7c6c93
--- /dev/null
+++ b/crates/nvisy-engine/src/analyzer/tabular.rs
@@ -0,0 +1,65 @@
+//! Compile the tabular-applicable parts of [`AnalyzerParams`]
+//! into an [`elide::detection::Analyzer<Tabular>`].
+//!
+//! Tabular runs Pattern and NER over each cell's text (cells
+//! are `TextRecognizable`). LLM is not implemented on Tabular in
+//! elide today — `RecognizerParams::llm` returns a `Validation`
+//! error. (ELIDE GAP: `impl LlmModality for Tabular` would let
+//! an LLM scan tables for PII.)
+//!
+//! [`AnalyzerParams`]: nvisy_core::plan::AnalyzerParams
+
+use elide::detection::Analyzer;
+use elide_core::modality::tabular::Tabular;
+use elide_core::recognition::Scope;
+use elide_core::{Error, ErrorKind};
+use nvisy_core::plan::AnalyzerParams;
+
+use super::common::{
+    attach_dedup, attach_ner, attach_pattern, build_catalog, reject_language_enricher,
+};
+use super::scope::compile_scope;
+
+/// Compile `spec` into a tabular-modality analyzer + its
+/// compiled [`Scope`].
+pub fn compile_tabular(
+    spec: &AnalyzerParams,
+) -> Result<(Analyzer<Tabular>, Scope<Tabular>), Error> {
+    let scope = compile_scope::<Tabular>(&spec.scope)?;
+    let catalog = build_catalog(spec);
+    let mut analyzer = Analyzer::<Tabular>::new();
+
+    if spec.enrichers.language.is_some() {
+        analyzer = reject_language_enricher::<Tabular>()?;
+    }
+    if spec.enrichers.ocr.is_some() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: OCR enricher is only valid on the image modality",
+        ));
+    }
+    if spec.enrichers.stt.is_some() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: STT enricher is only valid on the audio modality",
+        ));
+    }
+
+    if let Some(pattern) = &spec.recognizers.pattern {
+        analyzer = attach_pattern(analyzer, pattern)?;
+    }
+    for ner in &spec.recognizers.ner {
+        analyzer = attach_ner(analyzer, ner)?;
+    }
+    if !spec.recognizers.llm.is_empty() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: LLM recognizer is not available on the tabular \
+             modality (elide-llm has no LlmModality impl for Tabular today)",
+        ));
+    }
+
+    analyzer = attach_dedup(analyzer, &spec.deduplication);
+    let _ = catalog;
+    Ok((analyzer, scope))
+}
diff --git a/crates/nvisy-engine/src/analyzer/text.rs b/crates/nvisy-engine/src/analyzer/text.rs
new file mode 100644
index 000000000..91ba8c074
--- /dev/null
+++ b/crates/nvisy-engine/src/analyzer/text.rs
@@ -0,0 +1,97 @@
+//! Compile the text-applicable parts of [`AnalyzerParams`] into an
+//! [`elide::detection::Analyzer<Text>`].
+//!
+//! Text supports the full recognizer set: Pattern, NER, and LLM.
+//! Real LLM providers + custom prompts return a `Validation`
+//! error today (their credential / rate-limit wiring is not
+//! exposed through the compile surface yet).
+//!
+//! [`AnalyzerParams`]: nvisy_core::plan::AnalyzerParams
+
+use elide::detection::Analyzer;
+use elide::recognition::llm::LlmRecognizer;
+use elide_core::modality::text::Text;
+use elide_core::recognition::Scope;
+use elide_core::{Error, ErrorKind};
+use nvisy_core::plan::{AnalyzerParams, LlmBackendParams, LlmRecognizerParams};
+
+use super::common::{
+    attach_dedup, attach_ner, attach_pattern, build_catalog, reject_language_enricher,
+};
+use super::scope::compile_scope;
+
+/// Compile `spec` into a text-modality analyzer + its compiled
+/// [`Scope`].
+///
+/// The returned scope is **not** attached to the analyzer
+/// (elide's `Analyzer::analyze` takes `&Scope` per-call, not at
+/// build time); the caller pairs them: `analyzer.analyze(data,
+/// &scope)`.
+pub fn compile_text(spec: &AnalyzerParams) -> Result<(Analyzer<Text>, Scope<Text>), Error> {
+    let scope = compile_scope::<Text>(&spec.scope)?;
+    let catalog = build_catalog(spec);
+    let mut analyzer = Analyzer::<Text>::new();
+
+    // Enrichers: language is the only text-applicable kind; OCR
+    // / STT are modality-specific (image / audio).
+    if spec.enrichers.language.is_some() {
+        analyzer = reject_language_enricher::<Text>()?;
+    }
+    if spec.enrichers.ocr.is_some() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: OCR enricher is only valid on the image modality",
+        ));
+    }
+    if spec.enrichers.stt.is_some() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: STT enricher is only valid on the audio modality",
+        ));
+    }
+
+    // Recognizers: pattern (at-most-one), ner (list), llm (list).
+    if let Some(pattern) = &spec.recognizers.pattern {
+        analyzer = attach_pattern(analyzer, pattern)?;
+    }
+    for ner in &spec.recognizers.ner {
+        analyzer = attach_ner(analyzer, ner)?;
+    }
+    for llm in &spec.recognizers.llm {
+        analyzer = attach_llm(analyzer, llm)?;
+    }
+
+    analyzer = attach_dedup(analyzer, &spec.deduplication);
+    let _ = catalog; // catalog wiring lands when selectors learn to tag-match upstream
+    Ok((analyzer, scope))
+}
+
+fn attach_llm(
+    analyzer: Analyzer<Text>,
+    spec: &LlmRecognizerParams,
+) -> Result<Analyzer<Text>, Error> {
+    let mut builder = LlmRecognizer::<Text>::builder().with_name(spec.name.clone());
+    match &spec.backend {
+        LlmBackendParams::Mock => {
+            builder = builder.with_mock_backend();
+        }
+        LlmBackendParams::Openai { .. }
+        | LlmBackendParams::Anthropic { .. }
+        | LlmBackendParams::Google { .. } => {
+            return Err(Error::new(
+                ErrorKind::Validation,
+                "analyzer compile: real LLM providers need engine-side credential + \
+                 rate-limit wiring; not exposed through the compile surface yet",
+            ));
+        }
+    }
+    if spec.prompt.is_some() {
+        return Err(Error::new(
+            ErrorKind::Validation,
+            "analyzer compile: custom LLM prompts need the elide Prompt trait wiring; \
+             pass `prompt: null` to use the default prompt for now",
+        ));
+    }
+    builder = builder.with_default_prompt();
+    Ok(analyzer.with_recognizer(builder.build()?))
+}
diff --git a/crates/nvisy-engine/src/anonymizer/audio.rs b/crates/nvisy-engine/src/anonymizer/audio.rs
new file mode 100644
index 000000000..280234c5c
--- /dev/null
+++ b/crates/nvisy-engine/src/anonymizer/audio.rs
@@ -0,0 +1,126 @@
+//! Compile audio-modality rules to elide operators + attach to an
+//! [`Anonymizer<Audio>`].
+
+use elide::redaction::Anonymizer;
+use elide::redaction::operators::{Beep, Erase, Keep, Silence};
+use elide_core::Error;
+use elide_core::entity::LabelCatalog;
+use elide_core::modality::audio::Audio;
+use nvisy_core::policy::redaction::AudioRedaction;
+use nvisy_core::policy::{Policy, Rule, RuleAction};
+use uuid::Uuid;
+
+use super::selector::{attach, attach_override, fallback_attribution, rule_attribution};
+
+/// Compile every audio-applicable rule across `policies` into an
+/// audio-modality anonymizer.
+pub fn compile_audio(
+    policies: &[Policy],
+    catalog: LabelCatalog,
+) -> Result<Anonymizer<Audio>, Error> {
+    let anonymizer = Anonymizer::<Audio>::new().with_catalog(catalog);
+    Ok(attach_policies_audio(anonymizer, policies.iter()))
+}
+
+/// Attach every audio-applicable rule from `policies` onto an
+/// already-constructed anonymizer. Takes an iterator so the
+/// apply pipeline can pre-filter by [`Policy::applies_when`]
+/// without cloning.
+///
+/// [`Policy::applies_when`]: nvisy_core::policy::Policy::applies_when
+pub(crate) fn attach_policies_audio<'a>(
+    mut anonymizer: Anonymizer<Audio>,
+    policies: impl Iterator<Item = &'a Policy>,
+) -> Anonymizer<Audio> {
+    for policy in policies {
+        for rule in &policy.rules {
+            let RuleAction::Redact(redactions) = &rule.action else {
+                continue;
+            };
+            let Some(spec) = &redactions.audio else {
+                continue;
+            };
+            anonymizer = attach_rule(anonymizer, policy, rule, spec);
+        }
+        if let Some(RuleAction::Redact(redactions)) = &policy.fallback
+            && let Some(spec) = &redactions.audio
+        {
+            anonymizer = attach_fallback(anonymizer, policy, spec);
+        }
+    }
+    anonymizer
+}
+
+/// Attach a reviewer override for one entity. No-op when the
+/// override is not an audio-modality `Redact`.
+pub(crate) fn attach_override_audio(
+    anonymizer: Anonymizer<Audio>,
+    entity_id: Uuid,
+    action: &RuleAction,
+) -> Anonymizer<Audio> {
+    let RuleAction::Redact(redactions) = action else {
+        return anonymizer;
+    };
+    let Some(spec) = &redactions.audio else {
+        return anonymizer;
+    };
+    match build(spec) {
+        AudioOp::Erase => attach_override(anonymizer, entity_id, Erase),
+        AudioOp::Keep => attach_override(anonymizer, entity_id, Keep),
+        AudioOp::Silence => attach_override(anonymizer, entity_id, Silence),
+        AudioOp::Beep(op) => attach_override(anonymizer, entity_id, op),
+    }
+}
+
+fn attach_rule(
+    anonymizer: Anonymizer<Audio>,
+    policy: &Policy,
+    rule: &Rule,
+    spec: &AudioRedaction,
+) -> Anonymizer<Audio> {
+    let attribution = rule_attribution(policy, rule);
+    match build(spec) {
+        AudioOp::Erase => attach(anonymizer, &rule.predicate, Erase, attribution),
+        AudioOp::Keep => attach(anonymizer, &rule.predicate, Keep, attribution),
+        AudioOp::Silence => attach(anonymizer, &rule.predicate, Silence, attribution),
+        AudioOp::Beep(op) => attach(anonymizer, &rule.predicate, op, attribution),
+    }
+}
+
+fn attach_fallback(
+    anonymizer: Anonymizer<Audio>,
+    policy: &Policy,
+    spec: &AudioRedaction,
+) -> Anonymizer<Audio> {
+    let attribution = fallback_attribution(policy);
+    match build(spec) {
+        AudioOp::Erase => anonymizer.with_fallback(Erase).because(attribution),
+        AudioOp::Keep => anonymizer.with_fallback(Keep).because(attribution),
+        AudioOp::Silence => anonymizer.with_fallback(Silence).because(attribution),
+        AudioOp::Beep(op) => anonymizer.with_fallback(op).because(attribution),
+    }
+}
+
+enum AudioOp {
+    Erase,
+    Keep,
+    Silence,
+    Beep(Beep),
+}
+
+fn build(spec: &AudioRedaction) -> AudioOp {
+    match spec {
+        AudioRedaction::Erase => AudioOp::Erase,
+        AudioRedaction::Keep => AudioOp::Keep,
+        AudioRedaction::Silence => AudioOp::Silence,
+        AudioRedaction::Beep {
+            hz,
+            amplitude,
+            waveform,
+        } => AudioOp::Beep(
+            Beep::new(*hz)
+                .with_amplitude(*amplitude)
+                .with_waveform((*waveform).into()),
+        ),
+    }
+}
diff --git a/crates/nvisy-engine/src/anonymizer/image.rs b/crates/nvisy-engine/src/anonymizer/image.rs
new file mode 100644
index 000000000..c9e73838d
--- /dev/null
+++ b/crates/nvisy-engine/src/anonymizer/image.rs
@@ -0,0 +1,123 @@
+//! Compile image-modality rules to elide operators + attach to an
+//! [`Anonymizer<Image>`].
+
+use elide::redaction::Anonymizer;
+use elide::redaction::operators::{Blackbox, Blur, Erase, Keep, Pixelate};
+use elide_core::Error;
+use elide_core::entity::LabelCatalog;
+use elide_core::modality::image::Image;
+use nvisy_core::policy::redaction::ImageRedaction;
+use nvisy_core::policy::{Policy, Rule, RuleAction};
+use uuid::Uuid;
+
+use super::selector::{attach, attach_override, fallback_attribution, rule_attribution};
+
+/// Compile every image-applicable rule across `policies` into an
+/// image-modality anonymizer.
+pub fn compile_image(
+    policies: &[Policy],
+    catalog: LabelCatalog,
+) -> Result<Anonymizer<Image>, Error> {
+    let anonymizer = Anonymizer::<Image>::new().with_catalog(catalog);
+    Ok(attach_policies_image(anonymizer, policies.iter()))
+}
+
+/// Attach every image-applicable rule from `policies` onto an
+/// already-constructed anonymizer. Takes an iterator so the
+/// apply pipeline can pre-filter by [`Policy::applies_when`]
+/// without cloning.
+///
+/// [`Policy::applies_when`]: nvisy_core::policy::Policy::applies_when
+pub(crate) fn attach_policies_image<'a>(
+    mut anonymizer: Anonymizer<Image>,
+    policies: impl Iterator<Item = &'a Policy>,
+) -> Anonymizer<Image> {
+    for policy in policies {
+        for rule in &policy.rules {
+            let RuleAction::Redact(redactions) = &rule.action else {
+                continue;
+            };
+            let Some(spec) = &redactions.image else {
+                continue;
+            };
+            anonymizer = attach_rule(anonymizer, policy, rule, spec);
+        }
+        if let Some(RuleAction::Redact(redactions)) = &policy.fallback
+            && let Some(spec) = &redactions.image
+        {
+            anonymizer = attach_fallback(anonymizer, policy, spec);
+        }
+    }
+    anonymizer
+}
+
+/// Attach a reviewer override for one entity. No-op when the
+/// override is not an image-modality `Redact`.
+pub(crate) fn attach_override_image(
+    anonymizer: Anonymizer<Image>,
+    entity_id: Uuid,
+    action: &RuleAction,
+) -> Anonymizer<Image> {
+    let RuleAction::Redact(redactions) = action else {
+        return anonymizer;
+    };
+    let Some(spec) = &redactions.image else {
+        return anonymizer;
+    };
+    match build(spec) {
+        ImageOp::Erase => attach_override(anonymizer, entity_id, Erase),
+        ImageOp::Keep => attach_override(anonymizer, entity_id, Keep),
+        ImageOp::Blur(op) => attach_override(anonymizer, entity_id, op),
+        ImageOp::Pixelate(op) => attach_override(anonymizer, entity_id, op),
+        ImageOp::Blackbox(op) => attach_override(anonymizer, entity_id, op),
+    }
+}
+
+fn attach_rule(
+    anonymizer: Anonymizer<Image>,
+    policy: &Policy,
+    rule: &Rule,
+    spec: &ImageRedaction,
+) -> Anonymizer<Image> {
+    let attribution = rule_attribution(policy, rule);
+    match build(spec) {
+        ImageOp::Erase => attach(anonymizer, &rule.predicate, Erase, attribution),
+        ImageOp::Keep => attach(anonymizer, &rule.predicate, Keep, attribution),
+        ImageOp::Blur(op) => attach(anonymizer, &rule.predicate, op, attribution),
+        ImageOp::Pixelate(op) => attach(anonymizer, &rule.predicate, op, attribution),
+        ImageOp::Blackbox(op) => attach(anonymizer, &rule.predicate, op, attribution),
+    }
+}
+
+fn attach_fallback(
+    anonymizer: Anonymizer<Image>,
+    policy: &Policy,
+    spec: &ImageRedaction,
+) -> Anonymizer<Image> {
+    let attribution = fallback_attribution(policy);
+    match build(spec) {
+        ImageOp::Erase => anonymizer.with_fallback(Erase).because(attribution),
+        ImageOp::Keep => anonymizer.with_fallback(Keep).because(attribution),
+        ImageOp::Blur(op) => anonymizer.with_fallback(op).because(attribution),
+        ImageOp::Pixelate(op) => anonymizer.with_fallback(op).because(attribution),
+        ImageOp::Blackbox(op) => anonymizer.with_fallback(op).because(attribution),
+    }
+}
+
+enum ImageOp {
+    Erase,
+    Keep,
+    Blur(Blur),
+    Pixelate(Pixelate),
+    Blackbox(Blackbox),
+}
+
+fn build(spec: &ImageRedaction) -> ImageOp {
+    match spec {
+        ImageRedaction::Erase => ImageOp::Erase,
+        ImageRedaction::Keep => ImageOp::Keep,
+        ImageRedaction::Blur { sigma } => ImageOp::Blur(Blur::new(*sigma)),
+        ImageRedaction::Pixelate { block_size } => ImageOp::Pixelate(Pixelate::new(*block_size)),
+        ImageRedaction::Blackbox { color } => ImageOp::Blackbox(Blackbox::new((*color).into())),
+    }
+}
diff --git a/crates/nvisy-engine/src/anonymizer/mod.rs b/crates/nvisy-engine/src/anonymizer/mod.rs
new file mode 100644
index 000000000..5bd1c6ef7
--- /dev/null
+++ b/crates/nvisy-engine/src/anonymizer/mod.rs
@@ -0,0 +1,50 @@
+//! Compile a [`nvisy_core::policy::Policy`] set into an
+//! [`Anonymizer`] per modality at request time.
+//!
+//! Policy specs are serialisable and modality-agnostic; elide's
+//! [`Anonymizer`]`<M>` is a runtime, modality-typed value that
+//! drives actual redaction. This module bridges the two: it walks every
+//! enabled rule in precedence order, builds the matching elide
+//! operator from the spec, and attaches it to the anonymizer with a
+//! predicate built from the rule's selector.
+//!
+//! ## Layout
+//!
+//! - `text` / `tabular` consume the text-backed redaction specs
+//!   (the full elide built-in vocabulary: Erase, Keep, Mask,
+//!   Replace, Hash, Pseudonymize, Encrypt — plus the structural
+//!   DropRow / DropColumn on tabular).
+//! - `image` handles the image specs (Erase, Keep, Blur, Pixelate,
+//!   Blackbox).
+//! - `audio` handles the audio specs (Erase, Keep, Silence, Beep).
+//!
+//! Each per-modality `compile` entry walks `&[Policy]` in
+//! precedence order; within each policy, rules are tried in
+//! declared order; the first matching rule's operator wins. Each
+//! policy's `default_action`, if Redact with that modality's arm
+//! set, becomes the anonymizer's fallback.
+//!
+//! ## Audit decoration
+//!
+//! Each compiled operator will be wrapped in a thin decorator that
+//! stamps the policy/rule attribution onto the audit when the
+//! operator runs. The decoration lives outside the per-modality
+//! compile helpers — they assemble naked operators today; the audit
+//! pass wraps them in a follow-up.
+//!
+//! [`Anonymizer`]: elide::redaction::Anonymizer
+
+mod audio;
+mod image;
+mod selector;
+mod tabular;
+mod text;
+
+pub use self::audio::compile_audio;
+pub(crate) use self::audio::{attach_override_audio, attach_policies_audio};
+pub use self::image::compile_image;
+pub(crate) use self::image::{attach_override_image, attach_policies_image};
+pub use self::tabular::compile_tabular;
+pub(crate) use self::tabular::{attach_override_tabular, attach_policies_tabular};
+pub use self::text::compile_text;
+pub(crate) use self::text::{attach_override_text, attach_policies_text};
diff --git a/crates/nvisy-engine/src/anonymizer/selector.rs b/crates/nvisy-engine/src/anonymizer/selector.rs
new file mode 100644
index 000000000..3be86c9b7
--- /dev/null
+++ b/crates/nvisy-engine/src/anonymizer/selector.rs
@@ -0,0 +1,146 @@
+//! Attach a [`Rule`] / `fallback` onto an elide [`Anonymizer`],
+//! stamping the engine-internal attribution from policy + rule
+//! UUIDs.
+//!
+//! The wire now carries one composable [`Predicate`] per rule.
+//! [`attach`] pattern-matches that predicate against three
+//! degenerate shapes to keep elide's fast paths:
+//!
+//! - [`Predicate::LabelOneOf`] with a single label →
+//!   [`Anonymizer::with_label`] (audit records the matched
+//!   label).
+//! - [`Predicate::TagOneOf`] with a single tag →
+//!   [`Anonymizer::with_tag`] (audit records the matched tag).
+//! - Anything else → [`Anonymizer::with_catalog_predicate`]
+//!   (audit records `Predicate`).
+//!
+//! Attribution always sits in elide's [`Attribution`] slot:
+//! `policy_id` from the policy's UUID, `reason` from the rule's
+//! UUID (or `None` for the policy fallback).
+//!
+//! [`Anonymizer`]: elide::redaction::Anonymizer
+//! [`Rule`]: nvisy_core::policy::Rule
+//! [`Predicate`]: nvisy_core::policy::Predicate
+//! [`Attribution`]: elide_core::entity::provenance::Attribution
+
+use elide::redaction::Anonymizer;
+use elide_core::entity::provenance::Attribution;
+use elide_core::entity::{Entity, LabelCatalog, LabelRef};
+use elide_core::modality::Modality;
+use elide_core::operator::Operator;
+use nvisy_core::policy::{Policy, Predicate, Rule};
+use uuid::Uuid;
+
+/// Build an [`Attribution`] for a concrete rule that fired.
+pub(super) fn rule_attribution(policy: &Policy, rule: &Rule) -> Attribution {
+    Attribution {
+        policy_id: policy.id.to_string().into(),
+        reason: Some(rule.id.to_string().into()),
+    }
+}
+
+/// Build an [`Attribution`] for a policy's `fallback`.
+pub(super) fn fallback_attribution(policy: &Policy) -> Attribution {
+    Attribution {
+        policy_id: policy.id.to_string().into(),
+        reason: None,
+    }
+}
+
+/// Build an [`Attribution`] for a reviewer override. Uses the
+/// sentinel `policy_id = "override"` so audits can distinguish
+/// override-driven redactions from policy-driven ones; `reason`
+/// carries the overridden entity's id.
+pub(crate) fn override_attribution(entity_id: Uuid) -> Attribution {
+    Attribution {
+        policy_id: "override".into(),
+        reason: Some(entity_id.to_string().into()),
+    }
+}
+
+/// Attach an `operator` to `anonymizer` for the single entity
+/// identified by `entity_id`. Used by the apply pipeline to give
+/// reviewer overrides higher precedence than any policy-driven
+/// rule.
+pub(crate) fn attach_override<M, O>(
+    anonymizer: Anonymizer<M>,
+    entity_id: Uuid,
+    operator: O,
+) -> Anonymizer<M>
+where
+    M: Modality,
+    O: Operator<M> + 'static,
+{
+    let attribution = override_attribution(entity_id);
+    anonymizer
+        .with_predicate(move |entity: &Entity<M>| entity.id == entity_id, operator)
+        .because(attribution)
+}
+
+/// Attach `operator` to `anonymizer` for the rule's
+/// [`Predicate`], stamping `attribution` so every redaction it
+/// drives carries `policy_id` + `reason` on its provenance event.
+///
+/// Pattern-matches the predicate against degenerate single-label
+/// / single-tag shapes to keep elide's indexed fast paths. Any
+/// composite predicate falls through to
+/// [`Anonymizer::with_catalog_predicate`].
+pub(super) fn attach<M, O>(
+    anonymizer: Anonymizer<M>,
+    predicate: &Predicate,
+    operator: O,
+    attribution: Attribution,
+) -> Anonymizer<M>
+where
+    M: Modality,
+    O: Operator<M> + Clone + 'static,
+{
+    let anonymizer = match predicate {
+        Predicate::LabelOneOf { labels } if labels.len() == 1 => {
+            anonymizer.with_label(LabelRef::new(labels[0].clone()), operator)
+        }
+        Predicate::TagOneOf { tags } if tags.len() == 1 => {
+            anonymizer.with_tag(tags[0].clone(), operator)
+        }
+        other => anonymizer.with_catalog_predicate(compile_predicate::<M>(other.clone()), operator),
+    };
+    anonymizer.because(attribution)
+}
+
+/// Compile a [`Predicate`] tree into a closure consumed by
+/// [`Anonymizer::with_catalog_predicate`]. The closure receives
+/// the per-anonymizer [`LabelCatalog`] alongside the entity so
+/// [`Predicate::TagOneOf`] resolves correctly even inside
+/// composites ([`All`] / [`Any`] / [`Not`]).
+///
+/// [`Anonymizer::with_catalog_predicate`]: elide::redaction::Anonymizer::with_catalog_predicate
+/// [`All`]: Predicate::All
+/// [`Any`]: Predicate::Any
+/// [`Not`]: Predicate::Not
+pub(super) fn compile_predicate<M>(
+    predicate: Predicate,
+) -> impl Fn(&Entity<M>, &LabelCatalog) -> bool + Send + Sync + 'static
+where
+    M: Modality,
+{
+    move |entity, catalog| eval(&predicate, entity, catalog)
+}
+
+fn eval<M: Modality>(predicate: &Predicate, entity: &Entity<M>, catalog: &LabelCatalog) -> bool {
+    match predicate {
+        Predicate::Confidence { min } => f32::from(entity.confidence) >= *min,
+        Predicate::LabelOneOf { labels } => labels
+            .iter()
+            .any(|l| LabelRef::new(l.clone()) == entity.label),
+        Predicate::TagOneOf { tags } => catalog
+            .get(&entity.label)
+            .is_some_and(|label| tags.iter().any(|tag| label.has_tag(tag.as_str()))),
+        Predicate::CoRef { coref } => entity
+            .coref
+            .as_ref()
+            .is_some_and(|c| c.as_str() == coref.as_str()),
+        Predicate::All { all } => all.iter().all(|p| eval(p, entity, catalog)),
+        Predicate::Any { any } => any.iter().any(|p| eval(p, entity, catalog)),
+        Predicate::Not { not } => !eval(not, entity, catalog),
+    }
+}
diff --git a/crates/nvisy-engine/src/anonymizer/tabular.rs b/crates/nvisy-engine/src/anonymizer/tabular.rs
new file mode 100644
index 000000000..5be30c2ae
--- /dev/null
+++ b/crates/nvisy-engine/src/anonymizer/tabular.rs
@@ -0,0 +1,168 @@
+//! Compile tabular-modality rules to elide operators + attach to
+//! an [`Anonymizer<Tabular>`].
+//!
+//! Tabular shares the [`TextBacked`] vocabulary with text plus the
+//! structural [`DropRow`] / [`DropColumn`] operators.
+//!
+//! [`TextBacked`]: elide_core::modality::TextBacked
+//! [`DropRow`]: elide::redaction::operators::DropRow
+//! [`DropColumn`]: elide::redaction::operators::DropColumn
+
+use elide::redaction::Anonymizer;
+use elide::redaction::operators::{DropColumn, DropRow, Erase, Keep, Mask, Replace, Sha2Hash};
+use elide_core::entity::LabelCatalog;
+use elide_core::modality::tabular::Tabular;
+use elide_core::{Error, ErrorKind};
+use nvisy_core::policy::redaction::TabularRedaction;
+use nvisy_core::policy::{Policy, Rule, RuleAction};
+use uuid::Uuid;
+
+use super::selector::{attach, attach_override, fallback_attribution, rule_attribution};
+
+/// Compile every tabular-applicable rule across `policies` into a
+/// tabular-modality anonymizer.
+pub fn compile_tabular(
+    policies: &[Policy],
+    catalog: LabelCatalog,
+) -> Result<Anonymizer<Tabular>, Error> {
+    let anonymizer = Anonymizer::<Tabular>::new().with_catalog(catalog);
+    attach_policies_tabular(anonymizer, policies.iter())
+}
+
+/// Attach every tabular-applicable rule from `policies` onto an
+/// already-constructed anonymizer. Takes an iterator so the
+/// apply pipeline can pre-filter by [`Policy::applies_when`]
+/// without cloning.
+///
+/// [`Policy::applies_when`]: nvisy_core::policy::Policy::applies_when
+pub(crate) fn attach_policies_tabular<'a>(
+    mut anonymizer: Anonymizer<Tabular>,
+    policies: impl Iterator<Item = &'a Policy>,
+) -> Result<Anonymizer<Tabular>, Error> {
+    for policy in policies {
+        for rule in &policy.rules {
+            let RuleAction::Redact(redactions) = &rule.action else {
+                continue;
+            };
+            let Some(spec) = &redactions.tabular else {
+                continue;
+            };
+            anonymizer = attach_rule(anonymizer, policy, rule, spec)?;
+        }
+        if let Some(RuleAction::Redact(redactions)) = &policy.fallback
+            && let Some(spec) = &redactions.tabular
+        {
+            anonymizer = attach_fallback(anonymizer, policy, spec)?;
+        }
+    }
+    Ok(anonymizer)
+}
+
+/// Attach a reviewer override for one entity. No-op when the
+/// override is not a tabular-modality `Redact`.
+pub(crate) fn attach_override_tabular(
+    anonymizer: Anonymizer<Tabular>,
+    entity_id: Uuid,
+    action: &RuleAction,
+) -> Result<Anonymizer<Tabular>, Error> {
+    let RuleAction::Redact(redactions) = action else {
+        return Ok(anonymizer);
+    };
+    let Some(spec) = &redactions.tabular else {
+        return Ok(anonymizer);
+    };
+    Ok(match build(spec)? {
+        TabularOp::Erase => attach_override(anonymizer, entity_id, Erase),
+        TabularOp::Keep => attach_override(anonymizer, entity_id, Keep),
+        TabularOp::Mask(op) => attach_override(anonymizer, entity_id, op),
+        TabularOp::Replace(op) => attach_override(anonymizer, entity_id, op),
+        TabularOp::Hash(op) => attach_override(anonymizer, entity_id, op),
+        TabularOp::DropRow => attach_override(anonymizer, entity_id, DropRow),
+        TabularOp::DropColumn => attach_override(anonymizer, entity_id, DropColumn),
+    })
+}
+
+fn attach_rule(
+    anonymizer: Anonymizer<Tabular>,
+    policy: &Policy,
+    rule: &Rule,
+    spec: &TabularRedaction,
+) -> Result<Anonymizer<Tabular>, Error> {
+    let attribution = rule_attribution(policy, rule);
+    Ok(match build(spec)? {
+        TabularOp::Erase => attach(anonymizer, &rule.predicate, Erase, attribution),
+        TabularOp::Keep => attach(anonymizer, &rule.predicate, Keep, attribution),
+        TabularOp::Mask(op) => attach(anonymizer, &rule.predicate, op, attribution),
+        TabularOp::Replace(op) => attach(anonymizer, &rule.predicate, op, attribution),
+        TabularOp::Hash(op) => attach(anonymizer, &rule.predicate, op, attribution),
+        TabularOp::DropRow => attach(anonymizer, &rule.predicate, DropRow, attribution),
+        TabularOp::DropColumn => attach(anonymizer, &rule.predicate, DropColumn, attribution),
+    })
+}
+
+fn attach_fallback(
+    anonymizer: Anonymizer<Tabular>,
+    policy: &Policy,
+    spec: &TabularRedaction,
+) -> Result<Anonymizer<Tabular>, Error> {
+    let attribution = fallback_attribution(policy);
+    Ok(match build(spec)? {
+        TabularOp::Erase => anonymizer.with_fallback(Erase).because(attribution),
+        TabularOp::Keep => anonymizer.with_fallback(Keep).because(attribution),
+        TabularOp::Mask(op) => anonymizer.with_fallback(op).because(attribution),
+        TabularOp::Replace(op) => anonymizer.with_fallback(op).because(attribution),
+        TabularOp::Hash(op) => anonymizer.with_fallback(op).because(attribution),
+        TabularOp::DropRow => anonymizer.with_fallback(DropRow).because(attribution),
+        TabularOp::DropColumn => anonymizer.with_fallback(DropColumn).because(attribution),
+    })
+}
+
+enum TabularOp {
+    Erase,
+    Keep,
+    Mask(Mask),
+    Replace(Replace),
+    Hash(Sha2Hash),
+    DropRow,
+    DropColumn,
+}
+
+fn build(spec: &TabularRedaction) -> Result<TabularOp, Error> {
+    Ok(match spec {
+        TabularRedaction::Erase => TabularOp::Erase,
+        TabularRedaction::Keep => TabularOp::Keep,
+        TabularRedaction::Mask {
+            mask_char,
+            keep_prefix,
+            keep_suffix,
+        } => TabularOp::Mask(
+            Mask::new(*mask_char)
+                .with_keep_prefix(*keep_prefix)
+                .with_keep_suffix(*keep_suffix),
+        ),
+        TabularRedaction::Replace { template } => {
+            TabularOp::Replace(Replace::new(template.clone()))
+        }
+        TabularRedaction::Hash { algorithm, salt } => {
+            let mut op = Sha2Hash::new((*algorithm).into());
+            if let Some(s) = salt {
+                op = op.with_salt(s.as_bytes().to_vec());
+            }
+            TabularOp::Hash(op)
+        }
+        TabularRedaction::DropRow => TabularOp::DropRow,
+        TabularRedaction::DropColumn => TabularOp::DropColumn,
+        TabularRedaction::Pseudonymize => return Err(stateful_not_wired("pseudonymize")),
+        TabularRedaction::Encrypt => return Err(stateful_not_wired("encrypt")),
+    })
+}
+
+fn stateful_not_wired(operator: &'static str) -> Error {
+    Error::new(
+        ErrorKind::Validation,
+        format!(
+            "policy compile: `{operator}` needs engine-side stateful infrastructure (vault / \
+             key provider) that is not wired into the compile surface yet",
+        ),
+    )
+}
diff --git a/crates/nvisy-engine/src/anonymizer/text.rs b/crates/nvisy-engine/src/anonymizer/text.rs
new file mode 100644
index 000000000..77a868b0d
--- /dev/null
+++ b/crates/nvisy-engine/src/anonymizer/text.rs
@@ -0,0 +1,161 @@
+//! Compile text-modality rules to elide operators + attach to an
+//! [`Anonymizer<Text>`].
+
+use elide::redaction::Anonymizer;
+use elide::redaction::operators::{Erase, Keep, Mask, Replace, Sha2Hash};
+use elide_core::entity::LabelCatalog;
+use elide_core::modality::text::Text;
+use elide_core::{Error, ErrorKind};
+use nvisy_core::policy::redaction::TextRedaction;
+use nvisy_core::policy::{Policy, Rule, RuleAction};
+use uuid::Uuid;
+
+use super::selector::{attach, attach_override, fallback_attribution, rule_attribution};
+
+/// Compile every text-applicable rule across `policies` into a
+/// text-modality anonymizer. Walks policies in precedence order
+/// and rules in declared order; the first matching rule wins per
+/// entity at apply time.
+pub fn compile_text(policies: &[Policy], catalog: LabelCatalog) -> Result<Anonymizer<Text>, Error> {
+    let anonymizer = Anonymizer::<Text>::new().with_catalog(catalog);
+    attach_policies_text(anonymizer, policies.iter())
+}
+
+/// Attach every text-applicable rule from `policies` onto an
+/// already-constructed anonymizer. The apply pipeline calls this
+/// after layering per-entity override rules so the overrides
+/// keep first-match precedence. Takes an iterator (not a slice)
+/// so callers can pre-filter by [`Policy::applies_when`] without
+/// cloning the policy set.
+///
+/// [`Policy::applies_when`]: nvisy_core::policy::Policy::applies_when
+pub(crate) fn attach_policies_text<'a>(
+    mut anonymizer: Anonymizer<Text>,
+    policies: impl Iterator<Item = &'a Policy>,
+) -> Result<Anonymizer<Text>, Error> {
+    for policy in policies {
+        for rule in &policy.rules {
+            let RuleAction::Redact(redactions) = &rule.action else {
+                continue;
+            };
+            let Some(spec) = &redactions.text else {
+                continue;
+            };
+            anonymizer = attach_rule(anonymizer, policy, rule, spec)?;
+        }
+        if let Some(RuleAction::Redact(redactions)) = &policy.fallback
+            && let Some(spec) = &redactions.text
+        {
+            anonymizer = attach_fallback(anonymizer, policy, spec)?;
+        }
+    }
+    Ok(anonymizer)
+}
+
+/// Attach a reviewer override for one entity. The override only
+/// applies to `RuleAction::Redact` with a text-modality spec —
+/// other action shapes (or non-text overrides) are no-ops at the
+/// per-modality dispatch boundary.
+pub(crate) fn attach_override_text(
+    anonymizer: Anonymizer<Text>,
+    entity_id: Uuid,
+    action: &RuleAction,
+) -> Result<Anonymizer<Text>, Error> {
+    let RuleAction::Redact(redactions) = action else {
+        return Ok(anonymizer);
+    };
+    let Some(spec) = &redactions.text else {
+        return Ok(anonymizer);
+    };
+    Ok(match build(spec)? {
+        TextOp::Erase => attach_override(anonymizer, entity_id, Erase),
+        TextOp::Keep => attach_override(anonymizer, entity_id, Keep),
+        TextOp::Mask(op) => attach_override(anonymizer, entity_id, op),
+        TextOp::Replace(op) => attach_override(anonymizer, entity_id, op),
+        TextOp::Hash(op) => attach_override(anonymizer, entity_id, op),
+    })
+}
+
+fn attach_rule(
+    anonymizer: Anonymizer<Text>,
+    policy: &Policy,
+    rule: &Rule,
+    spec: &TextRedaction,
+) -> Result<Anonymizer<Text>, Error> {
+    let attribution = rule_attribution(policy, rule);
+    Ok(match build(spec)? {
+        TextOp::Erase => attach(anonymizer, &rule.predicate, Erase, attribution),
+        TextOp::Keep => attach(anonymizer, &rule.predicate, Keep, attribution),
+        TextOp::Mask(op) => attach(anonymizer, &rule.predicate, op, attribution),
+        TextOp::Replace(op) => attach(anonymizer, &rule.predicate, op, attribution),
+        TextOp::Hash(op) => attach(anonymizer, &rule.predicate, op, attribution),
+    })
+}
+
+fn attach_fallback(
+    anonymizer: Anonymizer<Text>,
+    policy: &Policy,
+    spec: &TextRedaction,
+) -> Result<Anonymizer<Text>, Error> {
+    let attribution = fallback_attribution(policy);
+    Ok(match build(spec)? {
+        TextOp::Erase => anonymizer.with_fallback(Erase).because(attribution),
+        TextOp::Keep => anonymizer.with_fallback(Keep).because(attribution),
+        TextOp::Mask(op) => anonymizer.with_fallback(op).because(attribution),
+        TextOp::Replace(op) => anonymizer.with_fallback(op).because(attribution),
+        TextOp::Hash(op) => anonymizer.with_fallback(op).because(attribution),
+    })
+}
+
+/// Discriminated builder result so [`attach`] /
+/// [`Anonymizer::with_fallback`] can attach the right concrete
+/// operator type. We can't return `Box<dyn Operator<_>>` because
+/// [`Anonymizer::with_label`] takes `O: Operator<M> + 'static` by
+/// value.
+enum TextOp {
+    Erase,
+    Keep,
+    Mask(Mask),
+    Replace(Replace),
+    Hash(Sha2Hash),
+}
+
+fn build(spec: &TextRedaction) -> Result<TextOp, Error> {
+    Ok(match spec {
+        TextRedaction::Erase => TextOp::Erase,
+        TextRedaction::Keep => TextOp::Keep,
+        TextRedaction::Mask {
+            mask_char,
+            keep_prefix,
+            keep_suffix,
+        } => TextOp::Mask(
+            Mask::new(*mask_char)
+                .with_keep_prefix(*keep_prefix)
+                .with_keep_suffix(*keep_suffix),
+        ),
+        TextRedaction::Replace { template } => TextOp::Replace(Replace::new(template.clone())),
+        TextRedaction::Hash { algorithm, salt } => {
+            let mut op = Sha2Hash::new((*algorithm).into());
+            if let Some(s) = salt {
+                op = op.with_salt(s.as_bytes().to_vec());
+            }
+            TextOp::Hash(op)
+        }
+        TextRedaction::Pseudonymize => {
+            return Err(stateful_not_wired("pseudonymize", "vault + generator"));
+        }
+        TextRedaction::Encrypt => {
+            return Err(stateful_not_wired("encrypt", "key provider"));
+        }
+    })
+}
+
+fn stateful_not_wired(operator: &'static str, infrastructure: &'static str) -> Error {
+    Error::new(
+        ErrorKind::Validation,
+        format!(
+            "policy compile: `{operator}` needs an engine-side {infrastructure}; \
+             stateful operator infrastructure is not wired into the compile surface yet",
+        ),
+    )
+}
diff --git a/crates/nvisy-engine/src/core/config.rs b/crates/nvisy-engine/src/core/config.rs
deleted file mode 100644
index fcc4cffc2..000000000
--- a/crates/nvisy-engine/src/core/config.rs
+++ /dev/null
@@ -1,183 +0,0 @@
-//! Engine-wide configuration, typically deserialized from TOML.
-//!
-//! [`RuntimeConfig`] is the top-level configuration object containing
-//! optional subsystem sections — [`EngineConfig`],
-//! [`ExtractionConfig`][ec], [`DetectionConfig`][dc],
-//! [`RedactionConfig`][rc].
-//!
-//! Per-request plan nodes (`Extraction`, `Redaction`,
-//! `DeduplicationParams`, `Validation`) live alongside their
-//! corresponding side's config in [`crate::detection`] or
-//! [`crate::redaction`]; the per-phase modules consume them at
-//! dispatch time.
-//!
-//! # Post-load steps
-//!
-//! After deserializing from TOML, callers should:
-//! 1. Call [`RuntimeConfig::resolve_env`] to fill empty `api_key`
-//!    fields from environment variables.
-//! 2. Call [`RuntimeConfig::validate`] to check structural
-//!    constraints.
-//!
-//! [ec]: crate::detection::ExtractionConfig
-//! [dc]: crate::detection::DetectionConfig
-//! [rc]: crate::redaction::RedactionConfig
-
-use std::num::NonZeroUsize;
-use std::time::Duration;
-
-use nvisy_core::{Error, Result};
-use nvisy_llm::backend::http::HttpConfig;
-use semver::Version;
-use serde::{Deserialize, Serialize};
-use validator::Validate;
-
-use crate::detection::{DetectionConfig, ExtractionConfig};
-use crate::redaction::RedactionConfig;
-
-fn default_config_version() -> Version {
-    Version::new(0, 1, 0)
-}
-
-/// Top-level pipeline configuration, typically deserialized from TOML.
-///
-/// Contains optional subsystem sections. The CLI layer owns the full
-/// TOML shape (including `[server]`) and passes this struct downstream
-/// to the engines.
-///
-/// Every section is load-once: the engines read this struct at
-/// startup, build the per-section state behind `Arc`s on their inner
-/// shared state, and never re-read it. Per-request override is not
-/// supported — plans tune behaviour through their own per-phase
-/// config nodes, not by resupplying a `RuntimeConfig`.
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct RuntimeConfig {
-    /// Configuration schema version.
-    #[serde(default = "default_config_version")]
-    pub version: Version,
-
-    /// Engine-level execution policies, networking, and resource
-    /// limits.
-    pub engine: Option<EngineConfig>,
-    /// Extraction registry — `[extractor.ocr]`, `[extractor.stt]`
-    /// sub-sections. Built once at engine startup; the `Extraction`
-    /// phase config carries per-call flags only.
-    pub extraction: Option<ExtractionConfig>,
-    /// Detection registry — `[detection.pattern]`,
-    /// `[detection.ner]` sub-sections. Built once at engine startup;
-    /// the `Detection` phase config only references these by kind.
-    pub detection: Option<DetectionConfig>,
-    /// Deployment-wide redaction defaults — `[redaction]` section.
-    /// Built once at engine startup; the per-plan `Redaction` node
-    /// falls back to these for any `None` fields. Per-request
-    /// override is not supported.
-    pub redaction: Option<RedactionConfig>,
-}
-
-impl Default for RuntimeConfig {
-    fn default() -> Self {
-        Self {
-            version: default_config_version(),
-            engine: None,
-            extraction: None,
-            detection: None,
-            redaction: None,
-        }
-    }
-}
-
-impl RuntimeConfig {
-    /// Resource limits from the engine section, or defaults.
-    #[must_use]
-    pub fn effective_limits(&self) -> ResourceLimits {
-        self.engine.as_ref().map(|e| e.limits).unwrap_or_default()
-    }
-
-    /// Concurrency limit from the engine section's resource limits,
-    /// if configured.
-    #[must_use]
-    pub fn effective_concurrency(&self) -> Option<NonZeroUsize> {
-        self.engine.as_ref().and_then(|e| e.limits.concurrency)
-    }
-
-    /// Validate all configuration sections.
-    ///
-    /// Checks structural constraints (e.g. retry/timeout ranges)
-    /// using the `validator` crate. Should be called once after
-    /// deserialization and after any merge.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error listing all constraint violations.
-    pub fn validate(&self) -> Result<(), Error> {
-        if let Some(ref engine) = self.engine {
-            engine
-                .validate()
-                .map_err(|e| Error::validation(format!("engine: {e}"), "config"))?;
-        }
-        Ok(())
-    }
-
-    /// Resolve `api_key` fields from environment variables.
-    ///
-    /// Placeholder: per-extractor/per-recognizer provider configs
-    /// will get their own env-var resolution path in a follow-up.
-    pub fn resolve_env(&mut self) {}
-}
-
-/// Engine-level configuration: networking + resource limits.
-///
-/// All settings are deployment-side — set once in `Nvisy.toml` by
-/// the operator. Per-request overrides apply only to fields
-/// explicitly noted as overridable.
-#[derive(Debug, Clone, Default, Validate, Serialize, Deserialize)]
-pub struct EngineConfig {
-    /// Shared HTTP client configuration for all downstream API calls.
-    ///
-    /// Applies to OCR providers, LLM agents, STT services, and any
-    /// other external HTTP dependencies. Controls timeouts, retries,
-    /// and connection pooling.
-    pub http: Option<HttpConfig>,
-
-    /// Run-level resource limits (concurrency cap + timeout).
-    ///
-    /// Nested under `[engine.limits]` in TOML.
-    #[validate(nested)]
-    #[serde(default)]
-    pub limits: ResourceLimits,
-}
-
-/// Hard limits on pipeline resource consumption.
-///
-/// Deployment-side caps applied to every
-/// [`DetectionEngine::detect`][de] / [`RedactionEngine::redact`][re]
-/// pass — not adjustable per-request.
-///
-/// [de]: crate::detection::DetectionEngine::detect
-/// [re]: crate::redaction::RedactionEngine::redact
-#[derive(Debug, Clone, Copy, Default, Validate, Serialize, Deserialize)]
-pub struct ResourceLimits {
-    /// Maximum number of documents processed in parallel via a
-    /// shared [`Semaphore`]. Server-wide; not overridable
-    /// per-request. `None` means unbounded.
-    ///
-    /// [`Semaphore`]: tokio::sync::Semaphore
-    #[serde(default)]
-    pub concurrency: Option<NonZeroUsize>,
-
-    /// Hard ceiling on total pipeline run duration.
-    ///
-    /// On expiry, the run-level cancellation token fires and the
-    /// run is marked as timed out. `None` means no run-level
-    /// timeout — rely on external supervision (k8s liveness, etc.)
-    /// instead.
-    ///
-    /// Parses from human-friendly strings via `humantime_serde`:
-    /// `"60s"`, `"5m"`, `"1h30m"`.
-    #[serde(
-        default,
-        with = "humantime_serde",
-        skip_serializing_if = "Option::is_none"
-    )]
-    pub run_timeout: Option<Duration>,
-}
diff --git a/crates/nvisy-engine/src/core/context.rs b/crates/nvisy-engine/src/core/context.rs
deleted file mode 100644
index d988fd440..000000000
--- a/crates/nvisy-engine/src/core/context.rs
+++ /dev/null
@@ -1,187 +0,0 @@
-//! Per-pass execution contexts: [`DetectionContext`] and
-//! [`RedactionContext`].
-//!
-//! Each side carries only the engine resources it actually consumes
-//! — detection needs the extractor and recognizer registries,
-//! redaction needs the redaction config and per-modality
-//! anonymizer registries. The fields they share — run-wide
-//! [`SharedData`], cancellation, concurrency cap — are exposed
-//! through the [`PhaseContext`] trait so modality-agnostic phases
-//! (deduplication, validation) can borrow either context type
-//! without duplication.
-//!
-//! Lives in `core/` because phases consume these types through
-//! their `ctx` parameter — they're part of the phase contract
-//! surface, not the orchestrator's private state.
-
-use std::num::NonZeroUsize;
-use std::sync::Arc;
-
-use nvisy_toolkit::detection::RecognizerRegistry;
-use nvisy_toolkit::extraction::ExtractorRegistry;
-use tokio_util::sync::CancellationToken;
-
-use super::SharedData;
-use crate::redaction::RedactionConfig;
-use crate::redaction::phases::RedactionRegistries;
-
-/// Shared surface every phase reads from regardless of which side
-/// (detection or redaction) it runs on. Implemented by both
-/// [`DetectionContext`] and [`RedactionContext`]; modality-agnostic
-/// phases (deduplication, validation) bound on this trait so they
-/// can borrow either context type.
-pub trait PhaseContext {
-    /// Run-wide shared state (policies, registry, key provider,
-    /// per-request entity-label catalog).
-    fn shared(&self) -> &Arc<SharedData>;
-    /// `true` when this run's cancellation token has fired.
-    fn is_cancelled(&self) -> bool;
-    /// Optional document-concurrency cap from `[engine.limits]`.
-    fn concurrency(&self) -> Option<NonZeroUsize>;
-}
-
-/// Per-pass detection context. Detection-side phases (extraction,
-/// detection, deduplication) consume it; redaction-side phases
-/// never see it.
-pub struct DetectionContext {
-    pub(crate) cancel: CancellationToken,
-    pub(crate) shared: Arc<SharedData>,
-    pub(crate) extraction_engine: ExtractorRegistry,
-    /// Per-request recognizer registry — built fresh from the
-    /// engine-side detection-config template plus the request's
-    /// label catalog.
-    pub(crate) recognizer_registry: Arc<RecognizerRegistry>,
-    pub(crate) concurrency: Option<NonZeroUsize>,
-}
-
-/// Engine resources the detection pipeline borrows when building a
-/// [`DetectionContext`]. Bundled into one type so adding new
-/// detection-side engines doesn't widen `DetectionContext::new`'s
-/// signature.
-pub(crate) struct DetectionEngines {
-    pub extraction_engine: ExtractorRegistry,
-    pub recognizer_registry: Arc<RecognizerRegistry>,
-}
-
-impl DetectionContext {
-    /// Construct a [`DetectionContext`] from its parts. Called once
-    /// per pass by the detection pipeline.
-    pub(crate) fn new(
-        cancel: CancellationToken,
-        shared: Arc<SharedData>,
-        engines: DetectionEngines,
-        concurrency: Option<NonZeroUsize>,
-    ) -> Self {
-        let DetectionEngines {
-            extraction_engine,
-            recognizer_registry,
-        } = engines;
-        Self {
-            cancel,
-            shared,
-            extraction_engine,
-            recognizer_registry,
-            concurrency,
-        }
-    }
-
-    /// Pre-built extraction engine borrowed by [`ExtractionPhase`].
-    ///
-    /// [`ExtractionPhase`]: crate::detection::phases::extraction::ExtractionPhase
-    pub(crate) fn extraction_engine(&self) -> &ExtractorRegistry {
-        &self.extraction_engine
-    }
-
-    /// Per-request recognizer registry borrowed by
-    /// [`DetectionPhase`].
-    ///
-    /// [`DetectionPhase`]: crate::detection::phases::detection::DetectionPhase
-    pub(crate) fn recognizer_registry(&self) -> &Arc<RecognizerRegistry> {
-        &self.recognizer_registry
-    }
-}
-
-impl PhaseContext for DetectionContext {
-    fn shared(&self) -> &Arc<SharedData> {
-        &self.shared
-    }
-
-    fn is_cancelled(&self) -> bool {
-        self.cancel.is_cancelled()
-    }
-
-    fn concurrency(&self) -> Option<NonZeroUsize> {
-        self.concurrency
-    }
-}
-
-/// Per-pass redaction context. Redaction-side phases (redaction,
-/// validation) consume it; detection-side phases never see it.
-pub struct RedactionContext {
-    pub(crate) cancel: CancellationToken,
-    pub(crate) shared: Arc<SharedData>,
-    pub(crate) redaction_config: RedactionConfig,
-    pub(crate) redaction_registries: RedactionRegistries,
-    pub(crate) concurrency: Option<NonZeroUsize>,
-}
-
-/// Engine resources the redaction pipeline borrows when building a
-/// [`RedactionContext`]. Bundled into one type so adding new
-/// redaction-side engines doesn't widen `RedactionContext::new`'s
-/// signature.
-pub(crate) struct RedactionEngines {
-    pub redaction_config: RedactionConfig,
-    pub redaction_registries: RedactionRegistries,
-}
-
-impl RedactionContext {
-    /// Construct a [`RedactionContext`] from its parts. Called once
-    /// per pass by the redaction pipeline.
-    pub(crate) fn new(
-        cancel: CancellationToken,
-        shared: Arc<SharedData>,
-        engines: RedactionEngines,
-        concurrency: Option<NonZeroUsize>,
-    ) -> Self {
-        let RedactionEngines {
-            redaction_config,
-            redaction_registries,
-        } = engines;
-        Self {
-            cancel,
-            shared,
-            redaction_config,
-            redaction_registries,
-            concurrency,
-        }
-    }
-
-    /// Server-wide redaction defaults the [`RedactionPhase`] reads.
-    ///
-    /// [`RedactionPhase`]: crate::redaction::phases::RedactionPhase
-    pub(crate) fn redaction_config(&self) -> &RedactionConfig {
-        &self.redaction_config
-    }
-
-    /// Per-modality custom-anonymizer registries the
-    /// [`RedactionPhase`] consults for `Custom`-arm lookups.
-    ///
-    /// [`RedactionPhase`]: crate::redaction::phases::RedactionPhase
-    pub(crate) fn redaction_registries(&self) -> &RedactionRegistries {
-        &self.redaction_registries
-    }
-}
-
-impl PhaseContext for RedactionContext {
-    fn shared(&self) -> &Arc<SharedData> {
-        &self.shared
-    }
-
-    fn is_cancelled(&self) -> bool {
-        self.cancel.is_cancelled()
-    }
-
-    fn concurrency(&self) -> Option<NonZeroUsize> {
-        self.concurrency
-    }
-}
diff --git a/crates/nvisy-engine/src/core/health.rs b/crates/nvisy-engine/src/core/health.rs
deleted file mode 100644
index 3d3703d33..000000000
--- a/crates/nvisy-engine/src/core/health.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-//! Composition helpers for [`Healthcheck`]-implementing components.
-//!
-//! The trait + data shapes live in [`nvisy_core::health`]. This
-//! module sits one layer up and aggregates: probe a set of
-//! components concurrently, collect their reports.
-//!
-//! Roll-up status and per-probe timeouts are caller concerns — wrap
-//! individual probes in [`tokio::time::timeout`] before passing
-//! them in, and compute a `Healthy / Degraded / Unhealthy`
-//! overall from the returned [`ComponentCheck`]s where the
-//! report is rendered.
-
-use futures::future::join_all;
-use nvisy_core::health::{ComponentCheck, Healthcheck};
-
-/// Probe every [`Healthcheck`] in `items` concurrently and collect
-/// their reports.
-///
-/// Each item's `name()` is read once and paired with its
-/// `healthcheck()` result. The output preserves the input order —
-/// no sorting, no roll-up, no overall status. The caller decides
-/// how to render the list (HTTP response body, CLI table,
-/// dashboard) and whether to compute a roll-up.
-pub async fn probe_all<'a, I>(items: I) -> Vec<ComponentCheck>
-where
-    I: IntoIterator<Item = &'a dyn Healthcheck>,
-{
-    let futures = items.into_iter().map(|item| async move {
-        let name = item.name();
-        let status = item.healthcheck().await;
-        ComponentCheck { name, status }
-    });
-    join_all(futures).await
-}
diff --git a/crates/nvisy-engine/src/core/ingestion/compression/mod.rs b/crates/nvisy-engine/src/core/ingestion/compression/mod.rs
deleted file mode 100644
index 53de4f181..000000000
--- a/crates/nvisy-engine/src/core/ingestion/compression/mod.rs
+++ /dev/null
@@ -1,72 +0,0 @@
-//! Content compression and decompression.
-//!
-//! Used as pre/post-processing steps within `ImportFile` and
-//! `ExportFile` operations, not as standalone pipeline operations.
-//!
-//! Gzip and Zstd are recognized but not yet implemented — selecting
-//! them returns a runtime error.
-
-use bytes::Bytes;
-use nvisy_core::{Error, Result};
-
-use crate::core::ingestion::CompressionAlgorithm;
-
-const TARGET: &str = "nvisy_engine::op::compression";
-
-/// Compression and decompression service.
-///
-/// Wraps a specific algorithm selected at construction time.
-pub struct CompressionService {
-    algorithm: CompressionAlgorithm,
-}
-
-impl CompressionService {
-    /// Create a new service for the given algorithm.
-    pub fn new(algorithm: CompressionAlgorithm) -> Self {
-        Self { algorithm }
-    }
-
-    /// Compress raw bytes.
-    pub fn compress(&self, data: &[u8]) -> Result<Bytes> {
-        tracing::debug!(
-            target: TARGET,
-            algorithm = ?self.algorithm,
-            input_len = data.len(),
-            "compressing content",
-        );
-        match self.algorithm {
-            CompressionAlgorithm::Gzip => Err(Error::runtime(
-                "gzip compression not yet implemented",
-                "compression",
-                false,
-            )),
-            CompressionAlgorithm::Zstd => Err(Error::runtime(
-                "zstd compression not yet implemented",
-                "compression",
-                false,
-            )),
-        }
-    }
-
-    /// Decompress raw bytes.
-    pub fn decompress(&self, data: &[u8]) -> Result<Bytes> {
-        tracing::debug!(
-            target: TARGET,
-            algorithm = ?self.algorithm,
-            input_len = data.len(),
-            "decompressing content",
-        );
-        match self.algorithm {
-            CompressionAlgorithm::Gzip => Err(Error::runtime(
-                "gzip decompression not yet implemented",
-                "compression",
-                false,
-            )),
-            CompressionAlgorithm::Zstd => Err(Error::runtime(
-                "zstd decompression not yet implemented",
-                "compression",
-                false,
-            )),
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/core/ingestion/encryption/mod.rs b/crates/nvisy-engine/src/core/ingestion/encryption/mod.rs
deleted file mode 100644
index d827bc2e4..000000000
--- a/crates/nvisy-engine/src/core/ingestion/encryption/mod.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-//! Content encryption and decryption.
-//!
-//! Provides `CryptoService` for AES-256-GCM encryption/decryption,
-//! [`KeyProvider`] for key resolution, and a self-describing wire format.
-
-mod provider;
-mod service;
-mod wire;
-
-pub use self::provider::{KeyProvider, SharedKeyProvider, StaticKeyProvider};
-pub(crate) use self::service::CryptoService;
-pub(crate) use self::wire::EncryptedContent;
diff --git a/crates/nvisy-engine/src/core/ingestion/encryption/provider.rs b/crates/nvisy-engine/src/core/ingestion/encryption/provider.rs
deleted file mode 100644
index 71b3304ce..000000000
--- a/crates/nvisy-engine/src/core/ingestion/encryption/provider.rs
+++ /dev/null
@@ -1,107 +0,0 @@
-//! Key provider abstraction for encryption key resolution.
-
-use std::collections::HashMap;
-use std::fmt;
-use std::sync::Arc;
-
-use base64::Engine;
-use base64::engine::general_purpose::STANDARD as BASE64;
-use bytes::Bytes;
-use nvisy_core::{Error, Result};
-/// Abstraction for resolving encryption keys by identifier.
-pub trait KeyProvider: Send + Sync {
-    /// Returns the raw key bytes for the given `key_id`, or an error if unknown.
-    fn resolve(&self, key_id: &str) -> Result<Bytes>;
-}
-
-/// Shared, cheaply-clonable handle to a [`KeyProvider`].
-#[derive(Clone)]
-pub struct SharedKeyProvider(Arc<dyn KeyProvider>);
-
-impl SharedKeyProvider {
-    /// Wrap a concrete [`KeyProvider`] implementation.
-    pub fn new(provider: impl KeyProvider + 'static) -> Self {
-        Self(Arc::new(provider))
-    }
-}
-
-impl KeyProvider for SharedKeyProvider {
-    fn resolve(&self, key_id: &str) -> Result<Bytes> {
-        self.0.resolve(key_id)
-    }
-}
-
-impl Default for SharedKeyProvider {
-    fn default() -> Self {
-        Self::new(StaticKeyProvider::default())
-    }
-}
-
-impl fmt::Debug for SharedKeyProvider {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_tuple("SharedKeyProvider").finish()
-    }
-}
-
-/// In-memory key store for tests and simple deployments.
-#[derive(Debug, Clone, Default)]
-pub struct StaticKeyProvider {
-    keys: HashMap<String, Bytes>,
-}
-
-impl StaticKeyProvider {
-    /// Creates a new provider from an iterator of `(key_id, key_bytes)` pairs.
-    pub fn new(keys: impl IntoIterator<Item = (String, Vec<u8>)>) -> Self {
-        Self {
-            keys: keys
-                .into_iter()
-                .map(|(id, bytes)| (id, Bytes::from(bytes)))
-                .collect(),
-        }
-    }
-
-    /// Insert a single key, replacing any existing entry with the same ID.
-    pub fn insert(&mut self, key_id: impl Into<String>, key: impl Into<Bytes>) {
-        self.keys.insert(key_id.into(), key.into());
-    }
-
-    /// Build from a JSON value.
-    ///
-    /// Expects an object whose keys are key IDs and values are
-    /// base64-encoded key bytes. Any other shape returns an empty provider.
-    pub fn from_json(value: &serde_json::Value) -> Result<Self> {
-        let Some(map) = value.as_object() else {
-            return Ok(Self::default());
-        };
-
-        let mut keys = HashMap::with_capacity(map.len());
-        for (id, val) in map {
-            let encoded = val.as_str().ok_or_else(|| {
-                Error::validation(
-                    format!("key \"{id}\" must be a base64-encoded string"),
-                    "StaticKeyProvider::from_json",
-                )
-            })?;
-            let bytes = BASE64.decode(encoded).map_err(|e| {
-                Error::validation(
-                    format!("key \"{id}\" is not valid base64: {e}"),
-                    "StaticKeyProvider::from_json",
-                )
-            })?;
-            keys.insert(id.clone(), Bytes::from(bytes));
-        }
-
-        Ok(Self { keys })
-    }
-}
-
-impl KeyProvider for StaticKeyProvider {
-    fn resolve(&self, key_id: &str) -> Result<Bytes> {
-        self.keys.get(key_id).cloned().ok_or_else(|| {
-            Error::validation(
-                format!("unknown key_id: {key_id}"),
-                "StaticKeyProvider::resolve",
-            )
-        })
-    }
-}
diff --git a/crates/nvisy-engine/src/core/ingestion/encryption/service.rs b/crates/nvisy-engine/src/core/ingestion/encryption/service.rs
deleted file mode 100644
index 4d67193ab..000000000
--- a/crates/nvisy-engine/src/core/ingestion/encryption/service.rs
+++ /dev/null
@@ -1,211 +0,0 @@
-//! AES-256-GCM encryption and decryption service.
-
-use aes_gcm::aead::Aead;
-use aes_gcm::{Aes256Gcm, KeyInit, Nonce};
-use bytes::Bytes;
-use nvisy_codec::content::ContentData;
-use nvisy_core::{Error, ErrorKind, Result};
-use rand::RngExt;
-
-use super::provider::{KeyProvider, SharedKeyProvider};
-use super::wire::{EncryptedContent, NONCE_SIZE, WireEnvelope};
-use crate::core::ingestion::EncryptionAlgorithm;
-
-const TARGET: &str = "nvisy_engine::op::encryption";
-
-/// AES-256-GCM encryption and decryption service.
-pub struct CryptoService {
-    key_id: String,
-    key_provider: SharedKeyProvider,
-}
-
-impl CryptoService {
-    /// Creates a new crypto service.
-    pub fn new(key_id: impl Into<String>, key_provider: SharedKeyProvider) -> Self {
-        Self {
-            key_id: key_id.into(),
-            key_provider,
-        }
-    }
-
-    /// Encrypt the supplied encoded content into an
-    /// [`EncryptedContent`] blob. Modality-agnostic — the caller has
-    /// already pulled bytes out of whatever handle they own.
-    pub async fn encrypt(&self, content: &ContentData) -> Result<EncryptedContent> {
-        let source = content.content_source;
-        let plaintext = content.as_bytes();
-
-        let raw_key = self.key_provider.resolve(&self.key_id)?;
-        let cipher = Aes256Gcm::new_from_slice(&raw_key).map_err(|e| {
-            Error::validation(
-                format!("invalid AES-256 key: {e}"),
-                "CryptoService::encrypt",
-            )
-        })?;
-
-        let mut nonce_bytes = [0u8; NONCE_SIZE];
-        rand::rng().fill(&mut nonce_bytes);
-        let nonce = Nonce::from_slice(&nonce_bytes);
-
-        let ciphertext = cipher.encrypt(nonce, plaintext).map_err(|e| {
-            Error::new(
-                ErrorKind::Internal,
-                format!("AES-256-GCM encryption failed: {e}"),
-            )
-        })?;
-
-        let wire = WireEnvelope {
-            algorithm: EncryptionAlgorithm::Aes256Gcm,
-            key_id: &self.key_id,
-            nonce: &nonce_bytes,
-            ciphertext: &ciphertext,
-        }
-        .build()?;
-
-        tracing::debug!(
-            target: TARGET,
-            key_id = %self.key_id,
-            plaintext_len = plaintext.len(),
-            wire_len = wire.len(),
-            "encrypted content",
-        );
-
-        Ok(EncryptedContent {
-            source,
-            ciphertext: wire,
-            key_id: self.key_id.clone(),
-            algorithm: EncryptionAlgorithm::Aes256Gcm,
-        })
-    }
-
-    /// Decrypt an [`EncryptedContent`] blob back to raw [`ContentData`].
-    pub async fn decrypt(&self, encrypted: EncryptedContent) -> Result<ContentData> {
-        let envelope = WireEnvelope::parse(&encrypted.ciphertext)?;
-
-        if envelope.algorithm != EncryptionAlgorithm::Aes256Gcm {
-            return Err(Error::validation(
-                format!("unsupported algorithm: {:?}", envelope.algorithm),
-                "CryptoService::decrypt",
-            ));
-        }
-
-        if envelope.key_id != self.key_id {
-            return Err(Error::validation(
-                format!(
-                    "key_id mismatch: wire envelope contains '{}' but service expects '{}'",
-                    envelope.key_id, self.key_id,
-                ),
-                "CryptoService::decrypt",
-            ));
-        }
-
-        tracing::debug!(
-            target: TARGET,
-            key_id = %encrypted.key_id,
-            algorithm = ?encrypted.algorithm,
-            "decrypting content",
-        );
-
-        let raw_key = self.key_provider.resolve(envelope.key_id)?;
-        let cipher = Aes256Gcm::new_from_slice(&raw_key).map_err(|e| {
-            Error::validation(
-                format!("invalid AES-256 key: {e}"),
-                "CryptoService::decrypt",
-            )
-        })?;
-
-        let nonce = Nonce::from_slice(envelope.nonce);
-        let plaintext = cipher.decrypt(nonce, envelope.ciphertext).map_err(|e| {
-            Error::validation(
-                format!("AES-256-GCM decryption failed (authentication error): {e}"),
-                "CryptoService::decrypt",
-            )
-        })?;
-
-        tracing::debug!(
-            target: TARGET,
-            key_id = %envelope.key_id,
-            ciphertext_len = envelope.ciphertext.len(),
-            plaintext_len = plaintext.len(),
-            "decrypted content",
-        );
-
-        let content = ContentData::new(encrypted.source, Bytes::from(plaintext));
-        Ok(content)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_codec::content::{ContentData, ContentSource};
-
-    use super::*;
-    use crate::core::ingestion::encryption::{SharedKeyProvider, StaticKeyProvider};
-
-    fn test_key_provider() -> SharedKeyProvider {
-        let key = vec![0xAB; 32];
-        SharedKeyProvider::new(StaticKeyProvider::new([("test-key".to_string(), key)]))
-    }
-
-    fn test_content() -> ContentData {
-        ContentData::from_text(ContentSource::new(), "Hello, world!")
-    }
-
-    #[tokio::test]
-    async fn round_trip_encrypt_decrypt() {
-        let provider = test_key_provider();
-        let content = test_content();
-        let original_bytes = content.as_bytes().to_vec();
-
-        let svc = CryptoService::new("test-key", provider);
-        let encrypted = svc.encrypt(&content).await.expect("encrypt");
-
-        assert_eq!(encrypted.key_id, "test-key");
-        assert_eq!(encrypted.algorithm, EncryptionAlgorithm::Aes256Gcm);
-        assert_ne!(encrypted.ciphertext, original_bytes);
-
-        let decrypted = svc.decrypt(encrypted).await.expect("decrypt");
-        assert_eq!(decrypted.as_bytes(), &original_bytes[..]);
-    }
-
-    #[tokio::test]
-    async fn wrong_key_fails_decryption() {
-        let provider = test_key_provider();
-        let content = test_content();
-
-        let enc_svc = CryptoService::new("test-key", provider);
-        let encrypted = enc_svc.encrypt(&content).await.expect("encrypt");
-
-        let wrong_provider = SharedKeyProvider::new(StaticKeyProvider::new([(
-            "test-key".to_string(),
-            vec![0xCD; 32],
-        )]));
-        let dec_svc = CryptoService::new("test-key", wrong_provider);
-        assert!(dec_svc.decrypt(encrypted).await.is_err());
-    }
-
-    #[tokio::test]
-    async fn unknown_key_id_fails() {
-        let empty_provider = SharedKeyProvider::new(StaticKeyProvider::new([]));
-        let content = test_content();
-
-        let svc = CryptoService::new("nonexistent", empty_provider);
-        assert!(svc.encrypt(&content).await.is_err());
-    }
-
-    #[tokio::test]
-    async fn tampered_ciphertext_fails() {
-        let provider = test_key_provider();
-        let content = test_content();
-
-        let svc = CryptoService::new("test-key", provider);
-        let mut encrypted = svc.encrypt(&content).await.expect("encrypt");
-
-        let mut tampered = encrypted.ciphertext.to_vec();
-        let last = tampered.len() - 1;
-        tampered[last] ^= 0xFF;
-        encrypted.ciphertext = Bytes::from(tampered);
-
-        assert!(svc.decrypt(encrypted).await.is_err());
-    }
-}
diff --git a/crates/nvisy-engine/src/core/ingestion/encryption/wire.rs b/crates/nvisy-engine/src/core/ingestion/encryption/wire.rs
deleted file mode 100644
index 83dbe4b96..000000000
--- a/crates/nvisy-engine/src/core/ingestion/encryption/wire.rs
+++ /dev/null
@@ -1,203 +0,0 @@
-//! Self-describing wire envelope for encrypted content.
-//!
-//! # Integrity
-//!
-//! The wire format relies on the AEAD construction's authentication
-//! tag (16 bytes appended to ciphertext for AES-256-GCM) to detect
-//! tampering. There is **no** separate MAC over the header bytes
-//! (magic / version / algorithm / key_id / nonce). This is safe
-//! because all header fields feed into either the AEAD's key
-//! derivation (key_id resolves to the actual key) or its nonce
-//! (the nonce is the AEAD's nonce, not a separate value), so a
-//! header-only tamper either produces a different decryption key
-//! or a wrong nonce — both surface as a tag-verification failure
-//! at decrypt time. Any future algorithm added via
-//! [`EncryptionAlgorithm`] must preserve this property (every
-//! header field must be either bound to the AEAD's key/nonce or
-//! covered by a separate MAC).
-
-use std::str;
-
-use bytes::Bytes;
-use nvisy_codec::content::ContentSource;
-use nvisy_core::{Error, Result};
-
-use crate::core::ingestion::EncryptionAlgorithm;
-
-/// Wire-format magic bytes identifying an Nvisy encrypted blob.
-const MAGIC: &[u8; 4] = b"NVSE";
-/// Wire-format version.
-const WIRE_VERSION: u8 = 0x01;
-/// AES-256-GCM nonce size in bytes.
-pub(crate) const NONCE_SIZE: usize = 12;
-/// Minimum wire envelope size: magic(4) + version(1) + algo(1) + key_id_len(2) + nonce(12).
-const MIN_HEADER_SIZE: usize = 4 + 1 + 1 + 2 + NONCE_SIZE;
-
-/// Encrypted content blob with metadata.
-#[derive(Debug, Clone)]
-pub struct EncryptedContent {
-    /// Source identity of the original content.
-    pub source: ContentSource,
-    /// Self-describing wire-format blob (header + ciphertext).
-    pub ciphertext: Bytes,
-    /// Identifier of the key used for encryption.
-    pub key_id: String,
-    /// Algorithm used for encryption.
-    pub algorithm: EncryptionAlgorithm,
-}
-
-/// Self-describing wire envelope.
-///
-/// Wire format:
-/// ```text
-/// [4B magic "NVSE"] [1B version] [1B algo]
-/// [2B key_id len BE] [N bytes key_id UTF-8]
-/// [12B nonce] [ciphertext + 16B GCM tag]
-/// ```
-pub(crate) struct WireEnvelope<'a> {
-    pub algorithm: EncryptionAlgorithm,
-    pub key_id: &'a str,
-    pub nonce: &'a [u8; NONCE_SIZE],
-    pub ciphertext: &'a [u8],
-}
-
-impl WireEnvelope<'_> {
-    /// Serialise to the on-wire byte sequence.
-    ///
-    /// # Errors
-    ///
-    /// Returns an error if `key_id` is longer than `u16::MAX` bytes
-    /// — the wire format prefixes it with a `u16` length. Callers
-    /// in this crate always pass short identifiers; this guard
-    /// exists so a future caller with an unbounded `key_id` source
-    /// fails loudly at build time instead of silently truncating.
-    pub fn build(&self) -> Result<Bytes> {
-        let key_id_bytes = self.key_id.as_bytes();
-        let key_id_len = u16::try_from(key_id_bytes.len()).map_err(|_| {
-            Error::validation(
-                format!(
-                    "key_id is {} bytes; wire format caps it at {}",
-                    key_id_bytes.len(),
-                    u16::MAX,
-                ),
-                "WireEnvelope::build",
-            )
-        })?;
-        let total = MIN_HEADER_SIZE + key_id_bytes.len() + self.ciphertext.len();
-
-        let mut buf = Vec::with_capacity(total);
-        buf.extend_from_slice(MAGIC);
-        buf.push(WIRE_VERSION);
-        buf.push(self.algorithm.wire_tag());
-        buf.extend_from_slice(&key_id_len.to_be_bytes());
-        buf.extend_from_slice(key_id_bytes);
-        buf.extend_from_slice(self.nonce);
-        buf.extend_from_slice(self.ciphertext);
-
-        Ok(Bytes::from(buf))
-    }
-
-    pub fn parse(data: &[u8]) -> Result<WireEnvelope<'_>> {
-        if data.len() < MIN_HEADER_SIZE {
-            return Err(Error::validation(
-                "wire envelope too short",
-                "WireEnvelope::parse",
-            ));
-        }
-
-        if &data[..4] != MAGIC {
-            return Err(Error::validation(
-                "invalid magic bytes",
-                "WireEnvelope::parse",
-            ));
-        }
-
-        let version = data[4];
-        if version != WIRE_VERSION {
-            return Err(Error::validation(
-                format!("unsupported wire version: {version}"),
-                "WireEnvelope::parse",
-            ));
-        }
-
-        let algorithm = EncryptionAlgorithm::from_wire_tag(data[5])?;
-        let key_id_len = u16::from_be_bytes([data[6], data[7]]) as usize;
-
-        let key_id_end = 8 + key_id_len;
-        let nonce_end = key_id_end + NONCE_SIZE;
-
-        if data.len() < nonce_end {
-            return Err(Error::validation(
-                "wire envelope truncated: missing key_id or nonce",
-                "WireEnvelope::parse",
-            ));
-        }
-
-        let key_id = str::from_utf8(&data[8..key_id_end]).map_err(|e| {
-            Error::validation(format!("invalid key_id UTF-8: {e}"), "WireEnvelope::parse")
-        })?;
-
-        // Length check above (`data.len() < nonce_end`) guarantees
-        // this slice is exactly NONCE_SIZE bytes.
-        let nonce: &[u8; NONCE_SIZE] = data[key_id_end..nonce_end]
-            .try_into()
-            .expect("nonce slice length verified by guard above");
-        let ciphertext = &data[nonce_end..];
-
-        Ok(WireEnvelope {
-            algorithm,
-            key_id,
-            nonce,
-            ciphertext,
-        })
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn wire_format_round_trip() {
-        let nonce = [1u8; NONCE_SIZE];
-        let ciphertext = b"some-ciphertext-data";
-        let key_id = "my-key-id";
-
-        let wire = WireEnvelope {
-            algorithm: EncryptionAlgorithm::Aes256Gcm,
-            key_id,
-            nonce: &nonce,
-            ciphertext,
-        }
-        .build()
-        .expect("build");
-
-        let parsed = WireEnvelope::parse(&wire).expect("parse");
-        assert_eq!(parsed.algorithm, EncryptionAlgorithm::Aes256Gcm);
-        assert_eq!(parsed.key_id, key_id);
-        assert_eq!(parsed.nonce, &nonce);
-        assert_eq!(parsed.ciphertext, ciphertext);
-    }
-
-    #[test]
-    fn wire_format_truncated_header() {
-        assert!(WireEnvelope::parse(b"NVS").is_err());
-    }
-
-    #[test]
-    fn wire_format_bad_magic() {
-        let mut data = vec![0u8; MIN_HEADER_SIZE];
-        data[..4].copy_from_slice(b"BAAD");
-        assert!(WireEnvelope::parse(&data).is_err());
-    }
-
-    #[test]
-    fn wire_format_truncated_key_id() {
-        let mut data = vec![0u8; MIN_HEADER_SIZE];
-        data[..4].copy_from_slice(MAGIC);
-        data[4] = WIRE_VERSION;
-        data[5] = 0x01;
-        data[6..8].copy_from_slice(&100u16.to_be_bytes());
-        assert!(WireEnvelope::parse(&data).is_err());
-    }
-}
diff --git a/crates/nvisy-engine/src/core/ingestion/export.rs b/crates/nvisy-engine/src/core/ingestion/export.rs
deleted file mode 100644
index 7e2fae090..000000000
--- a/crates/nvisy-engine/src/core/ingestion/export.rs
+++ /dev/null
@@ -1,33 +0,0 @@
-//! Export sink configuration.
-//!
-//! [`ExportFile`] delivers the processed (and optionally redacted)
-//! content to one or more destination content objects, applying
-//! encryption and compression as requested before writing the bytes
-//! out.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-use validator::Validate;
-
-use super::{CompressionAlgorithm, EncryptionConfig};
-
-/// Configuration for the [`ExportFile`] config.
-///
-/// Identifies the destination content objects and specifies any encoding
-/// steps that must be applied before the bytes are written out.
-///
-/// [`ExportFile`]: crate::core::ingestion::ExportFile
-#[derive(Debug, Clone, Default, PartialEq, Eq, Validate)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct ExportFile {
-    /// Identifiers of content destinations to export to.
-    #[serde(default)]
-    pub content_ids: Vec<Uuid>,
-    /// Encrypt the content before publishing.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub encryption: Option<EncryptionConfig>,
-    /// Compress the content before publishing.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub compression: Option<CompressionAlgorithm>,
-}
diff --git a/crates/nvisy-engine/src/core/ingestion/exporter.rs b/crates/nvisy-engine/src/core/ingestion/exporter.rs
deleted file mode 100644
index 95b83940e..000000000
--- a/crates/nvisy-engine/src/core/ingestion/exporter.rs
+++ /dev/null
@@ -1,102 +0,0 @@
-//! File export operation.
-//!
-//! Delivers processed content to the registry, optionally applying
-//! encryption and compression in order:
-//!
-//! 1. **Encryption** — encrypt content (if config specified)
-//! 2. **Compression** — compress for storage or transfer (if format specified)
-
-use std::sync::Arc;
-
-use nvisy_codec::content::{Content, ContentData, ContentSource};
-use nvisy_core::Result;
-use uuid::Uuid;
-
-use crate::core::ingestion::compression::CompressionService;
-use crate::core::ingestion::encryption::CryptoService;
-use crate::core::ingestion::{CompressionAlgorithm, EncryptionConfig};
-use crate::core::{AnyTree, DocumentTree, SharedData};
-use crate::modality::DocumentModality;
-
-const TARGET: &str = "nvisy_engine::op::export_file";
-
-/// Exports processed content, optionally applying encryption and
-/// compression afterward.
-#[derive(Default)]
-pub(crate) struct Exporter {
-    encryption: Option<EncryptionConfig>,
-    compression: Option<CompressionAlgorithm>,
-    content_ids: Vec<Uuid>,
-}
-
-impl Exporter {
-    pub(crate) fn new() -> Self {
-        Self::default()
-    }
-
-    pub(crate) fn with_encryption(mut self, config: Option<EncryptionConfig>) -> Self {
-        self.encryption = config;
-        self
-    }
-
-    pub(crate) fn with_compression(mut self, format: Option<CompressionAlgorithm>) -> Self {
-        self.compression = format;
-        self
-    }
-
-    pub(crate) fn with_content_ids(mut self, ids: Vec<Uuid>) -> Self {
-        self.content_ids = ids;
-        self
-    }
-
-    /// Encode the tree's codec handle back to bytes, optionally
-    /// encrypt + compress, and write to the registry under every
-    /// configured content id.
-    pub(crate) async fn export(&self, tree: &AnyTree, shared: &Arc<SharedData>) -> Result<()> {
-        match tree {
-            AnyTree::Text(t) => self.export_one(t, shared).await,
-            AnyTree::Tabular(t) => self.export_one(t, shared).await,
-            AnyTree::Image(t) => self.export_one(t, shared).await,
-            AnyTree::Audio(t) => self.export_one(t, shared).await,
-        }
-    }
-
-    async fn export_one<M>(&self, tree: &DocumentTree<M>, shared: &Arc<SharedData>) -> Result<()>
-    where
-        M: DocumentModality,
-    {
-        let content_data = tree.handle.handler().encode()?;
-        let mut output_bytes = bytes::Bytes::copy_from_slice(content_data.as_bytes());
-
-        if let Some(ref enc_cfg) = self.encryption {
-            tracing::debug!(target: TARGET, key_id = %enc_cfg.key_id, "encrypting export content");
-            let crypto = CryptoService::new(&enc_cfg.key_id, shared.key_provider.clone());
-            let encrypted = crypto.encrypt(&content_data).await?;
-            tracing::debug!(
-                target: TARGET,
-                ciphertext_len = encrypted.ciphertext.len(),
-                "content encrypted",
-            );
-            output_bytes = encrypted.ciphertext;
-        }
-
-        if let Some(algorithm) = self.compression {
-            tracing::debug!(target: TARGET, ?algorithm, "compressing export content");
-            output_bytes = CompressionService::new(algorithm).compress(&output_bytes)?;
-        }
-
-        for &content_id in &self.content_ids {
-            let source = ContentSource::from_uuid_unchecked(content_id);
-            let data = ContentData::new(source, output_bytes.clone());
-            let content = Content::new(data);
-            shared
-                .registry
-                .register_content(shared.actor_id, content, None)
-                .await?;
-            tracing::debug!(target: TARGET, %content_id, "wrote exported content to registry");
-        }
-
-        tracing::debug!(target: TARGET, "export complete");
-        Ok(())
-    }
-}
diff --git a/crates/nvisy-engine/src/core/ingestion/import.rs b/crates/nvisy-engine/src/core/ingestion/import.rs
deleted file mode 100644
index d7b56408e..000000000
--- a/crates/nvisy-engine/src/core/ingestion/import.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-//! Import node configuration.
-//!
-//! [`ImportFile`] is always the first processing step: it resolves
-//! previously uploaded content by UUID, optionally decompresses and
-//! decrypts it, then feeds the raw bytes into the pipeline envelope
-//! for downstream extraction.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-use validator::Validate;
-
-use super::{CompressionAlgorithm, EncryptionConfig};
-
-/// Configuration for the [`ImportFile`] config.
-///
-/// Identifies the content objects to load and specifies any decoding steps
-/// that must be applied before the bytes are passed to extraction nodes.
-///
-/// [`ImportFile`]: crate::core::ingestion::ImportFile
-#[derive(Debug, Clone, Default, PartialEq, Eq, Validate)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct ImportFile {
-    /// Identifiers of previously uploaded content to import. Must contain at least one.
-    #[validate(length(min = 1, message = "import requires at least one content_id"))]
-    #[serde(default)]
-    pub content_ids: Vec<Uuid>,
-    /// Decompress the content before decoding.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub decompression: Option<CompressionAlgorithm>,
-    /// Decrypt the content before decoding.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub decryption: Option<EncryptionConfig>,
-}
diff --git a/crates/nvisy-engine/src/core/ingestion/importer.rs b/crates/nvisy-engine/src/core/ingestion/importer.rs
deleted file mode 100644
index 0c6da15b0..000000000
--- a/crates/nvisy-engine/src/core/ingestion/importer.rs
+++ /dev/null
@@ -1,251 +0,0 @@
-//! File import operation.
-//!
-//! Decodes raw content into one [`AnyTree`], optionally applying
-//! pre-processing in order:
-//!
-//! 1. **Decompression** — decompress raw bytes (if format specified)
-//! 2. **Decryption** — decrypt content (if encryption config specified)
-//! 3. **Decode** — resolve a codec from the [`CodecRegistry`] by the
-//!    content's extension or MIME hint, then decode into an
-//!    [`UntypedDocumentHandle`].
-//! 4. **Dispatch** — match the untyped handle once, build the matching
-//!    typed [`DocumentTree<M>`], and wrap it in the [`AnyTree`]
-//!    variant the orchestrator dispatches on.
-//! 5. **Seed** — convert any [`Inclusion`] annotations supplied for
-//!    this content into pre-detected entities on the root document's
-//!    audit, and store the full annotation list on the document for
-//!    downstream exclusion filtering.
-//!
-//! [`AnyTree`]: crate::core::AnyTree
-//! [`DocumentTree<M>`]: crate::core::DocumentTree
-//! [`UntypedDocumentHandle`]: nvisy_codec::UntypedDocumentHandle
-//! [`CodecRegistry`]: nvisy_codec::CodecRegistry
-//! [`Inclusion`]: nvisy_core::entity::AnnotationKind::Inclusion
-
-use std::mem;
-use std::sync::Arc;
-
-use nvisy_codec::content::{Content, ContentData, ContentDescriptor, ContentSource};
-use nvisy_codec::{CodecRegistry, UntypedDocumentHandle};
-use nvisy_core::entity::{Annotation, LabelAnnotation};
-use nvisy_core::modality::{Audio, Image, Tabular, Text};
-use nvisy_core::{Error, Result};
-
-use crate::core::ingestion::compression::CompressionService;
-use crate::core::ingestion::encryption::{CryptoService, EncryptedContent};
-use crate::core::ingestion::{CompressionAlgorithm, EncryptionAlgorithm, EncryptionConfig};
-use crate::core::{AnyTree, DocumentTree, SharedData};
-use crate::document::{AnyAnnotations, Document};
-use crate::modality::{
-    AudioExtraction, AudioMetadata, DocumentModality, ImageExtraction, ImageMetadata,
-    TabularExtraction, TabularMetadata, TextExtraction, TextMetadata,
-};
-
-const TARGET: &str = "nvisy_engine::op::import_file";
-
-/// Decodes raw content into one [`AnyTree`], optionally applying
-/// decompression and decryption beforehand.
-#[derive(Default)]
-pub struct Importer {
-    decompression: Option<CompressionAlgorithm>,
-    decryption: Option<EncryptionConfig>,
-}
-
-impl Importer {
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    pub fn with_decompression(mut self, format: Option<CompressionAlgorithm>) -> Self {
-        self.decompression = format;
-        self
-    }
-
-    pub fn with_decryption(mut self, config: Option<EncryptionConfig>) -> Self {
-        self.decryption = config;
-        self
-    }
-
-    /// Decode `content` into a single-element `Vec<AnyTree>`,
-    /// attaching the supplied per-modality `annotations` to the
-    /// resulting `Document<M>`. The vector shape is preserved so the
-    /// orchestrator's collection loop stays uniform across import
-    /// sources.
-    pub async fn import(
-        &self,
-        content: Content,
-        annotations: AnyAnnotations,
-        shared: &Arc<SharedData>,
-    ) -> Result<Vec<AnyTree>> {
-        let mut content = content;
-
-        if let Some(algorithm) = self.decompression {
-            tracing::debug!(target: TARGET, ?algorithm, "decompressing content");
-            let decompressed = CompressionService::new(algorithm).decompress(content.as_bytes())?;
-            let source = content.content_source();
-            content = replace_data(content, ContentData::new(source, decompressed));
-        }
-
-        if let Some(ref enc_cfg) = self.decryption {
-            tracing::debug!(target: TARGET, key_id = %enc_cfg.key_id, "decrypting content");
-            let crypto = CryptoService::new(&enc_cfg.key_id, shared.key_provider.clone());
-            let encrypted = EncryptedContent {
-                source: content.content_source(),
-                ciphertext: bytes::Bytes::copy_from_slice(content.as_bytes()),
-                key_id: enc_cfg.key_id.clone(),
-                algorithm: EncryptionAlgorithm::Aes256Gcm,
-            };
-            let decrypted_data = crypto.decrypt(encrypted).await?;
-            content = replace_data(content, decrypted_data);
-        }
-
-        let untyped = decode(&shared.codec_registry, &content).await?;
-        tracing::debug!(
-            target: TARGET,
-            format = %untyped.format_id(),
-            "decoded document",
-        );
-
-        let (data, descriptor) = content.into_parts();
-        let descriptor = descriptor.unwrap_or_default();
-        let source = data.content_source;
-
-        let tree = build_tree(untyped, source, descriptor, annotations);
-        tracing::debug!(
-            target: TARGET,
-            modality = tree.modality_name(),
-            "produced tree",
-        );
-        Ok(vec![tree])
-    }
-}
-
-/// Resolve a format from the registry by extension (preferred) or
-/// MIME content type, then decode the raw bytes through its loader.
-async fn decode(registry: &CodecRegistry, content: &Content) -> Result<UntypedDocumentHandle> {
-    let format = content
-        .file_extension()
-        .and_then(|ext| registry.by_extension(ext))
-        .or_else(|| {
-            content
-                .content_type()
-                .and_then(|ct| registry.by_content_type(ct))
-        })
-        .ok_or_else(|| {
-            Error::validation(
-                format!(
-                    "no codec registered for extension `{:?}` / content-type `{:?}`",
-                    content.file_extension(),
-                    content.content_type(),
-                ),
-                TARGET,
-            )
-        })?;
-    format.decode(content.data().clone()).await
-}
-
-/// Build the per-modality typed tree from an [`UntypedDocumentHandle`],
-/// stamping the document's extraction provenance and seeding any
-/// inclusion annotations onto the root document's audit.
-fn build_tree(
-    untyped: UntypedDocumentHandle,
-    source: ContentSource,
-    descriptor: ContentDescriptor,
-    mut annotations: AnyAnnotations,
-) -> AnyTree {
-    match untyped {
-        UntypedDocumentHandle::Text(handle) => {
-            let mut doc = Document::<Text>::new(TextMetadata::from(TextExtraction::Native), source);
-            attach_annotations(
-                &mut doc,
-                mem::take(&mut annotations.text),
-                annotations.labels.clone(),
-            );
-            AnyTree::Text(DocumentTree::new(doc, handle, descriptor))
-        }
-        UntypedDocumentHandle::Tabular(handle) => {
-            let mut doc = Document::<Tabular>::new(
-                TabularMetadata::from(TabularExtraction::SchemaInferred),
-                source,
-            );
-            attach_annotations(
-                &mut doc,
-                mem::take(&mut annotations.tabular),
-                annotations.labels.clone(),
-            );
-            AnyTree::Tabular(DocumentTree::new(doc, handle, descriptor))
-        }
-        UntypedDocumentHandle::Image(handle) => {
-            let mut doc =
-                Document::<Image>::new(ImageMetadata::from(ImageExtraction::Pending), source);
-            attach_annotations(
-                &mut doc,
-                mem::take(&mut annotations.image),
-                annotations.labels.clone(),
-            );
-            AnyTree::Image(DocumentTree::new(doc, handle, descriptor))
-        }
-        UntypedDocumentHandle::Audio(handle) => {
-            let mut doc =
-                Document::<Audio>::new(AudioMetadata::from(AudioExtraction::Pending), source);
-            attach_annotations(
-                &mut doc,
-                mem::take(&mut annotations.audio),
-                annotations.labels.clone(),
-            );
-            AnyTree::Audio(DocumentTree::new(doc, handle, descriptor))
-        }
-    }
-}
-
-/// Store user annotations on the document and synthesise entities
-/// for every [`Assert`]-strength inclusion. [`Hint`]-strength
-/// inclusions stay on the document for downstream prompt-builder
-/// consumption; exclusions stay for the post-detection filter.
-///
-/// [`Assert`]: nvisy_core::entity::AnnotationStrength::Assert
-/// [`Hint`]: nvisy_core::entity::AnnotationStrength::Hint
-fn attach_annotations<M: DocumentModality>(
-    doc: &mut Document<M>,
-    annotations: Vec<Annotation<M>>,
-    labels: Vec<LabelAnnotation>,
-) {
-    if !labels.is_empty() {
-        tracing::debug!(
-            target: TARGET,
-            count = labels.len(),
-            "attaching labels to document",
-        );
-    }
-    if !annotations.is_empty() {
-        tracing::debug!(
-            target: TARGET,
-            count = annotations.len(),
-            "attaching annotations to document",
-        );
-        let mut synthesized = 0;
-        for ann in &annotations {
-            if let Some(entity) = ann.to_inclusion_entity() {
-                doc.add_entities(std::iter::once(entity));
-                synthesized += 1;
-            }
-        }
-        if synthesized > 0 {
-            tracing::debug!(
-                target: TARGET,
-                count = synthesized,
-                "synthesized entities from Assert inclusions",
-            );
-        }
-    }
-    doc.annotations = annotations;
-    doc.labels = labels;
-}
-
-/// Replace the data payload of a [`Content`] while preserving its descriptor.
-fn replace_data(content: Content, data: ContentData) -> Content {
-    match content.into_parts().1 {
-        Some(descriptor) => Content::with_descriptor(data, descriptor),
-        None => Content::new(data),
-    }
-}
diff --git a/crates/nvisy-engine/src/core/ingestion/mod.rs b/crates/nvisy-engine/src/core/ingestion/mod.rs
deleted file mode 100644
index 9e8262554..000000000
--- a/crates/nvisy-engine/src/core/ingestion/mod.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-//! Ingestion: the pipeline's edge — content I/O.
-//!
-//! These configs form the boundary between the outside world and the
-//! pipeline. [`ImportFile`] pulls content in and [`ExportFile`]
-//! pushes processed content out; both share the same
-//! [`CompressionAlgorithm`] / [`EncryptionConfig`] codec options.
-
-pub(crate) mod compression;
-pub mod encryption;
-mod export;
-mod exporter;
-mod import;
-mod importer;
-
-use nvisy_core::Error;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-pub use self::export::ExportFile;
-pub(crate) use self::exporter::Exporter;
-pub use self::import::ImportFile;
-pub(crate) use self::importer::Importer;
-
-/// Supported compression algorithms for import/export.
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum CompressionAlgorithm {
-    /// Gzip (.gz).
-    Gzip,
-    /// Zstandard (.zst).
-    Zstd,
-}
-
-/// Supported encryption algorithms for import/export.
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum EncryptionAlgorithm {
-    /// AES-256 in Galois/Counter Mode.
-    Aes256Gcm,
-}
-
-impl EncryptionAlgorithm {
-    /// Encode as a single-byte wire tag.
-    pub fn wire_tag(self) -> u8 {
-        match self {
-            Self::Aes256Gcm => 0x01,
-        }
-    }
-
-    /// Decode from a single-byte wire tag.
-    pub fn from_wire_tag(tag: u8) -> Result<Self, Error> {
-        match tag {
-            0x01 => Ok(Self::Aes256Gcm),
-            _ => Err(Error::validation(
-                format!("unknown encryption algorithm tag: 0x{tag:02x} (valid: 0x01 = aes256gcm)"),
-                "encryption-wire",
-            )),
-        }
-    }
-}
-
-/// Encryption configuration pairing an algorithm with a key identifier.
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
-pub struct EncryptionConfig {
-    /// Algorithm to use for encryption/decryption.
-    pub algorithm: EncryptionAlgorithm,
-    /// Identifier of the key to resolve via the engine's `KeyProvider`.
-    pub key_id: String,
-}
diff --git a/crates/nvisy-engine/src/core/mod.rs b/crates/nvisy-engine/src/core/mod.rs
deleted file mode 100644
index 080ac4af6..000000000
--- a/crates/nvisy-engine/src/core/mod.rs
+++ /dev/null
@@ -1,52 +0,0 @@
-//! Engine core: the contract every phase programs against.
-//!
-//! `core/` houses the shapes flowing through every phase's `apply`
-//! method. Phases (in `extraction/`, `detection/`, `deduplication/`,
-//! `redaction/`, `validation/`) depend on `core/` only; they never
-//! reach into `pipeline/` or each other. The execution layer
-//! (`pipeline/`) consumes `core/` plus every phase module.
-//!
-//! # Contents
-//!
-//! - [`RuntimeConfig`] + [`EngineConfig`] + [`ResourceLimits`] —
-//!   engine-wide deployment configuration shared between the
-//!   detection and redaction engines.
-//! - [`DetectionContext`] / [`RedactionContext`] — per-pass
-//!   execution contexts, each carrying only the engine resources
-//!   its side actually consumes. Both implement [`PhaseContext`]
-//!   for the shared surface modality-agnostic phases bound on.
-//! - [`SharedData`] — `Arc`-wrapped run-wide state (registry,
-//!   codecs, policies).
-//! - [`PolicyStore`] — per-modality policy storage + matching.
-//! - [`DocumentTree<M>`] — the typed per-document carrier phases
-//!   mutate.
-//! - [`AnyTree`] — the modality-erased tree at the import boundary;
-//!   the orchestrator matches once and dispatches into typed
-//!   pipelines.
-//! - [`TextAt`] / [`DataAt`] — implemented on [`DocumentTree<M>`]
-//!   so phases that need to resolve a location pass `&tree`
-//!   directly.
-//! - [`probe_all`] — concurrent [`Healthcheck`] composition helper.
-//!
-//! [`DataAt`]: nvisy_core::extraction::DataAt
-//! [`Healthcheck`]: nvisy_core::health::Healthcheck
-
-mod config;
-mod context;
-mod health;
-pub mod ingestion;
-mod policy_store;
-mod shared;
-mod target;
-mod tree;
-
-pub use nvisy_core::extraction::TextAt;
-
-pub use self::config::{EngineConfig, ResourceLimits, RuntimeConfig};
-pub use self::context::{DetectionContext, PhaseContext, RedactionContext};
-pub(crate) use self::context::{DetectionEngines, RedactionEngines};
-pub use self::health::probe_all;
-pub(crate) use self::policy_store::Decision;
-pub use self::policy_store::PolicyStore;
-pub use self::shared::SharedData;
-pub use self::tree::{AnyTree, DocumentTree};
diff --git a/crates/nvisy-engine/src/core/policy_store.rs b/crates/nvisy-engine/src/core/policy_store.rs
deleted file mode 100644
index 55357b103..000000000
--- a/crates/nvisy-engine/src/core/policy_store.rs
+++ /dev/null
@@ -1,278 +0,0 @@
-//! [`PolicyStore`]: per-request flat container of [`Arc<Policy>`],
-//! plus the per-entity decision resolver that walks it.
-//!
-//! Built from a `Vec<Policy>` submission via
-//! [`PolicyStore::from_policies`], which consumes the submitted
-//! policies and wraps each in an `Arc<Policy>` — no deep clones.
-//! Detection and redaction pipelines share a single store via
-//! `Arc<PolicyStore>`; per-call handoff is a refcount bump.
-//!
-//! Policies are flat (no per-modality bucketing); each rule's
-//! `Action::Redact(operators)` is itself modality-aware via
-//! [`ModalityRedactions::operator_for`]. `resolve::<M>` walks every
-//! policy and uses the per-modality projection to pick the operator
-//! for the entity's modality, falling back to the deployment-wide
-//! defaults when the rule didn't cover that modality.
-//!
-//! [`SharedData`]: super::SharedData
-//! [`ModalityRedactions::operator_for`]: crate::policy::redaction::ModalityRedactions::operator_for
-
-use std::fmt;
-use std::sync::Arc;
-
-use hipstr::HipStr;
-use nvisy_codec::content::ContentDescriptor;
-use nvisy_core::entity::{Entity, EntityLabelCatalog};
-
-use crate::modality::DocumentModality;
-use crate::policy::redaction::{ModalityRedactions, ProjectRedaction};
-use crate::policy::{Action, Condition, Policy, PolicyRule};
-
-/// Flat container of [`Arc<Policy>`] in precedence order — index
-/// `0` is highest precedence. Shared between detection and redaction
-/// pipelines via `Arc<PolicyStore>` so per-call handoff is a
-/// refcount bump.
-#[derive(Default)]
-pub struct PolicyStore {
-    policies: Vec<Arc<Policy>>,
-}
-
-impl PolicyStore {
-    /// Construct a store from a `Vec<Policy>` submission, taking
-    /// ownership so each policy moves straight into its [`Arc`] —
-    /// no deep clones.
-    pub(crate) fn from_policies(policies: Vec<Policy>) -> Self {
-        Self {
-            policies: policies.into_iter().map(Arc::new).collect(),
-        }
-    }
-
-    /// Union every stored policy's [`Policy::labels`] into a single
-    /// [`EntityLabelCatalog`]. Used at redaction time to rebuild
-    /// the per-request catalog the detection pass already validated.
-    /// Conflicts here are impossible because the same union was
-    /// validated at detection-time submission.
-    ///
-    /// [`EntityLabelCatalog`]: nvisy_core::entity::EntityLabelCatalog
-    pub(crate) fn catalog(&self) -> EntityLabelCatalog {
-        let mut catalog = EntityLabelCatalog::new();
-        for p in &self.policies {
-            for l in &p.labels {
-                catalog.insert(l.clone());
-            }
-        }
-        catalog
-    }
-
-    /// Borrow the full policy chain. Crate-internal because the
-    /// public surface is `resolve::<M>`.
-    #[cfg(test)]
-    fn chain(&self) -> &[Arc<Policy>] {
-        &self.policies
-    }
-
-    /// Resolve a single entity against the policy chain. Walks
-    /// policies in precedence order; within a policy, walks rules
-    /// in declaration order. First matching rule wins; when no
-    /// rule in a policy matches, falls back to that policy's
-    /// [`Policy::default_action`] (if any) before descending to
-    /// the next policy.
-    ///
-    /// `default_operators` is the deployment-wide fallback: when a
-    /// rule's `Redact` action doesn't cover the entity's modality
-    /// (per [`ModalityRedactions::operator_for`]), the resolver
-    /// tries the default before skipping the rule.
-    ///
-    /// Returns [`Decision::Fallthrough`] when no policy in the
-    /// chain produced a decision; the caller's default-threshold
-    /// path takes over.
-    pub(crate) fn resolve<M: DocumentModality + ProjectRedaction>(
-        &self,
-        entity: &Entity<M>,
-        catalog: &EntityLabelCatalog,
-        default_operators: &ModalityRedactions,
-        document_labels: &[&str],
-        descriptor: &ContentDescriptor,
-    ) -> Decision<M> {
-        for policy in &self.policies {
-            for rule in &policy.rules {
-                if !rule_matches(rule, entity, catalog, document_labels, descriptor) {
-                    continue;
-                }
-                if let Some(decision) = decide::<M>(
-                    &rule.action,
-                    &policy.name,
-                    Some(rule.name.clone()),
-                    default_operators,
-                ) {
-                    return decision;
-                }
-            }
-            if let Some(default) = policy.default_action.as_ref()
-                && let Some(decision) = decide::<M>(default, &policy.name, None, default_operators)
-            {
-                return decision;
-            }
-        }
-        Decision::Fallthrough
-    }
-}
-
-impl fmt::Debug for PolicyStore {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("PolicyStore")
-            .field("count", &self.policies.len())
-            .finish()
-    }
-}
-
-/// Outcome of walking the policy chain for one entity.
-///
-/// The winning rule's typed operator spec rides along on
-/// [`Decision::Redact`] so the apply phase can instantiate the
-/// operator (built-in arms) or look it up in the toolkit
-/// [`RedactionRegistry<M>`] (`Custom` arm) without re-walking
-/// the policy chain.
-///
-/// [`RedactionRegistry<M>`]: nvisy_toolkit::redaction::RedactionRegistry
-pub(crate) enum Decision<M: DocumentModality> {
-    /// A rule chose to redact. `operator` is the per-modality
-    /// operator spec the winning rule (or fallback) projected for
-    /// this entity's modality. `rule_name` is `None` when the
-    /// policy's `default_action` fallback fired.
-    Redact {
-        policy_name: HipStr<'static>,
-        rule_name: Option<HipStr<'static>>,
-        operator: M::Redaction,
-    },
-    /// A `Suppress` rule fired. `reason` is the author-supplied
-    /// suppression reason, if any.
-    Suppress {
-        policy_name: HipStr<'static>,
-        rule_name: Option<HipStr<'static>>,
-        reason: Option<HipStr<'static>>,
-    },
-    /// An `Audit` rule fired. The entity is left untouched; the
-    /// audit entry carries `severity` for downstream review
-    /// tooling.
-    Audit {
-        policy_name: HipStr<'static>,
-        rule_name: Option<HipStr<'static>>,
-        severity: Option<HipStr<'static>>,
-    },
-    /// No policy in the chain produced a decision. The caller
-    /// falls back to its default-threshold path.
-    Fallthrough,
-}
-
-fn decide<M: DocumentModality + ProjectRedaction>(
-    action: &Action,
-    policy_name: &HipStr<'static>,
-    rule_name: Option<HipStr<'static>>,
-    default_operators: &ModalityRedactions,
-) -> Option<Decision<M>> {
-    match action {
-        Action::Redact(operators) => {
-            let operator = operators
-                .operator_for::<M>()
-                .or_else(|| default_operators.operator_for::<M>())?;
-            Some(Decision::Redact {
-                policy_name: policy_name.clone(),
-                rule_name,
-                operator: operator.clone(),
-            })
-        }
-        Action::Suppress(opts) => Some(Decision::Suppress {
-            policy_name: policy_name.clone(),
-            rule_name,
-            reason: opts.reason.clone(),
-        }),
-        Action::Audit(opts) => Some(Decision::Audit {
-            policy_name: policy_name.clone(),
-            rule_name,
-            severity: opts.severity.clone(),
-        }),
-    }
-}
-
-fn rule_matches<M: DocumentModality>(
-    rule: &PolicyRule,
-    entity: &Entity<M>,
-    catalog: &EntityLabelCatalog,
-    document_labels: &[&str],
-    descriptor: &ContentDescriptor,
-) -> bool {
-    rule.enabled
-        && rule.selector.matches(entity, catalog)
-        && rule
-            .conditions
-            .iter()
-            .all(|c| condition_matches(c, document_labels, descriptor))
-}
-
-fn condition_matches(
-    condition: &Condition,
-    document_labels: &[&str],
-    descriptor: &ContentDescriptor,
-) -> bool {
-    match condition {
-        Condition::Labels { labels } => labels.iter().all(|label| {
-            document_labels
-                .iter()
-                .any(|doc| doc.eq_ignore_ascii_case(label))
-        }),
-        Condition::Metadata { key, value } => match descriptor.get_policy_metadata(key) {
-            Some(actual) => match value {
-                Some(expected) => actual.as_str().is_some_and(|s| s == expected),
-                None => true,
-            },
-            None => false,
-        },
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::Entity;
-    use nvisy_core::modality::Text;
-    use semver::Version;
-
-    use super::*;
-
-    fn empty_policy(name: &str) -> Policy {
-        Policy {
-            name: HipStr::from(name),
-            version: Version::new(1, 0, 0),
-            description: None,
-            labels: Vec::new(),
-            rules: Vec::new(),
-            default_action: None,
-            retention: Vec::new(),
-        }
-    }
-
-    #[test]
-    fn empty_chain_returns_fallthrough() {
-        let store = PolicyStore::default();
-        let entity = Entity::<Text>::test_builder(0, 4).test_build();
-        let descriptor = ContentDescriptor::new();
-        let catalog = EntityLabelCatalog::new();
-        let defaults = ModalityRedactions::default();
-        assert!(matches!(
-            store.resolve::<Text>(&entity, &catalog, &defaults, &[], &descriptor),
-            Decision::Fallthrough
-        ));
-    }
-
-    #[test]
-    fn from_policies_preserves_order() {
-        let store = PolicyStore::from_policies(vec![
-            empty_policy("a"),
-            empty_policy("b"),
-            empty_policy("c"),
-        ]);
-        assert_eq!(store.chain().len(), 3);
-        assert_eq!(store.chain()[0].name, "a");
-        assert_eq!(store.chain()[2].name, "c");
-    }
-}
diff --git a/crates/nvisy-engine/src/core/shared.rs b/crates/nvisy-engine/src/core/shared.rs
deleted file mode 100644
index 0d56bc8cd..000000000
--- a/crates/nvisy-engine/src/core/shared.rs
+++ /dev/null
@@ -1,63 +0,0 @@
-//! Run-wide shared state for pipeline operations.
-//!
-//! [`SharedData`] holds immutable run-wide state behind an `Arc` so
-//! that every envelope and operation can cheaply access the same actor
-//! identity, policies, and context cache.
-//!
-//! Policies are typed per modality; the [`PolicyStore`] in this struct
-//! holds the per-modality stacks heterogeneously, so a
-//! `DocumentEnvelope<M>` pulls its stack with
-//! `shared.policies.resolve::<M>(...)`.
-
-use std::fmt;
-use std::sync::Arc;
-
-use nvisy_codec::CodecRegistry;
-use nvisy_core::entity::EntityLabelCatalog;
-use uuid::Uuid;
-
-use super::PolicyStore;
-use crate::core::ingestion::encryption::SharedKeyProvider;
-use crate::registry::Registry;
-
-/// Immutable run-wide state shared across all envelopes via `Arc`.
-pub struct SharedData {
-    /// Unique identifier for this pipeline run.
-    pub run_id: Uuid,
-    /// Identity of the human or service account that initiated the run.
-    pub actor_id: Uuid,
-    /// Per-modality policy stacks built at submission time. Held as
-    /// `Arc<PolicyStore>` so detection and redaction passes share
-    /// the same store without rebuilding it.
-    pub policies: Arc<PolicyStore>,
-    /// Per-request entity-label catalog, unioned from every
-    /// submitted policy's [`Policy::labels`]. Drives recognizer
-    /// dispatch (NER label list, pattern filtering) and selector
-    /// tag matching.
-    ///
-    /// [`Policy::labels`]: crate::policy::Policy::labels
-    pub catalog: Arc<EntityLabelCatalog>,
-    /// Content and context storage.
-    pub registry: Registry,
-    /// Codec registry resolving file extensions / content types to
-    /// the appropriate per-format loader. Importers call into this
-    /// to decode raw bytes into a typed [`DocumentHandle<M>`].
-    ///
-    /// Built once at engine construction with
-    /// [`CodecRegistry::with_builtin`] so every importer in the run
-    /// shares the same set of registered formats.
-    ///
-    /// [`DocumentHandle<M>`]: nvisy_codec::DocumentHandle
-    pub codec_registry: CodecRegistry,
-    /// Key provider for encryption/decryption.
-    pub key_provider: SharedKeyProvider,
-}
-
-impl fmt::Debug for SharedData {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("SharedData")
-            .field("run_id", &self.run_id)
-            .field("actor_id", &self.actor_id)
-            .finish_non_exhaustive()
-    }
-}
diff --git a/crates/nvisy-engine/src/core/target.rs b/crates/nvisy-engine/src/core/target.rs
deleted file mode 100644
index 134e9ec49..000000000
--- a/crates/nvisy-engine/src/core/target.rs
+++ /dev/null
@@ -1,151 +0,0 @@
-//! [`TextAt<M>`], [`DataAt<M>`], and [`RedactAt<M>`] impls on
-//! [`DocumentTree<M>`].
-//!
-//! Phases that need to resolve a modality-typed location to its source
-//! text or wire payload take `&DocumentTree<M>` directly — there is no
-//! separate view type. Text and tabular reads go through the codec
-//! handle; image and audio reads walk the document's blocks. Write-back
-//! ([`RedactAt::redact_at`]) always goes through the codec handle.
-
-use nvisy_core::Result;
-use nvisy_core::extraction::{DataAt, TextAt};
-use nvisy_core::modality::{
-    Audio, AudioData, AudioLocation, Image, ImageData, ImageLocation, Tabular, TabularLocation,
-    Text, TextData, TextLocation,
-};
-use nvisy_core::redaction::{RedactAt, Redactions};
-
-use super::DocumentTree;
-use crate::modality::{AudioBlock, ImageBlock};
-
-#[async_trait::async_trait]
-impl TextAt<Text> for DocumentTree<Text> {
-    /// Resolve a [`Text`] location to its source text via the codec
-    /// handle. Returns `None` when the handle has no readable bytes
-    /// at the location.
-    async fn text_at(&self, location: &TextLocation) -> Option<String> {
-        self.handle
-            .handler()
-            .read(location)
-            .await
-            .ok()
-            .flatten()
-            .map(|d| d.into_string())
-    }
-}
-
-#[async_trait::async_trait]
-impl TextAt<Tabular> for DocumentTree<Tabular> {
-    /// Resolve a [`Tabular`] location to its source cell value via the
-    /// codec handle.
-    async fn text_at(&self, location: &TabularLocation) -> Option<String> {
-        self.handle
-            .handler()
-            .read(location)
-            .await
-            .ok()
-            .flatten()
-            .map(|d| d.into_string())
-    }
-}
-
-#[async_trait::async_trait]
-impl TextAt<Image> for DocumentTree<Image> {
-    /// Resolve an [`Image`] location to the OCR'd text at that region
-    /// by walking the document's blocks. Exact bounding-box match
-    /// against a block's `region` returns the whole block text;
-    /// sub-region matches consult the block's `spans`.
-    async fn text_at(&self, location: &ImageLocation) -> Option<String> {
-        for block in &self.root.blocks {
-            let (text, region) = match &block.kind {
-                ImageBlock::Text { text, region }
-                | ImageBlock::Heading { text, region }
-                | ImageBlock::Table { text, region } => (text, region),
-                _ => continue,
-            };
-            if region == location {
-                return Some(text.clone());
-            }
-            if let Some(s) = block.spans.iter().find(|s| s.source == *location) {
-                return Some(text[s.text_start..s.text_end].to_owned());
-            }
-        }
-        None
-    }
-}
-
-#[async_trait::async_trait]
-impl TextAt<Audio> for DocumentTree<Audio> {
-    /// Resolve an [`Audio`] location to the transcript at that time
-    /// span by walking the document's blocks.
-    async fn text_at(&self, location: &AudioLocation) -> Option<String> {
-        for block in &self.root.blocks {
-            let AudioBlock::Speech {
-                text, time_span, ..
-            } = &block.kind
-            else {
-                continue;
-            };
-            if time_span == &location.time_span {
-                return Some(text.clone());
-            }
-            if let Some(s) = block.spans.iter().find(|s| s.source == *location) {
-                return Some(text[s.text_start..s.text_end].to_owned());
-            }
-        }
-        None
-    }
-}
-
-#[async_trait::async_trait]
-impl DataAt<Text> for DocumentTree<Text> {
-    /// Resolve a [`Text`] location to its [`TextData`] payload via
-    /// the codec handle.
-    async fn data_at(&self, location: &TextLocation) -> Option<TextData> {
-        self.handle.handler().read(location).await.ok().flatten()
-    }
-}
-
-#[async_trait::async_trait]
-impl DataAt<Image> for DocumentTree<Image> {
-    /// Resolve an [`Image`] location to its [`ImageData`] payload via
-    /// the codec handle.
-    async fn data_at(&self, location: &ImageLocation) -> Option<ImageData> {
-        self.handle.handler().read(location).await.ok().flatten()
-    }
-}
-
-#[async_trait::async_trait]
-impl DataAt<Audio> for DocumentTree<Audio> {
-    /// Resolve an [`Audio`] location to its [`AudioData`] payload via
-    /// the codec handle.
-    async fn data_at(&self, location: &AudioLocation) -> Option<AudioData> {
-        self.handle.handler().read(location).await.ok().flatten()
-    }
-}
-
-#[async_trait::async_trait]
-impl RedactAt<Text> for DocumentTree<Text> {
-    /// Apply a batch of text replacements through the codec handle.
-    /// The handler decides per-format ordering (right-to-left for byte
-    /// streams, batched per page for PDF, …).
-    async fn redact_at(&mut self, redactions: Redactions<Text>) -> Result<()> {
-        self.handle.handler_mut().redact(redactions).await
-    }
-}
-
-#[async_trait::async_trait]
-impl RedactAt<Image> for DocumentTree<Image> {
-    /// Apply a batch of image replacements through the codec handle.
-    async fn redact_at(&mut self, redactions: Redactions<Image>) -> Result<()> {
-        self.handle.handler_mut().redact(redactions).await
-    }
-}
-
-#[async_trait::async_trait]
-impl RedactAt<Audio> for DocumentTree<Audio> {
-    /// Apply a batch of audio replacements through the codec handle.
-    async fn redact_at(&mut self, redactions: Redactions<Audio>) -> Result<()> {
-        self.handle.handler_mut().redact(redactions).await
-    }
-}
diff --git a/crates/nvisy-engine/src/core/tree.rs b/crates/nvisy-engine/src/core/tree.rs
deleted file mode 100644
index 8f336249e..000000000
--- a/crates/nvisy-engine/src/core/tree.rs
+++ /dev/null
@@ -1,132 +0,0 @@
-//! [`DocumentTree<M>`]: per-source, per-modality carrier the pipeline
-//! operates on. Owns its codec handle directly — no lock, no
-//! modality erasure.
-//!
-//! Modality erasure exists only at the import boundary in
-//! [`AnyTree`]. The orchestrator matches `AnyTree` once and dispatches
-//! into the typed [`DocumentTree<M>`]-shaped pipeline; from that
-//! point on every phase body is `<M: DocumentModality>`-generic.
-//!
-//! Rich sources (PDF/DOCX) land as [`AnyTree::Text`] whose `embeds`
-//! field carries the extracted image children as [`AnyTree::Image`]
-//! entries. The orchestrator walks `embeds` recursively after the
-//! root pipeline finishes.
-
-use nvisy_codec::DocumentHandle;
-use nvisy_codec::content::ContentDescriptor;
-use nvisy_core::modality::{Audio, Image, ModalityKind, Tabular, Text};
-
-use crate::document::Document;
-use crate::document::provenance::AnyAudit;
-use crate::modality::DocumentModality;
-
-/// Typed per-modality tree. One per imported file (or per embedded
-/// child of a rich import).
-pub struct DocumentTree<M: DocumentModality> {
-    /// The modality-typed document with its blocks + audit.
-    pub root: Document<M>,
-    /// Codec handle backing the underlying bytes. Phases borrow this
-    /// directly via the typed [`DocumentHandle<M>`] for reads
-    /// (via [`Handler::read`]) and redactions
-    /// (via [`Handler::redact`]).
-    ///
-    /// [`Handler::read`]: nvisy_codec::Handler::read
-    /// [`Handler::redact`]: nvisy_codec::Handler::redact
-    pub handle: DocumentHandle<M>,
-    /// Caller-supplied descriptor (filename, MIME hint, policy
-    /// metadata) from the original upload. Policy evaluation
-    /// matches rules against fields here (e.g.
-    /// `Condition::Metadata { key, value }` reads
-    /// `descriptor.get_policy_metadata(key)`).
-    pub descriptor: ContentDescriptor,
-    /// Nested embedded children. Populated for rich text sources
-    /// (PDF, DOCX) by the importer, who pulls embedded image
-    /// children out of the rich handler at import time. Empty for
-    /// non-rich modalities.
-    pub embeds: Vec<AnyTree>,
-}
-
-impl<M: DocumentModality> DocumentTree<M> {
-    /// Construct a new typed tree from its parts. `embeds` starts empty;
-    /// the importer is responsible for populating it for rich sources.
-    pub fn new(
-        root: Document<M>,
-        handle: DocumentHandle<M>,
-        descriptor: ContentDescriptor,
-    ) -> Self {
-        Self {
-            root,
-            handle,
-            descriptor,
-            embeds: Vec::new(),
-        }
-    }
-}
-
-/// Modality-erased tree, used only at the import boundary and during
-/// the orchestrator's modality dispatch. After dispatch, phases see
-/// [`DocumentTree<M>`] directly.
-#[non_exhaustive]
-pub enum AnyTree {
-    /// Text-modality tree.
-    Text(DocumentTree<Text>),
-    /// Tabular-modality tree.
-    Tabular(DocumentTree<Tabular>),
-    /// Image-modality tree.
-    Image(DocumentTree<Image>),
-    /// Audio-modality tree.
-    Audio(DocumentTree<Audio>),
-}
-
-impl AnyTree {
-    /// Runtime modality tag.
-    pub fn modality(&self) -> ModalityKind {
-        match self {
-            Self::Text(_) => ModalityKind::Text,
-            Self::Tabular(_) => ModalityKind::Tabular,
-            Self::Image(_) => ModalityKind::Image,
-            Self::Audio(_) => ModalityKind::Audio,
-        }
-    }
-
-    /// Stable lowercase name of the tree's modality, for telemetry.
-    pub fn modality_name(&self) -> &'static str {
-        match self {
-            Self::Text(_) => "text",
-            Self::Tabular(_) => "tabular",
-            Self::Image(_) => "image",
-            Self::Audio(_) => "audio",
-        }
-    }
-
-    /// Clone the root document's audit as a modality-erased
-    /// [`AnyAudit`] for the run summary collector.
-    pub fn audit_cloned(&self) -> AnyAudit {
-        match self {
-            Self::Text(t) => t.root.audit.clone().into(),
-            Self::Tabular(t) => t.root.audit.clone().into(),
-            Self::Image(t) => t.root.audit.clone().into(),
-            Self::Audio(t) => t.root.audit.clone().into(),
-        }
-    }
-
-    /// Iterate the root tree's embedded children in document order.
-    pub fn embeds(&self) -> &[AnyTree] {
-        match self {
-            Self::Text(t) => &t.embeds,
-            Self::Tabular(t) => &t.embeds,
-            Self::Image(t) => &t.embeds,
-            Self::Audio(t) => &t.embeds,
-        }
-    }
-
-    /// Mutable iterator over embedded children.
-    pub fn embeds_mut(&mut self) -> &mut Vec<AnyTree> {
-        match self {
-            Self::Text(t) => &mut t.embeds,
-            Self::Tabular(t) => &mut t.embeds,
-            Self::Image(t) => &mut t.embeds,
-            Self::Audio(t) => &mut t.embeds,
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/detection/config/mod.rs b/crates/nvisy-engine/src/detection/config/mod.rs
deleted file mode 100644
index ec937436f..000000000
--- a/crates/nvisy-engine/src/detection/config/mod.rs
+++ /dev/null
@@ -1,125 +0,0 @@
-//! Detection config: deployment-time `[detection.*]` configuration
-//! describing which recognizers are available. The per-request
-//! [`RecognizerRegistry`] is built from this template plus the
-//! request's policy-supplied [`EntityLabelCatalog`].
-//!
-//! Pattern detection is always-on; the registry built per request
-//! drops patterns whose label isn't in the request catalog. NER is
-//! opt-in via `[detection.ner]`. LLM and VLM sections are not
-//! currently wired — those modules are parked pending rework to
-//! implement [`EntityRecognizer<M>`] directly.
-//!
-//! Each recognizer owns its own post-recognition processing
-//! (boosting, deduplication-within-recognizer, validation post-pass).
-//! The engine orchestrates recognizers; it does not orchestrate
-//! recognizer-internal phases.
-//!
-//! [`RecognizerRegistry`]: nvisy_toolkit::detection::RecognizerRegistry
-//! [`EntityRecognizer<M>`]: nvisy_core::recognition::EntityRecognizer
-
-mod ner;
-mod pattern;
-
-#[cfg(not(feature = "bento"))]
-use nvisy_core::Error;
-use nvisy_core::Result;
-use nvisy_core::entity::EntityLabelCatalog;
-use nvisy_core::modality::Text;
-use nvisy_ner::NerRecognizer;
-use nvisy_ner::backend::NoopBackend;
-#[cfg(feature = "bento")]
-use nvisy_ner::backend::{BentoBackend, BentoParams};
-use nvisy_pattern::PatternRecognizer;
-use nvisy_toolkit::detection::RecognizerRegistry;
-
-pub use self::ner::{NerBackend, NerDetection};
-pub use self::pattern::PatternDetection;
-
-/// Stable name the NER recognizer registers under (surfaced in trail
-/// provenance on emitted entities).
-const NER_RECOGNIZER_NAME: &str = "ner";
-
-/// Configuration for the [`RecognizerRegistry`].
-///
-/// Each field maps to a `[detection.*]` section in `Nvisy.toml`.
-/// Every field is `Option<_>` so missing sections are valid — `None`
-/// means the recognizer is not loaded (or uses its always-on default
-/// for `pattern`).
-///
-/// [`RecognizerRegistry`]: nvisy_toolkit::detection::RecognizerRegistry
-#[derive(Debug, Clone, Default, PartialEq)]
-#[derive(serde::Serialize, serde::Deserialize)]
-pub struct DetectionConfig {
-    /// `[detection.pattern]` — pattern recognizer config. `None` uses
-    /// the shipped registry with default settings (the recognizer is
-    /// always-on).
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub pattern: Option<PatternDetection>,
-    /// `[detection.ner]` — NER recognizer config bundle.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub ner: Option<NerDetection>,
-}
-
-impl DetectionConfig {
-    /// Build a per-request [`RecognizerRegistry`] from each opted-in
-    /// section, wiring the request `catalog` into recognizers that
-    /// need it (NER's zero-shot label list, pattern filtering).
-    ///
-    /// Pattern detection is always-on; pattern emissions whose label
-    /// isn't registered in `catalog` are filtered out at registry
-    /// construction so unmatched patterns never run. NER is opt-in
-    /// via `cfg.ner`; when enabled, the request catalog's labels are
-    /// passed as the zero-shot label list.
-    ///
-    /// # Errors
-    ///
-    /// Returns the first construction error encountered — pattern
-    /// compile failure, NER backend init failure, or a
-    /// config-selected backend whose feature wasn't compiled in.
-    pub fn build_for_request(&self, catalog: &EntityLabelCatalog) -> Result<RecognizerRegistry> {
-        let mut reg = RecognizerRegistry::new();
-
-        let pattern_cfg = self.pattern.clone().unwrap_or_default();
-        if pattern_cfg.enabled {
-            let builder = PatternRecognizer::builder()
-                .with_builtin_patterns()
-                .with_builtin_dictionaries()
-                .filter_by_catalog(catalog);
-            if !builder.is_empty() {
-                reg = reg.with_recognizer::<Text>(builder.build_context_enhanced()?);
-            }
-        }
-
-        if let Some(ner_cfg) = self.ner.as_ref().filter(|c| c.enabled) {
-            let supported_labels = catalog.iter().map(|l| l.label_ref()).collect::<Vec<_>>();
-            let recognizer = match &ner_cfg.backend {
-                NerBackend::Noop => NerRecognizer::builder()
-                    .with_name(NER_RECOGNIZER_NAME)
-                    .with_engine(NoopBackend)
-                    .with_supported_labels(supported_labels)
-                    .build()?,
-
-                #[cfg(feature = "bento")]
-                NerBackend::Bento { base_url } => {
-                    let backend = BentoBackend::new(BentoParams::new(base_url.clone()))?;
-                    NerRecognizer::builder()
-                        .with_name(NER_RECOGNIZER_NAME)
-                        .with_engine(backend)
-                        .with_supported_labels(supported_labels)
-                        .build()?
-                }
-
-                #[cfg(not(feature = "bento"))]
-                NerBackend::Bento { .. } => {
-                    return Err(Error::validation(
-                        "NerBackend::Bento requires the `bento` feature",
-                        "ner",
-                    ));
-                }
-            };
-            reg = reg.with_recognizer::<Text>(recognizer);
-        }
-
-        Ok(reg)
-    }
-}
diff --git a/crates/nvisy-engine/src/detection/config/ner.rs b/crates/nvisy-engine/src/detection/config/ner.rs
deleted file mode 100644
index 1cb356ea3..000000000
--- a/crates/nvisy-engine/src/detection/config/ner.rs
+++ /dev/null
@@ -1,58 +0,0 @@
-//! [`NerDetection`] + [`NerBackend`]: NER-recognizer settings.
-
-use serde::{Deserialize, Serialize};
-
-/// Backend selection for the NER recognizer.
-///
-/// Parseable regardless of compiled features — selecting
-/// [`NerBackend::Bento`] on a build without the `bento` feature
-/// surfaces as a clear runtime error at construction rather than a
-/// deserialisation failure, so config files stay portable across
-/// deployments.
-#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
-#[serde(tag = "kind", rename_all = "kebab-case")]
-pub enum NerBackend {
-    /// Returns no entities. The default — used by tests and by
-    /// deployments that detect via patterns / LLM only.
-    #[default]
-    Noop,
-
-    /// Externalised [`BentoBackend`] — calls the `inference-gliner`
-    /// Bento over HTTP. Requires the runtime to be built with the
-    /// `bento` feature.
-    ///
-    /// [`BentoBackend`]: nvisy_ner::backend::BentoBackend
-    Bento {
-        /// Base URL of the `inference-gliner` Bento.
-        base_url: String,
-    },
-}
-
-/// NER-specific detection settings.
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
-pub struct NerDetection {
-    /// Enable this recognizer. When `false`, the recognizer is
-    /// neither built nor dispatched, but the config is preserved so
-    /// operators can toggle without losing it. Defaults to `true`.
-    #[serde(default = "default_true")]
-    pub enabled: bool,
-
-    /// Backend selection for the NER pipeline. Defaults to
-    /// [`NerBackend::Noop`] so a baseline runtime works without any
-    /// inference service configured.
-    #[serde(default)]
-    pub backend: NerBackend,
-}
-
-impl Default for NerDetection {
-    fn default() -> Self {
-        Self {
-            enabled: true,
-            backend: NerBackend::default(),
-        }
-    }
-}
-
-fn default_true() -> bool {
-    true
-}
diff --git a/crates/nvisy-engine/src/detection/config/pattern.rs b/crates/nvisy-engine/src/detection/config/pattern.rs
deleted file mode 100644
index 4e561adf6..000000000
--- a/crates/nvisy-engine/src/detection/config/pattern.rs
+++ /dev/null
@@ -1,24 +0,0 @@
-//! [`PatternDetection`]: pattern-recognizer settings.
-
-use serde::{Deserialize, Serialize};
-
-/// Pattern-recognizer settings.
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
-pub struct PatternDetection {
-    /// Enable this recognizer. When `false`, the recognizer is
-    /// neither built nor dispatched, but the config is preserved so
-    /// operators can toggle without losing it. Defaults to `true` —
-    /// pattern detection is always-on out of the box.
-    #[serde(default = "default_true")]
-    pub enabled: bool,
-}
-
-impl Default for PatternDetection {
-    fn default() -> Self {
-        Self { enabled: true }
-    }
-}
-
-fn default_true() -> bool {
-    true
-}
diff --git a/crates/nvisy-engine/src/detection/document.rs b/crates/nvisy-engine/src/detection/document.rs
deleted file mode 100644
index 6698dd266..000000000
--- a/crates/nvisy-engine/src/detection/document.rs
+++ /dev/null
@@ -1,100 +0,0 @@
-//! Per-document phase sequence for detection: extraction →
-//! detection → deduplication. No redaction, no validation, no
-//! export.
-
-use nvisy_core::modality::{Audio, Image, Tabular, Text};
-use nvisy_core::{Error, Result};
-
-use crate::core::{DetectionContext, DocumentTree, PhaseContext as _};
-use crate::detection::DetectionPlan;
-use crate::detection::phases::deduplication::DeduplicationPhase;
-use crate::detection::phases::detection::DetectionPhase;
-use crate::detection::phases::extraction::ExtractionPhase;
-
-const TARGET: &str = "nvisy_engine::pipeline::detection::document";
-
-/// Detection-only document pipeline. Constructed once per
-/// detection pass, shared across all per-document tasks via
-/// `Arc`.
-pub(super) struct DetectionDocumentPipeline {
-    extraction: ExtractionPhase,
-    detection: DetectionPhase,
-    deduplication: DeduplicationPhase,
-}
-
-impl DetectionDocumentPipeline {
-    pub(super) fn from_context(ctx: &DetectionContext) -> Self {
-        Self {
-            extraction: ExtractionPhase::new(ctx.extraction_engine().clone()),
-            detection: DetectionPhase::new(ctx.recognizer_registry().clone()),
-            deduplication: DeduplicationPhase::new(),
-        }
-    }
-
-    pub(super) async fn run_text(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Text>,
-    ) -> Result<()> {
-        check_cancelled(ctx)?;
-        self.extraction.apply_text(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.detection.apply_text(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.deduplication.apply_text(ctx, plan, tree).await?;
-        check_cancelled(ctx)
-    }
-
-    pub(super) async fn run_tabular(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Tabular>,
-    ) -> Result<()> {
-        check_cancelled(ctx)?;
-        self.extraction.apply_tabular(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.detection.apply_tabular(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.deduplication.apply_tabular(ctx, plan, tree).await?;
-        check_cancelled(ctx)
-    }
-
-    pub(super) async fn run_image(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Image>,
-    ) -> Result<()> {
-        check_cancelled(ctx)?;
-        self.extraction.apply_image(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.detection.apply_image(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.deduplication.apply_image(ctx, plan, tree).await?;
-        check_cancelled(ctx)
-    }
-
-    pub(super) async fn run_audio(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Audio>,
-    ) -> Result<()> {
-        check_cancelled(ctx)?;
-        self.extraction.apply_audio(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.detection.apply_audio(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.deduplication.apply_audio(ctx, plan, tree).await?;
-        check_cancelled(ctx)
-    }
-}
-
-fn check_cancelled(ctx: &DetectionContext) -> Result<()> {
-    if ctx.is_cancelled() {
-        return Err(Error::cancellation("detection cancelled", TARGET));
-    }
-    Ok(())
-}
diff --git a/crates/nvisy-engine/src/detection/engine.rs b/crates/nvisy-engine/src/detection/engine.rs
deleted file mode 100644
index ab644e50c..000000000
--- a/crates/nvisy-engine/src/detection/engine.rs
+++ /dev/null
@@ -1,365 +0,0 @@
-//! [`DetectionEngine`] — entry point for running detection passes.
-//!
-//! Owns the detection-side resources (extractor registry, detection
-//! config template, in-memory detection-pass tracker) plus the
-//! shared infrastructure both engines need (registry, runtime
-//! config, optional key provider, background task set). Per-pass
-//! recognizer registries are built fresh inside the pipeline from
-//! the detection-config template plus the request's policy-supplied
-//! label catalog.
-
-use std::path::Path;
-use std::sync::Arc;
-use std::{fmt, mem};
-
-use nvisy_core::Error;
-use nvisy_toolkit::extraction::ExtractorRegistry;
-use tokio::sync::Mutex;
-use tokio::task::JoinSet;
-use uuid::Uuid;
-
-use super::pipeline::{DetectionEngineState, DetectionPipeline};
-use super::result::DetectionResult;
-use super::state::DetectionState;
-use super::status::DetectionStatus;
-use super::{DetectionEntry, DetectionFilter, DetectionInput, DetectionSnapshot};
-use crate::core::RuntimeConfig;
-use crate::core::ingestion::encryption::SharedKeyProvider;
-use crate::detection::{DetectionConfig, ExtractionConfig};
-use crate::registry::Registry;
-
-/// Shared inner state for a [`DetectionEngine`], held behind an
-/// `Arc`. Cloning the engine is a refcount bump on this inner.
-pub(super) struct DetectionInner {
-    /// Configuration loaded once at engine startup.
-    pub runtime_config: RuntimeConfig,
-    /// Content + audit registry. Shared with [`RedactionEngine`]
-    /// when the two are paired through
-    /// [`RedactionEngine::from_detection`][rfd].
-    ///
-    /// [rfd]: super::super::redaction::RedactionEngine::from_detection
-    pub registry: Registry,
-    /// Pre-built extractor registry, constructed once from
-    /// `runtime_config.extraction` and shared across every pass.
-    pub extraction_engine: Arc<ExtractorRegistry>,
-    /// Detection config template. The per-request
-    /// [`RecognizerRegistry`][rr] is built fresh inside the
-    /// detection pipeline from this template plus the request's
-    /// policy-supplied label catalog.
-    ///
-    /// [rr]: nvisy_toolkit::detection::RecognizerRegistry
-    pub detection_config: Arc<DetectionConfig>,
-    /// Encryption key provider for import decrypt.
-    pub key_provider: Option<SharedKeyProvider>,
-    /// In-memory detection-pass tracker. Cloned (refcount bump)
-    /// into a paired [`RedactionEngine`][re] so its handoff path
-    /// can read prior detections without crossing the registry.
-    ///
-    /// [re]: super::super::redaction::RedactionEngine
-    pub detections: DetectionState,
-    /// Background pipeline tasks spawned by [`DetectionEngine::detect`].
-    pub background_tasks: Mutex<JoinSet<()>>,
-}
-
-/// Detection pipeline engine.
-///
-/// Thin facade over `Arc<DetectionInner>`; cloning is cheap.
-/// Builder methods ([`with_key_provider`]) require exclusive
-/// access via `Arc::get_mut` and must be called before the engine
-/// is cloned.
-///
-/// [`with_key_provider`]: DetectionEngine::with_key_provider
-#[derive(Clone)]
-pub struct DetectionEngine {
-    pub(super) inner: Arc<DetectionInner>,
-}
-
-impl fmt::Debug for DetectionEngine {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("DetectionEngine")
-            .field("runtime_config", &self.inner.runtime_config)
-            .finish()
-    }
-}
-
-impl DetectionEngine {
-    /// Open a new detection engine with the given configuration and
-    /// data directory.
-    ///
-    /// Constructs the registry and the extractor registry. Async
-    /// because future recognizer registrations may need to connect
-    /// to externalized inference services on first use.
-    ///
-    /// # Errors
-    ///
-    /// Returns an error if the registry database cannot be opened
-    /// or the extractor registry fails to construct from
-    /// `config.extraction`.
-    pub async fn open(data_dir: impl AsRef<Path>, config: RuntimeConfig) -> Result<Self, Error> {
-        let registry = Registry::open(data_dir.as_ref())?;
-        let extraction_engine = Arc::new(
-            config
-                .extraction
-                .as_ref()
-                .map(ExtractionConfig::build)
-                .transpose()?
-                .unwrap_or_default(),
-        );
-        let detection_config = Arc::new(config.detection.clone().unwrap_or_default());
-
-        Ok(Self {
-            inner: Arc::new(DetectionInner {
-                runtime_config: config,
-                registry,
-                extraction_engine,
-                detection_config,
-                key_provider: None,
-                detections: DetectionState::new(),
-                background_tasks: Mutex::new(JoinSet::new()),
-            }),
-        })
-    }
-
-    /// Create a temporary detection engine backed by a [`tempfile`]
-    /// directory.
-    ///
-    /// Uses default configuration. The directory is deleted when
-    /// the returned [`TempDir`] is dropped: keep it alive for the
-    /// duration of the test.
-    ///
-    /// # Panics
-    ///
-    /// Panics if the temp directory or registry cannot be created.
-    ///
-    /// [`tempfile`]: https://docs.rs/tempfile
-    /// [`TempDir`]: tempfile::TempDir
-    #[cfg(any(test, feature = "test-utils"))]
-    pub async fn temp() -> (Self, tempfile::TempDir) {
-        let dir = tempfile::tempdir().expect("failed to create temp directory");
-        let engine = Self::open(dir.path(), RuntimeConfig::default())
-            .await
-            .expect("failed to open detection engine in temp directory");
-        (engine, dir)
-    }
-
-    /// Set the key provider for encryption/decryption operations.
-    ///
-    /// # Panics
-    ///
-    /// Panics if the engine has already been cloned (Arc is shared).
-    pub fn with_key_provider(mut self, provider: SharedKeyProvider) -> Self {
-        Arc::get_mut(&mut self.inner)
-            .expect("detection engine must not be shared during construction")
-            .key_provider = Some(provider);
-        self
-    }
-
-    /// Returns the base runtime configuration before per-request
-    /// overrides.
-    pub fn config(&self) -> &RuntimeConfig {
-        &self.inner.runtime_config
-    }
-
-    /// Returns the content and audit registry.
-    pub fn registry(&self) -> &Registry {
-        &self.inner.registry
-    }
-
-    /// Returns the optional key provider. [`RedactionEngine`] reads
-    /// this through [`from_detection`][rfd] so both engines share
-    /// the same decrypt/encrypt material without re-installing it.
-    ///
-    /// [`RedactionEngine`]: super::super::redaction::RedactionEngine
-    /// [rfd]: super::super::redaction::RedactionEngine::from_detection
-    pub fn key_provider(&self) -> Option<&SharedKeyProvider> {
-        self.inner.key_provider.as_ref()
-    }
-
-    /// Returns the in-memory detection state. A paired
-    /// [`RedactionEngine`][re] clones this handle in
-    /// [`from_detection`][rfd] so its redact path can resolve a
-    /// prior detection by id without touching the persistent
-    /// registry.
-    ///
-    /// [re]: super::super::redaction::RedactionEngine
-    /// [rfd]: super::super::redaction::RedactionEngine::from_detection
-    pub fn detections(&self) -> &DetectionState {
-        &self.inner.detections
-    }
-
-    /// Returns the base data directory path backing the registry.
-    pub fn data_dir(&self) -> &Path {
-        self.inner.registry.base_dir()
-    }
-
-    /// Construct a new detection pipeline bound to this engine's
-    /// shared state.
-    fn pipeline(&self) -> DetectionPipeline {
-        DetectionPipeline::new(
-            self.inner.registry.clone(),
-            self.inner.key_provider.clone(),
-            self.inner.detections.clone(),
-            self.inner.runtime_config.clone(),
-            DetectionEngineState {
-                extraction_engine: Arc::clone(&self.inner.extraction_engine),
-                detection_config: Arc::clone(&self.inner.detection_config),
-            },
-        )
-    }
-
-    /// Submit a detection pass for background execution.
-    ///
-    /// Validates the submitted request synchronously (policy/rule
-    /// name uniqueness, label union with conflict detection,
-    /// selector-label resolution against the unioned catalog,
-    /// action validation). On success, registers the pass as
-    /// `Pending`, spawns execution on a background task, and
-    /// returns the detection id immediately. Poll
-    /// [`Self::get_detection`] for status and results.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when the submitted request fails
-    /// any of the four request-entry checks.
-    pub async fn detect(&self, input: DetectionInput) -> Result<Uuid, Error> {
-        let pipeline = self.pipeline();
-        let detection_id = pipeline.id();
-        let prepared = pipeline.register_pending(input).await?;
-
-        self.inner.background_tasks.lock().await.spawn(async move {
-            if let Err(e) = pipeline.execute(prepared).await {
-                tracing::error!(
-                    %detection_id,
-                    error = %e,
-                    "detection pass failed",
-                );
-            }
-        });
-        Ok(detection_id)
-    }
-
-    /// Look up a detection-pass snapshot.
-    ///
-    /// Reads the in-memory state first; falls back to the
-    /// persisted [`DetectionResult`] on disk when the in-memory
-    /// record is absent (which happens after a process restart or
-    /// after the in-memory record has been deleted but the
-    /// persisted blob retained).
-    ///
-    /// # Errors
-    ///
-    /// [`nvisy_core::ErrorKind::NotFound`] when neither the
-    /// in-memory state nor the registry has a matching entry for
-    /// this `(actor_id, id)`.
-    pub async fn get_detection(
-        &self,
-        actor_id: Uuid,
-        id: Uuid,
-    ) -> Result<DetectionSnapshot, Error> {
-        match self.inner.detections.snapshot(actor_id, id).await {
-            Ok(snap) => Ok(snap),
-            Err(e) if e.kind() == nvisy_core::ErrorKind::NotFound => {
-                match self.inner.registry.load_detection(actor_id, id).await {
-                    Ok(result) => Ok(hydrate_snapshot(result)),
-                    Err(_) => Err(e),
-                }
-            }
-            Err(e) => Err(e),
-        }
-    }
-
-    /// List the actor's detection passes.
-    pub async fn list_detections(
-        &self,
-        actor_id: Uuid,
-        filter: DetectionFilter,
-    ) -> Vec<DetectionEntry> {
-        self.inner.detections.list(actor_id, filter).await
-    }
-
-    /// Cancel an in-progress detection.
-    ///
-    /// # Errors
-    ///
-    /// - [`nvisy_core::ErrorKind::NotFound`] when missing or wrong
-    ///   actor.
-    /// - [`nvisy_core::ErrorKind::Validation`] when terminal.
-    pub async fn cancel_detection(&self, actor_id: Uuid, id: Uuid) -> Result<(), Error> {
-        self.inner.detections.cancel(actor_id, id).await
-    }
-
-    /// Delete a finished detection. Cascades to the registry so
-    /// the persisted result is removed atomically with the
-    /// in-memory entry.
-    ///
-    /// # Errors
-    ///
-    /// - [`nvisy_core::ErrorKind::NotFound`] when missing or wrong
-    ///   actor.
-    /// - [`nvisy_core::ErrorKind::Validation`] when still active.
-    pub async fn delete_detection(&self, actor_id: Uuid, id: Uuid) -> Result<(), Error> {
-        self.inner.detections.delete(actor_id, id).await?;
-        // In-memory delete succeeded; remove from disk. A failure
-        // here is logged but not propagated — the in-memory delete
-        // is the user-observable success criterion. A background
-        // reconciliation job (future work) cleans up orphaned
-        // blobs.
-        if let Err(e) = self.inner.registry.unregister_detection(actor_id, id).await {
-            tracing::warn!(
-                detection_id = %id,
-                error = %e,
-                "failed to remove persisted detection result",
-            );
-        }
-        Ok(())
-    }
-
-    /// Delete every finished detection for the actor. Active
-    /// passes are preserved. Returns the number of entries
-    /// removed. Cascades each deletion to the registry; cascade
-    /// failures are logged but don't unblock the result count.
-    pub async fn delete_all_detections(&self, actor_id: Uuid) -> usize {
-        let removed = self.inner.detections.delete_all(actor_id).await;
-        for id in &removed {
-            if let Err(e) = self
-                .inner
-                .registry
-                .unregister_detection(actor_id, *id)
-                .await
-            {
-                tracing::warn!(
-                    detection_id = %id,
-                    error = %e,
-                    "failed to remove persisted detection result during bulk delete",
-                );
-            }
-        }
-        removed.len()
-    }
-
-    /// Wait for all background detection tasks to complete.
-    ///
-    /// Call during graceful shutdown to ensure in-flight passes
-    /// finish and persist their results before the process exits.
-    pub async fn shutdown(&self) {
-        let tasks = mem::take(&mut *self.inner.background_tasks.lock().await);
-        tasks.join_all().await;
-    }
-}
-
-/// Build a synthetic [`DetectionSnapshot`] from a hydrated
-/// [`DetectionResult`] loaded from disk. Historical metadata
-/// (created_at, started_at) is lost on process restart — both
-/// timestamps collapse to "now" so the snapshot is non-empty.
-fn hydrate_snapshot(result: DetectionResult) -> DetectionSnapshot {
-    DetectionSnapshot {
-        id: result.id,
-        actor_id: result.actor_id,
-        status: DetectionStatus::Succeeded,
-        created_at: jiff::Timestamp::now(),
-        started_at: None,
-        completed_at: Some(jiff::Timestamp::now()),
-        result: Some(result),
-        error: None,
-    }
-}
diff --git a/crates/nvisy-engine/src/detection/extraction/mod.rs b/crates/nvisy-engine/src/detection/extraction/mod.rs
deleted file mode 100644
index 839adcc6a..000000000
--- a/crates/nvisy-engine/src/detection/extraction/mod.rs
+++ /dev/null
@@ -1,116 +0,0 @@
-//! Per-deployment `[extraction.*]` configuration that builds the
-//! toolkit-side [`ExtractorRegistry`] at engine startup.
-//!
-//! [`ExtractionConfig::build`] turns each opted-in `[extraction.*]`
-//! section into a concrete extractor and inserts it into the
-//! [`ExtractorRegistry`]. Per-technique sub-configs
-//! ([`OcrExtractorConfig`], [`SttExtractorConfig`]) live in their
-//! own files; the build path is centralised here so the two arms
-//! of each backend selector enum don't need to unify their
-//! concrete types via `dyn` erasure.
-//!
-//! Per-request extraction plan knobs ([`Extraction`] and the
-//! per-modality plan structs) live in [`super::plan`], not here.
-//!
-//! [`ExtractorRegistry`]: nvisy_toolkit::extraction::ExtractorRegistry
-//! [`Extraction`]: super::plan::Extraction
-
-#[cfg(feature = "image")]
-mod ocr;
-#[cfg(feature = "audio")]
-mod stt;
-
-#[cfg(all(feature = "image", not(feature = "bento")))]
-use nvisy_core::Error;
-use nvisy_core::Result;
-use nvisy_toolkit::extraction::ExtractorRegistry;
-use serde::{Deserialize, Serialize};
-
-#[cfg(feature = "image")]
-pub use self::ocr::{OcrBackend, OcrExtractorConfig};
-#[cfg(feature = "audio")]
-pub use self::stt::{SttBackend, SttExtractorConfig};
-
-/// Deployment-time configuration for the extractor registry.
-///
-/// Each field maps to a `[extraction.*]` section in `Nvisy.toml`.
-/// A `None` opts the technique out entirely; an opted-in section
-/// is built once at engine startup and inserted into the
-/// toolkit-side [`ExtractorRegistry`].
-///
-/// [`ExtractorRegistry`]: nvisy_toolkit::extraction::ExtractorRegistry
-#[derive(Debug, Clone, Default, Serialize, Deserialize)]
-pub struct ExtractionConfig {
-    /// `[extraction.ocr]` — OCR text extraction from images.
-    #[cfg(feature = "image")]
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub ocr: Option<OcrExtractorConfig>,
-    /// `[extraction.stt]` — speech-to-text transcription.
-    #[cfg(feature = "audio")]
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub stt: Option<SttExtractorConfig>,
-}
-
-impl ExtractionConfig {
-    /// Build the [`ExtractorRegistry`] from each opted-in section,
-    /// inserting the concrete extractor directly so no `dyn`
-    /// erasure is required outside the registry's internal
-    /// storage.
-    ///
-    /// # Errors
-    ///
-    /// Returns the first construction error encountered.
-    pub fn build(&self) -> Result<ExtractorRegistry> {
-        let mut reg = ExtractorRegistry::new();
-
-        #[cfg(feature = "image")]
-        if let Some(ocr_cfg) = self.ocr.as_ref().filter(|c| c.enabled) {
-            use nvisy_ocr::OcrExtractor;
-            use nvisy_ocr::backend::NoopBackend;
-            #[cfg(feature = "bento")]
-            use nvisy_ocr::backend::{BentoBackend, BentoParams};
-            reg = match &ocr_cfg.backend {
-                OcrBackend::Noop => reg.with_image_extractor(OcrExtractor::new(NoopBackend)),
-
-                #[cfg(feature = "bento")]
-                OcrBackend::Bento { base_url } => {
-                    let backend = BentoBackend::new(BentoParams::new(base_url.clone()))?;
-                    reg.with_image_extractor(OcrExtractor::new(backend))
-                }
-
-                #[cfg(not(feature = "bento"))]
-                OcrBackend::Bento { .. } => {
-                    return Err(Error::validation(
-                        "OcrBackend::Bento requires the `bento` feature",
-                        "ocr",
-                    ));
-                }
-            };
-        }
-
-        #[cfg(feature = "audio")]
-        if let Some(stt_cfg) = self.stt.as_ref().filter(|c| c.enabled) {
-            use nvisy_stt::SttExtractor;
-            use nvisy_stt::backend::NoopBackend;
-            reg = match &stt_cfg.backend {
-                SttBackend::Noop => reg.with_audio_extractor(SttExtractor::new(NoopBackend)),
-            };
-        }
-
-        Ok(reg)
-    }
-
-    /// `true` when every technique is `None`.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        #[cfg(feature = "image")]
-        let ocr_empty = self.ocr.is_none();
-        #[cfg(not(feature = "image"))]
-        let ocr_empty = true;
-        #[cfg(feature = "audio")]
-        let stt_empty = self.stt.is_none();
-        #[cfg(not(feature = "audio"))]
-        let stt_empty = true;
-        ocr_empty && stt_empty
-    }
-}
diff --git a/crates/nvisy-engine/src/detection/extraction/ocr.rs b/crates/nvisy-engine/src/detection/extraction/ocr.rs
deleted file mode 100644
index 8b5fb0168..000000000
--- a/crates/nvisy-engine/src/detection/extraction/ocr.rs
+++ /dev/null
@@ -1,50 +0,0 @@
-//! [`OcrExtractorConfig`] + the [`OcrBackend`] selector enum.
-//!
-//! `[extraction.ocr]` is the TOML section the engine reads at startup;
-//! it picks one of the backends [`nvisy_ocr`] ships. The build path
-//! lives in the parent [`ExtractionConfig::build`].
-//!
-//! [`ExtractionConfig::build`]: super::ExtractionConfig::build
-
-use serde::{Deserialize, Serialize};
-
-/// `[extraction.ocr]` config bundle.
-#[derive(Debug, Clone, Default, Serialize, Deserialize)]
-pub struct OcrExtractorConfig {
-    /// Enable this extractor. When `false`, the extractor is neither
-    /// built nor dispatched, but the config is preserved so operators
-    /// can toggle without losing it. Defaults to `true`.
-    #[serde(default = "default_true")]
-    pub enabled: bool,
-    /// OCR backend selection + connection settings.
-    #[serde(default)]
-    pub backend: OcrBackend,
-}
-
-/// Config-side selection of which OCR backend to construct.
-///
-/// The enum is always parseable regardless of compiled features —
-/// selecting [`OcrBackend::Bento`] on a build without the `bento`
-/// feature surfaces as a clear runtime error at construction instead
-/// of a deserialisation failure, so config files stay portable across
-/// deployments.
-#[derive(Debug, Clone, Default, Serialize, Deserialize)]
-#[serde(tag = "kind", rename_all = "kebab-case")]
-pub enum OcrBackend {
-    /// No-op backend — produces zero OCR results. The default; used
-    /// in tests and in deployments that don't need OCR.
-    #[default]
-    Noop,
-
-    /// Externalised Bento backend — calls the `inference-ocr`
-    /// Bento over HTTP. Requires the runtime to be built with the
-    /// `bento` feature.
-    Bento {
-        /// Base URL of the `inference-ocr` Bento.
-        base_url: String,
-    },
-}
-
-fn default_true() -> bool {
-    true
-}
diff --git a/crates/nvisy-engine/src/detection/extraction/stt.rs b/crates/nvisy-engine/src/detection/extraction/stt.rs
deleted file mode 100644
index 1f72df471..000000000
--- a/crates/nvisy-engine/src/detection/extraction/stt.rs
+++ /dev/null
@@ -1,42 +0,0 @@
-//! [`SttExtractorConfig`] + the closed [`SttBackend`] selector enum.
-//!
-//! `[extraction.stt]` is the TOML section the engine reads at startup;
-//! it picks one of the backends [`nvisy_stt`] ships. The build path
-//! lives in the parent [`ExtractionConfig::build`].
-//!
-//! [`ExtractionConfig::build`]: super::ExtractionConfig::build
-
-use serde::{Deserialize, Serialize};
-
-/// `[extraction.stt]` config bundle.
-#[derive(Debug, Clone, Default, Serialize, Deserialize)]
-pub struct SttExtractorConfig {
-    /// Enable this extractor. When `false`, the extractor is neither
-    /// built nor dispatched, but the config is preserved so operators
-    /// can toggle without losing it. Defaults to `true`.
-    #[serde(default = "default_true")]
-    pub enabled: bool,
-    /// STT backend selection + connection settings.
-    #[serde(default)]
-    pub backend: SttBackend,
-}
-
-/// Config-side selection of which STT [`Backend`] to construct.
-///
-/// The enum is always parseable regardless of compiled features;
-/// future hosted backends (Whisper, Deepgram, AssemblyAI) plug in here
-/// as additional variants.
-///
-/// [`Backend`]: nvisy_stt::backend::SttBackend
-#[derive(Debug, Clone, Default, Serialize, Deserialize)]
-#[serde(tag = "kind", rename_all = "kebab-case")]
-pub enum SttBackend {
-    /// No-op backend — produces zero transcription segments. The
-    /// default; used in tests and in deployments that don't need STT.
-    #[default]
-    Noop,
-}
-
-fn default_true() -> bool {
-    true
-}
diff --git a/crates/nvisy-engine/src/detection/input.rs b/crates/nvisy-engine/src/detection/input.rs
deleted file mode 100644
index 07b06bf52..000000000
--- a/crates/nvisy-engine/src/detection/input.rs
+++ /dev/null
@@ -1,407 +0,0 @@
-//! [`DetectionInput`]: arguments to [`DetectionEngine::detect`].
-//!
-//! [`DetectionEngine::detect`]: crate::detection::DetectionEngine::detect
-
-use std::collections::{HashMap, HashSet};
-
-use hipstr::HipStr;
-use nvisy_core::Error;
-use nvisy_core::entity::EntityLabelCatalog;
-use uuid::Uuid;
-
-use crate::core::ingestion::ImportFile;
-use crate::detection::DetectionPlan;
-use crate::policy::{Action, Policy};
-
-const TARGET: &str = "nvisy_engine::pipeline::detection::input";
-
-/// Input required to execute a detection pass.
-///
-/// A detection pass runs imports → extraction → recognition →
-/// deduplication → policy evaluation. It stops before applying
-/// any redaction — the [`DetectionResult`] holds the audits with
-/// their `Execution::Pending` decisions for the caller to review.
-///
-/// Export sinks live on the matching [`RedactionInput`] instead.
-///
-/// [`DetectionResult`]: super::DetectionResult
-/// [`RedactionInput`]: super::super::redaction::RedactionInput
-#[derive(Clone)]
-pub struct DetectionInput {
-    /// Identity of the human or service account initiating the run.
-    pub actor_id: Uuid,
-    /// Policies to apply, in precedence order: index `0` is highest
-    /// precedence. Submitted inline with their full rule bodies —
-    /// the engine does not persist policies as a resource. Callers
-    /// reuse policies by re-submitting the same bytes.
-    pub policies: Vec<Policy>,
-    /// Content sources to ingest at the start of the run.
-    pub imports: Vec<ImportFile>,
-    /// Per-phase behaviour knobs the detection pipeline reads for
-    /// each document.
-    pub plan: DetectionPlan,
-}
-
-impl DetectionInput {
-    /// Validate the name uniqueness invariants the audit's
-    /// [`PolicyDecisionRef`] depends on: no two submitted policies
-    /// share a name, no two rules inside one policy share a name.
-    /// Audit decisions reference rules by policy + rule name, so
-    /// duplicates would make those references ambiguous.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error naming the offending duplicate.
-    ///
-    /// [`PolicyDecisionRef`]: crate::policy::PolicyDecisionRef
-    pub fn validate_namespace(&self) -> Result<(), Error> {
-        let mut seen_policies: HashSet<&str> = HashSet::with_capacity(self.policies.len());
-        for policy in &self.policies {
-            let policy_name = policy.name.as_str();
-            if !seen_policies.insert(policy_name) {
-                return Err(Error::validation(
-                    format!(
-                        "duplicate policy name `{policy_name}` in detection submission; \
-                         audit references rules by policy + rule name, so policy \
-                         names must be unique within a single detect call",
-                    ),
-                    TARGET,
-                ));
-            }
-            let mut seen_rules: HashSet<&str> = HashSet::with_capacity(policy.rules.len());
-            for rule in &policy.rules {
-                let rule_name = rule.name.as_str();
-                if !seen_rules.insert(rule_name) {
-                    return Err(Error::validation(
-                        format!(
-                            "policy `{policy_name}` has duplicate rule name `{rule_name}`; \
-                             rule names must be unique within their owning policy",
-                        ),
-                        TARGET,
-                    ));
-                }
-            }
-        }
-        Ok(())
-    }
-
-    /// Union every submitted policy's [`Policy::labels`] into a
-    /// per-request [`EntityLabelCatalog`]. The returned catalog
-    /// drives recognizer dispatch (NER label list, pattern
-    /// filtering) and selector tag matching for this request.
-    ///
-    /// Two policies declaring the same label name with non-equal
-    /// `(description, tags)` are a conflict — the engine cannot
-    /// pick a winner without violating one author's intent.
-    /// Identical re-declarations of the same label are allowed
-    /// (idempotent).
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error naming both policies on the
-    /// first conflict encountered.
-    pub fn unify_labels(&self) -> Result<EntityLabelCatalog, Error> {
-        let mut catalog = EntityLabelCatalog::new();
-        let mut origin: HashMap<HipStr<'static>, HipStr<'static>> = HashMap::new();
-
-        for policy in &self.policies {
-            for label in &policy.labels {
-                if let Some(existing) = catalog.lookup(label.name.as_str()) {
-                    if existing != label {
-                        let prior = origin
-                            .get(&label.name)
-                            .map(HipStr::as_str)
-                            .unwrap_or("<unknown>");
-                        return Err(Error::validation(
-                            format!(
-                                "policy `{}` redeclares label `{}` with a different body \
-                                 than policy `{prior}`; resolve the divergence so both \
-                                 policies share one definition",
-                                policy.name, label.name,
-                            ),
-                            TARGET,
-                        ));
-                    }
-                    continue;
-                }
-                origin.insert(label.name.clone(), policy.name.clone());
-                catalog.insert(label.clone());
-            }
-        }
-        Ok(catalog)
-    }
-
-    /// Validate that every [`EntitySelector::labels`] entry across
-    /// every rule references a label registered in `catalog`.
-    /// Selectors targeting unregistered names would never match —
-    /// silently dropping them would let a typo bypass redaction.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error naming the policy, rule, and
-    /// missing label on the first violation encountered.
-    ///
-    /// [`EntitySelector::labels`]: crate::policy::EntitySelector::labels
-    pub fn validate_selector_labels(&self, catalog: &EntityLabelCatalog) -> Result<(), Error> {
-        for policy in &self.policies {
-            for rule in &policy.rules {
-                for label in &rule.selector.labels {
-                    if catalog.lookup(label.as_str()).is_none() {
-                        return Err(Error::validation(
-                            format!(
-                                "policy `{}` rule `{}` selects label `{}` \
-                                 which no policy declares; declare it on `policy.labels` \
-                                 or remove the selector entry",
-                                policy.name,
-                                rule.name,
-                                label.as_str(),
-                            ),
-                            TARGET,
-                        ));
-                    }
-                }
-            }
-        }
-        Ok(())
-    }
-
-    /// Validate every `Action::Redact` action carries at least one
-    /// per-modality operator. A rule that says "redact" but
-    /// specifies no operators is an author bug — the apply phase
-    /// would either fall through to the default operator (which
-    /// might not exist either) or silently skip the entity.
-    /// Either way, the author meant *something* and didn't say
-    /// what — reject loudly.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error naming the policy and rule on
-    /// the first empty `Redact` action encountered.
-    pub fn validate_actions(&self) -> Result<(), Error> {
-        for policy in &self.policies {
-            for rule in &policy.rules {
-                check_action(&rule.action, &policy.name, Some(&rule.name))?;
-            }
-            if let Some(action) = policy.default_action.as_ref() {
-                check_action(action, &policy.name, None)?;
-            }
-        }
-        Ok(())
-    }
-}
-
-fn check_action(action: &Action, policy_name: &str, rule_name: Option<&str>) -> Result<(), Error> {
-    if let Action::Redact(operators) = action
-        && operators.is_empty()
-    {
-        let location = match rule_name {
-            Some(r) => format!("rule `{r}`"),
-            None => "default_action".to_string(),
-        };
-        return Err(Error::validation(
-            format!(
-                "policy `{policy_name}` {location} declares `redact` but specifies no \
-                 per-modality operators; add at least one of `text`, `tabular`, \
-                 `image`, `audio`",
-            ),
-            TARGET,
-        ));
-    }
-    Ok(())
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{EntityLabel, EntityLabelRef};
-    use semver::Version;
-
-    use super::*;
-    use crate::policy::redaction::{ModalityRedactions, TextRedaction};
-    use crate::policy::{EntitySelector, PolicyRule, SuppressAction};
-
-    fn label(name: &str, tags: &[&'static str]) -> EntityLabel {
-        EntityLabel::new(name.to_owned()).with_tags(tags.iter().copied())
-    }
-
-    fn policy(name: &str, rules: &[&str], labels: Vec<EntityLabel>) -> Policy {
-        let rules = rules
-            .iter()
-            .map(|n| PolicyRule {
-                name: HipStr::from(*n),
-                selector: EntitySelector::default(),
-                action: Action::Suppress(SuppressAction::default()),
-                conditions: Vec::new(),
-                enabled: true,
-            })
-            .collect();
-        Policy {
-            name: HipStr::from(name),
-            version: Version::new(1, 0, 0),
-            description: None,
-            labels,
-            rules,
-            default_action: None,
-            retention: Vec::new(),
-        }
-    }
-
-    fn policy_with_selector(name: &str, selector_labels: &[&str]) -> Policy {
-        let labels: Vec<EntityLabelRef> = selector_labels
-            .iter()
-            .map(|s| EntityLabelRef::new(HipStr::from(s.to_owned())))
-            .collect();
-        Policy {
-            name: HipStr::from(name),
-            version: Version::new(1, 0, 0),
-            description: None,
-            labels: Vec::new(),
-            rules: vec![PolicyRule {
-                name: HipStr::from("r"),
-                selector: EntitySelector {
-                    labels,
-                    ..EntitySelector::default()
-                },
-                action: Action::Suppress(SuppressAction::default()),
-                conditions: Vec::new(),
-                enabled: true,
-            }],
-            default_action: None,
-            retention: Vec::new(),
-        }
-    }
-
-    fn input(policies: Vec<Policy>) -> DetectionInput {
-        DetectionInput {
-            actor_id: Uuid::nil(),
-            policies,
-            imports: Vec::new(),
-            plan: DetectionPlan::default(),
-        }
-    }
-
-    #[test]
-    fn validate_namespace_empty_passes() {
-        assert!(input(Vec::new()).validate_namespace().is_ok());
-    }
-
-    #[test]
-    fn validate_namespace_distinct_policies_pass() {
-        let r = input(vec![
-            policy("gdpr", &["redact-ssn", "redact-email"], Vec::new()),
-            policy("hipaa", &["redact-mrn"], Vec::new()),
-        ])
-        .validate_namespace();
-        assert!(r.is_ok());
-    }
-
-    #[test]
-    fn validate_namespace_duplicate_policy_name_fails() {
-        let err = input(vec![
-            policy("gdpr", &["a"], Vec::new()),
-            policy("gdpr", &["b"], Vec::new()),
-        ])
-        .validate_namespace()
-        .unwrap_err();
-        assert!(err.to_string().contains("duplicate policy name `gdpr`"));
-    }
-
-    #[test]
-    fn validate_namespace_duplicate_rule_name_fails() {
-        let err = input(vec![policy("gdpr", &["dup", "dup"], Vec::new())])
-            .validate_namespace()
-            .unwrap_err();
-        assert!(err.to_string().contains("duplicate rule name `dup`"));
-    }
-
-    #[test]
-    fn unify_labels_empty_input_is_empty_catalog() {
-        let cat = input(Vec::new()).unify_labels().unwrap();
-        assert!(cat.is_empty());
-    }
-
-    #[test]
-    fn unify_labels_unions_disjoint_policies() {
-        let cat = input(vec![
-            policy("gdpr", &[], vec![label("email_address", &["pii"])]),
-            policy("hipaa", &[], vec![label("diagnosis", &["phi"])]),
-        ])
-        .unify_labels()
-        .unwrap();
-        assert_eq!(cat.len(), 2);
-    }
-
-    #[test]
-    fn unify_labels_idempotent_redeclaration_passes() {
-        let lbl = label("email_address", &["pii"]);
-        let cat = input(vec![
-            policy("gdpr", &[], vec![lbl.clone()]),
-            policy("ccpa", &[], vec![lbl]),
-        ])
-        .unify_labels()
-        .unwrap();
-        assert_eq!(cat.len(), 1);
-    }
-
-    #[test]
-    fn unify_labels_conflicting_redeclaration_fails() {
-        let err = input(vec![
-            policy("gdpr", &[], vec![label("email_address", &["pii"])]),
-            policy(
-                "ccpa",
-                &[],
-                vec![label("email_address", &["pii", "contact_info"])],
-            ),
-        ])
-        .unify_labels()
-        .unwrap_err();
-        let msg = err.to_string();
-        assert!(msg.contains("ccpa") && msg.contains("gdpr") && msg.contains("email_address"));
-    }
-
-    #[test]
-    fn validate_selector_labels_passes_when_label_in_catalog() {
-        let catalog = EntityLabelCatalog::new().with_label(label("email_address", &["pii"]));
-        let i = input(vec![policy_with_selector("gdpr", &["email_address"])]);
-        assert!(i.validate_selector_labels(&catalog).is_ok());
-    }
-
-    #[test]
-    fn validate_selector_labels_fails_on_missing_label() {
-        let catalog = EntityLabelCatalog::new();
-        let i = input(vec![policy_with_selector("gdpr", &["email_address"])]);
-        let err = i.validate_selector_labels(&catalog).unwrap_err();
-        let msg = err.to_string();
-        assert!(msg.contains("email_address") && msg.contains("gdpr"));
-    }
-
-    #[test]
-    fn validate_actions_passes_with_non_empty_redact() {
-        let mut p = policy("gdpr", &[], Vec::new());
-        p.rules.push(PolicyRule {
-            name: HipStr::from("r"),
-            selector: EntitySelector::default(),
-            action: Action::Redact(ModalityRedactions {
-                text: Some(TextRedaction::Redact),
-                ..Default::default()
-            }),
-            conditions: Vec::new(),
-            enabled: true,
-        });
-        assert!(input(vec![p]).validate_actions().is_ok());
-    }
-
-    #[test]
-    fn validate_actions_fails_on_empty_redact() {
-        let mut p = policy("gdpr", &[], Vec::new());
-        p.rules.push(PolicyRule {
-            name: HipStr::from("empty"),
-            selector: EntitySelector::default(),
-            action: Action::Redact(ModalityRedactions::default()),
-            conditions: Vec::new(),
-            enabled: true,
-        });
-        let err = input(vec![p]).validate_actions().unwrap_err();
-        let msg = err.to_string();
-        assert!(msg.contains("empty") && msg.contains("gdpr"));
-    }
-}
diff --git a/crates/nvisy-engine/src/detection/mod.rs b/crates/nvisy-engine/src/detection/mod.rs
deleted file mode 100644
index 31e508844..000000000
--- a/crates/nvisy-engine/src/detection/mod.rs
+++ /dev/null
@@ -1,44 +0,0 @@
-//! Detection subsystem: standalone recognise + evaluate pipeline.
-//!
-//! [`DetectionEngine::detect`] runs imports → extraction →
-//! detection → deduplication → policy evaluation and stops. The
-//! result is an immutable [`DetectionResult`] holding the
-//! per-document audits (with `Execution::Pending` decisions) plus
-//! the original `ImportFile` references so a follow-up
-//! [`RedactionEngine::redact`][rr] call can re-open the same
-//! content for byte rewriting.
-//!
-//! Detection results are first-class addressable artifacts: one
-//! detection can feed multiple redaction passes (e.g. preview
-//! with `Mask`, then commit with `Fake`), and the user can review
-//! + override the decisions between detect and redact.
-//!
-//! [rr]: crate::redaction::RedactionEngine::redact
-
-mod config;
-mod document;
-mod engine;
-mod extraction;
-mod input;
-mod orchestrator;
-pub mod phases;
-mod pipeline;
-mod plan;
-mod result;
-mod state;
-mod status;
-
-pub use self::config::{DetectionConfig, NerBackend, NerDetection, PatternDetection};
-pub use self::engine::DetectionEngine;
-pub use self::extraction::ExtractionConfig;
-#[cfg(feature = "image")]
-pub use self::extraction::{OcrBackend, OcrExtractorConfig};
-#[cfg(feature = "audio")]
-pub use self::extraction::{SttBackend, SttExtractorConfig};
-pub use self::input::DetectionInput;
-pub use self::plan::{
-    AudioPlan, DeduplicationParams, DetectionPlan, Extraction, ImagePlan, TabularPlan, TextPlan,
-};
-pub use self::result::{DetectionEntry, DetectionFilter, DetectionResult, DetectionSnapshot};
-pub use self::state::DetectionState;
-pub use self::status::DetectionStatus;
diff --git a/crates/nvisy-engine/src/detection/orchestrator.rs b/crates/nvisy-engine/src/detection/orchestrator.rs
deleted file mode 100644
index 75f366ada..000000000
--- a/crates/nvisy-engine/src/detection/orchestrator.rs
+++ /dev/null
@@ -1,183 +0,0 @@
-//! Detection orchestrator: imports content, fans out per-document
-//! detection tasks, collects audits.
-
-use std::sync::Arc;
-
-use nvisy_core::Error;
-use tokio::sync::Semaphore;
-use tokio::task::JoinSet;
-
-use super::document::DetectionDocumentPipeline;
-use crate::core::ingestion::{ImportFile, Importer};
-use crate::core::{AnyTree, DetectionContext, PhaseContext as _};
-use crate::detection::DetectionPlan;
-use crate::document::provenance::AnyAudit;
-
-const TARGET: &str = "nvisy_engine::pipeline::detection::orchestrator";
-
-pub(super) struct DetectionDocumentResult {
-    pub tree: Option<AnyTree>,
-    pub error: Option<String>,
-}
-
-pub(super) struct DetectionOutput {
-    pub results: Vec<DetectionDocumentResult>,
-}
-
-impl DetectionOutput {
-    pub(super) fn into_audits(self) -> (Vec<AnyAudit>, u64, bool, bool) {
-        let mut audits = Vec::new();
-        let mut entities = 0u64;
-        let mut any_ok = false;
-        let mut any_failed = false;
-        for result in self.results {
-            if result.error.is_some() {
-                any_failed = true;
-                continue;
-            }
-            any_ok = true;
-            if let Some(tree) = result.tree {
-                let audit = tree.audit_cloned();
-                entities += audit.entities_count() as u64;
-                audits.push(audit);
-            }
-        }
-        (audits, entities, any_ok, any_failed)
-    }
-}
-
-pub(super) struct DetectionOrchestrator {
-    ctx: Arc<DetectionContext>,
-    semaphore: Option<Arc<Semaphore>>,
-}
-
-impl DetectionOrchestrator {
-    pub(super) fn new(ctx: DetectionContext) -> Self {
-        let semaphore = ctx.concurrency().map(|c| Arc::new(Semaphore::new(c.get())));
-        Self {
-            ctx: Arc::new(ctx),
-            semaphore,
-        }
-    }
-
-    pub(super) async fn run(
-        &self,
-        imports: &[ImportFile],
-        plan: &DetectionPlan,
-    ) -> Result<DetectionOutput, Error> {
-        let trees = self.run_imports(imports).await?;
-
-        let pipeline = Arc::new(DetectionDocumentPipeline::from_context(&self.ctx));
-        let plan = Arc::new(plan.clone());
-        let mut join_set: JoinSet<DetectionDocumentResult> = JoinSet::new();
-        for tree in trees {
-            let ctx = Arc::clone(&self.ctx);
-            let sem = self.semaphore.clone();
-            let plan = Arc::clone(&plan);
-            let pipeline = Arc::clone(&pipeline);
-            join_set.spawn(run_one(pipeline, tree, ctx, sem, plan));
-        }
-
-        let mut results: Vec<DetectionDocumentResult> = Vec::new();
-        while let Some(joined) = join_set.join_next().await {
-            match joined {
-                Ok(doc) => results.push(doc),
-                Err(e) => {
-                    let msg = if e.is_panic() {
-                        let payload = e.into_panic();
-                        let panic_msg = payload
-                            .downcast_ref::<&str>()
-                            .copied()
-                            .or_else(|| payload.downcast_ref::<String>().map(|s| s.as_str()))
-                            .unwrap_or("unknown panic");
-                        tracing::error!(
-                            target: TARGET,
-                            panic = panic_msg,
-                            "detection document task panicked"
-                        );
-                        format!("Detection task panicked: {panic_msg}")
-                    } else {
-                        tracing::error!(
-                            target: TARGET,
-                            error = %e,
-                            "detection document task failed"
-                        );
-                        format!("Detection task failed: {e}")
-                    };
-                    results.push(DetectionDocumentResult {
-                        tree: None,
-                        error: Some(msg),
-                    });
-                }
-            }
-        }
-        Ok(DetectionOutput { results })
-    }
-
-    async fn run_imports(&self, imports: &[ImportFile]) -> Result<Vec<AnyTree>, Error> {
-        let mut trees = Vec::new();
-        for cfg in imports {
-            let importer = Importer::new()
-                .with_decompression(cfg.decompression)
-                .with_decryption(cfg.decryption.clone());
-            let shared = self.ctx.shared();
-            for &content_id in &cfg.content_ids {
-                tracing::debug!(target: TARGET, %content_id, "importing content");
-                if self.ctx.is_cancelled() {
-                    return Err(Error::cancellation("detection cancelled", TARGET));
-                }
-                let handle = shared
-                    .registry
-                    .read_content(shared.actor_id, content_id)
-                    .await?;
-                let content = handle.content().await?;
-                let annotations = shared
-                    .registry
-                    .load_annotations(shared.actor_id, content_id)
-                    .await?;
-                trees.extend(importer.import(content, annotations, shared).await?);
-            }
-        }
-        tracing::info!(target: TARGET, count = trees.len(), "trees imported");
-        Ok(trees)
-    }
-}
-
-async fn run_one(
-    pipeline: Arc<DetectionDocumentPipeline>,
-    tree: AnyTree,
-    ctx: Arc<DetectionContext>,
-    sem: Option<Arc<Semaphore>>,
-    plan: Arc<DetectionPlan>,
-) -> DetectionDocumentResult {
-    let _permit = match sem {
-        Some(s) => match Arc::clone(&s).acquire_owned().await {
-            Ok(permit) => Some(permit),
-            Err(_) => {
-                return DetectionDocumentResult {
-                    tree: None,
-                    error: Some("concurrency semaphore closed".to_owned()),
-                };
-            }
-        },
-        None => None,
-    };
-
-    let mut tree = tree;
-    let result = match &mut tree {
-        AnyTree::Text(t) => pipeline.run_text(&ctx, &plan, t).await,
-        AnyTree::Tabular(t) => pipeline.run_tabular(&ctx, &plan, t).await,
-        AnyTree::Image(t) => pipeline.run_image(&ctx, &plan, t).await,
-        AnyTree::Audio(t) => pipeline.run_audio(&ctx, &plan, t).await,
-    };
-    if let Err(e) = result {
-        return DetectionDocumentResult {
-            tree: None,
-            error: Some(e.to_string()),
-        };
-    }
-    DetectionDocumentResult {
-        tree: Some(tree),
-        error: None,
-    }
-}
diff --git a/crates/nvisy-engine/src/detection/phases/deduplication.rs b/crates/nvisy-engine/src/detection/phases/deduplication.rs
deleted file mode 100644
index 1a4019f48..000000000
--- a/crates/nvisy-engine/src/detection/phases/deduplication.rs
+++ /dev/null
@@ -1,128 +0,0 @@
-//! [`DeduplicationPhase`]: per-document dedup driver.
-//!
-//! Runs the canonical dedup pipeline (calibrate → filter → fuse →
-//! resolve) via [`LayerPipeline::from_params`] against each
-//! [`DocumentTree<M>`]'s audit records. Stateless; per-run config
-//! comes from `plan` each call.
-//!
-//! [`LayerPipeline::from_params`]: nvisy_toolkit::deduplication::LayerPipeline::from_params
-
-use std::mem;
-
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::{Audio, Image, Overlap, Tabular, Text};
-use nvisy_toolkit::deduplication::{LayerContext, LayerPipeline, SpanSize};
-use tracing::Instrument;
-use uuid::Uuid;
-
-use crate::core::{DetectionContext, DocumentTree, PhaseContext as _};
-use crate::detection::{DeduplicationParams, DetectionPlan};
-use crate::document::provenance::EntityRecord;
-use crate::modality::DocumentModality;
-
-const TARGET: &str = "nvisy_engine::deduplication";
-
-/// Deduplication phase orchestrator.
-///
-/// Stateless. Reads [`DeduplicationParams`] from `plan` each call.
-pub struct DeduplicationPhase;
-
-impl DeduplicationPhase {
-    /// Build a fresh dedup phase. The phase is stateless; instances
-    /// are interchangeable.
-    pub fn new() -> Self {
-        Self
-    }
-
-    pub(crate) async fn apply_text(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Text>,
-    ) -> Result<()> {
-        self.run(ctx, plan, tree).await
-    }
-
-    pub(crate) async fn apply_tabular(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Tabular>,
-    ) -> Result<()> {
-        self.run(ctx, plan, tree).await
-    }
-
-    pub(crate) async fn apply_image(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Image>,
-    ) -> Result<()> {
-        self.run(ctx, plan, tree).await
-    }
-
-    pub(crate) async fn apply_audio(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Audio>,
-    ) -> Result<()> {
-        self.run(ctx, plan, tree).await
-    }
-
-    async fn run<M>(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<M>,
-    ) -> Result<()>
-    where
-        M: DocumentModality,
-        M::Location: Overlap + SpanSize,
-        DocumentTree<M>: TextAt<M>,
-    {
-        let span = tracing::info_span!(target: TARGET, "phase", name = "deduplication");
-        let run_id = ctx.shared().run_id;
-        async move { dedup_one(tree, &plan.deduplication, run_id).await }
-            .instrument(span)
-            .await
-    }
-}
-
-impl Default for DeduplicationPhase {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
-async fn dedup_one<M>(
-    tree: &mut DocumentTree<M>,
-    dedup: &DeduplicationParams,
-    run_id: Uuid,
-) -> Result<()>
-where
-    M: DocumentModality,
-    M::Location: Overlap + SpanSize,
-    DocumentTree<M>: TextAt<M>,
-{
-    if tree.root.audit.records.is_empty() {
-        return Ok(());
-    }
-    tracing::debug!(
-        target: TARGET,
-        entities = tree.root.audit.records.len(),
-        "running deduplication",
-    );
-    // Dedup runs before redaction evaluation, so every record's
-    // `audit` is still None; we can pull entities out, dedup, and
-    // rewrap without losing audit state.
-    let records = mem::take(&mut tree.root.audit.records);
-    let entities: Vec<Entity<M>> = records.into_iter().map(|r| r.entity).collect();
-    let pipeline: LayerPipeline<M, _> = LayerPipeline::from_params(dedup)?;
-    let ctx = LayerContext::new(&*tree).with_correlation_id(run_id);
-    let deduped = pipeline.run(entities, &ctx).await;
-    tree.root.audit.records = deduped.into_iter().map(EntityRecord::new).collect();
-    Ok(())
-}
diff --git a/crates/nvisy-engine/src/detection/phases/detection.rs b/crates/nvisy-engine/src/detection/phases/detection.rs
deleted file mode 100644
index 8e1f1f581..000000000
--- a/crates/nvisy-engine/src/detection/phases/detection.rs
+++ /dev/null
@@ -1,306 +0,0 @@
-//! [`DetectionPhase`]: per-modality recognition driver.
-//!
-//! Walks each [`Document<M>`]'s blocks, runs the matching recognizers
-//! through [`RecognizerRegistry`], filters by the plan's entity-kind
-//! allowlist, and lifts block-local offsets back to absolute modality
-//! coordinates. For image trees, additionally walks every image chunk
-//! and runs image-side recognizers against the raw bytes.
-//!
-//! [`Document<M>`]: crate::document::Document
-//! [`RecognizerRegistry`]: nvisy_toolkit::detection::RecognizerRegistry
-
-use std::sync::Arc;
-
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{
-    Audio, AudioLocation, Image, ImageLocation, Modality, Overlap, Tabular, TabularLocation, Text,
-    TextData, TextLocation,
-};
-use nvisy_core::recognition::RecognizerInput;
-use nvisy_toolkit::detection::RecognizerRegistry;
-use tracing::Instrument;
-
-use crate::core::{DetectionContext, DocumentTree, PhaseContext as _};
-use crate::detection::DetectionPlan;
-use crate::document::{Document, Span};
-use crate::modality::{DocumentModality, ModalityBlock};
-
-const TARGET: &str = "nvisy_engine::detection";
-
-/// Detection phase: runs every registered recognizer over each
-/// document's blocks and writes [`EntityRecord`]s to `doc.audit`.
-///
-/// Holds an `Arc<RecognizerRegistry>` so the registry is shared
-/// cheaply across per-document phases without cloning the
-/// underlying recognizer lists. Recognizers own any post-detection
-/// work they need (boosting, dedup, validation) — the engine just
-/// orchestrates the registry.
-///
-/// [`EntityRecord`]: crate::document::provenance::EntityRecord
-pub struct DetectionPhase {
-    registry: Arc<RecognizerRegistry>,
-}
-
-impl DetectionPhase {
-    /// Build the phase from the shared recognizer registry. Called
-    /// once per pipeline by the pipeline orchestrator.
-    pub fn new(registry: Arc<RecognizerRegistry>) -> Self {
-        Self { registry }
-    }
-
-    pub(crate) async fn apply_text(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Text>,
-    ) -> Result<()> {
-        self.run_text_only(ctx, plan, &mut tree.root).await
-    }
-
-    pub(crate) async fn apply_tabular(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Tabular>,
-    ) -> Result<()> {
-        self.run_text_only(ctx, plan, &mut tree.root).await
-    }
-
-    pub(crate) async fn apply_audio(
-        &self,
-        ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Audio>,
-    ) -> Result<()> {
-        self.run_text_only(ctx, plan, &mut tree.root).await
-    }
-
-    pub(crate) async fn apply_image(
-        &self,
-        ctx: &DetectionContext,
-        _plan: &DetectionPlan,
-        tree: &mut DocumentTree<Image>,
-    ) -> Result<()> {
-        let span = tracing::info_span!(target: TARGET, "phase", name = "detection.image");
-        let run_id = ctx.shared().run_id;
-        async move {
-            detect_text_blocks(&self.registry, &mut tree.root, run_id).await?;
-            detect_image_chunks(
-                &self.registry,
-                &mut tree.root,
-                tree.handle.handler_mut(),
-                run_id,
-            )
-            .await?;
-            Ok(())
-        }
-        .instrument(span)
-        .await
-    }
-
-    async fn run_text_only<M>(
-        &self,
-        ctx: &DetectionContext,
-        _plan: &DetectionPlan,
-        doc: &mut Document<M>,
-    ) -> Result<()>
-    where
-        M: DocumentModality + LiftFromBlock,
-        M::Location: Overlap,
-    {
-        let span = tracing::info_span!(target: TARGET, "phase", name = "detection.text_only");
-        let run_id = ctx.shared().run_id;
-        async move {
-            detect_text_blocks(&self.registry, doc, run_id).await?;
-            Ok(())
-        }
-        .instrument(span)
-        .await
-    }
-}
-
-/// Shared text-side block loop. Used by every modality that exposes
-/// text via [`ModalityBlock::scan_text`] (today: every modality).
-async fn detect_text_blocks<M>(
-    registry: &RecognizerRegistry,
-    doc: &mut Document<M>,
-    run_id: uuid::Uuid,
-) -> Result<()>
-where
-    M: DocumentModality + LiftFromBlock,
-    M::Location: Overlap,
-{
-    if doc.blocks.is_empty() {
-        return Ok(());
-    }
-
-    let mut lifted: Vec<Entity<M>> = Vec::new();
-    let mut scanned_blocks = 0usize;
-
-    for block in &doc.blocks {
-        let Some(text) = block.kind.scan_text() else {
-            continue;
-        };
-        if text.is_empty() {
-            continue;
-        }
-        scanned_blocks += 1;
-
-        let mut input = RecognizerInput::new(TextData::new(text.to_owned()));
-        input.correlation_id = Some(run_id);
-
-        let detected = registry.run::<Text>(input).await?;
-        for entity in detected {
-            let Some(location) =
-                M::lift_from_block(&block.spans, entity.location.start, entity.location.end)
-            else {
-                tracing::debug!(
-                    target: TARGET,
-                    label = %entity.label,
-                    "dropping entity with no overlapping span",
-                );
-                continue;
-            };
-            lifted.push(Entity {
-                id: entity.id,
-                entity_id: entity.entity_id,
-                label: entity.label,
-                location,
-                confidence: entity.confidence,
-                trail: entity.trail,
-                language: entity.language,
-            });
-        }
-    }
-
-    tracing::debug!(
-        target: TARGET,
-        detected = lifted.len(),
-        blocks = scanned_blocks,
-        "appending text-detected entities",
-    );
-    doc.add_entities(lifted);
-    Ok(())
-}
-
-/// Walk every image chunk reachable through the handle, run the
-/// registry's image recognizers against the raw bytes, and append
-/// resulting entities to the document's audit.
-#[cfg(feature = "image")]
-async fn detect_image_chunks(
-    registry: &RecognizerRegistry,
-    doc: &mut Document<Image>,
-    handle: &mut dyn nvisy_codec::Handler<Image>,
-    run_id: uuid::Uuid,
-) -> Result<()> {
-    if registry.count::<Image>() == 0 {
-        return Ok(());
-    }
-
-    let mut detected_total = 0usize;
-    while let Some(chunk) = handle.next_chunk().await? {
-        let mut input = RecognizerInput::new(chunk.data);
-        input.correlation_id = Some(run_id);
-
-        let detected = registry.run::<Image>(input).await?;
-        detected_total += detected.len();
-        doc.add_entities(detected);
-    }
-
-    tracing::debug!(
-        target: TARGET,
-        detected = detected_total,
-        "appending image-detected entities",
-    );
-    Ok(())
-}
-
-#[cfg(not(feature = "image"))]
-async fn detect_image_chunks(
-    _registry: &RecognizerRegistry,
-    _doc: &mut Document<Image>,
-    _handle: &mut dyn nvisy_codec::Handler<Image>,
-    _run_id: uuid::Uuid,
-) -> Result<()> {
-    Ok(())
-}
-
-/// Map a block-text byte range to an absolute `M` location using the
-/// block's spans.
-pub trait LiftFromBlock: DocumentModality + Sized {
-    /// Translate a block-text byte range `[start, end)` to the
-    /// source-coordinate `Self::Location` carried by the block's
-    /// spans. Returns `None` when the range has no matching pre-image
-    /// in `spans` (e.g. a recognizer match that straddled a span gap).
-    fn lift_from_block(
-        spans: &[Span<Self>],
-        start: usize,
-        end: usize,
-    ) -> Option<<Self as Modality>::Location>;
-}
-
-impl LiftFromBlock for Text {
-    fn lift_from_block(spans: &[Span<Self>], start: usize, end: usize) -> Option<TextLocation> {
-        let span = spans.iter().find(|s| s.overlaps(start, end))?;
-        let span_text_start = span.text_start;
-        let source_base = span.source.start;
-        let lifted_start = source_base + start.saturating_sub(span_text_start);
-        let lifted_end = source_base + end.saturating_sub(span_text_start);
-        Some(TextLocation::new(lifted_start, lifted_end))
-    }
-}
-
-impl LiftFromBlock for Tabular {
-    fn lift_from_block(spans: &[Span<Self>], start: usize, end: usize) -> Option<TabularLocation> {
-        let span = spans.iter().find(|s| s.overlaps(start, end))?;
-        let local_start = start.saturating_sub(span.text_start);
-        let local_end = end.saturating_sub(span.text_start);
-        let cell = &span.source;
-        Some(TabularLocation {
-            row_index: cell.row_index,
-            column_index: cell.column_index,
-            start_offset: Some(local_start),
-            end_offset: Some(local_end),
-            column_name: cell.column_name.clone(),
-            sheet_name: cell.sheet_name.clone(),
-        })
-    }
-}
-
-impl LiftFromBlock for Image {
-    fn lift_from_block(spans: &[Span<Self>], start: usize, end: usize) -> Option<ImageLocation> {
-        let mut iter = spans.iter().filter(|s| s.overlaps(start, end));
-        let first = iter.next()?;
-        let mut bbox = first.source.bounding_box;
-        let image_id = first.source.image_id;
-        let page_number = first.source.page_number;
-        for s in iter {
-            bbox = bbox.union(&s.source.bounding_box);
-        }
-        Some(ImageLocation {
-            bounding_box: bbox,
-            polygon: None,
-            image_id,
-            page_number,
-        })
-    }
-}
-
-impl LiftFromBlock for Audio {
-    fn lift_from_block(spans: &[Span<Self>], start: usize, end: usize) -> Option<AudioLocation> {
-        let mut iter = spans.iter().filter(|s| s.overlaps(start, end));
-        let first = iter.next()?;
-        let mut time_span = first.source.time_span;
-        let speaker_id = first.source.speaker_id.clone();
-        let audio_id = first.source.audio_id;
-        for s in iter {
-            time_span = time_span.union(&s.source.time_span);
-        }
-        Some(AudioLocation {
-            time_span,
-            speaker_id,
-            audio_id,
-        })
-    }
-}
diff --git a/crates/nvisy-engine/src/detection/phases/extraction.rs b/crates/nvisy-engine/src/detection/phases/extraction.rs
deleted file mode 100644
index 244cb2e82..000000000
--- a/crates/nvisy-engine/src/detection/phases/extraction.rs
+++ /dev/null
@@ -1,358 +0,0 @@
-//! [`ExtractionPhase`]: pulls chunks through the codec and writes
-//! per-modality [`Block<M>`] values onto each [`DocumentTree<M>`].
-//!
-//! Text and tabular extraction drive [`Handler::next_chunk`] — one
-//! pass through the codec yields `(location, data)` pairs the phase
-//! turns into blocks. Image and audio still go through the toolkit
-//! extractor slots (OCR / STT) because the codec only carries raw
-//! bytes for those modalities; the per-extractor output is what
-//! becomes a [`Block<Image>`] or [`Block<Audio>`].
-//!
-//! [`Block<M>`]: crate::document::Block
-//! [`DocumentTree<M>`]: crate::core::DocumentTree
-//! [`Handler::next_chunk`]: nvisy_codec::Handler::next_chunk
-
-use std::collections::BTreeMap;
-
-use nvisy_codec::content::ContentDescriptor;
-use nvisy_core::Result;
-#[cfg(any(feature = "image", feature = "audio"))]
-use nvisy_core::extraction::Span as ExtractionSpan;
-use nvisy_core::modality::{
-    Audio, Image, ImageExtraction, ImageLocation, Tabular, TabularLocation, Text,
-};
-#[cfg(feature = "image")]
-use nvisy_core::primitive::BoundingBox;
-use nvisy_ocr::types::RawOcrBlock;
-use nvisy_toolkit::extraction::{Extractor, ExtractorRegistry, ImageExtractorOutput};
-use tracing::Instrument;
-
-use crate::core::{DetectionContext, DocumentTree};
-use crate::detection::{DetectionPlan, Extraction};
-use crate::document::{Block, Document, Span};
-use crate::modality::{ImageBlock, TabularBlock, TextBlock, TextContent};
-
-const TARGET: &str = "nvisy_engine::extraction";
-
-/// Cell separator used by [`populate_tabular_blocks`] to concatenate
-/// per-row text. Tab is chosen because cell values rarely contain it,
-/// so the resulting flat text round-trips back to per-cell ranges
-/// without ambiguity.
-const TABULAR_CELL_SEPARATOR: &str = "\t";
-
-/// Extraction phase: pulls bytes through the codec, runs the matching
-/// extractor, writes [`Block<M>`] values into each document.
-///
-/// Holds an [`ExtractorRegistry`] by value — the engine's per-slot
-/// `Option<Arc<…>>` keeps the underlying OCR/STT services shared
-/// across runs without an outer wrap.
-///
-/// [`Block<M>`]: crate::document::Block
-/// [`ExtractorRegistry`]: nvisy_toolkit::extraction::ExtractorRegistry
-pub struct ExtractionPhase {
-    engine: ExtractorRegistry,
-}
-
-impl ExtractionPhase {
-    /// Build the phase from the shared extraction engine. Called once
-    /// per pipeline by the pipeline orchestrator.
-    pub fn new(engine: ExtractorRegistry) -> Self {
-        Self { engine }
-    }
-
-    /// Apply to a [`Text`] tree: walk every text chunk, append one
-    /// block per chunk with a single span mapping flat block text
-    /// back to the codec's [`Text`] coordinates.
-    pub(crate) async fn apply_text(
-        &self,
-        _ctx: &DetectionContext,
-        _plan: &DetectionPlan,
-        tree: &mut DocumentTree<Text>,
-    ) -> Result<()> {
-        let span = tracing::info_span!(target: TARGET, "phase", name = "extraction.text");
-        async move { populate_text_blocks(&mut tree.root, tree.handle.handler_mut()).await }
-            .instrument(span)
-            .await
-    }
-
-    /// Apply to a [`Tabular`] tree: walk every cell chunk and group
-    /// cells into one block per row, with one span per cell mapping
-    /// the cell's substring back to its `(row, col)` coordinates.
-    pub(crate) async fn apply_tabular(
-        &self,
-        _ctx: &DetectionContext,
-        _plan: &DetectionPlan,
-        tree: &mut DocumentTree<Tabular>,
-    ) -> Result<()> {
-        let span = tracing::info_span!(target: TARGET, "phase", name = "extraction.tabular");
-        async move { populate_tabular_blocks(&mut tree.root, tree.handle.handler_mut()).await }
-            .instrument(span)
-            .await
-    }
-
-    /// Apply to an [`Image`] tree: run the OCR extractor against every
-    /// image chunk and write the extractor's output as
-    /// [`Block<Image>`] values.
-    pub(crate) async fn apply_image(
-        &self,
-        _ctx: &DetectionContext,
-        _plan: &DetectionPlan,
-        tree: &mut DocumentTree<Image>,
-    ) -> Result<()> {
-        let span = tracing::info_span!(target: TARGET, "phase", name = "extraction.image");
-        async move { populate_image_doc(&self.engine, &mut tree.root, tree.handle.handler_mut()).await }
-            .instrument(span)
-            .await
-    }
-
-    /// Apply to an [`Audio`] tree: pull the audio chunk's bytes, run
-    /// the STT extractor, write the transcript as a [`Block<Audio>`].
-    pub(crate) async fn apply_audio(
-        &self,
-        _ctx: &DetectionContext,
-        plan: &DetectionPlan,
-        tree: &mut DocumentTree<Audio>,
-    ) -> Result<()> {
-        let span = tracing::info_span!(target: TARGET, "phase", name = "extraction.audio");
-        let descriptor = tree.descriptor.clone();
-        async move {
-            populate_audio_doc(
-                &self.engine,
-                &mut tree.root,
-                tree.handle.handler_mut(),
-                &descriptor,
-                &plan.extraction,
-            )
-            .await
-        }
-        .instrument(span)
-        .await
-    }
-}
-
-/// Walk the codec's text chunks and append one [`Block<Text>`] per
-/// chunk. Each block's single span maps `0..text.len()` back to the
-/// codec's [`Text`] coordinates so downstream detection can resolve
-/// entity offsets to source locations uniformly across modalities.
-async fn populate_text_blocks(
-    doc: &mut Document<Text>,
-    handle: &mut dyn nvisy_codec::Handler<Text>,
-) -> Result<()> {
-    let mut blocks = Vec::new();
-    while let Some(chunk) = handle.next_chunk().await? {
-        let text = chunk.data.into_string();
-        let span = Span {
-            text_start: 0,
-            text_end: text.len(),
-            confidence: None,
-            source: chunk.location,
-        };
-        let block =
-            Block::new(TextBlock::Text(TextContent::Paragraph { text })).with_spans(vec![span]);
-        blocks.push(block);
-    }
-
-    if blocks.is_empty() {
-        return Ok(());
-    }
-    tracing::debug!(
-        target: TARGET,
-        blocks = blocks.len(),
-        "populated text document",
-    );
-    doc.blocks.extend(blocks);
-    Ok(())
-}
-
-/// Walk the codec's tabular chunks and group cells into one block per
-/// row. Each block carries the row's concatenated text and one span
-/// per cell mapping the substring range back to the codec's
-/// [`Tabular`] coordinates.
-async fn populate_tabular_blocks(
-    doc: &mut Document<Tabular>,
-    handle: &mut dyn nvisy_codec::Handler<Tabular>,
-) -> Result<()> {
-    let mut rows: BTreeMap<u32, Vec<(TabularLocation, String)>> = BTreeMap::new();
-    while let Some(chunk) = handle.next_chunk().await? {
-        let value = chunk.data.into_string();
-        rows.entry(chunk.location.row_index)
-            .or_default()
-            .push((chunk.location, value));
-    }
-    if rows.is_empty() {
-        return Ok(());
-    }
-    for cells in rows.values_mut() {
-        cells.sort_by_key(|(loc, _)| loc.column_index);
-    }
-
-    let mut blocks = Vec::with_capacity(rows.len());
-    for cells in rows.into_values() {
-        let mut text = String::new();
-        let mut spans = Vec::with_capacity(cells.len());
-        for (i, (cell, value)) in cells.into_iter().enumerate() {
-            if i > 0 {
-                text.push_str(TABULAR_CELL_SEPARATOR);
-            }
-            let start = text.len();
-            text.push_str(&value);
-            let end = text.len();
-            spans.push(Span {
-                text_start: start,
-                text_end: end,
-                confidence: None,
-                source: cell,
-            });
-        }
-        let block = Block::new(TabularBlock::Row { text }).with_spans(spans);
-        blocks.push(block);
-    }
-
-    tracing::debug!(
-        target: TARGET,
-        rows = blocks.len(),
-        "populated tabular document",
-    );
-    doc.blocks.extend(blocks);
-    Ok(())
-}
-
-/// Run image extraction over `doc` using the supplied codec handle.
-/// Today that's a single OCR pass when an extractor is configured.
-#[cfg(feature = "image")]
-async fn populate_image_doc(
-    engine: &ExtractorRegistry,
-    doc: &mut Document<Image>,
-    handle: &mut dyn nvisy_codec::Handler<Image>,
-) -> Result<()> {
-    if let Some(ref ocr) = engine.image {
-        run_ocr_into(ocr.as_ref(), doc, handle).await?;
-    }
-    Ok(())
-}
-
-#[cfg(not(feature = "image"))]
-async fn populate_image_doc(
-    _engine: &ExtractorRegistry,
-    _doc: &mut Document<Image>,
-    _handle: &mut dyn nvisy_codec::Handler<Image>,
-) -> Result<()> {
-    Ok(())
-}
-
-/// Drive a single OCR extractor over every image chunk reachable via
-/// `handle`, populating `doc` with the resulting blocks and stamping
-/// the extractor's provenance on the document metadata.
-#[cfg(feature = "image")]
-async fn run_ocr_into(
-    ocr: &dyn Extractor<Image, Output = ImageExtractorOutput>,
-    doc: &mut Document<Image>,
-    handle: &mut dyn nvisy_codec::Handler<Image>,
-) -> Result<()> {
-    while let Some(chunk) = handle.next_chunk().await? {
-        let dims = chunk.data.dims;
-        let location = ImageLocation::new(BoundingBox::new(
-            0.0,
-            0.0,
-            f64::from(dims.width),
-            f64::from(dims.height),
-        ));
-        let span = ExtractionSpan::new(chunk.data, location);
-        let output = ocr.extract(&span).await?;
-        doc.meta.extraction = output.extraction;
-        for block in output.value.blocks {
-            doc.blocks.push(ocr_block_to_block(block));
-        }
-    }
-    if doc.blocks.is_empty() && matches!(doc.meta.extraction, ImageExtraction::Pending) {
-        // No chunks streamed and no extractor output stamped — leave
-        // the document in its `Pending` extraction state.
-        tracing::debug!(target: TARGET, "no image chunks to OCR");
-    }
-    Ok(())
-}
-
-/// Convert a backend-shaped [`RawOcrBlock`] to a document-shaped
-/// [`Block<Image>`].
-#[cfg(feature = "image")]
-fn ocr_block_to_block(raw: RawOcrBlock) -> Block<Image> {
-    use nvisy_ocr::types::OcrBlockKind;
-    let kind = match raw.kind {
-        OcrBlockKind::Text { region, text } => ImageBlock::Text { region, text },
-        OcrBlockKind::Heading { region, text } => ImageBlock::Heading { region, text },
-        OcrBlockKind::Table { region, text } => ImageBlock::Table { region, text },
-        _ => unreachable!("OcrBlockKind has no further variants"),
-    };
-    let spans: Vec<Span<Image>> = raw
-        .spans
-        .into_iter()
-        .map(|s| Span {
-            text_start: s.text_start,
-            text_end: s.text_end,
-            source: s.source,
-            confidence: Some(s.confidence),
-        })
-        .collect();
-    Block {
-        kind,
-        spans,
-        confidence: Some(raw.confidence),
-    }
-}
-
-/// Transcribe the audio reachable via `handle` into `doc`, optionally
-/// diarising when the plan requests it.
-#[cfg(feature = "audio")]
-async fn populate_audio_doc(
-    engine: &ExtractorRegistry,
-    doc: &mut Document<Audio>,
-    handle: &mut dyn nvisy_codec::Handler<Audio>,
-    _descriptor: &ContentDescriptor,
-    plan: &Extraction,
-) -> Result<()> {
-    use nvisy_core::modality::AudioLocation;
-
-    use crate::modality::AudioBlock;
-
-    let Some(ref stt_arc) = engine.audio else {
-        return Ok(());
-    };
-
-    if plan.audio.diarization {
-        tracing::warn!(target: TARGET, "diarization not yet supported, skipping");
-    }
-
-    while let Some(chunk) = handle.next_chunk().await? {
-        let chunk_span = chunk.location.time_span;
-        let span = ExtractionSpan::new(chunk.data, AudioLocation::new(chunk_span));
-        let output = stt_arc.extract(&span).await?;
-        doc.meta.extraction = output.extraction;
-        let transcription = output.value;
-
-        if transcription.segments.is_empty() {
-            tracing::debug!(target: TARGET, "transcription returned no segments");
-            continue;
-        }
-        for seg in transcription.segments {
-            doc.blocks.push(Block::new(AudioBlock::Speech {
-                time_span: seg.time_span,
-                text: seg.text,
-                speaker_id: seg.speaker_id,
-            }));
-        }
-    }
-
-    tracing::debug!(target: TARGET, "audio extraction complete");
-    Ok(())
-}
-
-#[cfg(not(feature = "audio"))]
-async fn populate_audio_doc(
-    _engine: &ExtractorRegistry,
-    _doc: &mut Document<Audio>,
-    _handle: &mut dyn nvisy_codec::Handler<Audio>,
-    _descriptor: &ContentDescriptor,
-    _plan: &Extraction,
-) -> Result<()> {
-    Ok(())
-}
diff --git a/crates/nvisy-engine/src/detection/phases/mod.rs b/crates/nvisy-engine/src/detection/phases/mod.rs
deleted file mode 100644
index dac8700f7..000000000
--- a/crates/nvisy-engine/src/detection/phases/mod.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-//! Detection-side per-document phase orchestrators.
-//!
-//! Each phase is a document-walking driver around its toolkit-side
-//! subsystem (extractor registry, recognizer registry, dedup
-//! layers). Phases own the per-document state, the sequencing
-//! through `Document<M>` blocks / nodes, and the conversion
-//! between toolkit-shaped library calls and the typed document
-//! audit.
-//!
-//! - [`ExtractionPhase`] runs phase 1: pulls chunks through the
-//!   codec, runs the matching toolkit-side extractor (OCR / STT
-//!   for image / audio), and writes per-modality `Block`s onto
-//!   each `Document<M>`.
-//! - [`DetectionPhase`] runs phase 2: walks each `Document<M>`'s
-//!   blocks, dispatches text + image content through the
-//!   per-request `RecognizerRegistry`, and appends detected
-//!   entities to the document's audit.
-//! - [`DeduplicationPhase`] runs phase 3: merges co-referent
-//!   detections across recognizers using the toolkit's
-//!   [`LayerPipeline`].
-//!
-//! [`LayerPipeline`]: nvisy_toolkit::deduplication::LayerPipeline
-
-pub mod deduplication;
-pub mod detection;
-pub mod extraction;
-
-pub use self::deduplication::DeduplicationPhase;
-pub use self::detection::DetectionPhase;
-pub use self::extraction::ExtractionPhase;
diff --git a/crates/nvisy-engine/src/detection/pipeline.rs b/crates/nvisy-engine/src/detection/pipeline.rs
deleted file mode 100644
index 6ff0a9dd1..000000000
--- a/crates/nvisy-engine/src/detection/pipeline.rs
+++ /dev/null
@@ -1,241 +0,0 @@
-//! Per-pass detection pipeline lifecycle.
-//!
-//! [`DetectionPipeline`] is created per [`DetectionEngine::detect`] call.
-//! Owns the pass id, registry handle, run context, and
-//! [`DetectionState`] accessors.
-//!
-//! [`DetectionEngine::detect`]: crate::detection::DetectionEngine::detect
-
-use std::sync::Arc;
-
-use jiff::Timestamp;
-use nvisy_codec::CodecRegistry;
-use nvisy_core::Error;
-use nvisy_core::entity::EntityLabelCatalog;
-use nvisy_toolkit::extraction::ExtractorRegistry;
-use tokio_util::sync::CancellationToken;
-use uuid::Uuid;
-
-use super::input::DetectionInput;
-use super::orchestrator::DetectionOrchestrator;
-use super::result::DetectionResult;
-use super::state::{DetectionRecord, DetectionState};
-use super::status::DetectionStatus;
-use crate::core::ingestion::ImportFile;
-use crate::core::ingestion::encryption::SharedKeyProvider;
-use crate::core::{DetectionContext, DetectionEngines, PolicyStore, RuntimeConfig, SharedData};
-use crate::detection::DetectionConfig;
-use crate::document::provenance::AnyAudit;
-use crate::policy::{Policy, PolicyDigest};
-use crate::registry::Registry;
-
-const TARGET: &str = "nvisy_engine::pipeline::detection::pipeline";
-
-/// Pre-built engine resources the detection pipeline borrows for
-/// the duration of one pass.
-pub(crate) struct DetectionEngineState {
-    pub extraction_engine: Arc<ExtractorRegistry>,
-    pub detection_config: Arc<DetectionConfig>,
-}
-
-/// A single detection-pass lifecycle.
-pub(crate) struct DetectionPipeline {
-    detection_id: Uuid,
-    registry: Registry,
-    key_provider: Option<SharedKeyProvider>,
-    detections: DetectionState,
-    base_config: RuntimeConfig,
-    state: DetectionEngineState,
-}
-
-impl DetectionPipeline {
-    pub(crate) fn new(
-        registry: Registry,
-        key_provider: Option<SharedKeyProvider>,
-        detections: DetectionState,
-        base_config: RuntimeConfig,
-        state: DetectionEngineState,
-    ) -> Self {
-        Self {
-            detection_id: Uuid::now_v7(),
-            registry,
-            key_provider,
-            detections,
-            base_config,
-            state,
-        }
-    }
-
-    pub(crate) fn id(&self) -> Uuid {
-        self.detection_id
-    }
-
-    /// Validate the submitted request, build the per-pass
-    /// [`PolicyStore`] + digests, and register a pending detection
-    /// record.
-    ///
-    /// Validation chains the three [`DetectionInput`] checks:
-    /// policy/rule name uniqueness, label union with conflict
-    /// detection, and selector-label resolution against the
-    /// unioned catalog.
-    ///
-    /// # Errors
-    ///
-    /// Returns the first validation error raised by any of the
-    /// three checks.
-    pub(crate) async fn register_pending(
-        &self,
-        input: DetectionInput,
-    ) -> Result<PreparedDetection, Error> {
-        input.validate_namespace()?;
-        let catalog = input.unify_labels()?;
-        input.validate_selector_labels(&catalog)?;
-        input.validate_actions()?;
-
-        let actor_id = input.actor_id;
-        let policy_digests: Vec<PolicyDigest> = input.policies.iter().map(Policy::digest).collect();
-        let policies = Arc::new(PolicyStore::from_policies(input.policies));
-
-        self.detections
-            .insert(
-                self.detection_id,
-                DetectionRecord {
-                    actor_id,
-                    policies: Arc::clone(&policies),
-                    policy_digests: policy_digests.clone(),
-                    imports: input.imports.clone(),
-                    status: DetectionStatus::Pending,
-                    created_at: Timestamp::now(),
-                    started_at: None,
-                    completed_at: None,
-                    cancel: CancellationToken::new(),
-                    audits: Vec::new(),
-                    entities_detected: 0,
-                    error: None,
-                },
-            )
-            .await;
-
-        Ok(PreparedDetection {
-            actor_id,
-            policies,
-            policy_digests,
-            catalog,
-            imports: input.imports,
-            plan: input.plan,
-        })
-    }
-
-    /// Run the detection pass to completion.
-    pub(crate) async fn execute(
-        &self,
-        prepared: PreparedDetection,
-    ) -> Result<(Vec<AnyAudit>, u64, DetectionStatus), Error> {
-        let actor_id = prepared.actor_id;
-
-        let recognizer_registry = match self
-            .state
-            .detection_config
-            .build_for_request(&prepared.catalog)
-        {
-            Ok(r) => Arc::new(r),
-            Err(e) => {
-                self.detections.fail(self.detection_id, e.to_string()).await;
-                return Err(e);
-            }
-        };
-
-        let mut shared_data = SharedData {
-            run_id: self.detection_id,
-            actor_id,
-            policies: prepared.policies,
-            catalog: Arc::new(prepared.catalog),
-            registry: self.registry.clone(),
-            codec_registry: CodecRegistry::with_builtin(),
-            key_provider: SharedKeyProvider::default(),
-        };
-        if let Some(ref kp) = self.key_provider {
-            shared_data.key_provider = kp.clone();
-        }
-
-        let cancel = CancellationToken::new();
-        let engines = DetectionEngines {
-            extraction_engine: (*self.state.extraction_engine).clone(),
-            recognizer_registry,
-        };
-        let concurrency = self.base_config.effective_concurrency();
-        let ctx = DetectionContext::new(cancel, Arc::new(shared_data), engines, concurrency);
-
-        self.detections.set_started_at(self.detection_id).await;
-
-        let orchestrator = DetectionOrchestrator::new(ctx);
-        let output = match orchestrator.run(&prepared.imports, &prepared.plan).await {
-            Ok(out) => out,
-            Err(e) => {
-                self.detections.fail(self.detection_id, e.to_string()).await;
-                return Err(e);
-            }
-        };
-
-        let (audits, entities_detected, any_ok, any_failed) = output.into_audits();
-        let status = if !any_failed {
-            DetectionStatus::Succeeded
-        } else if any_ok {
-            DetectionStatus::PartialFailure
-        } else {
-            DetectionStatus::Failed
-        };
-
-        if matches!(
-            status,
-            DetectionStatus::Succeeded | DetectionStatus::PartialFailure
-        ) {
-            let result = DetectionResult {
-                id: self.detection_id,
-                actor_id,
-                policies: prepared.policy_digests,
-                imports: prepared.imports.clone(),
-                audits: audits.clone(),
-                entities_detected,
-            };
-            if let Err(e) = self
-                .registry
-                .store_detection(actor_id, self.detection_id, &result)
-                .await
-            {
-                tracing::error!(
-                    target: TARGET,
-                    detection_id = %self.detection_id,
-                    error = %e,
-                    "failed to persist detection result; in-memory state stays valid",
-                );
-            }
-        }
-
-        self.detections
-            .finalize(self.detection_id, status, audits.clone(), entities_detected)
-            .await;
-
-        tracing::info!(
-            target: TARGET,
-            detection_id = %self.detection_id,
-            ?status,
-            entities_detected,
-            "detection pass finalized",
-        );
-
-        Ok((audits, entities_detected, status))
-    }
-}
-
-/// Output of [`DetectionPipeline::register_pending`] — the
-/// per-pass state that has to flow to [`DetectionPipeline::execute`]
-/// without re-touching the original `Vec<Policy>`.
-pub(crate) struct PreparedDetection {
-    actor_id: Uuid,
-    policies: Arc<PolicyStore>,
-    policy_digests: Vec<PolicyDigest>,
-    catalog: EntityLabelCatalog,
-    imports: Vec<ImportFile>,
-    plan: crate::detection::DetectionPlan,
-}
diff --git a/crates/nvisy-engine/src/detection/plan.rs b/crates/nvisy-engine/src/detection/plan.rs
deleted file mode 100644
index 77777a811..000000000
--- a/crates/nvisy-engine/src/detection/plan.rs
+++ /dev/null
@@ -1,85 +0,0 @@
-//! Per-pass plan and per-modality knobs the detection pipeline
-//! reads against each imported document.
-//!
-//! Three concerns share this file:
-//!
-//! - [`DetectionPlan`] — top-level per-request bundle the
-//!   detection pipeline reads once per document.
-//! - [`Extraction`] + per-modality plans ([`TextPlan`],
-//!   [`TabularPlan`], [`ImagePlan`], [`AudioPlan`]) — the
-//!   `Extraction` aggregate carries one plan struct per modality;
-//!   the extraction phase dispatches `&plan.extraction.<modality>`
-//!   to the matching backend.
-//! - [`DeduplicationParams`] — re-export of the toolkit
-//!   `LayerParams` type the deduplication phase consumes.
-
-pub use nvisy_toolkit::deduplication::LayerParams as DeduplicationParams;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Per-request bundle of detection-side phase configs.
-///
-/// The detection pipeline reads this once per document and routes
-/// each phase (extraction, then detection, then deduplication) to
-/// the matching field. Detection itself has no plan node — its
-/// per-request behaviour is driven by the policy-supplied label
-/// catalog rather than a config object.
-#[derive(Debug, Clone, Default)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase", default)]
-pub struct DetectionPlan {
-    /// Extraction settings per modality.
-    pub extraction: Extraction,
-    /// Deduplication settings applied to combined detection results.
-    pub deduplication: DeduplicationParams,
-}
-
-/// Unified extraction plan.
-///
-/// Carries one per-modality plan struct per modality. The
-/// orchestrator dispatches `&plan.extraction.<modality>` to the
-/// matching `ExtractDispatch<M>` impl.
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Extraction {
-    /// Text-modality plan.
-    #[serde(default)]
-    pub text: TextPlan,
-    /// Tabular-modality plan.
-    #[serde(default)]
-    pub tabular: TabularPlan,
-    /// Image-modality plan (OCR).
-    #[serde(default)]
-    pub image: ImagePlan,
-    /// Audio-modality plan (STT + diarization).
-    #[serde(default)]
-    pub audio: AudioPlan,
-}
-
-/// Text-modality plan knobs. No tunables today; reserved for future
-/// per-call settings (e.g. whitespace normalization).
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct TextPlan {}
-
-/// Tabular-modality plan knobs. No tunables today.
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct TabularPlan {}
-
-/// Image-modality plan knobs. No tunables today; reserved for
-/// future OCR tuning (e.g. language hint, page subset).
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct ImagePlan {}
-
-/// Audio-modality plan knobs (speech-to-text).
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct AudioPlan {
-    /// Segment the audio by speaker identity.
-    ///
-    /// Currently silently degrades to flat transcription — see #239.
-    #[serde(default)]
-    pub diarization: bool,
-}
diff --git a/crates/nvisy-engine/src/detection/result.rs b/crates/nvisy-engine/src/detection/result.rs
deleted file mode 100644
index a2d832e74..000000000
--- a/crates/nvisy-engine/src/detection/result.rs
+++ /dev/null
@@ -1,119 +0,0 @@
-//! [`DetectionResult`] and friends: the artifact a detection pass
-//! produces plus the lightweight listing/filter shapes.
-
-use jiff::Timestamp;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-use super::status::DetectionStatus;
-use crate::core::ingestion::ImportFile;
-use crate::document::provenance::AnyAudit;
-use crate::policy::PolicyDigest;
-
-/// Immutable artifact produced by one detection pass.
-///
-/// Contains the per-document audits with `Execution::Pending`
-/// decisions (so the caller can see what would be redacted under
-/// the policy chain) plus the original [`ImportFile`] references —
-/// a later [`RedactionEngine::redact`] call needs them to re-open content
-/// for byte rewriting.
-///
-/// [`RedactionEngine::redact`]: crate::redaction::RedactionEngine::redact
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct DetectionResult {
-    /// Unique identifier for this detection pass.
-    pub id: Uuid,
-    /// Identity of the actor who initiated the pass.
-    pub actor_id: Uuid,
-    /// Header cards for the policies that were applied, in
-    /// precedence order. The full rule bodies live only in the
-    /// original [`DetectionInput::policies`] submission; the
-    /// digests here let a reader render `"<name> v<version>"` and
-    /// correlate audit entries by [`PolicyDigest::name`] without
-    /// re-uploading anything.
-    ///
-    /// [`DetectionInput::policies`]: super::DetectionInput::policies
-    pub policies: Vec<PolicyDigest>,
-    /// Original import references — needed by [`RedactionEngine::redact`].
-    ///
-    /// [`RedactionEngine::redact`]: crate::redaction::RedactionEngine::redact
-    pub imports: Vec<ImportFile>,
-    /// Per-document audit trails. One entry per modality the
-    /// pipeline produced.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub audits: Vec<AnyAudit>,
-    /// Total entities detected across all documents.
-    pub entities_detected: u64,
-}
-
-/// Full point-in-time snapshot of a detection pass.
-///
-/// `DetectionResult` is only present once the pass has reached a
-/// terminal state with at least one successfully processed
-/// document. While running, the caller still gets timestamps and
-/// status.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct DetectionSnapshot {
-    /// Unique identifier for this detection pass.
-    pub id: Uuid,
-    /// Identity of the actor who initiated the pass.
-    pub actor_id: Uuid,
-    /// Current lifecycle state.
-    pub status: DetectionStatus,
-    /// When the pass was created.
-    #[schemars(with = "String")]
-    pub created_at: Timestamp,
-    /// When recognisers actually started executing.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub started_at: Option<Timestamp>,
-    /// When the pass reached a terminal state.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub completed_at: Option<Timestamp>,
-    /// Detection result. `Some` once the pass has produced at
-    /// least one audit (terminal status with `Succeeded` /
-    /// `PartialFailure`).
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub result: Option<DetectionResult>,
-    /// Error description for failed passes.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub error: Option<String>,
-}
-
-/// Lightweight summary of a detection pass for listing endpoints.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct DetectionEntry {
-    /// Unique identifier for this detection pass.
-    pub id: Uuid,
-    /// Identity of the actor who initiated the pass.
-    pub actor_id: Uuid,
-    /// Current overall status.
-    pub status: DetectionStatus,
-    /// When the pass was created.
-    #[schemars(with = "String")]
-    pub created_at: Timestamp,
-    /// When recognisers started, if applicable.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub started_at: Option<Timestamp>,
-    /// When the pass finished, if applicable.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub completed_at: Option<Timestamp>,
-    /// Number of imports processed.
-    pub import_count: usize,
-    /// Total entities detected across all documents.
-    pub entities_detected: u64,
-}
-
-/// Filter criteria for listing detection passes.
-#[derive(Debug, Clone, Default)]
-pub struct DetectionFilter {
-    /// If set, only return passes with this status.
-    pub status: Option<DetectionStatus>,
-}
diff --git a/crates/nvisy-engine/src/detection/state.rs b/crates/nvisy-engine/src/detection/state.rs
deleted file mode 100644
index e89651468..000000000
--- a/crates/nvisy-engine/src/detection/state.rs
+++ /dev/null
@@ -1,355 +0,0 @@
-//! Volatile in-memory state for active detection passes.
-//!
-//! Analog of the deleted `RunState`. Holds per-pass timestamps,
-//! status, cancellation token, and a running audit accumulator
-//! that's frozen into a [`DetectionResult`] once the pass reaches
-//! a terminal state.
-
-use std::cmp::Reverse;
-use std::collections::HashMap;
-use std::sync::Arc;
-
-use jiff::Timestamp;
-use nvisy_core::Error;
-use tokio::sync::RwLock;
-use tokio_util::sync::CancellationToken;
-use uuid::Uuid;
-
-use super::result::{DetectionEntry, DetectionFilter, DetectionResult, DetectionSnapshot};
-use super::status::DetectionStatus;
-use crate::core::PolicyStore;
-use crate::core::ingestion::ImportFile;
-use crate::document::provenance::AnyAudit;
-use crate::policy::PolicyDigest;
-
-const TARGET: &str = "nvisy_engine::detection::state";
-
-/// In-memory detection-pass tracker backed by `Arc<RwLock<HashMap>>`.
-///
-/// Cheaply clonable (Arc bump). Both the [`DetectionEngine`][de] and
-/// a paired [`RedactionEngine`][re] (via
-/// [`RedactionEngine::from_detection`][rfd]) hold clones — the
-/// redaction side reads it for the detect→redact handoff path,
-/// avoiding a registry round-trip.
-///
-/// [de]: super::DetectionEngine
-/// [re]: super::super::redaction::RedactionEngine
-/// [rfd]: super::super::redaction::RedactionEngine::from_detection
-#[derive(Clone, Default)]
-pub struct DetectionState {
-    inner: Arc<RwLock<HashMap<Uuid, DetectionRecord>>>,
-}
-
-/// Internal handoff shape the redaction pipeline reads from the
-/// detection record. Carries the shared per-modality policy chains
-/// the redact phase needs to re-evaluate operator specs, alongside
-/// the per-document imports and the pending audits. Lives only
-/// between detection-state lookup and redaction-pipeline
-/// construction; never serialised.
-pub(crate) struct DetectionHandoff {
-    pub policies: Arc<PolicyStore>,
-    pub imports: Vec<ImportFile>,
-    pub audits: Vec<AnyAudit>,
-}
-
-/// Mutable per-pass state.
-pub(crate) struct DetectionRecord {
-    pub actor_id: Uuid,
-    /// Shared per-modality policy chains built once from the
-    /// submitted inline bodies. Detection and redaction pipelines
-    /// both `Arc::clone` this — no deep policy clones after
-    /// submission.
-    pub policies: Arc<PolicyStore>,
-    /// Header cards for the policies that ran, computed once at
-    /// submission time. Surfaced verbatim in [`DetectionResult`].
-    pub policy_digests: Vec<PolicyDigest>,
-    pub imports: Vec<ImportFile>,
-    pub status: DetectionStatus,
-    pub created_at: Timestamp,
-    pub started_at: Option<Timestamp>,
-    pub completed_at: Option<Timestamp>,
-    pub cancel: CancellationToken,
-    pub audits: Vec<AnyAudit>,
-    pub entities_detected: u64,
-    pub error: Option<String>,
-}
-
-impl DetectionRecord {
-    fn to_snapshot(&self, id: Uuid) -> DetectionSnapshot {
-        let result = match self.status {
-            DetectionStatus::Succeeded | DetectionStatus::PartialFailure => Some(DetectionResult {
-                id,
-                actor_id: self.actor_id,
-                policies: self.policy_digests.clone(),
-                imports: self.imports.clone(),
-                audits: self.audits.clone(),
-                entities_detected: self.entities_detected,
-            }),
-            _ => None,
-        };
-        DetectionSnapshot {
-            id,
-            actor_id: self.actor_id,
-            status: self.status,
-            created_at: self.created_at,
-            started_at: self.started_at,
-            completed_at: self.completed_at,
-            result,
-            error: self.error.clone(),
-        }
-    }
-
-    fn to_entry(&self, id: Uuid) -> DetectionEntry {
-        DetectionEntry {
-            id,
-            actor_id: self.actor_id,
-            status: self.status,
-            created_at: self.created_at,
-            started_at: self.started_at,
-            completed_at: self.completed_at,
-            import_count: self.imports.len(),
-            entities_detected: self.entities_detected,
-        }
-    }
-}
-
-impl DetectionState {
-    pub(crate) fn new() -> Self {
-        Self::default()
-    }
-
-    pub(crate) async fn insert(&self, id: Uuid, record: DetectionRecord) {
-        self.inner.write().await.insert(id, record);
-    }
-
-    /// Look up a snapshot scoped by actor.
-    ///
-    /// # Errors
-    ///
-    /// [`ErrorKind::NotFound`] when the detection does not exist
-    /// or belongs to a different actor (the two cases share an
-    /// error kind to avoid leaking existence to unauthorised
-    /// callers).
-    ///
-    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
-    pub(crate) async fn snapshot(
-        &self,
-        actor_id: Uuid,
-        id: Uuid,
-    ) -> Result<DetectionSnapshot, Error> {
-        let guard = self.inner.read().await;
-        let Some(record) = guard.get(&id) else {
-            return Err(Error::not_found(
-                format!("detection {id} not found"),
-                TARGET,
-            ));
-        };
-        if record.actor_id != actor_id {
-            return Err(Error::not_found(
-                format!("detection {id} not found"),
-                TARGET,
-            ));
-        }
-        Ok(record.to_snapshot(id))
-    }
-
-    /// Resolve a [`DetectionResult`] for redaction.
-    ///
-    /// # Errors
-    ///
-    /// - [`ErrorKind::NotFound`] when the detection does not
-    ///   exist or belongs to a different actor. The two cases
-    ///   share an error kind on purpose — actor-scoping leaks
-    ///   information if "exists for other actor" is
-    ///   distinguishable from "does not exist."
-    /// - [`ErrorKind::Conflict`] when the detection exists for
-    ///   this actor but is not in a terminal succeeded /
-    ///   partial-failure state. Includes the current status in
-    ///   the message so the caller can poll.
-    ///
-    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
-    /// [`ErrorKind::Conflict`]: nvisy_core::ErrorKind::Conflict
-    /// Internal hand-off shape used by the redaction pipeline: full
-    /// inline policy bodies plus the rest of the detection record's
-    /// payload. Replaces the public-facing `DetectionResult` for the
-    /// detect→redact wiring so the redact phase can re-evaluate
-    /// operator specs against the actual rule bodies.
-    pub(crate) async fn handoff(
-        &self,
-        actor_id: Uuid,
-        id: Uuid,
-    ) -> Result<DetectionHandoff, Error> {
-        let guard = self.inner.read().await;
-        let Some(record) = guard.get(&id) else {
-            return Err(Error::not_found(
-                format!("detection {id} not found"),
-                TARGET,
-            ));
-        };
-        if record.actor_id != actor_id {
-            return Err(Error::not_found(
-                format!("detection {id} not found"),
-                TARGET,
-            ));
-        }
-        if !matches!(
-            record.status,
-            DetectionStatus::Succeeded | DetectionStatus::PartialFailure
-        ) {
-            return Err(Error::conflict(
-                format!(
-                    "detection {id} not ready for redaction (status: {:?})",
-                    record.status
-                ),
-                TARGET,
-            ));
-        }
-        Ok(DetectionHandoff {
-            policies: record.policies.clone(),
-            imports: record.imports.clone(),
-            audits: record.audits.clone(),
-        })
-    }
-
-    pub(crate) async fn list(
-        &self,
-        actor_id: Uuid,
-        filter: DetectionFilter,
-    ) -> Vec<DetectionEntry> {
-        let guard = self.inner.read().await;
-        let mut out: Vec<DetectionEntry> = guard
-            .iter()
-            .filter(|(_, r)| r.actor_id == actor_id)
-            .filter(|(_, r)| filter.status.is_none_or(|s| r.status == s))
-            .map(|(id, r)| r.to_entry(*id))
-            .collect();
-        out.sort_by_key(|e| Reverse(e.created_at));
-        out
-    }
-
-    pub(crate) async fn set_started_at(&self, id: Uuid) {
-        let mut guard = self.inner.write().await;
-        if let Some(record) = guard.get_mut(&id)
-            && record.started_at.is_none()
-        {
-            record.started_at = Some(Timestamp::now());
-            record.status = DetectionStatus::Running;
-        }
-    }
-
-    pub(crate) async fn fail(&self, id: Uuid, error: impl Into<String>) {
-        let mut guard = self.inner.write().await;
-        if let Some(record) = guard.get_mut(&id) {
-            record.status = DetectionStatus::Failed;
-            record.completed_at = Some(Timestamp::now());
-            record.error = Some(error.into());
-        }
-    }
-
-    pub(crate) async fn finalize(
-        &self,
-        id: Uuid,
-        status: DetectionStatus,
-        audits: Vec<AnyAudit>,
-        entities_detected: u64,
-    ) {
-        let mut guard = self.inner.write().await;
-        if let Some(record) = guard.get_mut(&id) {
-            record.status = status;
-            record.completed_at = Some(Timestamp::now());
-            record.audits = audits;
-            record.entities_detected = entities_detected;
-        }
-    }
-
-    /// Cancel an in-progress detection.
-    ///
-    /// # Errors
-    ///
-    /// - [`ErrorKind::NotFound`] when the detection does not
-    ///   exist or belongs to a different actor.
-    /// - [`ErrorKind::Conflict`] when the detection is already
-    ///   in a terminal state.
-    ///
-    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
-    /// [`ErrorKind::Conflict`]: nvisy_core::ErrorKind::Conflict
-    pub(crate) async fn cancel(&self, actor_id: Uuid, id: Uuid) -> Result<(), Error> {
-        let mut guard = self.inner.write().await;
-        let Some(record) = guard.get_mut(&id) else {
-            return Err(Error::not_found(
-                format!("detection {id} not found"),
-                TARGET,
-            ));
-        };
-        if record.actor_id != actor_id {
-            return Err(Error::not_found(
-                format!("detection {id} not found"),
-                TARGET,
-            ));
-        }
-        if record.status.is_terminal() {
-            return Err(Error::conflict(
-                format!(
-                    "detection {id} already in terminal state {:?}",
-                    record.status
-                ),
-                TARGET,
-            ));
-        }
-        record.cancel.cancel();
-        record.status = DetectionStatus::Cancelled;
-        record.completed_at = Some(Timestamp::now());
-        Ok(())
-    }
-
-    /// Delete a finished detection.
-    ///
-    /// # Errors
-    ///
-    /// - [`ErrorKind::NotFound`] when the detection does not
-    ///   exist or belongs to a different actor.
-    /// - [`ErrorKind::Conflict`] when the detection is still
-    ///   active.
-    ///
-    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
-    /// [`ErrorKind::Conflict`]: nvisy_core::ErrorKind::Conflict
-    pub(crate) async fn delete(&self, actor_id: Uuid, id: Uuid) -> Result<(), Error> {
-        let mut guard = self.inner.write().await;
-        let Some(record) = guard.get(&id) else {
-            return Err(Error::not_found(
-                format!("detection {id} not found"),
-                TARGET,
-            ));
-        };
-        if record.actor_id != actor_id {
-            return Err(Error::not_found(
-                format!("detection {id} not found"),
-                TARGET,
-            ));
-        }
-        if !record.status.is_terminal() {
-            return Err(Error::conflict(
-                format!("detection {id} cannot be deleted while {:?}", record.status),
-                TARGET,
-            ));
-        }
-        guard.remove(&id);
-        Ok(())
-    }
-
-    /// Delete every terminal-status detection for `actor_id`.
-    /// Returns the ids removed so the caller can cascade the
-    /// delete to other stores (registry persistence).
-    pub(crate) async fn delete_all(&self, actor_id: Uuid) -> Vec<Uuid> {
-        let mut guard = self.inner.write().await;
-        let to_delete: Vec<Uuid> = guard
-            .iter()
-            .filter(|(_, r)| r.actor_id == actor_id && r.status.is_terminal())
-            .map(|(id, _)| *id)
-            .collect();
-        for id in &to_delete {
-            guard.remove(id);
-        }
-        to_delete
-    }
-}
diff --git a/crates/nvisy-engine/src/detection/status.rs b/crates/nvisy-engine/src/detection/status.rs
deleted file mode 100644
index 413ee66bd..000000000
--- a/crates/nvisy-engine/src/detection/status.rs
+++ /dev/null
@@ -1,37 +0,0 @@
-//! [`DetectionStatus`]: lifecycle states for a detection pass.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Lifecycle states for a single detection pass.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum DetectionStatus {
-    /// Submitted, not yet started.
-    Pending,
-    /// Imports loaded; recognisers / evaluators running.
-    Running,
-    /// All documents processed without error.
-    Succeeded,
-    /// Some documents detected entities; others errored.
-    PartialFailure,
-    /// Every document errored or the pass failed before any
-    /// document was processed.
-    Failed,
-    /// The caller cancelled the detection pass before it finished.
-    Cancelled,
-}
-
-impl DetectionStatus {
-    /// `true` for [`Self::Succeeded`] / [`Self::PartialFailure`] /
-    /// [`Self::Failed`] / [`Self::Cancelled`] — anything past the
-    /// running phase.
-    #[must_use]
-    pub fn is_terminal(self) -> bool {
-        matches!(
-            self,
-            Self::Succeeded | Self::PartialFailure | Self::Failed | Self::Cancelled
-        )
-    }
-}
diff --git a/crates/nvisy-engine/src/document/annotations.rs b/crates/nvisy-engine/src/document/annotations.rs
deleted file mode 100644
index 5423bd226..000000000
--- a/crates/nvisy-engine/src/document/annotations.rs
+++ /dev/null
@@ -1,45 +0,0 @@
-//! Per-content user-supplied annotations, bucketed per modality.
-
-use nvisy_core::entity::{Annotation, LabelAnnotation};
-use nvisy_core::modality::{Audio, Image, Tabular, Text};
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Per-modality buckets of user-supplied annotations on a piece of
-/// content.
-///
-/// Each modality-typed [`Annotation<M>`] targets a `Document<M>`
-/// envelope of the same modality; document-level
-/// [`LabelAnnotation`]s apply to every envelope spawned from the
-/// source.
-#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct AnyAnnotations {
-    /// Annotations targeting text-modality content.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub text: Vec<Annotation<Text>>,
-    /// Annotations targeting tabular-modality content.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub tabular: Vec<Annotation<Tabular>>,
-    /// Annotations targeting image-modality content.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub image: Vec<Annotation<Image>>,
-    /// Annotations targeting audio-modality content.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub audio: Vec<Annotation<Audio>>,
-    /// Document-level classification labels.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub labels: Vec<LabelAnnotation>,
-}
-
-impl AnyAnnotations {
-    /// `true` when every bucket is empty.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.text.is_empty()
-            && self.tabular.is_empty()
-            && self.image.is_empty()
-            && self.audio.is_empty()
-            && self.labels.is_empty()
-    }
-}
diff --git a/crates/nvisy-engine/src/document/block.rs b/crates/nvisy-engine/src/document/block.rs
deleted file mode 100644
index 14cccd443..000000000
--- a/crates/nvisy-engine/src/document/block.rs
+++ /dev/null
@@ -1,56 +0,0 @@
-//! [`Block`] — universal wrapper around a per-modality block payload.
-
-use nvisy_core::primitive::Confidence;
-
-use super::Span;
-use crate::modality::DocumentModality;
-
-/// One block of a [`Document<M>`].
-///
-/// Universal across modalities: `kind` carries the modality-specific
-/// payload (variant + its location-specific data) via `M::Block`,
-/// while `spans` and `confidence` are the common per-block
-/// bookkeeping.
-///
-/// `spans` is empty for blocks that don't carry text (e.g. an image
-/// `Figure` or `Logo`).
-///
-/// Detected entities don't live on the block — they're a run-scoped
-/// finding and live on the engine's [`DocumentEnvelope`], not on the
-/// structural document.
-///
-/// [`Document<M>`]: super::Document
-/// [`DocumentEnvelope`]: # "predecessor of DocumentTree"
-#[derive(Debug, Clone, PartialEq)]
-pub struct Block<M: DocumentModality> {
-    /// Modality-specific payload (variant + its data).
-    pub kind: M::Block,
-    /// Source-mapped spans into the block's text. Carries the
-    /// `block-local offset -> M::Location` map every modality uses
-    /// to lift recognizer findings (which run on the block's text
-    /// string) back to source coordinates. Empty only for layout-
-    /// only variants that have no text (e.g. `ImageBlock::Figure`,
-    /// `ImageBlock::Logo`).
-    pub spans: Vec<Span<M>>,
-    /// Recognition confidence for the block as a whole. Absent for
-    /// native text-layer extraction where the source already provides
-    /// the text directly.
-    pub confidence: Option<Confidence>,
-}
-
-impl<M: DocumentModality> Block<M> {
-    /// Construct a new block with empty spans and no confidence.
-    pub fn new(kind: M::Block) -> Self {
-        Self {
-            kind,
-            spans: Vec::new(),
-            confidence: None,
-        }
-    }
-
-    /// Set the source-mapped spans (builder-style).
-    pub fn with_spans(mut self, spans: Vec<Span<M>>) -> Self {
-        self.spans = spans;
-        self
-    }
-}
diff --git a/crates/nvisy-engine/src/document/mod.rs b/crates/nvisy-engine/src/document/mod.rs
deleted file mode 100644
index e1c15ab10..000000000
--- a/crates/nvisy-engine/src/document/mod.rs
+++ /dev/null
@@ -1,107 +0,0 @@
-//! [`Document`] — the unified addressable view of any processed
-//! document, parameterised by modality.
-//!
-//! Native text (PDF text layers, DOCX runs, plain text), recognized
-//! text (OCR'd images, transcribed audio), and tabular cells all
-//! flow into the same shape: a [`Document<M>`] holding ordered
-//! [`Block<M>`]s plus user annotations and the run-scoped
-//! [`Audit<M>`] that accrues detected entities and redaction entries
-//! as the pipeline progresses.
-//!
-//! `Block<M>` is the universal wrapper carrying the common per-block
-//! fields (spans, confidence). The modality-specific payload
-//! (text+spans, region, time span, row coordinates) lives in
-//! `M::Block` inside `block.kind`.
-//!
-//! Rich sources (PDFs with both text and image layers) decompose
-//! into a single recursive `Document<Text>` whose blocks can host
-//! nested `Document<Image>` / `Document<Tabular>` children via
-//! [`TextBlock::Embed`]. Each nested document carries its own typed
-//! audit; the per-source audit is the tree of nested audits, with
-//! the wire shape ([`AnyAudit`]) being one projection of it.
-//!
-//! `Document` is an in-memory pipeline carrier — it is intentionally
-//! not `Serialize`/`Deserialize` and has no `Default`.
-//!
-//! [`Audit<M>`]: self::provenance::Audit
-//! [`AnyAudit`]: self::provenance::AnyAudit
-//! [`TextBlock::Embed`]: crate::modality::TextBlock::Embed
-
-mod annotations;
-mod block;
-pub mod provenance;
-mod span;
-
-use nvisy_core::entity::{Annotation, ContentSource, Entity, LabelAnnotation};
-
-pub use self::annotations::AnyAnnotations;
-pub use self::block::Block;
-use self::provenance::Audit;
-pub use self::span::Span;
-use crate::modality::DocumentModality;
-
-/// Unified addressable view of a parsed document for modality `M`.
-///
-/// Combines structural content (blocks, annotations, labels) with
-/// the run-scoped [`Audit<M>`] that accumulates detected entities and
-/// redaction records as the pipeline progresses. The audit lives
-/// here so the tree-shaped recursive document model (a
-/// `Document<Text>` whose blocks embed `Document<Image>` /
-/// `Document<Tabular>`) keeps every nested doc's audit attached to
-/// the doc it describes — the per-source audit is the tree of these.
-///
-/// [`Audit<M>`]: self::provenance::Audit
-#[derive(Debug, Clone)]
-pub struct Document<M: DocumentModality> {
-    /// Per-modality document-level metadata.
-    pub meta: M::Metadata,
-    /// Ordered blocks. One per page, paragraph, speaker turn, row,
-    /// or just one for documents with no inherent block structure.
-    pub blocks: Vec<Block<M>>,
-
-    /// User-supplied region annotations (inclusions and exclusions)
-    /// attached at upload time. Annotation locations target source
-    /// coordinates within the document.
-    pub annotations: Vec<Annotation<M>>,
-    /// Document-level classification labels. Modality-agnostic and
-    /// propagated to every envelope spawned from the same source so
-    /// policy rules that condition on labels can fire uniformly.
-    pub labels: Vec<LabelAnnotation>,
-
-    /// Run-scoped provenance for this document: detected entities
-    /// and per-redaction audit entries. Opens empty against the
-    /// document's source at construction; each pipeline phase
-    /// (detection, redaction, validation) reads from and writes to
-    /// it as it executes.
-    pub audit: Audit<M>,
-}
-
-impl<M: DocumentModality> Document<M> {
-    /// Construct an empty [`Document`] with explicit metadata and a
-    /// fresh audit opened against `source`. Blocks, annotations, and
-    /// labels start empty; producers push onto them directly.
-    ///
-    /// The importer always knows which extraction path produced the
-    /// document, so the metadata is required at construction time
-    /// rather than defaulted.
-    pub fn new(meta: M::Metadata, source: ContentSource) -> Self {
-        Self {
-            meta,
-            blocks: Vec::new(),
-            annotations: Vec::new(),
-            labels: Vec::new(),
-            audit: Audit::new(source),
-        }
-    }
-
-    /// Append detected entities to the document's audit, wrapping
-    /// each into a fresh [`EntityRecord`]. Used by detection phases
-    /// at both the root and nested-document level.
-    ///
-    /// [`EntityRecord`]: self::provenance::EntityRecord
-    pub fn add_entities(&mut self, entities: impl IntoIterator<Item = Entity<M>>) {
-        for entity in entities {
-            self.audit.push_entity(entity);
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/document/provenance/entry.rs b/crates/nvisy-engine/src/document/provenance/entry.rs
deleted file mode 100644
index 5dab34970..000000000
--- a/crates/nvisy-engine/src/document/provenance/entry.rs
+++ /dev/null
@@ -1,194 +0,0 @@
-//! Audit entry: per-entity redaction record.
-//!
-//! [`AuditEntry<M>`] bundles three sub-records into one row of the
-//! audit:
-//!
-//! - [`Decision<M>`] — what the policy evaluator chose (the rule
-//!   that fired, where it ranks in the chain, and — for `Redact` —
-//!   the operator spec the rule carried). Immutable after evaluation.
-//! - [`Execution<M>`] — what the codec applicator did (still
-//!   pending, applied with an `M::Replacement`, failed with a
-//!   reason, or explicitly suppressed). Mutated by the applicator.
-//! - [`EntryMetadata`] — when, correlation, optional review.
-
-use derive_builder::Builder;
-use hipstr::HipStr;
-use jiff::Timestamp;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-use super::override_decision::RedactionDecision;
-use crate::modality::DocumentModality;
-use crate::policy::PolicyDecisionRef;
-
-/// A per-entity redaction record produced during a pipeline run.
-///
-/// The entry lives next to its [`Entity<M>`] inside an
-/// [`EntityRecord<M>`]; the entity's `id` and `location` are read
-/// directly through the record rather than duplicated here.
-///
-/// [`Entity<M>`]: nvisy_core::entity::Entity
-/// [`EntityRecord<M>`]: super::EntityRecord
-#[derive(Debug, Clone, Builder, Serialize, Deserialize, JsonSchema)]
-#[builder(
-    name = "AuditEntryBuilder",
-    pattern = "owned",
-    setter(into, strip_option, prefix = "with")
-)]
-#[serde(rename_all = "camelCase")]
-pub struct AuditEntry<M: DocumentModality> {
-    /// What the policy evaluator chose for this entity.
-    pub decision: Decision<M>,
-    /// What the codec applicator did (or didn't).
-    #[builder(default = "Execution::Pending")]
-    pub execution: Execution<M>,
-    /// Timestamp, correlation, optional review.
-    #[builder(default)]
-    pub metadata: EntryMetadata,
-}
-
-impl<M: DocumentModality> AuditEntry<M> {
-    /// Start building a new audit entry.
-    pub fn builder() -> AuditEntryBuilder<M> {
-        AuditEntryBuilder::default()
-    }
-}
-
-/// What the policy evaluator chose for an entity. Immutable after
-/// evaluation; the applicator only writes to [`AuditEntry::execution`].
-///
-/// The resolved per-entity [`ResolvedAction<M>`] is carried verbatim so
-/// the audit record names the operator the apply phase ran (or
-/// would have run). Callers reading audits can render the operator
-/// without re-resolving the policy chain.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct Decision<M: DocumentModality> {
-    /// Reference to the rule that produced this decision. `None`
-    /// when the decision came from a source outside the policy chain
-    /// (e.g. the default-threshold fallback path).
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub policy_ref: Option<PolicyDecisionRef>,
-    /// The resolved per-entity action — the [`ResolvedAction::Redact`]
-    /// arm carries the single operator picked for this entity's
-    /// modality (after fallback to the deployment-wide defaults).
-    pub action: ResolvedAction<M>,
-}
-
-/// Per-entity resolved action recorded on an [`AuditEntry`].
-///
-/// Differs from the policy-side [`Action`][crate::policy::Action]:
-/// the policy author's `Redact` carries operators for every
-/// modality the rule wants to cover, but a given audit entry is
-/// always for one specific entity — so the audit-side `Redact`
-/// carries just the typed operator that got picked.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum ResolvedAction<M: DocumentModality> {
-    /// Redact the entity using the named operator.
-    Redact {
-        /// Operator spec resolved for this entity's modality.
-        operator: M::Redaction,
-    },
-    /// Suppress the entity (treat as false positive). Optional
-    /// author-supplied reason rides along on the audit entry.
-    Suppress {
-        /// Reason the rule supplied for the suppression, if any.
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        #[schemars(with = "Option<String>")]
-        reason: Option<HipStr<'static>>,
-    },
-    /// Flag the entity for human review without transforming it.
-    /// Optional severity hint rides along on the audit entry.
-    Audit {
-        /// Severity hint the rule supplied, if any.
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        #[schemars(with = "Option<String>")]
-        severity: Option<HipStr<'static>>,
-    },
-}
-
-/// State machine for what the codec applicator did with a
-/// [`Decision`]. The discriminant is the single source of truth for
-/// "did the redaction run, and if so what happened" — there is no
-/// parallel `is_applied` flag.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "status", rename_all = "snake_case")]
-pub enum Execution<M: DocumentModality> {
-    /// Decision recorded; applicator hasn't run yet.
-    Pending,
-    /// Applicator ran successfully. `replacement` records *what the
-    /// operator wrote* at the entity's location, in the modality's
-    /// per-`M::Replacement` shape.
-    Applied {
-        /// What the operator wrote at the entity's location.
-        replacement: M::Replacement,
-    },
-    /// Operator dispatch or codec apply errored. The decision stays
-    /// on the entry; this variant records why no bytes were written.
-    Failed {
-        /// Human-readable failure description from the operator or codec.
-        reason: String,
-    },
-    /// A `Suppress` rule fired: the entity was deliberately not
-    /// redacted. The decision is recorded for completeness but no
-    /// codec work was scheduled.
-    Suppressed,
-}
-
-impl<M: DocumentModality> Execution<M> {
-    /// `true` when the applicator finished and wrote a replacement.
-    pub fn is_applied(&self) -> bool {
-        matches!(self, Self::Applied { .. })
-    }
-
-    /// `true` when no apply work has been attempted yet.
-    pub fn is_pending(&self) -> bool {
-        matches!(self, Self::Pending)
-    }
-}
-
-/// Per-entry timestamp + correlation. Separate from [`Decision`] /
-/// [`Execution`] so consumers reading the "what happened" axis
-/// don't pay for fields they don't need.
-#[derive(Debug, Clone, Default, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct EntryMetadata {
-    /// When this entry was created.
-    #[serde(default)]
-    #[schemars(with = "Option<String>")]
-    pub timestamp: Option<Timestamp>,
-    /// Correlation identifier for tracing across services.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub correlation_id: Option<Uuid>,
-    /// Provenance of the decision on this entry. `None` means
-    /// the decision came from the policy chain with no override
-    /// involvement (legacy unified runs). [`RedactionEngine::redact`]
-    /// stamps every entry with an explicit tag — `PolicyChain`
-    /// when no override touched the entity, or the matching
-    /// `Override*` variant when one did.
-    ///
-    /// [`RedactionEngine::redact`]: crate::redaction::RedactionEngine::redact
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub override_decision: Option<RedactionDecision>,
-}
-
-impl EntryMetadata {
-    /// Empty metadata stamped with the current wall-clock time.
-    pub fn now() -> Self {
-        Self {
-            timestamp: Some(Timestamp::now()),
-            correlation_id: None,
-            override_decision: None,
-        }
-    }
-
-    /// Stamp the metadata with an override-provenance tag.
-    /// Returns `self` for chained construction.
-    #[must_use]
-    pub fn with_override(mut self, decision: RedactionDecision) -> Self {
-        self.override_decision = Some(decision);
-        self
-    }
-}
diff --git a/crates/nvisy-engine/src/document/provenance/mod.rs b/crates/nvisy-engine/src/document/provenance/mod.rs
deleted file mode 100644
index 27ca45afe..000000000
--- a/crates/nvisy-engine/src/document/provenance/mod.rs
+++ /dev/null
@@ -1,162 +0,0 @@
-//! Provenance: audit trails and per-file processing logs.
-//!
-//! This module is the single source of truth for all pipeline audit
-//! types:
-//!
-//! - [`AuditEntry<M>`]: per-entity redaction record with strategy,
-//!   original/replacement values, and optional review.
-//! - [`EntityRecord<M>`]: an [`Entity<M>`] bundled with its
-//!   optional [`AuditEntry<M>`].
-//! - [`Audit<M>`]: per-document container holding the records
-//!   produced during a pipeline run.
-//!
-//! All provenance types are typed per modality. Cross-modality
-//! aggregation (rich documents that process as multiple typed
-//! envelopes) is the engine's responsibility — provenance stays
-//! per-envelope.
-//!
-//! [`Entity<M>`]: nvisy_core::entity::Entity
-
-mod entry;
-mod override_decision;
-mod record;
-
-use derive_more::{From, IsVariant};
-use nvisy_core::entity::{ContentSource, Entity};
-pub use nvisy_toolkit::redaction::{TabularReplacement, TextReplacement};
-use schemars::JsonSchema;
-use serde::de::DeserializeOwned;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-pub use self::entry::{
-    AuditEntry, AuditEntryBuilder, Decision, EntryMetadata, Execution, ResolvedAction,
-};
-pub use self::override_decision::RedactionDecision;
-pub use self::record::EntityRecord;
-use crate::modality::{Audio, DocumentModality, Image, Tabular, Text};
-
-/// A per-document audit trail: per-entity records bundling the
-/// detected entity with the audit entry (if any) produced for it
-/// during the pipeline run.
-///
-/// `Audit<M>` is the compliance artifact for one typed document.
-/// Each [`EntityRecord<M>`] in [`records`] represents one
-/// detection; the record's `audit` field is `Some` when a
-/// redaction or suppression rule matched.
-///
-/// [`records`]: Self::records
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(
-    rename_all = "camelCase",
-    bound(
-        serialize = "M: Serialize, M::Replacement: Serialize",
-        deserialize = "M: DeserializeOwned, M::Replacement: DeserializeOwned",
-    )
-)]
-#[schemars(bound = "M: JsonSchema, M::Replacement: JsonSchema")]
-pub struct Audit<M: DocumentModality> {
-    /// Content source this audit belongs to.
-    pub source: ContentSource,
-    /// Identifier of the pipeline run that produced this audit.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub run_id: Option<Uuid>,
-    /// Identifier of the human or service account that triggered the
-    /// run.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub actor_id: Option<Uuid>,
-    /// Per-entity records produced during the pipeline run.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub records: Vec<EntityRecord<M>>,
-}
-
-impl<M: DocumentModality> Audit<M> {
-    /// Create a new empty audit for the given source.
-    pub fn new(source: ContentSource) -> Self {
-        Self {
-            source,
-            run_id: None,
-            actor_id: None,
-            records: Vec::new(),
-        }
-    }
-
-    /// Append a record for `entity` with no audit decision yet.
-    pub fn push_entity(&mut self, entity: Entity<M>) {
-        self.records.push(EntityRecord::new(entity));
-    }
-
-    /// Iterate over the audit entries that have been recorded.
-    pub fn entries(&self) -> impl Iterator<Item = &AuditEntry<M>> {
-        self.records.iter().filter_map(|r| r.audit.as_ref())
-    }
-
-    /// Number of detected entities.
-    pub fn entities_count(&self) -> usize {
-        self.records.len()
-    }
-
-    /// Number of redaction audit entries whose redaction was
-    /// actually applied (as opposed to suppressed or skipped).
-    pub fn applied_redactions_count(&self) -> usize {
-        self.records
-            .iter()
-            .filter(|r| r.audit.as_ref().is_some_and(|a| a.execution.is_applied()))
-            .count()
-    }
-}
-
-/// A modality-erased [`Audit`].
-///
-/// The persistence layer ([`RegistryStore::store_audits`]) and the
-/// engine's public output ([`EngineOutput::audits`]) work with
-/// `Vec<AnyAudit>` so that a single multi-modal pipeline run can
-/// surface audits for every modality it produced (e.g. a PDF that
-/// fanned out into a `Text` envelope and an `Image` envelope returns
-/// two audits — one of each modality).
-///
-/// Wire format: tagged by `modality`, flattening the audit's own
-/// fields into the same JSON object:
-///
-/// ```json
-/// { "modality": "text", "source": {...}, "records": [...] }
-/// ```
-///
-/// [`RegistryStore::store_audits`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/ingestion/struct.RegistryStore.html#method.store_audits
-/// [`EngineOutput::audits`]: https://docs.rs/nvisy-engine/latest/nvisy_engine/pipeline/struct.EngineOutput.html#structfield.audits
-#[derive(Debug, Clone, From, IsVariant, Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "modality", rename_all = "snake_case")]
-pub enum AnyAudit {
-    /// Text-modality audit.
-    Text(Audit<Text>),
-    /// Tabular-modality audit.
-    Tabular(Audit<Tabular>),
-    /// Image-modality audit.
-    Image(Audit<Image>),
-    /// Audio-modality audit.
-    Audio(Audit<Audio>),
-}
-
-impl AnyAudit {
-    /// Number of detected entities on the contained audit.
-    pub fn entities_count(&self) -> usize {
-        match self {
-            Self::Text(a) => a.entities_count(),
-            Self::Tabular(a) => a.entities_count(),
-            Self::Image(a) => a.entities_count(),
-            Self::Audio(a) => a.entities_count(),
-        }
-    }
-
-    /// Number of redaction audit entries whose redaction was actually
-    /// applied (as opposed to suppressed or skipped). Used by the
-    /// engine's run-summary counters.
-    pub fn applied_redactions_count(&self) -> usize {
-        match self {
-            Self::Text(a) => a.applied_redactions_count(),
-            Self::Tabular(a) => a.applied_redactions_count(),
-            Self::Image(a) => a.applied_redactions_count(),
-            Self::Audio(a) => a.applied_redactions_count(),
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/document/provenance/override_decision.rs b/crates/nvisy-engine/src/document/provenance/override_decision.rs
deleted file mode 100644
index b5e49d909..000000000
--- a/crates/nvisy-engine/src/document/provenance/override_decision.rs
+++ /dev/null
@@ -1,49 +0,0 @@
-//! [`RedactionDecision`]: per-entity provenance recorded on an
-//! [`AuditEntry`] when a redaction pass applied a human override.
-//!
-//! The value is stamped onto [`EntryMetadata::override_decision`]
-//! during [`RedactionEngine::redact`]. Reviewers inspecting the final
-//! audit can distinguish a policy-chain decision from a
-//! human-mediated one.
-//!
-//! Lives in `provenance/` rather than `pipeline/redaction/` because
-//! it's a property of the audit, not of the redaction-request
-//! shape. The redaction module re-exports it for ergonomic access.
-//!
-//! [`AuditEntry`]: super::AuditEntry
-//! [`EntryMetadata::override_decision`]: super::EntryMetadata::override_decision
-//! [`RedactionEngine::redact`]: crate::redaction::RedactionEngine::redact
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Provenance tag for an [`AuditEntry`] decision.
-///
-/// `PolicyChain` means the policy chain's choice stands and no
-/// human touched this entity. The four `Override*` variants
-/// correspond 1:1 with the override types
-/// [`RedactionOverride`] accepts.
-///
-/// [`AuditEntry`]: super::AuditEntry
-/// [`RedactionOverride`]: crate::redaction::RedactionOverride
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum RedactionDecision {
-    /// Default — no override touched this entity; the policy
-    /// chain's choice stands.
-    PolicyChain,
-    /// A `RedactionOverride::Accept` explicitly approved the
-    /// policy chain's choice. Distinct from `PolicyChain` so
-    /// reviewers can see "this was reviewed" vs "this was
-    /// untouched."
-    OverrideAccept,
-    /// A `RedactionOverride::Reject` suppressed this entity.
-    OverrideReject,
-    /// A `RedactionOverride::Replace` swapped the operator.
-    OverrideReplace,
-    /// A `RedactionOverride::Add` introduced this entity; the
-    /// recognisers did not detect it. Policy evaluation ran
-    /// against the synthesised entity exactly as it would for a
-    /// recogniser-detected one.
-    OverrideAdd,
-}
diff --git a/crates/nvisy-engine/src/document/provenance/record.rs b/crates/nvisy-engine/src/document/provenance/record.rs
deleted file mode 100644
index 64ef2e243..000000000
--- a/crates/nvisy-engine/src/document/provenance/record.rs
+++ /dev/null
@@ -1,56 +0,0 @@
-//! [`EntityRecord`]: a detected entity bundled with the optional
-//! audit entry produced for it during redaction.
-
-use nvisy_core::entity::Entity;
-use schemars::JsonSchema;
-use serde::de::DeserializeOwned;
-use serde::{Deserialize, Serialize};
-
-use super::entry::AuditEntry;
-use crate::modality::DocumentModality;
-
-/// One per-entity record on an [`Audit<M>`]: the detected entity
-/// plus the audit entry produced for it during redaction (if any).
-///
-/// `audit` is `None` when the entity was detected but no
-/// redaction or suppression rule matched — the detection still
-/// shows up in the compliance trail but no redaction decision
-/// was made. Suppressed-and-recorded ends up as `Some(entry)`
-/// with `entry.execution == Execution::Suppressed`.
-///
-/// Bundling the two together replaces the previous
-/// `Vec<Entity> + Vec<AuditEntry>` pair indexed by `entity_id`:
-/// every consumer that needed both (dedup, applicator, validator,
-/// exporter) used to join the two vectors via a hash map. The
-/// bundled shape removes the join, the duplicated identifier, and
-/// the entire `RedactionMap` denormalization that used to mirror
-/// `entity.location`.
-///
-/// [`Audit<M>`]: super::Audit
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(
-    rename_all = "camelCase",
-    bound(
-        serialize = "M: Serialize, M::Replacement: Serialize",
-        deserialize = "M: DeserializeOwned, M::Replacement: DeserializeOwned",
-    )
-)]
-#[schemars(bound = "M: JsonSchema, M::Replacement: JsonSchema")]
-pub struct EntityRecord<M: DocumentModality> {
-    /// The detected entity.
-    pub entity: Entity<M>,
-    /// The redaction record for this entity, if a strategy decided
-    /// to act on it.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub audit: Option<AuditEntry<M>>,
-}
-
-impl<M: DocumentModality> EntityRecord<M> {
-    /// Wrap an [`Entity<M>`] with no audit decision yet.
-    pub fn new(entity: Entity<M>) -> Self {
-        Self {
-            entity,
-            audit: None,
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/document/span.rs b/crates/nvisy-engine/src/document/span.rs
deleted file mode 100644
index a63a49a41..000000000
--- a/crates/nvisy-engine/src/document/span.rs
+++ /dev/null
@@ -1,55 +0,0 @@
-//! [`Span`] — a range within a block's flat text, tagged with its
-//! source coordinates.
-
-use nvisy_core::modality::Modality;
-use nvisy_core::primitive::Confidence;
-
-use crate::modality::DocumentModality;
-
-/// A range of text within a block's flat `text`, paired with the
-/// source coordinates of where that text came from.
-///
-/// Source-mapped granularity is typically one span per recognized
-/// unit:
-/// - one span per OCR word for image blocks,
-/// - one span per cell for tabular rows,
-/// - one span per transcribed word for audio,
-/// - one span per text run for natively-extracted text.
-#[derive(Debug, Clone, PartialEq)]
-pub struct Span<M: DocumentModality> {
-    /// Byte offset into the block's `text` where this span starts.
-    pub text_start: usize,
-    /// Byte offset into the block's `text` where this span ends
-    /// (exclusive).
-    pub text_end: usize,
-    /// Recognition confidence in this span. Populated for OCR and STT
-    /// spans; absent for native text-layer extractions where the
-    /// source already provides the text directly.
-    pub confidence: Option<Confidence>,
-    /// Source coordinates of where this text came from.
-    pub source: <M as Modality>::Location,
-}
-
-impl<M: DocumentModality> Span<M> {
-    /// Byte length of the span in the block's `text`.
-    pub fn len(&self) -> usize {
-        self.text_end.saturating_sub(self.text_start)
-    }
-
-    /// Returns `true` when the span covers zero bytes.
-    pub fn is_empty(&self) -> bool {
-        self.text_end <= self.text_start
-    }
-
-    /// Returns `true` when `offset` falls within
-    /// `text_start..text_end`.
-    pub fn contains(&self, offset: usize) -> bool {
-        offset >= self.text_start && offset < self.text_end
-    }
-
-    /// Returns `true` when this span's text range overlaps
-    /// `[start, end)`.
-    pub fn overlaps(&self, start: usize, end: usize) -> bool {
-        self.text_start < end && self.text_end > start
-    }
-}
diff --git a/crates/nvisy-engine/src/engine.rs b/crates/nvisy-engine/src/engine.rs
new file mode 100644
index 000000000..0421b90b8
--- /dev/null
+++ b/crates/nvisy-engine/src/engine.rs
@@ -0,0 +1,74 @@
+//! [`EngineHandle`]: the single dependency entry point every
+//! orchestrator call takes.
+//!
+//! Bundles the two long-lived runtime registries:
+//!
+//! - The [`RegistryHandle`] over [`fjall`] (policies, contexts,
+//!   runs).
+//! - The [`FormatRegistry`] over elide's codec set (decode bytes
+//!   → modality-typed handle at analyze + apply time).
+//!
+//! Cheaply cloneable — both fields are `Arc`-backed under the
+//! hood — so the server hands a clone to every HTTP handler
+//! without coordinating lifetime.
+//!
+//! [`FormatRegistry`]: elide::codec::FormatRegistry
+//! [`fjall`]: ::fjall
+
+use std::path::Path;
+use std::sync::Arc;
+
+use elide::codec::FormatRegistry;
+use nvisy_core::Result;
+
+use crate::registry::RegistryHandle;
+
+/// Cheaply-cloneable bundle of the persistence registry + codec
+/// registry every orchestrator entry point consumes.
+#[derive(Clone)]
+pub struct EngineHandle {
+    registry: RegistryHandle,
+    formats: Arc<FormatRegistry>,
+}
+
+impl EngineHandle {
+    /// Open (or create) the engine database at `path` and pair it
+    /// with elide's built-in codec set
+    /// ([`FormatRegistry::with_builtin`]).
+    pub fn open(path: &Path) -> Result<Self> {
+        let registry = RegistryHandle::open(path)?;
+        let formats = Arc::new(FormatRegistry::with_builtin());
+        Ok(Self { registry, formats })
+    }
+
+    /// Open (or create) the engine database at `path` and pair it
+    /// with a caller-supplied `formats` registry. Useful for tests
+    /// that need to register fake codecs, or for deployments that
+    /// extend the built-in set.
+    pub fn with_formats(path: &Path, formats: FormatRegistry) -> Result<Self> {
+        let registry = RegistryHandle::open(path)?;
+        Ok(Self {
+            registry,
+            formats: Arc::new(formats),
+        })
+    }
+
+    /// The persistence registry. Holds the fjall keyspaces every
+    /// resource module reads and writes.
+    pub fn registry(&self) -> &RegistryHandle {
+        &self.registry
+    }
+
+    /// The codec registry. Pipeline calls reach for it to decode
+    /// raw bytes into a modality-typed
+    /// [`DocumentHandle`](elide::codec::DocumentHandle).
+    pub fn formats(&self) -> &FormatRegistry {
+        &self.formats
+    }
+
+    /// Flush pending writes to disk. The server's HTTP layer
+    /// calls this on graceful shutdown.
+    pub fn sync(&self) -> Result<()> {
+        self.registry.sync()
+    }
+}
diff --git a/crates/nvisy-engine/src/keyspace/context.rs b/crates/nvisy-engine/src/keyspace/context.rs
new file mode 100644
index 000000000..132db1099
--- /dev/null
+++ b/crates/nvisy-engine/src/keyspace/context.rs
@@ -0,0 +1,174 @@
+//! Context resource API over the engine's fjall registry.
+//!
+//! Surfaced as the [`ContextRegistry`] extension trait on
+//! [`RegistryHandle`]. Mirrors [`super::policy::PolicyRegistry`]
+//! in shape: contexts are immutable per
+//! `(actor_id, context_id, version)`. To "edit" a context, write
+//! a new version. The keyspace stores every version side-by-side
+//! so old runs can replay against the exact bytes they were
+//! submitted under.
+
+use nvisy_core::context::Context;
+use nvisy_core::{Error, Result};
+use semver::Version;
+use uuid::Uuid;
+
+use crate::registry::{PagedResult, RegistryHandle, VersionedKey, blocking, not_found};
+
+const COMPONENT: &str = "contexts";
+const KIND: &str = "context";
+
+/// Extension trait adding context-resource CRUD to
+/// [`RegistryHandle`].
+///
+/// Implemented for `RegistryHandle` itself; bring the trait into
+/// scope (`use nvisy_engine::ContextRegistry;`) to call its
+/// methods.
+pub trait ContextRegistry {
+    /// Write a new context version. Fails with
+    /// [`ErrorKind::Conflict`] when an entry already exists at
+    /// `(actor_id, context.id, version)`.
+    ///
+    /// [`ErrorKind::Conflict`]: nvisy_core::ErrorKind::Conflict
+    fn put_context(
+        &self,
+        actor_id: Uuid,
+        context: &Context,
+    ) -> impl Future<Output = Result<()>> + Send;
+
+    /// Read a specific context version.
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    fn get_context(
+        &self,
+        actor_id: Uuid,
+        context_id: Uuid,
+        version: Version,
+    ) -> impl Future<Output = Result<Context>> + Send;
+
+    /// Read the highest-version context for
+    /// `(actor_id, context_id)`.
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    fn latest_context(
+        &self,
+        actor_id: Uuid,
+        context_id: Uuid,
+    ) -> impl Future<Output = Result<Context>> + Send;
+
+    /// List every `(context_id, version)` pair stored for
+    /// `actor_id`.
+    fn list_contexts(
+        &self,
+        actor_id: Uuid,
+    ) -> impl Future<Output = Result<PagedResult<(Uuid, Version)>>> + Send;
+
+    /// Remove one specific context version.
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    fn delete_context(
+        &self,
+        actor_id: Uuid,
+        context_id: Uuid,
+        version: Version,
+    ) -> impl Future<Output = Result<()>> + Send;
+}
+
+impl ContextRegistry for RegistryHandle {
+    async fn put_context(&self, actor_id: Uuid, context: &Context) -> Result<()> {
+        let key = VersionedKey::new(actor_id, context.id, &context.version);
+        let value = serde_json::to_vec(context)?;
+        let contexts = self.contexts().clone();
+        let context_id = context.id;
+        let version = context.version.clone();
+        blocking(move || {
+            if contexts.contains_key(key.as_bytes()).map_err(fjall_err)? {
+                return Err(Error::conflict(
+                    format!(
+                        "context already exists at actor_id={actor_id}, \
+                         context_id={context_id}, version={version} — bump the version to \
+                         write again",
+                    ),
+                    COMPONENT,
+                ));
+            }
+            contexts.insert(key.as_bytes(), value).map_err(fjall_err)?;
+            Ok(())
+        })
+        .await
+    }
+
+    async fn get_context(
+        &self,
+        actor_id: Uuid,
+        context_id: Uuid,
+        version: Version,
+    ) -> Result<Context> {
+        let key = VersionedKey::new(actor_id, context_id, &version);
+        let contexts = self.contexts().clone();
+        blocking(move || {
+            let value = contexts
+                .get(key.as_bytes())
+                .map_err(fjall_err)?
+                .ok_or_else(|| not_found(KIND, actor_id, context_id))?;
+            serde_json::from_slice(&value).map_err(Into::into)
+        })
+        .await
+    }
+
+    async fn latest_context(&self, actor_id: Uuid, context_id: Uuid) -> Result<Context> {
+        let contexts = self.contexts().clone();
+        blocking(move || {
+            let prefix = VersionedKey::prefix(actor_id, context_id);
+            let guard = contexts
+                .prefix(prefix)
+                .next_back()
+                .ok_or_else(|| not_found(KIND, actor_id, context_id))?;
+            let (_, value) = guard.into_inner().map_err(fjall_err)?;
+            serde_json::from_slice(&value).map_err(Into::into)
+        })
+        .await
+    }
+
+    async fn list_contexts(&self, actor_id: Uuid) -> Result<PagedResult<(Uuid, Version)>> {
+        let contexts = self.contexts().clone();
+        blocking(move || {
+            let prefix = VersionedKey::actor_prefix(actor_id);
+            let mut items: Vec<(Uuid, Version)> = Vec::new();
+            for guard in contexts.prefix(prefix) {
+                let key = guard.key().map_err(fjall_err)?;
+                let bytes = key.as_ref();
+                let id = VersionedKey::resource_id_from_bytes(bytes)
+                    .ok_or_else(|| Error::internal("unexpected context key length", COMPONENT))?;
+                let (major, minor, patch) = VersionedKey::version_from_bytes(bytes)
+                    .ok_or_else(|| Error::internal("unexpected context key length", COMPONENT))?;
+                items.push((id, Version::new(major, minor, patch)));
+            }
+            let total = items.len();
+            Ok(PagedResult { items, total })
+        })
+        .await
+    }
+
+    async fn delete_context(
+        &self,
+        actor_id: Uuid,
+        context_id: Uuid,
+        version: Version,
+    ) -> Result<()> {
+        let key = VersionedKey::new(actor_id, context_id, &version);
+        let contexts = self.contexts().clone();
+        blocking(move || {
+            if !contexts.contains_key(key.as_bytes()).map_err(fjall_err)? {
+                return Err(not_found(KIND, actor_id, context_id));
+            }
+            contexts.remove(key.as_bytes()).map_err(fjall_err)?;
+            Ok(())
+        })
+        .await
+    }
+}
+
+fn fjall_err(err: impl std::error::Error + Send + Sync + 'static) -> Error {
+    Error::internal("fjall operation failed", COMPONENT).with_source(err)
+}
diff --git a/crates/nvisy-engine/src/keyspace/file.rs b/crates/nvisy-engine/src/keyspace/file.rs
new file mode 100644
index 000000000..69e02c882
--- /dev/null
+++ b/crates/nvisy-engine/src/keyspace/file.rs
@@ -0,0 +1,284 @@
+//! File resource API over the engine's fjall registry.
+//!
+//! Surfaced as the [`FileRegistry`] extension trait on
+//! [`RegistryHandle`] so file calls flow through the same handle
+//! every other resource API takes —
+//! `handle.put_file(actor, descriptor, bytes)`, etc.
+//!
+//! A file is a `(metadata, bytes)` pair: the bytes live in a
+//! blob-separated `files_content` keyspace; the metadata
+//! ([`FileMetadata`]) lives in a small JSON `files_metadata`
+//! keyspace. The split lets [`FileRegistry::list_files`] enumerate
+//! every file for an actor without paying the cost of loading the
+//! bytes.
+//!
+//! Writes are content-first, metadata-second. A failure between
+//! the two leaves orphan bytes (garbage-collectable later by a
+//! sweep over `files_content` keys without a matching metadata
+//! entry); the opposite order would leave a metadata row that
+//! lists a file whose bytes can't be downloaded — strictly worse.
+//!
+//! Surface:
+//!
+//! - [`put_file`](FileRegistry::put_file) — write bytes +
+//!   metadata; mints a UUIDv7, computes the SHA-256 digest.
+//! - [`get_file`](FileRegistry::get_file) — read the metadata
+//!   blob.
+//! - [`get_file_bytes`](FileRegistry::get_file_bytes) — read the
+//!   raw bytes.
+//! - [`list_files`](FileRegistry::list_files) — list every
+//!   metadata blob for an actor.
+//! - [`delete_file`](FileRegistry::delete_file) — remove one
+//!   file (bytes + metadata).
+//! - [`delete_all_files`](FileRegistry::delete_all_files) — wipe
+//!   every file for an actor.
+
+use std::collections::HashMap;
+
+use bytes::Bytes;
+use hipstr::HipStr;
+use jiff::Timestamp;
+use nvisy_core::file::{FileLineage, FileMetadata};
+use nvisy_core::{Error, Result};
+use sha2::{Digest, Sha256};
+use uuid::Uuid;
+
+use crate::registry::{CompositeKey, RegistryHandle, blocking, not_found};
+
+const COMPONENT: &str = "files";
+const KIND: &str = "file";
+
+/// Caller-supplied descriptor for a new upload. The engine
+/// derives [`FileMetadata::id`], [`uploaded_at`], [`size`], and
+/// [`digest`] at write time.
+///
+/// [`uploaded_at`]: FileMetadata::uploaded_at
+/// [`size`]: FileMetadata::size
+/// [`digest`]: FileMetadata::digest
+#[derive(Debug, Clone, Default)]
+pub struct FileDescriptor {
+    /// Original filename, if any (e.g. parsed from
+    /// `Content-Disposition`).
+    pub filename: Option<HipStr<'static>>,
+    /// Caller-supplied MIME hint.
+    pub content_type: Option<HipStr<'static>>,
+    /// File extension the codec resolves on (case-insensitive,
+    /// no leading dot).
+    pub extension: HipStr<'static>,
+    /// Where the file came from. Uploads pass `None`; engine
+    /// callers (today: the redaction apply path) stamp a
+    /// [`FileLineage::RedactedFrom`] so the output is traceable
+    /// back to its source.
+    pub lineage: Option<FileLineage>,
+    /// Doc-level labels (gate `DocumentPredicate::HasLabel`
+    /// policies when a run later references this file).
+    pub descriptor_labels: Vec<String>,
+    /// Doc-level metadata (gate `DocumentPredicate::HasMetadata`
+    /// policies).
+    pub descriptor_metadata: HashMap<String, String>,
+}
+
+/// Extension trait adding file-resource CRUD to
+/// [`RegistryHandle`].
+///
+/// Implemented for `RegistryHandle` itself; bring the trait into
+/// scope (`use nvisy_engine::FileRegistry;`) to call its
+/// methods.
+pub trait FileRegistry {
+    /// Write bytes + metadata for a new file. Engine mints a
+    /// UUIDv7 id, computes the SHA-256 digest, stamps
+    /// `uploaded_at`, and returns the assembled
+    /// [`FileMetadata`]. Bytes are persisted before metadata so
+    /// a partial failure leaves recoverable orphan bytes rather
+    /// than a metadata row pointing at missing content.
+    fn put_file(
+        &self,
+        actor_id: Uuid,
+        descriptor: FileDescriptor,
+        bytes: Bytes,
+    ) -> impl Future<Output = Result<FileMetadata>> + Send;
+
+    /// Read one file's metadata. Returns
+    /// [`ErrorKind::NotFound`] when no entry exists at
+    /// `(actor_id, file_id)`.
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    fn get_file(
+        &self,
+        actor_id: Uuid,
+        file_id: Uuid,
+    ) -> impl Future<Output = Result<FileMetadata>> + Send;
+
+    /// Read one file's raw bytes. Returns
+    /// [`ErrorKind::NotFound`] when no entry exists at
+    /// `(actor_id, file_id)`.
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    fn get_file_bytes(
+        &self,
+        actor_id: Uuid,
+        file_id: Uuid,
+    ) -> impl Future<Output = Result<Bytes>> + Send;
+
+    /// List every file metadata blob for `actor_id`. Bytes are
+    /// not loaded.
+    fn list_files(&self, actor_id: Uuid) -> impl Future<Output = Result<Vec<FileMetadata>>> + Send;
+
+    /// Remove one file (bytes + metadata). Returns
+    /// [`ErrorKind::NotFound`] if the metadata entry was already
+    /// absent. Best-effort on the bytes — silently skips if the
+    /// bytes were already gone.
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    fn delete_file(&self, actor_id: Uuid, file_id: Uuid)
+    -> impl Future<Output = Result<()>> + Send;
+
+    /// Remove every file (bytes + metadata) belonging to
+    /// `actor_id`. Returns the total count removed across both
+    /// keyspaces.
+    fn delete_all_files(&self, actor_id: Uuid) -> impl Future<Output = Result<usize>> + Send;
+}
+
+impl FileRegistry for RegistryHandle {
+    async fn put_file(
+        &self,
+        actor_id: Uuid,
+        descriptor: FileDescriptor,
+        bytes: Bytes,
+    ) -> Result<FileMetadata> {
+        let file_id = Uuid::now_v7();
+        let size = bytes.len() as u64;
+        let digest = hex_sha256(&bytes);
+        let metadata = FileMetadata {
+            id: file_id,
+            filename: descriptor.filename,
+            content_type: descriptor.content_type,
+            extension: descriptor.extension,
+            size,
+            digest,
+            uploaded_at: Timestamp::now(),
+            lineage: descriptor.lineage,
+            descriptor_labels: descriptor.descriptor_labels,
+            descriptor_metadata: descriptor.descriptor_metadata,
+        };
+
+        // Content-first, metadata-second. A partial failure here
+        // leaves orphan bytes that a future sweep can reap; the
+        // opposite order would advertise a file we can't serve.
+        let key = CompositeKey::new(actor_id, file_id);
+        let content_ks = self.files_content().clone();
+        let metadata_ks = self.files_metadata().clone();
+        let metadata_value = serde_json::to_vec(&metadata)?;
+        let raw_bytes = bytes.to_vec();
+        blocking(move || {
+            content_ks.insert(*key, raw_bytes).map_err(fjall_err)?;
+            metadata_ks
+                .insert(*key, metadata_value)
+                .map_err(fjall_err)?;
+            Ok(())
+        })
+        .await?;
+
+        Ok(metadata)
+    }
+
+    async fn get_file(&self, actor_id: Uuid, file_id: Uuid) -> Result<FileMetadata> {
+        let key = CompositeKey::new(actor_id, file_id);
+        let metadata_ks = self.files_metadata().clone();
+        blocking(move || {
+            let value = metadata_ks
+                .get(*key)
+                .map_err(fjall_err)?
+                .ok_or_else(|| not_found(KIND, actor_id, file_id))?;
+            serde_json::from_slice(&value).map_err(Into::into)
+        })
+        .await
+    }
+
+    async fn get_file_bytes(&self, actor_id: Uuid, file_id: Uuid) -> Result<Bytes> {
+        let key = CompositeKey::new(actor_id, file_id);
+        let content_ks = self.files_content().clone();
+        blocking(move || {
+            let value = content_ks
+                .get(*key)
+                .map_err(fjall_err)?
+                .ok_or_else(|| not_found(KIND, actor_id, file_id))?;
+            Ok(Bytes::from(value.to_vec()))
+        })
+        .await
+    }
+
+    async fn list_files(&self, actor_id: Uuid) -> Result<Vec<FileMetadata>> {
+        let metadata_ks = self.files_metadata().clone();
+        blocking(move || {
+            let prefix = CompositeKey::actor_prefix(actor_id);
+            let mut out: Vec<FileMetadata> = Vec::new();
+            for guard in metadata_ks.prefix(prefix) {
+                let (_, value) = guard.into_inner().map_err(fjall_err)?;
+                let metadata: FileMetadata = serde_json::from_slice(&value)?;
+                out.push(metadata);
+            }
+            Ok(out)
+        })
+        .await
+    }
+
+    async fn delete_file(&self, actor_id: Uuid, file_id: Uuid) -> Result<()> {
+        let key = CompositeKey::new(actor_id, file_id);
+        let metadata_ks = self.files_metadata().clone();
+        let content_ks = self.files_content().clone();
+        blocking(move || {
+            if !metadata_ks.contains_key(*key).map_err(fjall_err)? {
+                return Err(not_found(KIND, actor_id, file_id));
+            }
+            // Metadata-first, content-second on delete: removing
+            // metadata makes the file invisible to subsequent
+            // `list_files` / `get_file` calls; if the content
+            // remove fails the file is logically gone, the bytes
+            // are reapable later.
+            metadata_ks.remove(*key).map_err(fjall_err)?;
+            content_ks.remove(*key).map_err(fjall_err)?;
+            Ok(())
+        })
+        .await
+    }
+
+    async fn delete_all_files(&self, actor_id: Uuid) -> Result<usize> {
+        let metadata_ks = self.files_metadata().clone();
+        let content_ks = self.files_content().clone();
+        blocking(move || {
+            let prefix = CompositeKey::actor_prefix(actor_id);
+            let mut removed = 0usize;
+            for keyspace in [&metadata_ks, &content_ks] {
+                let keys: Vec<Vec<u8>> = keyspace
+                    .prefix(prefix)
+                    .map(|guard| {
+                        let key = guard.key().map_err(fjall_err)?;
+                        Ok(key.as_ref().to_vec())
+                    })
+                    .collect::<Result<Vec<_>>>()?;
+                removed += keys.len();
+                for key in keys {
+                    keyspace.remove(&key).map_err(fjall_err)?;
+                }
+            }
+            Ok(removed)
+        })
+        .await
+    }
+}
+
+fn hex_sha256(bytes: &[u8]) -> HipStr<'static> {
+    let mut hasher = Sha256::new();
+    hasher.update(bytes);
+    let digest = hasher.finalize();
+    let mut out = String::with_capacity(64);
+    for byte in digest {
+        out.push_str(&format!("{byte:02x}"));
+    }
+    HipStr::from(out)
+}
+
+fn fjall_err(err: impl std::error::Error + Send + Sync + 'static) -> Error {
+    Error::internal("fjall operation failed", COMPONENT).with_source(err)
+}
diff --git a/crates/nvisy-engine/src/keyspace/mod.rs b/crates/nvisy-engine/src/keyspace/mod.rs
new file mode 100644
index 000000000..421d9712c
--- /dev/null
+++ b/crates/nvisy-engine/src/keyspace/mod.rs
@@ -0,0 +1,41 @@
+//! User-owned resource CRUD over the engine's fjall keyspaces.
+//!
+//! Three resource families, all keyed by `(actor_id, …)`:
+//!
+//! - [`mod@policy`] — versioned [`Policy`] blobs keyed by
+//!   `(actor_id, policy_id, version)`. Immutable per
+//!   `(id, version)`; "edit" by writing a new version.
+//! - [`mod@context`] — versioned [`Context`] blobs, same shape.
+//! - [`mod@file`] — opaque `(metadata, bytes)` pair keyed by
+//!   `(actor_id, file_id)`. Metadata + content live in
+//!   separate keyspaces so `list_files` is cheap; bytes are
+//!   loaded on demand.
+//!
+//! Each submodule exposes a `pub` extension trait on
+//! [`RegistryHandle`] (`PolicyRegistry`, `ContextRegistry`,
+//! `FileRegistry`). Method names are namespaced
+//! (`put_policy`, `get_file`, `list_contexts`, …) so traits
+//! can be brought into scope together without collisions.
+//!
+//! **Why these three live together.** They're the resources
+//! the user directly owns — clients freely create, read,
+//! modify, and delete them; the engine has no opinion about
+//! when they may be written. Contrast with engine-managed
+//! state (the run lifecycle), which lives in
+//! [`crate::runs`] behind a `pub(crate)` storage trait. The
+//! orchestrator coordinates run writes with state-machine
+//! transitions, so writing one directly from outside the
+//! engine would corrupt the invariants — hence the visibility
+//! split.
+//!
+//! [`Policy`]: nvisy_core::policy::Policy
+//! [`Context`]: nvisy_core::context::Context
+//! [`RegistryHandle`]: crate::registry::RegistryHandle
+
+pub mod context;
+pub mod file;
+pub mod policy;
+
+pub use self::context::ContextRegistry;
+pub use self::file::{FileDescriptor, FileRegistry};
+pub use self::policy::PolicyRegistry;
diff --git a/crates/nvisy-engine/src/keyspace/policy.rs b/crates/nvisy-engine/src/keyspace/policy.rs
new file mode 100644
index 000000000..92757d2ed
--- /dev/null
+++ b/crates/nvisy-engine/src/keyspace/policy.rs
@@ -0,0 +1,189 @@
+//! Policy resource API over the engine's fjall registry.
+//!
+//! Surfaced as the [`PolicyRegistry`] extension trait on
+//! [`RegistryHandle`] so resource calls flow through the same
+//! handle every other engine entry point takes —
+//! `handle.get_policy(...)`, `handle.put_policy(...)`, etc.
+//!
+//! Policies are immutable per `(actor_id, policy_id, version)`.
+//! To "edit" a policy, write a new version. The keyspace stores
+//! every version side-by-side so old runs can replay against the
+//! exact bytes they were submitted under.
+//!
+//! Surface:
+//!
+//! - [`put_policy`](PolicyRegistry::put_policy) — write a new
+//!   version (rejects writes that would overwrite an existing
+//!   `(actor, id, version)`).
+//! - [`get_policy`](PolicyRegistry::get_policy) — read one
+//!   specific version.
+//! - [`latest_policy`](PolicyRegistry::latest_policy) — find the
+//!   highest version for `(actor, id)`.
+//! - [`list_policies`](PolicyRegistry::list_policies) — list
+//!   `(id, version)` pairs for an actor.
+//! - [`delete_policy`](PolicyRegistry::delete_policy) — remove
+//!   one specific version.
+
+use nvisy_core::policy::Policy;
+use nvisy_core::{Error, Result};
+use semver::Version;
+use uuid::Uuid;
+
+use crate::registry::{PagedResult, RegistryHandle, VersionedKey, blocking, not_found};
+
+const COMPONENT: &str = "policies";
+const KIND: &str = "policy";
+
+/// Extension trait adding policy-resource CRUD to
+/// [`RegistryHandle`].
+///
+/// Implemented for `RegistryHandle` itself; bring the trait into
+/// scope (`use nvisy_engine::PolicyRegistry;`) to call its
+/// methods.
+pub trait PolicyRegistry {
+    /// Write a new policy version. Fails with
+    /// [`ErrorKind::Conflict`] when an entry already exists at
+    /// `(actor_id, policy.id, version)`.
+    ///
+    /// [`ErrorKind::Conflict`]: nvisy_core::ErrorKind::Conflict
+    fn put_policy(
+        &self,
+        actor_id: Uuid,
+        policy: &Policy,
+    ) -> impl Future<Output = Result<()>> + Send;
+
+    /// Read a specific policy version. Returns
+    /// [`ErrorKind::NotFound`] when no entry exists at
+    /// `(actor_id, policy_id, version)`.
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    fn get_policy(
+        &self,
+        actor_id: Uuid,
+        policy_id: Uuid,
+        version: Version,
+    ) -> impl Future<Output = Result<Policy>> + Send;
+
+    /// Read the highest-version policy for
+    /// `(actor_id, policy_id)`. Returns
+    /// [`ErrorKind::NotFound`] when no version exists.
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    fn latest_policy(
+        &self,
+        actor_id: Uuid,
+        policy_id: Uuid,
+    ) -> impl Future<Output = Result<Policy>> + Send;
+
+    /// List every `(policy_id, version)` pair stored for
+    /// `actor_id`. Versions land in lex order (matches semver
+    /// order for the encoded `(major, minor, patch)` triple).
+    fn list_policies(
+        &self,
+        actor_id: Uuid,
+    ) -> impl Future<Output = Result<PagedResult<(Uuid, Version)>>> + Send;
+
+    /// Remove one specific policy version. Returns
+    /// [`ErrorKind::NotFound`] if the entry was already absent.
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    fn delete_policy(
+        &self,
+        actor_id: Uuid,
+        policy_id: Uuid,
+        version: Version,
+    ) -> impl Future<Output = Result<()>> + Send;
+}
+
+impl PolicyRegistry for RegistryHandle {
+    async fn put_policy(&self, actor_id: Uuid, policy: &Policy) -> Result<()> {
+        let key = VersionedKey::new(actor_id, policy.id, &policy.version);
+        let value = serde_json::to_vec(policy)?;
+        let policies = self.policies().clone();
+        let policy_id = policy.id;
+        let version = policy.version.clone();
+        blocking(move || {
+            if policies.contains_key(key.as_bytes()).map_err(fjall_err)? {
+                return Err(Error::conflict(
+                    format!(
+                        "policy already exists at actor_id={actor_id}, policy_id={policy_id}, \
+                         version={version} — bump the version to write again",
+                    ),
+                    COMPONENT,
+                ));
+            }
+            policies.insert(key.as_bytes(), value).map_err(fjall_err)?;
+            Ok(())
+        })
+        .await
+    }
+
+    async fn get_policy(
+        &self,
+        actor_id: Uuid,
+        policy_id: Uuid,
+        version: Version,
+    ) -> Result<Policy> {
+        let key = VersionedKey::new(actor_id, policy_id, &version);
+        let policies = self.policies().clone();
+        blocking(move || {
+            let value = policies
+                .get(key.as_bytes())
+                .map_err(fjall_err)?
+                .ok_or_else(|| not_found(KIND, actor_id, policy_id))?;
+            serde_json::from_slice(&value).map_err(Into::into)
+        })
+        .await
+    }
+
+    async fn latest_policy(&self, actor_id: Uuid, policy_id: Uuid) -> Result<Policy> {
+        let policies = self.policies().clone();
+        blocking(move || {
+            let prefix = VersionedKey::prefix(actor_id, policy_id);
+            let guard = policies
+                .prefix(prefix)
+                .next_back()
+                .ok_or_else(|| not_found(KIND, actor_id, policy_id))?;
+            let (_, value) = guard.into_inner().map_err(fjall_err)?;
+            serde_json::from_slice(&value).map_err(Into::into)
+        })
+        .await
+    }
+
+    async fn list_policies(&self, actor_id: Uuid) -> Result<PagedResult<(Uuid, Version)>> {
+        let policies = self.policies().clone();
+        blocking(move || {
+            let prefix = VersionedKey::actor_prefix(actor_id);
+            let mut items: Vec<(Uuid, Version)> = Vec::new();
+            for guard in policies.prefix(prefix) {
+                let key = guard.key().map_err(fjall_err)?;
+                let bytes = key.as_ref();
+                let id = VersionedKey::resource_id_from_bytes(bytes)
+                    .ok_or_else(|| Error::internal("unexpected policy key length", COMPONENT))?;
+                let (major, minor, patch) = VersionedKey::version_from_bytes(bytes)
+                    .ok_or_else(|| Error::internal("unexpected policy key length", COMPONENT))?;
+                items.push((id, Version::new(major, minor, patch)));
+            }
+            let total = items.len();
+            Ok(PagedResult { items, total })
+        })
+        .await
+    }
+
+    async fn delete_policy(&self, actor_id: Uuid, policy_id: Uuid, version: Version) -> Result<()> {
+        let key = VersionedKey::new(actor_id, policy_id, &version);
+        let policies = self.policies().clone();
+        blocking(move || {
+            if !policies.contains_key(key.as_bytes()).map_err(fjall_err)? {
+                return Err(not_found(KIND, actor_id, policy_id));
+            }
+            policies.remove(key.as_bytes()).map_err(fjall_err)?;
+            Ok(())
+        })
+        .await
+    }
+}
+
+fn fjall_err(err: impl std::error::Error + Send + Sync + 'static) -> Error {
+    Error::internal("fjall operation failed", COMPONENT).with_source(err)
+}
diff --git a/crates/nvisy-engine/src/lib.rs b/crates/nvisy-engine/src/lib.rs
index 1bb32defd..b5a718cfb 100644
--- a/crates/nvisy-engine/src/lib.rs
+++ b/crates/nvisy-engine/src/lib.rs
@@ -2,14 +2,32 @@
 #![cfg_attr(docsrs, feature(doc_cfg))]
 #![doc = include_str!("../README.md")]
 
-pub mod core;
-pub mod detection;
-pub mod document;
-pub mod modality;
-pub mod policy;
-pub mod redaction;
+//! The runtime adapter over [`elide`]. Engine is what makes elide a
+//! long-running, multi-tenant, multi-document, two-phase redaction
+//! service:
+//!
+//! - [`analyzer`] compiles a request's recognition plan into an
+//!   [`elide::detection::Analyzer`] per modality.
+//! - [`anonymizer`] compiles a request's [`nvisy_core::policy::Policy`]
+//!   set into an [`elide::redaction::Anonymizer`] per modality, stamping each
+//!   resolved decision with the policy + rule provenance the audit
+//!   trail requires.
+//! - [`registry`] holds the multi-tenant, actor-scoped storage
+//!   primitives ([`fjall`] keyspaces keyed by `[actor_id | …]`).
+//! - [`keyspace`] hosts the user-owned resource CRUD —
+//!   [`PolicyRegistry`], [`FileRegistry`], [`ContextRegistry`] —
+//!   all extension traits on [`registry::RegistryHandle`].
+//! - [`runs`] is the orchestrator over the engine-managed run
+//!   lifecycle, with its persistence trait kept `pub(crate)` so
+//!   external code can't write malformed runs.
+
+mod engine;
+
+pub mod analyzer;
+pub mod anonymizer;
+pub mod keyspace;
 pub mod registry;
+pub mod runs;
 
-pub use nvisy_codec::content::{
-    Content, ContentData, ContentDescriptor, ContentDigest, ContentRecord, ContentSource,
-};
+pub use self::engine::EngineHandle;
+pub use self::keyspace::{ContextRegistry, FileDescriptor, FileRegistry, PolicyRegistry};
diff --git a/crates/nvisy-engine/src/modality/any.rs b/crates/nvisy-engine/src/modality/any.rs
deleted file mode 100644
index deb63fefb..000000000
--- a/crates/nvisy-engine/src/modality/any.rs
+++ /dev/null
@@ -1,48 +0,0 @@
-//! Type-erased [`Modality::Location`] enum.
-
-use nvisy_core::modality::{
-    AudioLocation, ImageLocation, ModalityKind, TabularLocation, TextLocation,
-};
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use strum::EnumTryAs;
-
-/// Type-erased [`Modality::Location`] carrying both the modality
-/// tag and the typed coordinate value. Used by the redaction
-/// override surface where the caller hasn't yet pinned a
-/// `M: Modality` type at the API boundary.
-///
-/// Wire shape matches the [`ModalityKind`] tag plus a flattened
-/// location object:
-///
-/// ```json
-/// { "modality": "text", "start": 0, "end": 10 }
-/// ```
-///
-/// [`Modality::Location`]: nvisy_core::modality::Modality::Location
-#[derive(Debug, Clone, PartialEq, EnumTryAs)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "modality", rename_all = "snake_case")]
-pub enum AnyLocation {
-    /// Text-modality location.
-    Text(TextLocation),
-    /// Tabular-modality location.
-    Tabular(TabularLocation),
-    /// Image-modality location.
-    Image(ImageLocation),
-    /// Audio-modality location.
-    Audio(AudioLocation),
-}
-
-impl AnyLocation {
-    /// The modality this location belongs to.
-    #[must_use]
-    pub fn kind(&self) -> ModalityKind {
-        match self {
-            Self::Text(_) => ModalityKind::Text,
-            Self::Tabular(_) => ModalityKind::Tabular,
-            Self::Image(_) => ModalityKind::Image,
-            Self::Audio(_) => ModalityKind::Audio,
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/modality/audio.rs b/crates/nvisy-engine/src/modality/audio.rs
deleted file mode 100644
index c96cb7cff..000000000
--- a/crates/nvisy-engine/src/modality/audio.rs
+++ /dev/null
@@ -1,100 +0,0 @@
-//! Audio-modality document shape: [`AudioBlock`], [`AudioMetadata`].
-
-use nvisy_core::modality::AudioExtraction;
-use nvisy_core::primitive::{LanguageDetection, TimeSpan};
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::ModalityBlock;
-
-/// Per-modality block payload for [`Audio`].
-/// [`Speech`] carries the transcript text and optional speaker;
-/// per-word source spans live on the wrapping [`Block<Audio>`].
-/// [`Silence`] carries no payload. Every variant carries the segment
-/// `time_span`.
-///
-/// [`Audio`]: nvisy_core::modality::Audio
-/// [`Speech`]: Self::Speech
-/// [`Silence`]: Self::Silence
-/// [`Block<Audio>`]: crate::document::Block
-#[derive(Debug, Clone, PartialEq)]
-#[non_exhaustive]
-pub enum AudioBlock {
-    /// A transcribed speech segment (typically one speaker turn).
-    Speech {
-        /// Segment time interval.
-        time_span: TimeSpan,
-        /// Transcribed text content.
-        text: String,
-        /// Speaker identifier from diarization, when available.
-        speaker_id: Option<String>,
-    },
-    /// A silence or non-speech segment surfaced for completeness.
-    Silence {
-        /// Segment time interval.
-        time_span: TimeSpan,
-    },
-}
-
-impl AudioBlock {
-    /// Segment time interval.
-    pub fn time_span(&self) -> &TimeSpan {
-        match self {
-            Self::Speech { time_span, .. } | Self::Silence { time_span } => time_span,
-        }
-    }
-
-    /// Transcribed text for [`Speech`], `None` for silence.
-    ///
-    /// [`Speech`]: Self::Speech
-    pub fn text(&self) -> Option<&str> {
-        match self {
-            Self::Speech { text, .. } => Some(text),
-            Self::Silence { .. } => None,
-        }
-    }
-}
-
-impl ModalityBlock for AudioBlock {
-    fn scan_text(&self) -> Option<&str> {
-        self.text()
-    }
-}
-
-/// Document-level metadata for [`Document<Audio>`].
-///
-/// [`Document<Audio>`]: crate::document::Document
-// TODO(#226): wire an importer that stamps `extraction` (Transcription
-// or Diarization). No audio importer exists today; the field is
-// type-required so that lands as a real value the moment one does.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct AudioMetadata {
-    /// How this document's audio content was processed (STT
-    /// transcription, speaker diarization).
-    pub extraction: AudioExtraction,
-    /// Languages detected (or asserted) for the transcribed content.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    #[schemars(skip)]
-    pub languages: Vec<LanguageDetection>,
-    /// Sample rate of the source audio, in Hz, when known.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub sample_rate_hz: Option<u32>,
-    /// Number of channels in the source audio, when known.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub channels: Option<u16>,
-}
-
-impl From<AudioExtraction> for AudioMetadata {
-    /// Build [`AudioMetadata`] carrying only the importer-known
-    /// extraction tag. Languages, sample rate, and channel count
-    /// start empty; downstream stages fill them in.
-    fn from(extraction: AudioExtraction) -> Self {
-        Self {
-            extraction,
-            languages: Vec::new(),
-            sample_rate_hz: None,
-            channels: None,
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/modality/image.rs b/crates/nvisy-engine/src/modality/image.rs
deleted file mode 100644
index e0d4d5cd9..000000000
--- a/crates/nvisy-engine/src/modality/image.rs
+++ /dev/null
@@ -1,145 +0,0 @@
-//! Image-modality document shape: [`ImageBlock`], [`ImageMetadata`],
-//! [`PageDimensions`].
-
-use nvisy_core::modality::{ImageExtraction, ImageLocation};
-use nvisy_core::primitive::LanguageDetection;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::ModalityBlock;
-
-/// Per-modality block payload for [`Image`]. Text-bearing variants
-/// carry recognized text; per-word source spans live on the wrapping
-/// [`Block<Image>`]. Non-textual variants ([`Figure`], [`Separator`],
-/// [`Background`], [`Logo`]) carry no text. Every variant carries the
-/// bounding `region` since image blocks are always spatially located.
-///
-/// [`Image`]: nvisy_core::modality::Image
-/// [`Figure`]: Self::Figure
-/// [`Separator`]: Self::Separator
-/// [`Background`]: Self::Background
-/// [`Logo`]: Self::Logo
-/// [`Block<Image>`]: crate::document::Block
-#[derive(Debug, Clone, PartialEq)]
-#[non_exhaustive]
-pub enum ImageBlock {
-    /// A region of recognized text (paragraph, line, OCR text block).
-    Text {
-        /// Bounding region in image coordinates.
-        region: ImageLocation,
-        /// Recognized text content.
-        text: String,
-    },
-    /// A heading.
-    Heading {
-        /// Bounding region in image coordinates.
-        region: ImageLocation,
-        /// Recognized text content.
-        text: String,
-    },
-    /// A tabular region recognized in the image.
-    Table {
-        /// Bounding region in image coordinates.
-        region: ImageLocation,
-        /// Recognized text content (cells flattened in reading order).
-        text: String,
-    },
-    /// A figure, illustration or photograph.
-    Figure {
-        /// Bounding region in image coordinates.
-        region: ImageLocation,
-    },
-    /// A separator (rule, line, divider).
-    Separator {
-        /// Bounding region in image coordinates.
-        region: ImageLocation,
-    },
-    /// A background element (watermark, fill, decoration).
-    Background {
-        /// Bounding region in image coordinates.
-        region: ImageLocation,
-    },
-    /// A logo or brand mark.
-    Logo {
-        /// Bounding region in image coordinates.
-        region: ImageLocation,
-    },
-}
-
-impl ImageBlock {
-    /// The image region this block occupies.
-    pub fn region(&self) -> &ImageLocation {
-        match self {
-            Self::Text { region, .. }
-            | Self::Heading { region, .. }
-            | Self::Table { region, .. }
-            | Self::Figure { region }
-            | Self::Separator { region }
-            | Self::Background { region }
-            | Self::Logo { region } => region,
-        }
-    }
-
-    /// Recognized text for text-bearing kinds, `None` for non-textual.
-    pub fn text(&self) -> Option<&str> {
-        match self {
-            Self::Text { text, .. } | Self::Heading { text, .. } | Self::Table { text, .. } => {
-                Some(text)
-            }
-            Self::Figure { .. }
-            | Self::Separator { .. }
-            | Self::Background { .. }
-            | Self::Logo { .. } => None,
-        }
-    }
-}
-
-impl ModalityBlock for ImageBlock {
-    fn scan_text(&self) -> Option<&str> {
-        self.text()
-    }
-}
-
-/// Document-level metadata for [`Document<Image>`].
-///
-/// [`Document<Image>`]: crate::document::Document
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct ImageMetadata {
-    /// How this document's image content was processed (OCR, scene
-    /// text, object detection, layout analysis).
-    pub extraction: ImageExtraction,
-    /// Languages detected (or asserted) for the document content.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    #[schemars(skip)]
-    pub languages: Vec<LanguageDetection>,
-    /// Page dimensions for multi-page sources (PDFs, multi-image
-    /// uploads). Empty for single-page sources.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub pages: Vec<PageDimensions>,
-}
-
-impl From<ImageExtraction> for ImageMetadata {
-    /// Build [`ImageMetadata`] carrying only the importer-known
-    /// extraction tag. Languages and page dimensions start empty;
-    /// downstream stages fill them in.
-    fn from(extraction: ImageExtraction) -> Self {
-        Self {
-            extraction,
-            languages: Vec::new(),
-            pages: Vec::new(),
-        }
-    }
-}
-
-/// Dimensions of a single page in an image document.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct PageDimensions {
-    /// 1-based page number.
-    pub number: u32,
-    /// Page width in pixels.
-    pub width: f64,
-    /// Page height in pixels.
-    pub height: f64,
-}
diff --git a/crates/nvisy-engine/src/modality/mod.rs b/crates/nvisy-engine/src/modality/mod.rs
deleted file mode 100644
index acc828191..000000000
--- a/crates/nvisy-engine/src/modality/mod.rs
+++ /dev/null
@@ -1,128 +0,0 @@
-//! Document-shape per-modality types: block payloads, metadata,
-//! and the [`DocumentModality`] extension trait.
-//!
-//! Document-level shape lives here, not in `nvisy-core`. Core owns the
-//! atomic [`Modality`] marker trait + the per-modality coordinate
-//! structs (`Text`, `Image`, `Audio`, `Tabular`); this module adds:
-//!
-//! - per-modality block payloads (`TextBlock`, `ImageBlock`,
-//!   `AudioBlock`, `TabularBlock`) describing what a block of that
-//!   modality carries;
-//! - per-modality document-level metadata (`TextMetadata`, etc.)
-//!   carrying extraction tags + per-modality fields (languages, page
-//!   dimensions, column headers);
-//! - the [`ModalityBlock`] trait every block payload implements
-//!   (shared per-block queries);
-//! - the [`DocumentModality`] extension trait that binds `Block` +
-//!   `Metadata` per modality and is the trait every document-shape
-//!   generic actually bounds on.
-//!
-//! [`Modality`]: nvisy_core::modality::Modality
-
-mod any;
-mod audio;
-mod image;
-mod tabular;
-mod text;
-
-use std::fmt::Debug;
-
-pub use nvisy_core::modality::{
-    Audio, AudioExtraction, Image, ImageExtraction, Modality, Tabular, TabularExtraction, Text,
-    TextExtraction,
-};
-use schemars::JsonSchema;
-use serde::Serialize;
-use serde::de::DeserializeOwned;
-
-pub use self::any::AnyLocation;
-pub use self::audio::{AudioBlock, AudioMetadata};
-pub use self::image::{ImageBlock, ImageMetadata, PageDimensions};
-pub use self::tabular::{ColumnHeader, TabularBlock, TabularMetadata};
-pub use self::text::{EmbeddedDocument, TextBlock, TextContent, TextMetadata};
-use crate::policy::redaction::{AudioRedaction, ImageRedaction, TabularRedaction, TextRedaction};
-
-/// Shared per-block surface every modality's block payload exposes.
-///
-/// This is the trait every [`DocumentModality::Block`] type
-/// implements; it collects the modality-agnostic queries the engine
-/// needs to drive pipeline stages without matching on the concrete
-/// block variant.
-///
-/// Today the only method is [`scan_text`], used by the detection
-/// driver to decide whether a block carries text the text-typed
-/// recognizers should scan. Future shared block queries (block id,
-/// kind tag, block-level confidence accessor) land here.
-///
-/// [`scan_text`]: Self::scan_text
-pub trait ModalityBlock {
-    /// The text a text-typed recognizer should scan over this block,
-    /// or `None` when the block carries no scannable text (image
-    /// figures/logos, audio silences).
-    fn scan_text(&self) -> Option<&str>;
-}
-
-/// Document-shape extension of [`Modality`].
-///
-/// Adds the document-level shape (block payload, document metadata,
-/// policy redaction enum) the document carrier and its phases need.
-/// Implemented for the four modalities in this module — code that
-/// just needs a marker can bound on bare `Modality`; code that
-/// needs any document-level associated type bounds on
-/// `DocumentModality`.
-pub trait DocumentModality: Modality {
-    /// The modality's block payload. See [`TextBlock`], [`ImageBlock`],
-    /// [`AudioBlock`], [`TabularBlock`].
-    type Block: ModalityBlock + Clone + Debug + Send + Sync + 'static;
-
-    /// Document-level metadata: extraction tag plus modality-specific
-    /// fields (languages, page dimensions, column headers).
-    type Metadata: Clone + Debug + PartialEq + Send + Sync + 'static;
-
-    /// Operator-spec enum a `redact` policy rule of this modality
-    /// carries. Per-call instantiated for built-ins; `Custom(id)`
-    /// arms resolve through the toolkit-side
-    /// [`RedactionRegistry<Self>`].
-    ///
-    /// The serde + schemars bounds let [`Policy<Self>`] /
-    /// [`PolicyRule<Self>`] derive Serialize / Deserialize /
-    /// JsonSchema transparently across all four modalities — the
-    /// per-modality enum types meet them naturally.
-    ///
-    /// [`RedactionRegistry<Self>`]: nvisy_toolkit::redaction::RedactionRegistry
-    /// [`Policy<Self>`]: crate::policy::Policy
-    /// [`PolicyRule<Self>`]: crate::policy::PolicyRule
-    type Redaction: Clone
-        + Debug
-        + PartialEq
-        + Send
-        + Sync
-        + Serialize
-        + DeserializeOwned
-        + JsonSchema
-        + 'static;
-}
-
-impl DocumentModality for Text {
-    type Block = TextBlock;
-    type Metadata = TextMetadata;
-    type Redaction = TextRedaction;
-}
-
-impl DocumentModality for Image {
-    type Block = ImageBlock;
-    type Metadata = ImageMetadata;
-    type Redaction = ImageRedaction;
-}
-
-impl DocumentModality for Audio {
-    type Block = AudioBlock;
-    type Metadata = AudioMetadata;
-    type Redaction = AudioRedaction;
-}
-
-impl DocumentModality for Tabular {
-    type Block = TabularBlock;
-    type Metadata = TabularMetadata;
-    type Redaction = TabularRedaction;
-}
diff --git a/crates/nvisy-engine/src/modality/tabular.rs b/crates/nvisy-engine/src/modality/tabular.rs
deleted file mode 100644
index ee168c995..000000000
--- a/crates/nvisy-engine/src/modality/tabular.rs
+++ /dev/null
@@ -1,86 +0,0 @@
-//! Tabular-modality document shape: [`TabularBlock`],
-//! [`TabularMetadata`], [`ColumnHeader`].
-
-use nvisy_core::modality::TabularExtraction;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::ModalityBlock;
-
-/// Per-modality block payload for [`Tabular`]. Today only [`Row`] —
-/// carries the flat row text and its index. Per-cell source spans
-/// live on the wrapping [`Block<Tabular>`].
-///
-/// [`Tabular`]: nvisy_core::modality::Tabular
-/// [`Row`]: Self::Row
-/// [`Block<Tabular>`]: crate::document::Block
-#[derive(Debug, Clone, PartialEq)]
-#[non_exhaustive]
-pub enum TabularBlock {
-    /// A row. The row index is carried on each
-    /// [`Span<Tabular>::source`] (every span maps a sub-range of
-    /// `text` back to its originating cell — the cell's `row_index`
-    /// and `column_index` live there), so it's not duplicated at
-    /// the block level.
-    ///
-    /// [`Span<Tabular>::source`]: crate::document::Span::source
-    Row {
-        /// Flattened row text (cells joined per the source delimiter).
-        text: String,
-    },
-}
-
-impl TabularBlock {
-    /// The row's flat text.
-    pub fn text(&self) -> &str {
-        match self {
-            Self::Row { text } => text,
-        }
-    }
-}
-
-impl ModalityBlock for TabularBlock {
-    fn scan_text(&self) -> Option<&str> {
-        Some(self.text())
-    }
-}
-
-/// Document-level metadata for [`Document<Tabular>`].
-///
-/// [`Document<Tabular>`]: crate::document::Document
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct TabularMetadata {
-    /// How this document's tabular structure was produced
-    /// (schema-typed from source, inferred, or recovered from an image).
-    pub extraction: TabularExtraction,
-    /// Column headers indexed by 0-based column position.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub headers: Vec<ColumnHeader>,
-    /// Sheet names for multi-sheet documents, in source order.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub sheet_names: Vec<String>,
-}
-
-impl From<TabularExtraction> for TabularMetadata {
-    /// Build [`TabularMetadata`] carrying only the importer-known
-    /// extraction tag. Headers and sheet names start empty;
-    /// downstream stages fill them in.
-    fn from(extraction: TabularExtraction) -> Self {
-        Self {
-            extraction,
-            headers: Vec::new(),
-            sheet_names: Vec::new(),
-        }
-    }
-}
-
-/// A column header in a tabular document.
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct ColumnHeader {
-    /// 0-based column index.
-    pub column_index: u32,
-    /// Header text.
-    pub text: String,
-}
diff --git a/crates/nvisy-engine/src/modality/text.rs b/crates/nvisy-engine/src/modality/text.rs
deleted file mode 100644
index ac616a820..000000000
--- a/crates/nvisy-engine/src/modality/text.rs
+++ /dev/null
@@ -1,146 +0,0 @@
-//! Text-modality document shape: [`TextBlock`], [`TextContent`],
-//! [`EmbeddedDocument`], [`TextMetadata`].
-
-use nvisy_core::modality::{Image, Tabular, TextExtraction};
-use nvisy_core::primitive::LanguageDetection;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::ModalityBlock;
-use crate::document::Document;
-
-/// Per-modality block payload for [`Text`].
-///
-/// Splits into two variants:
-///
-/// - [`Text`] wraps a structural text-shaped kind
-///   (paragraph, heading, list item, code, quote) — see [`TextContent`].
-/// - [`Embed`] hosts a nested [`Document`] of another
-///   modality (image, tabular) for sources like PDF, DOCX, or HTML
-///   that mix text with non-text content in one flow.
-///
-/// Per-word source spans live on the wrapping [`Block<Text>`].
-///
-/// [`Text`]: Self::Text
-/// [`Embed`]: Self::Embed
-/// [`Block<Text>`]: crate::document::Block
-#[derive(Debug, Clone)]
-#[non_exhaustive]
-pub enum TextBlock {
-    /// Text-shaped content (paragraph, heading, list item, code, quote).
-    Text(TextContent),
-    /// A nested document of another modality (image, tabular) embedded
-    /// in the text flow. Boxed because [`EmbeddedDocument`] wraps a
-    /// full [`Document<M>`] (blocks + audit) which dwarfs the
-    /// [`TextContent`] variant; without the box every `TextBlock`
-    /// pays the embed footprint.
-    Embed(Box<EmbeddedDocument>),
-}
-
-/// Text-shaped block content — paragraphs, headings, and other
-/// structural variants that carry flat text.
-///
-/// Split out from [`TextBlock`] so the embed variants stay distinct
-/// and text-only recognizers can match on `TextBlock::Text(_)` in one
-/// arm without forgetting embed variants.
-#[derive(Debug, Clone, PartialEq)]
-#[non_exhaustive]
-pub enum TextContent {
-    /// A regular paragraph or text run.
-    Paragraph {
-        /// Paragraph text.
-        text: String,
-    },
-    /// A heading.
-    Heading {
-        /// Heading text.
-        text: String,
-        /// Heading depth (1 = h1, 2 = h2, …). `None` when the source
-        /// doesn't expose a level.
-        level: Option<u8>,
-    },
-    /// A list item.
-    ListItem {
-        /// List item text.
-        text: String,
-    },
-    /// A code block or pre-formatted text.
-    Code {
-        /// Code-block content.
-        text: String,
-    },
-    /// A blockquote.
-    Quote {
-        /// Quoted text.
-        text: String,
-    },
-}
-
-impl TextContent {
-    /// The content's text.
-    pub fn text(&self) -> &str {
-        match self {
-            Self::Paragraph { text }
-            | Self::Heading { text, .. }
-            | Self::ListItem { text }
-            | Self::Code { text }
-            | Self::Quote { text } => text,
-        }
-    }
-}
-
-/// A nested document of another modality embedded inside a
-/// [`TextBlock::Embed`] variant.
-///
-/// PDFs and other rich containers can host images and tables inline
-/// with their text flow. The engine processes each nested document
-/// through its own per-modality phase chain; the recursion is
-/// orchestrator-driven so individual phases stay single-doc.
-///
-/// Recursion is one-directional: only text can host other modalities.
-/// `ImageBlock`, `AudioBlock`, and `TabularBlock` have no embed
-/// variants by construction, so the maximum nesting depth is 2 and
-/// no termination check is needed at runtime.
-#[derive(Debug, Clone)]
-pub enum EmbeddedDocument {
-    /// A nested image document (e.g. a PDF figure or DOCX picture).
-    Image(Document<Image>),
-    /// A nested tabular document (e.g. a PDF table or DOCX table).
-    Tabular(Document<Tabular>),
-}
-
-impl ModalityBlock for TextBlock {
-    fn scan_text(&self) -> Option<&str> {
-        match self {
-            Self::Text(content) => Some(content.text()),
-            Self::Embed(_) => None,
-        }
-    }
-}
-
-/// Document-level metadata for [`Document<Text>`].
-///
-/// [`Document<Text>`]: crate::document::Document
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct TextMetadata {
-    /// How this document's text was produced (native text-layer parse
-    /// vs OCR'd image-backed page).
-    pub extraction: TextExtraction,
-    /// Languages detected (or asserted) for the document content.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    #[schemars(skip)]
-    pub languages: Vec<LanguageDetection>,
-}
-
-impl From<TextExtraction> for TextMetadata {
-    /// Build [`TextMetadata`] carrying only the importer-known
-    /// extraction tag. All other fields start empty; downstream
-    /// stages fill them in as they discover the document's content.
-    fn from(extraction: TextExtraction) -> Self {
-        Self {
-            extraction,
-            languages: Vec::new(),
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/policy/condition.rs b/crates/nvisy-engine/src/policy/condition.rs
deleted file mode 100644
index 852640509..000000000
--- a/crates/nvisy-engine/src/policy/condition.rs
+++ /dev/null
@@ -1,25 +0,0 @@
-//! Conditions that gate when a strategy policy applies.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// A condition that must be met for a strategy policy to apply.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "type", rename_all = "snake_case")]
-#[non_exhaustive]
-pub enum Condition {
-    /// All specified document labels must be present.
-    Labels {
-        /// Document labels that must all be present.
-        labels: Vec<String>,
-    },
-    /// A metadata key-value pair that must be present on the envelope.
-    Metadata {
-        /// Metadata key to match.
-        key: String,
-        /// Expected value. If `None`, any value for the key satisfies
-        /// the condition.
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        value: Option<String>,
-    },
-}
diff --git a/crates/nvisy-engine/src/policy/mod.rs b/crates/nvisy-engine/src/policy/mod.rs
deleted file mode 100644
index 50725a741..000000000
--- a/crates/nvisy-engine/src/policy/mod.rs
+++ /dev/null
@@ -1,175 +0,0 @@
-//! Policy types: authored vocabulary for redaction governance.
-//!
-//! A [`Policy`] is a named, versioned governance artefact: an
-//! ordered list of [`PolicyRule`]s plus an optional fallback
-//! [`Policy::default_action`] plus a retention configuration.
-//! Policies are reusable — the same policy can participate in many
-//! runs.
-//!
-//! Per-run composition (which policies apply to *this* run, in what
-//! order) lives in the engine; the ontology does not model it.
-//! Precedence is positional: in a run, the first policy in the
-//! caller-supplied list is highest precedence; within a policy, the
-//! first matching rule wins; the policy's `default_action` fires
-//! only when no rule in that policy matched.
-
-mod audit;
-mod condition;
-pub mod redaction;
-mod retention;
-mod rule;
-mod selector;
-mod suppress;
-
-use derive_builder::Builder;
-use hipstr::HipStr;
-use nvisy_core::entity::EntityLabel;
-use schemars::JsonSchema;
-use semver::Version;
-use serde::{Deserialize, Serialize};
-
-pub use self::audit::AuditAction;
-pub use self::condition::Condition;
-pub use self::redaction::AnyRedaction;
-pub use self::retention::{Retention, RetentionPolicy, RetentionScope};
-pub use self::rule::{Action, PolicyRule};
-pub use self::selector::EntitySelector;
-pub use self::suppress::SuppressAction;
-
-/// A named, versioned governance policy.
-///
-/// Identified by [`name`] + [`version`]; the name must be unique
-/// within a single [`DetectionInput::policies`] submission. Held as
-/// a [`HipStr<'static>`] so per-decision audit stamps and per-run
-/// snapshots share refcounts rather than allocating.
-///
-/// Modality is not part of the type. Each rule's action carries
-/// per-modality operator specs ([`ModalityRedactions`]); the apply
-/// phase picks the operator matching the entity's modality.
-///
-/// [`name`]: Self::name
-/// [`version`]: Self::version
-/// [`DetectionInput::policies`]: crate::detection::DetectionInput::policies
-/// [`ModalityRedactions`]: crate::policy::redaction::ModalityRedactions
-#[derive(Debug, Clone, Builder, Serialize, Deserialize, JsonSchema)]
-#[builder(
-    name = "PolicyBuilder",
-    pattern = "owned",
-    setter(into, strip_option, prefix = "with")
-)]
-#[serde(rename_all = "camelCase")]
-pub struct Policy {
-    /// Author-supplied policy name. Must be unique across every
-    /// policy in a detection submission; audit entries reference
-    /// this string verbatim.
-    #[schemars(with = "String")]
-    pub name: HipStr<'static>,
-    /// Policy version.
-    #[schemars(with = "String")]
-    pub version: Version,
-    /// Description of the policy's purpose.
-    #[builder(default)]
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub description: Option<String>,
-    /// Entity labels this policy operates over. Every label name a
-    /// [`PolicyRule::selector`] references must appear here. The
-    /// engine unions every submitted policy's `labels` into a
-    /// per-request [`EntityLabelCatalog`] used to drive recognizer
-    /// dispatch and tag-based selector matching. Two policies
-    /// declaring the same label name with different
-    /// `(description, tags)` are a conflict and fail the request.
-    ///
-    /// [`EntityLabelCatalog`]: nvisy_core::entity::EntityLabelCatalog
-    #[builder(default = "Vec::new()")]
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub labels: Vec<EntityLabel>,
-    /// Ordered list of rules. First matching rule wins.
-    #[builder(default = "Vec::new()")]
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub rules: Vec<PolicyRule>,
-    /// Fallback action for entities that no [`PolicyRule`] in this
-    /// policy matched. Consulted only after every rule in this
-    /// policy has been considered; the engine then moves to the next
-    /// policy in the per-run chain. `None` means "this policy has no
-    /// opinion for unmatched entities; let the next policy decide."
-    #[builder(default, setter(into = false))]
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub default_action: Option<Action>,
-    /// Data retention lifecycle rules.
-    #[builder(default = "Vec::new()")]
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub retention: Vec<RetentionPolicy>,
-}
-
-impl Policy {
-    /// Start building a new policy.
-    pub fn builder() -> PolicyBuilder {
-        PolicyBuilder::default()
-    }
-
-    /// Header-card summary suitable for storing on the detection
-    /// record without inlining the full rule body.
-    pub fn digest(&self) -> PolicyDigest {
-        PolicyDigest {
-            name: self.name.clone(),
-            version: self.version.clone(),
-        }
-    }
-}
-
-/// Header card identifying a policy submitted to a run, persisted on
-/// the detection record so a reader can render
-/// `"<name> v<version>"` without the caller having to keep the
-/// original [`Policy`] bytes around.
-///
-/// Carries name + version only — no rules. The rule body lives only
-/// in the caller's submission; the engine remembers which policies
-/// ran by digest, and stamps every audit decision with a
-/// [`PolicyDecisionRef`] pointing at one of them.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct PolicyDigest {
-    /// Policy name; matches `Policy::name` of the submitted policy.
-    #[schemars(with = "String")]
-    pub name: HipStr<'static>,
-    /// Policy version.
-    #[schemars(with = "String")]
-    pub version: Version,
-}
-
-/// Reference to the specific rule (or fallback) that produced a
-/// decision, stamped onto every policy-driven [`AuditEntry`].
-///
-/// Names map back into the [`Policy`] / [`PolicyRule`] structures
-/// the caller submitted. `rule_name` is `None` when the producing
-/// rule was the policy's [`default_action`] fallback rather than a
-/// concrete named rule.
-///
-/// Both fields hold [`HipStr<'static>`] clones so audit-heavy passes
-/// share refcounts rather than allocating per-entity.
-///
-/// [`AuditEntry`]: crate::document::provenance::AuditEntry
-/// [`default_action`]: Policy::default_action
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct PolicyDecisionRef {
-    /// Name of the policy that made the decision. Matches the
-    /// `name` of one [`PolicyDigest`] on the detection record.
-    #[schemars(with = "String")]
-    pub policy_name: HipStr<'static>,
-    /// Name of the producing rule inside its policy. `None` means
-    /// the policy's [`Policy::default_action`] fallback fired.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub rule_name: Option<HipStr<'static>>,
-}
-
-impl PolicyDecisionRef {
-    /// Construct a reference from a policy + rule name.
-    pub fn new(policy_name: HipStr<'static>, rule_name: Option<HipStr<'static>>) -> Self {
-        Self {
-            policy_name,
-            rule_name,
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/policy/redaction/any.rs b/crates/nvisy-engine/src/policy/redaction/any.rs
deleted file mode 100644
index 3b2f5e0c5..000000000
--- a/crates/nvisy-engine/src/policy/redaction/any.rs
+++ /dev/null
@@ -1,59 +0,0 @@
-//! [`AnyRedaction`]: type-erased redaction operator spec.
-//!
-//! Carries the modality tag and the typed [`*Redaction`] enum that
-//! a policy author would pick. Used at API boundaries where the
-//! caller hasn't pinned an `M: Modality` type yet — most notably
-//! [`RedactionOverride::Replace`] and [`RedactionAddEntity`] in the
-//! redaction subsystem.
-//!
-//! Wire shape:
-//!
-//! ```json
-//! { "modality": "text", "kind": "mask", ... }
-//! ```
-//!
-//! [`RedactionOverride::Replace`]: crate::redaction::RedactionOverride::Replace
-//! [`RedactionAddEntity`]: crate::redaction::RedactionAddEntity
-
-use nvisy_core::modality::ModalityKind;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use strum::EnumTryAs;
-
-use super::audio::AudioRedaction;
-use super::image::ImageRedaction;
-use super::tabular::TabularRedaction;
-use super::text::TextRedaction;
-
-/// Type-erased redaction operator spec.
-///
-/// Use the typed [`TextRedaction`] / [`ImageRedaction`] /
-/// [`AudioRedaction`] / [`TabularRedaction`] enums directly inside
-/// policy rules where the modality is known at compile time;
-/// reach for `AnyRedaction` only at the dynamic API boundary.
-#[derive(Debug, Clone, PartialEq, Eq, EnumTryAs)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "modality", rename_all = "snake_case")]
-pub enum AnyRedaction {
-    /// Text-modality redaction operator.
-    Text(TextRedaction),
-    /// Tabular-modality redaction operator.
-    Tabular(TabularRedaction),
-    /// Image-modality redaction operator.
-    Image(ImageRedaction),
-    /// Audio-modality redaction operator.
-    Audio(AudioRedaction),
-}
-
-impl AnyRedaction {
-    /// The modality this redaction operator belongs to.
-    #[must_use]
-    pub fn modality(&self) -> ModalityKind {
-        match self {
-            Self::Text(_) => ModalityKind::Text,
-            Self::Tabular(_) => ModalityKind::Tabular,
-            Self::Image(_) => ModalityKind::Image,
-            Self::Audio(_) => ModalityKind::Audio,
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/policy/redaction/audio.rs b/crates/nvisy-engine/src/policy/redaction/audio.rs
deleted file mode 100644
index e1a8b1c4b..000000000
--- a/crates/nvisy-engine/src/policy/redaction/audio.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-//! [`AudioRedaction`]: the operator spec an audio-modality policy
-//! rule carries.
-//!
-//! The toolkit ships no built-in audio operators yet, so the only
-//! variant today is [`AudioRedaction::Custom`] — deployments
-//! register their own audio anonymisers (silence, white-noise,
-//! beep, …) on the [`RedactionRegistry<Audio>`] and reference them
-//! by id.
-//!
-//! [`RedactionRegistry<Audio>`]: nvisy_toolkit::redaction::RedactionRegistry
-
-use nvisy_core::modality::Audio;
-use nvisy_toolkit::redaction::AnonymizerId;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Operator spec a `redact` audio rule carries.
-#[derive(Debug, Clone, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum AudioRedaction {
-    /// Look up a deployment-registered custom operator by id.
-    Custom {
-        /// Id under which the operator was registered in the
-        /// [`RedactionRegistry<Audio>`].
-        ///
-        /// [`RedactionRegistry<Audio>`]: nvisy_toolkit::redaction::RedactionRegistry
-        id: AnonymizerId<Audio>,
-    },
-}
diff --git a/crates/nvisy-engine/src/policy/redaction/image.rs b/crates/nvisy-engine/src/policy/redaction/image.rs
deleted file mode 100644
index d59534396..000000000
--- a/crates/nvisy-engine/src/policy/redaction/image.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-//! [`ImageRedaction`]: the operator spec an image-modality policy
-//! rule carries.
-//!
-//! The toolkit ships no built-in image operators yet, so the only
-//! variant today is [`ImageRedaction::Custom`] — deployments
-//! register their own image anonymisers (blur, pixelate,
-//! blackbox, …) on the [`RedactionRegistry<Image>`] and reference
-//! them by id.
-//!
-//! [`RedactionRegistry<Image>`]: nvisy_toolkit::redaction::RedactionRegistry
-
-use nvisy_core::modality::Image;
-use nvisy_toolkit::redaction::AnonymizerId;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Operator spec a `redact` image rule carries.
-#[derive(Debug, Clone, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum ImageRedaction {
-    /// Look up a deployment-registered custom operator by id.
-    Custom {
-        /// Id under which the operator was registered in the
-        /// [`RedactionRegistry<Image>`].
-        ///
-        /// [`RedactionRegistry<Image>`]: nvisy_toolkit::redaction::RedactionRegistry
-        id: AnonymizerId<Image>,
-    },
-}
diff --git a/crates/nvisy-engine/src/policy/redaction/mod.rs b/crates/nvisy-engine/src/policy/redaction/mod.rs
deleted file mode 100644
index 254e59f62..000000000
--- a/crates/nvisy-engine/src/policy/redaction/mod.rs
+++ /dev/null
@@ -1,157 +0,0 @@
-//! Per-modality redaction operator specs — the closed wire vocabulary
-//! a policy author can choose from inside a `redact` rule — plus the
-//! [`ModalityRedactions`] map a rule carries to wire one operator per
-//! modality, and the [`AnyRedaction`] erasure used at the dynamic
-//! override boundary.
-//!
-//! Each modality has its own enum because the operator catalogue
-//! differs by modality. Text gets the full toolkit built-in set
-//! (`Replace`, `Mask`, `Hash`, `Redact`, `Keep`) plus a `Custom`
-//! escape hatch. Image / Audio / Tabular currently expose only
-//! `Custom` — the toolkit ships no built-in operators for those
-//! modalities yet.
-//!
-//! The split between these enums (the spec) and the toolkit's
-//! [`Anonymizer<M>`] trait is intentional: the spec is the
-//! serialisable, author-facing wire shape; instantiating it at
-//! apply time produces the runtime operator instance.
-//!
-//! [`Anonymizer<M>`]: nvisy_toolkit::redaction::Anonymizer
-
-mod any;
-mod audio;
-mod image;
-mod tabular;
-mod text;
-
-use nvisy_core::modality::{Audio, Image, Tabular, Text};
-pub use nvisy_toolkit::redaction::anonymizer::HashAlgorithm;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-pub use self::any::AnyRedaction;
-pub use self::audio::AudioRedaction;
-pub use self::image::ImageRedaction;
-pub use self::tabular::TabularRedaction;
-pub use self::text::TextRedaction;
-use crate::modality::DocumentModality;
-
-/// Per-modality operator specs carried by a `redact` rule.
-///
-/// A single rule can name an operator for every modality the
-/// workspace supports. At apply time the redaction phase picks the
-/// operator matching the entity's modality via
-/// [`ModalityRedactions::operator_for`]; modalities the rule didn't
-/// cover fall through to the deployment-wide default
-/// (`RedactionConfig::default_operators`), and entities with no
-/// operator from either source are skipped.
-#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub struct ModalityRedactions {
-    /// Operator for text-modality entities.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub text: Option<TextRedaction>,
-    /// Operator for tabular-modality entities.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub tabular: Option<TabularRedaction>,
-    /// Operator for image-modality entities.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub image: Option<ImageRedaction>,
-    /// Operator for audio-modality entities.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub audio: Option<AudioRedaction>,
-}
-
-impl ModalityRedactions {
-    /// `true` when no operator is set for any modality. A rule
-    /// whose `redact` field is empty after deserialisation is an
-    /// author error — the request validator rejects it.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.text.is_none()
-            && self.tabular.is_none()
-            && self.image.is_none()
-            && self.audio.is_none()
-    }
-
-    /// Borrow the operator spec for modality `M`, if one is set.
-    ///
-    /// The apply phase uses this to pick the right operator for an
-    /// entity. Returns `None` when this rule didn't cover the
-    /// modality.
-    #[must_use]
-    pub fn operator_for<M: ProjectRedaction>(&self) -> Option<&M::Redaction> {
-        M::project(self)
-    }
-}
-
-/// Sealed projection trait that picks the typed `*Redaction` field
-/// from a [`ModalityRedactions`] for the implementing modality.
-///
-/// One impl per workspace modality; not extensible by downstream
-/// crates today. When user-defined modalities land, this becomes
-/// the seam to widen.
-pub trait ProjectRedaction: DocumentModality {
-    /// Borrow this modality's operator spec out of `redactions`.
-    fn project(redactions: &ModalityRedactions) -> Option<&Self::Redaction>;
-}
-
-impl ProjectRedaction for Text {
-    fn project(r: &ModalityRedactions) -> Option<&Self::Redaction> {
-        r.text.as_ref()
-    }
-}
-
-impl ProjectRedaction for Tabular {
-    fn project(r: &ModalityRedactions) -> Option<&Self::Redaction> {
-        r.tabular.as_ref()
-    }
-}
-
-impl ProjectRedaction for Image {
-    fn project(r: &ModalityRedactions) -> Option<&Self::Redaction> {
-        r.image.as_ref()
-    }
-}
-
-impl ProjectRedaction for Audio {
-    fn project(r: &ModalityRedactions) -> Option<&Self::Redaction> {
-        r.audio.as_ref()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_toolkit::redaction::anonymizer::HashAlgorithm;
-
-    use super::*;
-
-    #[test]
-    fn empty_when_no_modality_set() {
-        let r = ModalityRedactions::default();
-        assert!(r.is_empty());
-    }
-
-    #[test]
-    fn non_empty_when_any_modality_set() {
-        let r = ModalityRedactions {
-            text: Some(TextRedaction::Redact),
-            ..Default::default()
-        };
-        assert!(!r.is_empty());
-    }
-
-    #[test]
-    fn operator_for_text_round_trips() {
-        let r = ModalityRedactions {
-            text: Some(TextRedaction::Hash {
-                algorithm: HashAlgorithm::Sha256,
-                salt: None,
-            }),
-            ..Default::default()
-        };
-        let op = r.operator_for::<Text>().expect("text operator set");
-        assert!(matches!(op, TextRedaction::Hash { .. }));
-        assert!(r.operator_for::<Image>().is_none());
-    }
-}
diff --git a/crates/nvisy-engine/src/policy/redaction/tabular.rs b/crates/nvisy-engine/src/policy/redaction/tabular.rs
deleted file mode 100644
index 439db2117..000000000
--- a/crates/nvisy-engine/src/policy/redaction/tabular.rs
+++ /dev/null
@@ -1,24 +0,0 @@
-//! [`TabularRedaction`]: the operator spec a tabular-modality policy
-//! rule carries.
-//!
-//! No tabular operators exist yet in the toolkit, but the enum
-//! exists so policy schemas can be parameterised uniformly over all
-//! four modalities. Only [`TabularRedaction::Custom`] is
-//! implementable today.
-
-use nvisy_core::modality::Tabular;
-use nvisy_toolkit::redaction::AnonymizerId;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Operator spec a `redact` tabular rule carries.
-#[derive(Debug, Clone, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum TabularRedaction {
-    /// Reserved for a future tabular registry; not invokable today.
-    Custom {
-        /// Id under which the operator would be registered.
-        id: AnonymizerId<Tabular>,
-    },
-}
diff --git a/crates/nvisy-engine/src/policy/redaction/text.rs b/crates/nvisy-engine/src/policy/redaction/text.rs
deleted file mode 100644
index bca0b984f..000000000
--- a/crates/nvisy-engine/src/policy/redaction/text.rs
+++ /dev/null
@@ -1,98 +0,0 @@
-//! [`TextRedaction`]: the operator spec a text-modality policy rule
-//! carries.
-//!
-//! Built-in variants ([`Replace`], [`Mask`], [`Hash`],
-//! [`Redact`], [`Keep`]) are instantiated per-call from the
-//! params on the variant — no registry round-trip.
-//! [`TextRedaction::Custom`] names a deployment-registered operator
-//! looked up in the [`RedactionRegistry<Text>`] at apply time.
-//!
-//! ## Why no `Encrypt` variant?
-//!
-//! [`Encrypt`] needs 32 bytes of raw key material plus an
-//! [`Aes256Gcm`] instance. Putting either in a TOML / JSON wire
-//! shape is unsafe (key leakage in version control, log dumps) and
-//! awkward (binary blobs in declarative config). Deployments that
-//! want reversible AES-256-GCM redaction build an [`Encrypt`] in
-//! Rust code, register it as a custom anonymizer on the
-//! [`RedactionRegistry<Text>`], and reference it from policy with
-//! `{ kind = "custom", id = "..." }`.
-//!
-//! [`Replace`]: nvisy_toolkit::redaction::anonymizer::Replace
-//! [`Mask`]: nvisy_toolkit::redaction::anonymizer::Mask
-//! [`Hash`]: nvisy_toolkit::redaction::anonymizer::Hash
-//! [`Redact`]: nvisy_toolkit::redaction::anonymizer::Redact
-//! [`Keep`]: nvisy_toolkit::redaction::anonymizer::Keep
-//! [`Encrypt`]: nvisy_toolkit::redaction::anonymizer::Encrypt
-//! [`Aes256Gcm`]: https://docs.rs/aes-gcm
-//! [`RedactionRegistry<Text>`]: nvisy_toolkit::redaction::RedactionRegistry
-
-use nvisy_core::modality::Text;
-use nvisy_toolkit::redaction::AnonymizerId;
-use nvisy_toolkit::redaction::anonymizer::HashAlgorithm;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Operator spec a `redact` text rule carries.
-#[derive(Debug, Clone, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum TextRedaction {
-    /// Substitute the span with a fixed template. Supports
-    /// `{entity_kind}` / `{value}` placeholders. Default template
-    /// is `[{entity_kind}]`.
-    Replace {
-        /// Template string. Default `[{entity_kind}]`.
-        #[serde(default = "default_replace_template")]
-        template: String,
-    },
-    /// Character-replacement masking. Leaves `keep_prefix` leading
-    /// and `keep_suffix` trailing characters visible; masks the
-    /// rest with `mask_char`.
-    Mask {
-        /// The character that replaces masked positions.
-        #[serde(default = "default_mask_char")]
-        mask_char: char,
-        /// Characters to leave unmasked at the start of the value.
-        /// `0` (the default) masks from the start.
-        #[serde(default, skip_serializing_if = "is_zero")]
-        keep_prefix: usize,
-        /// Characters to leave unmasked at the end of the value.
-        /// `0` (the default) masks through to the end.
-        #[serde(default, skip_serializing_if = "is_zero")]
-        keep_suffix: usize,
-    },
-    /// One-way SHA-2 hash with optional salt.
-    Hash {
-        /// SHA-256 (default) or SHA-512.
-        #[serde(default)]
-        algorithm: HashAlgorithm,
-        /// Salt prepended to the value before hashing.
-        #[serde(default, skip_serializing_if = "Option::is_none")]
-        salt: Option<String>,
-    },
-    /// Delete the matched span entirely.
-    Redact,
-    /// Pass the value through unchanged.
-    Keep,
-    /// Look up a deployment-registered custom operator by id.
-    Custom {
-        /// Id under which the operator was registered in the
-        /// [`RedactionRegistry<Text>`].
-        ///
-        /// [`RedactionRegistry<Text>`]: nvisy_toolkit::redaction::RedactionRegistry
-        id: AnonymizerId<Text>,
-    },
-}
-
-fn default_replace_template() -> String {
-    "[{entity_kind}]".to_string()
-}
-
-fn default_mask_char() -> char {
-    '*'
-}
-
-fn is_zero(n: &usize) -> bool {
-    *n == 0
-}
diff --git a/crates/nvisy-engine/src/policy/rule.rs b/crates/nvisy-engine/src/policy/rule.rs
deleted file mode 100644
index 4d8c56f8f..000000000
--- a/crates/nvisy-engine/src/policy/rule.rs
+++ /dev/null
@@ -1,92 +0,0 @@
-//! [`Action`] and [`PolicyRule`] — what a policy rule does when it
-//! matches an entity, plus the rule wrapper that binds a selector,
-//! action, conditions, and enabled flag.
-
-use hipstr::HipStr;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::audit::AuditAction;
-use super::condition::Condition;
-use super::redaction::ModalityRedactions;
-use super::selector::EntitySelector;
-use super::suppress::SuppressAction;
-
-/// What a policy rule does when its selector matches an entity.
-///
-/// Three verbs:
-///
-/// - [`Action::Redact`] transforms matching entities. Carries one
-///   [`ModalityRedactions`] map: a rule can name operators for
-///   multiple modalities at once, and the apply phase picks the
-///   one matching the entity's modality.
-/// - [`Action::Suppress`] drops matching entities (treat as a
-///   false positive). Carries an optional `reason` propagated into
-///   the audit entry.
-/// - [`Action::Audit`] flags matching entities for human review
-///   without transforming them. The detection pass already emits
-///   an audit entry per matched entity; this action tags those
-///   entries with a severity hint downstream review tooling can
-///   prioritise on.
-///
-/// Wire shape uses an externally-tagged enum with the variant name
-/// as the discriminator key — `[policies.rules.redact]`,
-/// `[policies.rules.suppress]`, `[policies.rules.audit]` — flattened
-/// onto the parent [`PolicyRule`] via `#[serde(flatten)]`. Exactly
-/// one of the three keys must be present per rule; serde rejects
-/// duplicates at deserialise time.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum Action {
-    /// Redact matching entities. The carried map names operators
-    /// per modality; modalities the rule didn't cover fall through
-    /// to the deployment-wide default.
-    Redact(ModalityRedactions),
-    /// Suppress matching entities — treat as a false positive and
-    /// skip redaction. Audit entry records the suppression.
-    Suppress(SuppressAction),
-    /// Flag matching entities for human review. The entity is
-    /// left untouched; the audit entry carries the severity hint
-    /// for downstream tooling.
-    Audit(AuditAction),
-}
-
-/// One rule inside a [`Policy`]: a name, a selector, an action,
-/// optional conditions, and an enabled flag.
-///
-/// Rules are ordered inside their owning policy; the first matching
-/// rule wins. There is no separate `priority` field — re-ordering
-/// rules in the policy file is how authors change priority. The
-/// [`name`] field is what audit entries reference; it must be
-/// unique within the owning policy.
-///
-/// [`Policy`]: super::Policy
-/// [`name`]: Self::name
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct PolicyRule {
-    /// Author-supplied rule name. Must be unique within the
-    /// owning policy; audit entries reference this string verbatim.
-    #[schemars(with = "String")]
-    pub name: HipStr<'static>,
-    /// Which entities this rule applies to. Authored as
-    /// `match = { ... }` for parity with the policy-DSL phrasing.
-    #[serde(rename = "match")]
-    pub selector: EntitySelector,
-    /// What this rule does when it matches. Flattened so the
-    /// variant key (`redact` / `suppress` / `audit`) sits at the
-    /// rule's top level — the shape author-facing TOML reads most
-    /// naturally.
-    #[serde(flatten)]
-    pub action: Action,
-    /// Conditions that must all be met for this rule to apply.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub conditions: Vec<Condition>,
-    /// Whether this rule is active.
-    #[serde(default = "default_enabled")]
-    pub enabled: bool,
-}
-
-fn default_enabled() -> bool {
-    true
-}
diff --git a/crates/nvisy-engine/src/policy/selector.rs b/crates/nvisy-engine/src/policy/selector.rs
deleted file mode 100644
index b751f85bf..000000000
--- a/crates/nvisy-engine/src/policy/selector.rs
+++ /dev/null
@@ -1,72 +0,0 @@
-//! Entity selection criteria for policy rules.
-
-use hipstr::HipStr;
-use nvisy_core::entity::{Entity, EntityLabelCatalog, EntityLabelRef};
-use nvisy_core::primitive::ConfidenceThreshold;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use crate::modality::DocumentModality;
-
-/// Criteria for selecting which entities a policy rule applies to.
-///
-/// All fields use "empty means all" semantics: an empty `labels`
-/// list matches every label, an empty `tags` list matches every
-/// tag, and so on. When multiple fields are set, they are combined
-/// with AND logic.
-#[derive(Debug, Clone, Default, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct EntitySelector {
-    /// Specific entity labels this selector matches. Empty means
-    /// all labels.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub labels: Vec<EntityLabelRef>,
-    /// Tags this selector matches. An entity matches when its
-    /// label — looked up in the per-request catalog — carries any
-    /// of the listed tags. Labels not registered in the catalog
-    /// never match a tag selector and must be matched by name via
-    /// [`labels`].
-    ///
-    /// [`labels`]: Self::labels
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    #[schemars(with = "Vec<String>")]
-    pub tags: Vec<HipStr<'static>>,
-    /// Minimum detection confidence required. Entities below this
-    /// threshold are not matched. `None` means no threshold
-    /// (matches any confidence).
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub confidence_threshold: Option<ConfidenceThreshold>,
-}
-
-impl EntitySelector {
-    /// Create a selector that matches all entities.
-    pub fn all() -> Self {
-        Self::default()
-    }
-
-    /// Returns `true` if the given entity matches this selector,
-    /// resolving tag-based filters against `catalog`.
-    pub fn matches<M: DocumentModality>(
-        &self,
-        entity: &Entity<M>,
-        catalog: &EntityLabelCatalog,
-    ) -> bool {
-        if let Some(threshold) = self.confidence_threshold
-            && !threshold.admits(entity.confidence)
-        {
-            return false;
-        }
-        if !self.tags.is_empty() {
-            let Some(catalog_entry) = catalog.lookup(entity.label.as_str()) else {
-                return false;
-            };
-            if !self.tags.iter().any(|t| catalog_entry.has_tag(t)) {
-                return false;
-            }
-        }
-        if !self.labels.is_empty() && !self.labels.contains(&entity.label) {
-            return false;
-        }
-        true
-    }
-}
diff --git a/crates/nvisy-engine/src/redaction/applicator.rs b/crates/nvisy-engine/src/redaction/applicator.rs
deleted file mode 100644
index 9cb646efe..000000000
--- a/crates/nvisy-engine/src/redaction/applicator.rs
+++ /dev/null
@@ -1,352 +0,0 @@
-//! [`OverrideApplicator`]: apply human overrides to a detection's
-//! audit before the redaction phase consumes it.
-//!
-//! Inputs: an immutable [`AnyAudit`] from the detection +
-//! every override targeting it. Output: a mutated `AnyAudit`
-//! ready for the redaction phase. Provenance for each override
-//! is stamped onto the produced [`EntryMetadata::override_decision`]
-//! so reviewers see "what happened and why."
-//!
-//! Order of operations per audit:
-//!
-//! 1. Apply `Accept` / `Reject` / `Replace` overrides against
-//!    existing records (lookup by `Entity::id`).
-//! 2. For each `Add` override whose `location.kind()` matches
-//!    this audit's modality, synthesise an
-//!    [`EntityRecord<M>`] with a fresh UUID and append it. The
-//!    redaction phase's normal policy-resolution path runs the
-//!    chain for it; if the override pinned an operator, the
-//!    record arrives with a pre-resolved [`AuditEntry`] and the
-//!    redaction phase honours it directly.
-//!
-//! [`AnyAudit`]: crate::document::provenance::AnyAudit
-//! [`AuditEntry`]: crate::document::provenance::AuditEntry
-//! [`EntityRecord<M>`]: crate::document::provenance::EntityRecord
-//! [`EntryMetadata::override_decision`]: crate::document::provenance::EntryMetadata::override_decision
-
-use std::collections::{HashMap, HashSet};
-
-use jiff::Timestamp;
-use nvisy_core::Error;
-use nvisy_core::entity::{Entity, EntityLabelRef};
-use nvisy_core::modality::ModalityKind;
-use nvisy_core::primitive::Confidence;
-use uuid::Uuid;
-
-use super::override_::{RedactionAddEntity, RedactionOverride};
-use crate::document::provenance::{
-    AnyAudit, Audit, AuditEntry, Decision, EntityRecord, EntryMetadata, Execution,
-    RedactionDecision, ResolvedAction,
-};
-use crate::modality::DocumentModality;
-use crate::policy::AnyRedaction;
-
-const TARGET: &str = "nvisy_engine::pipeline::redaction::applicator";
-
-/// Apply a batch of overrides to a vector of audits.
-///
-/// Mutates `audits` in place. Returns an error when any override
-/// references an entity-id that isn't present in any of the
-/// audits — silently dropping such an override would let typos
-/// bypass redaction without anyone noticing.
-///
-/// # Errors
-///
-/// - [`ErrorKind::Validation`] when an override's
-///   `entity_id` doesn't resolve to any entity in the detection.
-/// - [`ErrorKind::Validation`] when a `Replace` operator's
-///   modality differs from the targeted entity's modality.
-/// - [`ErrorKind::Validation`] when an `Add` override's pinned
-///   operator modality differs from the location's modality
-///   (also caught earlier by `validate_overrides`; double-checked
-///   here as defence-in-depth).
-///
-/// [`ErrorKind::Validation`]: nvisy_core::ErrorKind::Validation
-pub(crate) fn apply_overrides(
-    audits: &mut [AnyAudit],
-    overrides: Vec<RedactionOverride>,
-) -> Result<(), Error> {
-    // Bucket overrides by target so we can validate "every
-    // entity-id-targeted override resolved" cheaply at the end.
-    let mut by_target: HashMap<Uuid, RedactionOverride> = HashMap::new();
-    let mut adds: Vec<RedactionAddEntity> = Vec::new();
-    for ov in overrides {
-        match ov {
-            RedactionOverride::Add(add) => adds.push(add),
-            other => {
-                let Some(target) = other.target() else {
-                    unreachable!("only Add has no target; matched above");
-                };
-                if let Some(existing) = by_target.insert(target, other) {
-                    return Err(Error::validation(
-                        format!(
-                            "duplicate override for entity {target} (had {existing:?})",
-                            existing = std::mem::discriminant(&existing),
-                        ),
-                        TARGET,
-                    ));
-                }
-            }
-        }
-    }
-
-    // Apply per-audit. Track which targeted overrides were
-    // consumed so unresolved ones surface as errors.
-    let mut consumed: HashSet<Uuid> = HashSet::new();
-    for any in audits.iter_mut() {
-        match any {
-            AnyAudit::Text(a) => apply_to(a, &by_target, &mut consumed, AnyRedaction::try_as_text)?,
-            AnyAudit::Tabular(a) => {
-                apply_to(a, &by_target, &mut consumed, AnyRedaction::try_as_tabular)?
-            }
-            AnyAudit::Image(a) => {
-                apply_to(a, &by_target, &mut consumed, AnyRedaction::try_as_image)?
-            }
-            AnyAudit::Audio(a) => {
-                apply_to(a, &by_target, &mut consumed, AnyRedaction::try_as_audio)?
-            }
-        }
-    }
-
-    // Any targeted override that wasn't consumed is a typo or
-    // stale id — fail loudly rather than silently drop.
-    for target in by_target.keys() {
-        if !consumed.contains(target) {
-            return Err(Error::validation(
-                format!("override targets entity {target} not present in detection"),
-                TARGET,
-            ));
-        }
-    }
-
-    // Apply `Add` overrides last. Match each `AnyLocation` variant
-    // to the corresponding audit (first audit of matching
-    // modality) so the synthesised entity lands in the right
-    // document. If no audit of matching modality exists, fail —
-    // we can't redact bytes that aren't in the run.
-    for add in adds {
-        append_add(audits, add)?;
-    }
-    Ok(())
-}
-
-/// Apply Accept/Reject/Replace overrides to one typed audit.
-/// Records consumed override targets in `consumed`.
-///
-/// `project_operator` is the per-modality `AnyRedaction::try_as_*`
-/// the caller supplies based on `M`; keeps the helper generic
-/// without a `TryFrom<AnyRedaction>` trait bound on
-/// `M::Redaction`.
-fn apply_to<M, F>(
-    audit: &mut Audit<M>,
-    overrides: &HashMap<Uuid, RedactionOverride>,
-    consumed: &mut HashSet<Uuid>,
-    project_operator: F,
-) -> Result<(), Error>
-where
-    M: DocumentModality,
-    F: Fn(AnyRedaction) -> Option<M::Redaction>,
-{
-    for record in &mut audit.records {
-        let id = record.entity.id;
-        let Some(ov) = overrides.get(&id) else {
-            continue;
-        };
-        consumed.insert(id);
-        match ov {
-            RedactionOverride::Accept { .. } => {
-                stamp_provenance(record, RedactionDecision::OverrideAccept);
-            }
-            RedactionOverride::Reject { .. } => {
-                let entry = record.audit.take().unwrap_or_else(|| AuditEntry {
-                    decision: Decision {
-                        policy_ref: None,
-                        action: ResolvedAction::Suppress { reason: None },
-                    },
-                    execution: Execution::Suppressed,
-                    metadata: EntryMetadata::now().with_override(RedactionDecision::OverrideReject),
-                });
-                let mut entry = entry;
-                entry.execution = Execution::Suppressed;
-                entry.metadata = entry
-                    .metadata
-                    .with_override(RedactionDecision::OverrideReject);
-                record.audit = Some(entry);
-            }
-            RedactionOverride::Replace {
-                entity_id: _,
-                operator,
-            } => {
-                let typed = project_operator(operator.clone()).ok_or_else(|| {
-                    Error::validation(
-                        format!(
-                            "override Replace for entity {id} carries operator of modality {:?} but entity is modality {:?}",
-                            operator.modality(),
-                            ModalityKind::of::<M>(),
-                        ),
-                        TARGET,
-                    )
-                })?;
-                let prior = record.audit.take();
-                let policy_ref = prior.as_ref().and_then(|e| e.decision.policy_ref.clone());
-                record.audit = Some(AuditEntry {
-                    decision: Decision {
-                        policy_ref,
-                        action: ResolvedAction::Redact { operator: typed },
-                    },
-                    execution: Execution::Pending,
-                    metadata: EntryMetadata::now()
-                        .with_override(RedactionDecision::OverrideReplace),
-                });
-            }
-            RedactionOverride::Add(_) => {
-                unreachable!("Add overrides handled separately in append_add")
-            }
-        }
-    }
-    Ok(())
-}
-
-/// Append a synthesised entity to the matching audit.
-fn append_add(audits: &mut [AnyAudit], add: RedactionAddEntity) -> Result<(), Error> {
-    let kind = add.location.kind();
-
-    // Optional pinned operator: if it disagrees with the
-    // location's modality, fail.
-    if let Some(op) = &add.operator
-        && op.modality() != kind
-    {
-        return Err(Error::validation(
-            format!(
-                "override Add operator modality {:?} differs from location modality {kind:?}",
-                op.modality(),
-            ),
-            TARGET,
-        ));
-    }
-
-    // Find a target audit of matching modality.
-    let target = audits.iter_mut().find(|a| {
-        matches!(
-            (a, kind),
-            (AnyAudit::Text(_), ModalityKind::Text)
-                | (AnyAudit::Tabular(_), ModalityKind::Tabular)
-                | (AnyAudit::Image(_), ModalityKind::Image)
-                | (AnyAudit::Audio(_), ModalityKind::Audio)
-        )
-    });
-    let Some(target) = target else {
-        return Err(Error::validation(
-            format!(
-                "override Add for modality {kind:?} has no audit of that modality in the detection",
-            ),
-            TARGET,
-        ));
-    };
-
-    match target {
-        AnyAudit::Text(a) => {
-            let loc = add.location.try_as_text().ok_or_else(modality_mismatch)?;
-            let op = add
-                .operator
-                .map(|o| o.try_as_text().ok_or_else(modality_mismatch))
-                .transpose()?;
-            append_typed(a, add.label, loc, op);
-        }
-        AnyAudit::Tabular(a) => {
-            let loc = add
-                .location
-                .try_as_tabular()
-                .ok_or_else(modality_mismatch)?;
-            let op = add
-                .operator
-                .map(|o| o.try_as_tabular().ok_or_else(modality_mismatch))
-                .transpose()?;
-            append_typed(a, add.label, loc, op);
-        }
-        AnyAudit::Image(a) => {
-            let loc = add.location.try_as_image().ok_or_else(modality_mismatch)?;
-            let op = add
-                .operator
-                .map(|o| o.try_as_image().ok_or_else(modality_mismatch))
-                .transpose()?;
-            append_typed(a, add.label, loc, op);
-        }
-        AnyAudit::Audio(a) => {
-            let loc = add.location.try_as_audio().ok_or_else(modality_mismatch)?;
-            let op = add
-                .operator
-                .map(|o| o.try_as_audio().ok_or_else(modality_mismatch))
-                .transpose()?;
-            append_typed(a, add.label, loc, op);
-        }
-    }
-    Ok(())
-}
-
-fn modality_mismatch() -> Error {
-    // Caught earlier by `validate_overrides` and the kind-match
-    // above; double-checked here as defence-in-depth.
-    Error::validation(
-        "internal: append_typed called with mismatched modality",
-        TARGET,
-    )
-}
-
-/// Append a synthesised entity to a typed audit.
-fn append_typed<M>(
-    audit: &mut Audit<M>,
-    label: EntityLabelRef,
-    location: M::Location,
-    operator: Option<M::Redaction>,
-) where
-    M: DocumentModality,
-{
-    let entity = Entity {
-        id: Uuid::now_v7(),
-        entity_id: None,
-        label,
-        location,
-        confidence: Confidence::clamped(1.0),
-        trail: Vec::new(),
-        language: None,
-    };
-    let prebuilt = operator.map(|op| AuditEntry {
-        decision: Decision {
-            policy_ref: None,
-            action: ResolvedAction::Redact { operator: op },
-        },
-        execution: Execution::Pending,
-        metadata: EntryMetadata {
-            timestamp: Some(Timestamp::now()),
-            correlation_id: None,
-            override_decision: Some(RedactionDecision::OverrideAdd),
-        },
-    });
-    audit.records.push(EntityRecord {
-        entity,
-        audit: prebuilt,
-    });
-}
-
-/// Stamp `record.audit.metadata.override_decision` with `tag`.
-/// If the record has no audit entry yet, none is created — the
-/// redaction phase will fill it in via policy resolution; the
-/// override tag stays "OverrideAccept" applied at that point.
-fn stamp_provenance<M>(record: &mut EntityRecord<M>, tag: RedactionDecision)
-where
-    M: DocumentModality,
-{
-    if let Some(entry) = record.audit.as_mut() {
-        entry.metadata = entry.metadata.clone().with_override(tag);
-    } else {
-        record.audit = Some(AuditEntry {
-            decision: Decision {
-                policy_ref: None,
-                action: ResolvedAction::Suppress { reason: None },
-            },
-            execution: Execution::Pending,
-            metadata: EntryMetadata::now().with_override(tag),
-        });
-    }
-}
diff --git a/crates/nvisy-engine/src/redaction/config.rs b/crates/nvisy-engine/src/redaction/config.rs
deleted file mode 100644
index 63c3c8d3b..000000000
--- a/crates/nvisy-engine/src/redaction/config.rs
+++ /dev/null
@@ -1,66 +0,0 @@
-//! `[redaction]` config section: deployment-wide fallback values
-//! for the per-plan [`Redaction`] node's optional fields.
-//!
-//! Matches the `*Config` naming used by [`EngineConfig`],
-//! [`ExtractionConfig`], [`DetectionConfig`] so every
-//! [`RuntimeConfig`] subsystem section shares one shape: a struct
-//! deserialized from one TOML section, built once at engine
-//! startup, shared across runs via `Arc`.
-//!
-//! Per-request redaction plan knobs ([`Redaction`], [`Validation`])
-//! live in [`super::plan`].
-//!
-//! [`Redaction`]: super::plan::Redaction
-//! [`Validation`]: super::plan::Validation
-//! [`EngineConfig`]: crate::core::EngineConfig
-//! [`ExtractionConfig`]: crate::detection::ExtractionConfig
-//! [`DetectionConfig`]: crate::detection::DetectionConfig
-//! [`RuntimeConfig`]: crate::core::RuntimeConfig
-
-use nvisy_core::primitive::ConfidenceThreshold;
-use serde::{Deserialize, Serialize};
-
-use crate::policy::redaction::ModalityRedactions;
-
-/// Deployment-wide redaction defaults.
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct RedactionConfig {
-    /// Default minimum confidence threshold for entities that
-    /// don't match a policy rule. The plan
-    /// [`Redaction::confidence_threshold`][rct] overrides this
-    /// when set. Defaults to `0.5`.
-    ///
-    /// [rct]: super::plan::Redaction::confidence_threshold
-    #[serde(default = "default_confidence_threshold")]
-    pub confidence_threshold: ConfidenceThreshold,
-    /// Default for whether to strip embedded document metadata
-    /// (EXIF, PDF properties). The plan
-    /// [`Redaction::process_metadata`][rpm] overrides this when
-    /// set. Defaults to `false`.
-    ///
-    /// [rpm]: super::plan::Redaction::process_metadata
-    #[serde(default)]
-    pub process_metadata: bool,
-    /// Deployment-wide fallback operators: when a policy rule's
-    /// [`Action::Redact`][crate::policy::Action::Redact] doesn't
-    /// cover the entity's modality, the redaction phase falls
-    /// back to whatever is set here for that modality. Missing
-    /// entries at both levels cause the rule to silently skip —
-    /// operators should be configured at one of the two layers.
-    #[serde(default)]
-    pub default_operators: ModalityRedactions,
-}
-
-impl Default for RedactionConfig {
-    fn default() -> Self {
-        Self {
-            confidence_threshold: default_confidence_threshold(),
-            process_metadata: false,
-            default_operators: ModalityRedactions::default(),
-        }
-    }
-}
-
-fn default_confidence_threshold() -> ConfidenceThreshold {
-    ConfidenceThreshold::clamped(0.5)
-}
diff --git a/crates/nvisy-engine/src/redaction/document.rs b/crates/nvisy-engine/src/redaction/document.rs
deleted file mode 100644
index f5c802bb8..000000000
--- a/crates/nvisy-engine/src/redaction/document.rs
+++ /dev/null
@@ -1,88 +0,0 @@
-//! Per-document phase sequence for redaction: redaction →
-//! validation.
-
-use nvisy_core::modality::{Audio, Image, Tabular, Text};
-use nvisy_core::{Error, Result};
-
-use crate::core::{DocumentTree, PhaseContext as _, RedactionContext};
-use crate::redaction::RedactionPlan;
-use crate::redaction::phases::RedactionPhase;
-use crate::redaction::phases::validation::ValidationPhase;
-
-const TARGET: &str = "nvisy_engine::pipeline::redaction::document";
-
-pub(super) struct RedactionDocumentPipeline {
-    redaction: RedactionPhase,
-    validation: ValidationPhase,
-}
-
-impl RedactionDocumentPipeline {
-    pub(super) fn from_context(ctx: &RedactionContext) -> Self {
-        Self {
-            redaction: RedactionPhase::new(
-                ctx.redaction_config().clone(),
-                ctx.redaction_registries().clone(),
-            ),
-            validation: ValidationPhase::new(),
-        }
-    }
-
-    pub(super) async fn run_text(
-        &self,
-        ctx: &RedactionContext,
-        plan: &RedactionPlan,
-        tree: &mut DocumentTree<Text>,
-    ) -> Result<()> {
-        check_cancelled(ctx)?;
-        self.redaction.apply_text(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.validation.apply_text(ctx, plan, tree).await?;
-        check_cancelled(ctx)
-    }
-
-    pub(super) async fn run_tabular(
-        &self,
-        ctx: &RedactionContext,
-        plan: &RedactionPlan,
-        tree: &mut DocumentTree<Tabular>,
-    ) -> Result<()> {
-        check_cancelled(ctx)?;
-        self.redaction.apply_tabular(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.validation.apply_tabular(ctx, plan, tree).await?;
-        check_cancelled(ctx)
-    }
-
-    pub(super) async fn run_image(
-        &self,
-        ctx: &RedactionContext,
-        plan: &RedactionPlan,
-        tree: &mut DocumentTree<Image>,
-    ) -> Result<()> {
-        check_cancelled(ctx)?;
-        self.redaction.apply_image(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.validation.apply_image(ctx, plan, tree).await?;
-        check_cancelled(ctx)
-    }
-
-    pub(super) async fn run_audio(
-        &self,
-        ctx: &RedactionContext,
-        plan: &RedactionPlan,
-        tree: &mut DocumentTree<Audio>,
-    ) -> Result<()> {
-        check_cancelled(ctx)?;
-        self.redaction.apply_audio(ctx, plan, tree).await?;
-        check_cancelled(ctx)?;
-        self.validation.apply_audio(ctx, plan, tree).await?;
-        check_cancelled(ctx)
-    }
-}
-
-fn check_cancelled(ctx: &RedactionContext) -> Result<()> {
-    if ctx.is_cancelled() {
-        return Err(Error::cancellation("redaction cancelled", TARGET));
-    }
-    Ok(())
-}
diff --git a/crates/nvisy-engine/src/redaction/engine.rs b/crates/nvisy-engine/src/redaction/engine.rs
deleted file mode 100644
index 75ba09ba9..000000000
--- a/crates/nvisy-engine/src/redaction/engine.rs
+++ /dev/null
@@ -1,303 +0,0 @@
-//! [`RedactionEngine`] — entry point for running redaction passes.
-//!
-//! Constructed from a paired [`DetectionEngine`]: the redaction
-//! engine borrows the detection engine's [`Registry`], runtime
-//! config, optional key provider, and an in-memory
-//! [`DetectionState`] read-handle so the redact path can resolve a
-//! prior detection by id without crossing the persistent
-//! registry. Owns the redaction-side resources independently
-//! (redaction config, custom-anonymizer registries, in-memory
-//! redaction-pass tracker, background task set).
-//!
-//! [`DetectionEngine`]: super::super::detection::DetectionEngine
-//! [`Registry`]: crate::registry::Registry
-//! [`DetectionState`]: super::super::detection::DetectionState
-
-use std::sync::Arc;
-use std::{fmt, mem};
-
-use nvisy_core::Error;
-use tokio::sync::Mutex;
-use tokio::task::JoinSet;
-use uuid::Uuid;
-
-use super::pipeline::{RedactionEngineState, RedactionPipeline};
-use super::result::RedactionResult;
-use super::state::RedactionState;
-use super::status::RedactionStatus;
-use super::{RedactionEntry, RedactionFilter, RedactionInput, RedactionSnapshot};
-use crate::core::RuntimeConfig;
-use crate::core::ingestion::encryption::SharedKeyProvider;
-use crate::detection::{DetectionEngine, DetectionState};
-use crate::redaction::RedactionConfig;
-use crate::redaction::phases::RedactionRegistries;
-use crate::registry::Registry;
-
-/// Shared inner state for a [`RedactionEngine`], held behind an
-/// `Arc`. Cloning the engine is a refcount bump on this inner.
-pub(super) struct RedactionInner {
-    pub runtime_config: RuntimeConfig,
-    pub registry: Registry,
-    pub key_provider: Option<SharedKeyProvider>,
-    pub redaction_config: Arc<RedactionConfig>,
-    pub redaction_registries: Arc<RedactionRegistries>,
-    /// Read-handle to the paired [`DetectionEngine`][de]'s in-memory
-    /// detection state. The redact path's handoff step resolves a
-    /// prior detection-id through this without crossing the
-    /// persistent registry.
-    ///
-    /// [de]: super::super::detection::DetectionEngine
-    pub detections: DetectionState,
-    /// In-memory redaction-pass tracker.
-    pub redactions: RedactionState,
-    /// Background pipeline tasks spawned by [`RedactionEngine::redact`].
-    pub background_tasks: Mutex<JoinSet<()>>,
-}
-
-/// Redaction pipeline engine.
-///
-/// Thin facade over `Arc<RedactionInner>`; cloning is cheap.
-/// Builder methods ([`with_redaction_registries`]) require
-/// exclusive access via `Arc::get_mut` and must be called before
-/// the engine is cloned.
-///
-/// [`with_redaction_registries`]: RedactionEngine::with_redaction_registries
-#[derive(Clone)]
-pub struct RedactionEngine {
-    pub(super) inner: Arc<RedactionInner>,
-}
-
-impl fmt::Debug for RedactionEngine {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("RedactionEngine")
-            .field("runtime_config", &self.inner.runtime_config)
-            .finish()
-    }
-}
-
-impl RedactionEngine {
-    /// Build a redaction engine paired with `detection`.
-    ///
-    /// Shares the detection engine's [`Registry`], runtime
-    /// configuration, optional key provider, and in-memory
-    /// [`DetectionState`] (so the redact handoff path is a fast
-    /// in-memory lookup, not a disk round-trip). Owns the
-    /// redaction-side state (config, registries, in-memory
-    /// tracker, background tasks) independently — calling
-    /// [`shutdown`] on this engine drains only the redaction
-    /// background tasks.
-    ///
-    /// [`Registry`]: crate::registry::Registry
-    /// [`DetectionState`]: crate::detection::DetectionState
-    /// [`shutdown`]: Self::shutdown
-    pub fn from_detection(detection: &DetectionEngine) -> Self {
-        let redaction_config = Arc::new(detection.config().redaction.clone().unwrap_or_default());
-        Self {
-            inner: Arc::new(RedactionInner {
-                runtime_config: detection.config().clone(),
-                registry: detection.registry().clone(),
-                key_provider: detection.key_provider().cloned(),
-                redaction_config,
-                redaction_registries: Arc::new(RedactionRegistries::default()),
-                detections: detection.detections().clone(),
-                redactions: RedactionState::new(),
-                background_tasks: Mutex::new(JoinSet::new()),
-            }),
-        }
-    }
-
-    /// Install the per-modality custom-anonymizer registries the
-    /// redaction phase consults for `Custom`-arm operator lookups.
-    ///
-    /// Replaces whatever registries the engine was constructed
-    /// with (an empty default). Deployments call this once at
-    /// startup after building their [`RedactionRegistries`].
-    ///
-    /// # Panics
-    ///
-    /// Panics if the engine has already been cloned (Arc is shared).
-    pub fn with_redaction_registries(mut self, registries: RedactionRegistries) -> Self {
-        Arc::get_mut(&mut self.inner)
-            .expect("redaction engine must not be shared during construction")
-            .redaction_registries = Arc::new(registries);
-        self
-    }
-
-    /// Returns the base runtime configuration before per-request
-    /// overrides.
-    pub fn config(&self) -> &RuntimeConfig {
-        &self.inner.runtime_config
-    }
-
-    /// Returns the content and audit registry.
-    pub fn registry(&self) -> &Registry {
-        &self.inner.registry
-    }
-
-    /// Construct a new redaction pipeline bound to this engine's
-    /// shared state.
-    fn pipeline(&self) -> RedactionPipeline {
-        RedactionPipeline::new(
-            self.inner.registry.clone(),
-            self.inner.key_provider.clone(),
-            self.inner.detections.clone(),
-            self.inner.redactions.clone(),
-            self.inner.runtime_config.clone(),
-            RedactionEngineState {
-                redaction_config: Arc::clone(&self.inner.redaction_config),
-                redaction_registries: Arc::clone(&self.inner.redaction_registries),
-            },
-        )
-    }
-
-    /// Submit a redaction pass for background execution.
-    ///
-    /// Registers the pass synchronously as `Pending`, then spawns
-    /// execution on a background task. Returns the redaction id
-    /// immediately. Poll [`Self::get_redaction`] for status.
-    ///
-    /// Validation that runs synchronously before the task spawns:
-    /// nothing. The detection lookup, override validation, and
-    /// override application happen inside the background task so
-    /// they're observable via `get_redaction` rather than returned
-    /// from `redact`. This keeps the submission shape symmetric
-    /// with [`DetectionEngine::detect`][de].
-    ///
-    /// # Errors
-    ///
-    /// Currently infallible at submit time. Pass-level failures
-    /// (detection not found, override validation, override
-    /// application, document errors) are recorded on the
-    /// in-memory state and observable via
-    /// [`Self::get_redaction`].
-    ///
-    /// [de]: super::super::detection::DetectionEngine::detect
-    pub async fn redact(&self, input: RedactionInput) -> Result<Uuid, Error> {
-        let pipeline = self.pipeline();
-        let redaction_id = pipeline.id();
-        pipeline.register_pending(&input).await;
-
-        self.inner.background_tasks.lock().await.spawn(async move {
-            if let Err(e) = pipeline.execute(input).await {
-                tracing::error!(
-                    %redaction_id,
-                    error = %e,
-                    "redaction pass failed",
-                );
-            }
-        });
-        Ok(redaction_id)
-    }
-
-    /// Look up a redaction-pass snapshot. Hydrates from the
-    /// registry when the in-memory record is absent.
-    ///
-    /// # Errors
-    ///
-    /// [`nvisy_core::ErrorKind::NotFound`] when neither store has
-    /// a matching entry.
-    pub async fn get_redaction(
-        &self,
-        actor_id: Uuid,
-        id: Uuid,
-    ) -> Result<RedactionSnapshot, Error> {
-        match self.inner.redactions.snapshot(actor_id, id).await {
-            Ok(snap) => Ok(snap),
-            Err(e) if e.kind() == nvisy_core::ErrorKind::NotFound => {
-                match self.inner.registry.load_redaction(actor_id, id).await {
-                    Ok(result) => Ok(hydrate_snapshot(result)),
-                    Err(_) => Err(e),
-                }
-            }
-            Err(e) => Err(e),
-        }
-    }
-
-    /// List the actor's redaction passes.
-    pub async fn list_redactions(
-        &self,
-        actor_id: Uuid,
-        filter: RedactionFilter,
-    ) -> Vec<RedactionEntry> {
-        self.inner.redactions.list(actor_id, filter).await
-    }
-
-    /// Cancel an in-progress redaction.
-    ///
-    /// # Errors
-    ///
-    /// - [`nvisy_core::ErrorKind::NotFound`] when missing or wrong
-    ///   actor.
-    /// - [`nvisy_core::ErrorKind::Validation`] when terminal.
-    pub async fn cancel_redaction(&self, actor_id: Uuid, id: Uuid) -> Result<(), Error> {
-        self.inner.redactions.cancel(actor_id, id).await
-    }
-
-    /// Delete a finished redaction. Cascades to the registry.
-    ///
-    /// # Errors
-    ///
-    /// - [`nvisy_core::ErrorKind::NotFound`] when missing or wrong
-    ///   actor.
-    /// - [`nvisy_core::ErrorKind::Validation`] when still active.
-    pub async fn delete_redaction(&self, actor_id: Uuid, id: Uuid) -> Result<(), Error> {
-        self.inner.redactions.delete(actor_id, id).await?;
-        if let Err(e) = self.inner.registry.unregister_redaction(actor_id, id).await {
-            tracing::warn!(
-                redaction_id = %id,
-                error = %e,
-                "failed to remove persisted redaction result",
-            );
-        }
-        Ok(())
-    }
-
-    /// Delete every finished redaction for the actor. Active
-    /// passes are preserved. Returns the number of entries
-    /// removed. Cascade failures are logged but don't unblock the
-    /// result count.
-    pub async fn delete_all_redactions(&self, actor_id: Uuid) -> usize {
-        let removed = self.inner.redactions.delete_all(actor_id).await;
-        for id in &removed {
-            if let Err(e) = self
-                .inner
-                .registry
-                .unregister_redaction(actor_id, *id)
-                .await
-            {
-                tracing::warn!(
-                    redaction_id = %id,
-                    error = %e,
-                    "failed to remove persisted redaction result during bulk delete",
-                );
-            }
-        }
-        removed.len()
-    }
-
-    /// Wait for all background redaction tasks to complete.
-    ///
-    /// Call during graceful shutdown to ensure in-flight passes
-    /// finish and persist their results before the process exits.
-    pub async fn shutdown(&self) {
-        let tasks = mem::take(&mut *self.inner.background_tasks.lock().await);
-        tasks.join_all().await;
-    }
-}
-
-/// Build a synthetic [`RedactionSnapshot`] from a hydrated
-/// [`RedactionResult`] loaded from disk. Historical metadata
-/// collapses to "now" so the snapshot is non-empty.
-fn hydrate_snapshot(result: RedactionResult) -> RedactionSnapshot {
-    RedactionSnapshot {
-        id: result.id,
-        detection_id: result.detection_id,
-        actor_id: result.actor_id,
-        status: RedactionStatus::Succeeded,
-        created_at: jiff::Timestamp::now(),
-        started_at: None,
-        completed_at: Some(jiff::Timestamp::now()),
-        result: Some(result),
-        error: None,
-    }
-}
diff --git a/crates/nvisy-engine/src/redaction/input.rs b/crates/nvisy-engine/src/redaction/input.rs
deleted file mode 100644
index a1328e5fb..000000000
--- a/crates/nvisy-engine/src/redaction/input.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-//! [`RedactionInput`]: arguments to [`RedactionEngine::redact`].
-//!
-//! [`RedactionEngine::redact`]: crate::redaction::RedactionEngine::redact
-
-use uuid::Uuid;
-
-use super::override_::RedactionOverride;
-use crate::core::ingestion::ExportFile;
-use crate::redaction::RedactionPlan;
-
-/// Input required to execute a redaction pass.
-///
-/// References a previously completed detection by id. The
-/// detection's audits supply the entities and the operator each
-/// policy chain picked; [`overrides`] lets the caller approve /
-/// reject / replace specific decisions or add entities the
-/// recognisers missed.
-///
-/// [`overrides`]: Self::overrides
-#[derive(Clone)]
-pub struct RedactionInput {
-    /// Identity of the actor initiating the redaction. Must match
-    /// the actor that owns the referenced detection.
-    pub actor_id: Uuid,
-    /// Detection pass to redact against.
-    pub detection_id: Uuid,
-    /// Per-entity decision overrides + added entities the
-    /// recognisers missed. Empty means "apply the detection
-    /// verbatim."
-    pub overrides: Vec<RedactionOverride>,
-    /// Per-phase behaviour knobs for the redaction / validation
-    /// phases (validation thresholds, severity filters, etc.).
-    pub plan: RedactionPlan,
-    /// Sinks to write redacted content to.
-    pub exports: Vec<ExportFile>,
-}
diff --git a/crates/nvisy-engine/src/redaction/mod.rs b/crates/nvisy-engine/src/redaction/mod.rs
deleted file mode 100644
index 58e3e7cec..000000000
--- a/crates/nvisy-engine/src/redaction/mod.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-//! Redaction subsystem: standalone apply + validate + export
-//! pipeline.
-//!
-//! [`RedactionEngine::redact`] takes a [`RedactionInput`]
-//! referencing a prior [`DetectionResult`] plus optional
-//! [`RedactionOverride`]s and exports. It re-opens the original
-//! imports, applies the (overridden) decisions, runs validation,
-//! and writes to the configured exports.
-//!
-//! [`DetectionResult`]: crate::detection::DetectionResult
-
-mod applicator;
-mod config;
-mod document;
-mod engine;
-mod input;
-mod orchestrator;
-mod override_;
-pub mod phases;
-mod pipeline;
-mod plan;
-mod result;
-mod state;
-mod status;
-
-pub use self::config::RedactionConfig;
-pub use self::engine::RedactionEngine;
-pub use self::input::RedactionInput;
-pub use self::override_::{
-    RedactionAddEntity, RedactionDecision, RedactionOverride, validate_overrides,
-};
-pub use self::plan::{Redaction, RedactionPlan, Validation};
-pub use self::result::{RedactionEntry, RedactionFilter, RedactionResult, RedactionSnapshot};
-pub use self::status::RedactionStatus;
diff --git a/crates/nvisy-engine/src/redaction/orchestrator.rs b/crates/nvisy-engine/src/redaction/orchestrator.rs
deleted file mode 100644
index ceb521112..000000000
--- a/crates/nvisy-engine/src/redaction/orchestrator.rs
+++ /dev/null
@@ -1,273 +0,0 @@
-//! Redaction orchestrator: re-imports content from the
-//! detection's import refs, replays audits onto the freshly-built
-//! trees, fans out per-document redaction tasks, runs export.
-//!
-//! The orchestrator deliberately does no recognition / detection /
-//! deduplication — the audit from the detection IS the source of
-//! truth for what entities to redact. Overrides have already been
-//! applied to that audit by [`apply_overrides`] before this
-//! orchestrator sees it.
-//!
-//! [`apply_overrides`]: super::applicator::apply_overrides
-
-use std::collections::HashMap;
-use std::sync::Arc;
-
-use nvisy_core::Error;
-use nvisy_core::entity::ContentSource;
-use tokio::sync::Semaphore;
-use tokio::task::JoinSet;
-
-use super::document::RedactionDocumentPipeline;
-use crate::core::ingestion::{ExportFile, Exporter, ImportFile, Importer};
-use crate::core::{AnyTree, PhaseContext as _, RedactionContext};
-use crate::document::provenance::AnyAudit;
-use crate::redaction::RedactionPlan;
-
-const TARGET: &str = "nvisy_engine::pipeline::redaction::orchestrator";
-
-pub(super) struct RedactionDocumentResult {
-    pub tree: Option<AnyTree>,
-    pub error: Option<String>,
-}
-
-pub(super) struct RedactionOutput {
-    pub results: Vec<RedactionDocumentResult>,
-}
-
-impl RedactionOutput {
-    pub(super) fn into_audits(self) -> (Vec<AnyAudit>, u64, bool, bool) {
-        let mut audits = Vec::new();
-        let mut applied = 0u64;
-        let mut any_ok = false;
-        let mut any_failed = false;
-        for r in self.results {
-            if r.error.is_some() {
-                any_failed = true;
-                continue;
-            }
-            any_ok = true;
-            if let Some(tree) = r.tree {
-                let audit = tree.audit_cloned();
-                applied += audit.applied_redactions_count() as u64;
-                audits.push(audit);
-            }
-        }
-        (audits, applied, any_ok, any_failed)
-    }
-}
-
-pub(super) struct RedactionOrchestrator {
-    ctx: Arc<RedactionContext>,
-    semaphore: Option<Arc<Semaphore>>,
-}
-
-impl RedactionOrchestrator {
-    pub(super) fn new(ctx: RedactionContext) -> Self {
-        let semaphore = ctx.concurrency().map(|c| Arc::new(Semaphore::new(c.get())));
-        Self {
-            ctx: Arc::new(ctx),
-            semaphore,
-        }
-    }
-
-    /// Run the redaction pass.
-    ///
-    /// `imports` are re-opened from the registry; `exports` are
-    /// the sinks the caller wants written; `plan` carries the
-    /// per-phase knobs; `prepared_audits` is the detection's
-    /// audit with overrides already applied.
-    pub(super) async fn run(
-        &self,
-        imports: &[ImportFile],
-        exports: &[ExportFile],
-        plan: &RedactionPlan,
-        prepared_audits: Vec<AnyAudit>,
-    ) -> Result<RedactionOutput, Error> {
-        let trees = self.run_imports(imports).await?;
-        let trees = replay_audits_into_trees(trees, prepared_audits)?;
-
-        let pipeline = Arc::new(RedactionDocumentPipeline::from_context(&self.ctx));
-        let plan = Arc::new(plan.clone());
-        let exports = Arc::new(exports.to_vec());
-        let mut join_set: JoinSet<RedactionDocumentResult> = JoinSet::new();
-        for tree in trees {
-            let ctx = Arc::clone(&self.ctx);
-            let sem = self.semaphore.clone();
-            let plan = Arc::clone(&plan);
-            let exports = Arc::clone(&exports);
-            let pipeline = Arc::clone(&pipeline);
-            join_set.spawn(run_one(pipeline, tree, ctx, sem, plan, exports));
-        }
-
-        let mut results = Vec::new();
-        while let Some(joined) = join_set.join_next().await {
-            match joined {
-                Ok(r) => results.push(r),
-                Err(e) => {
-                    let msg = if e.is_panic() {
-                        let payload = e.into_panic();
-                        let panic_msg = payload
-                            .downcast_ref::<&str>()
-                            .copied()
-                            .or_else(|| payload.downcast_ref::<String>().map(|s| s.as_str()))
-                            .unwrap_or("unknown panic");
-                        tracing::error!(
-                            target: TARGET,
-                            panic = panic_msg,
-                            "redaction document task panicked"
-                        );
-                        format!("Redaction task panicked: {panic_msg}")
-                    } else {
-                        tracing::error!(target: TARGET, error = %e, "redaction document task failed");
-                        format!("Redaction task failed: {e}")
-                    };
-                    results.push(RedactionDocumentResult {
-                        tree: None,
-                        error: Some(msg),
-                    });
-                }
-            }
-        }
-        Ok(RedactionOutput { results })
-    }
-
-    async fn run_imports(&self, imports: &[ImportFile]) -> Result<Vec<AnyTree>, Error> {
-        let mut trees = Vec::new();
-        for cfg in imports {
-            let importer = Importer::new()
-                .with_decompression(cfg.decompression)
-                .with_decryption(cfg.decryption.clone());
-            let shared = self.ctx.shared();
-            for &content_id in &cfg.content_ids {
-                if self.ctx.is_cancelled() {
-                    return Err(Error::cancellation("redaction cancelled", TARGET));
-                }
-                tracing::debug!(target: TARGET, %content_id, "re-importing content for redaction");
-                let handle = shared
-                    .registry
-                    .read_content(shared.actor_id, content_id)
-                    .await?;
-                let content = handle.content().await?;
-                let annotations = shared
-                    .registry
-                    .load_annotations(shared.actor_id, content_id)
-                    .await?;
-                trees.extend(importer.import(content, annotations, shared).await?);
-            }
-        }
-        Ok(trees)
-    }
-}
-
-/// Replay a list of `AnyAudit` onto a list of `AnyTree`, matching
-/// each audit to a tree of matching modality and source.
-///
-/// # Errors
-///
-/// [`ErrorKind::Validation`] when an audit's source has no
-/// matching tree (the detection saw an envelope the redaction
-/// re-import didn't produce — content drifted or import
-/// semantics changed).
-///
-/// [`ErrorKind::Validation`]: nvisy_core::ErrorKind::Validation
-fn replay_audits_into_trees(
-    trees: Vec<AnyTree>,
-    audits: Vec<AnyAudit>,
-) -> Result<Vec<AnyTree>, Error> {
-    let mut by_key: HashMap<(ContentSource, &'static str), usize> = HashMap::new();
-    for (idx, tree) in trees.iter().enumerate() {
-        let source = match tree {
-            AnyTree::Text(t) => t.root.audit.source,
-            AnyTree::Tabular(t) => t.root.audit.source,
-            AnyTree::Image(t) => t.root.audit.source,
-            AnyTree::Audio(t) => t.root.audit.source,
-        };
-        by_key.insert((source, tree.modality_name()), idx);
-    }
-
-    let mut trees = trees;
-    for audit in audits {
-        let (source, modality_name) = match &audit {
-            AnyAudit::Text(a) => (a.source, "text"),
-            AnyAudit::Tabular(a) => (a.source, "tabular"),
-            AnyAudit::Image(a) => (a.source, "image"),
-            AnyAudit::Audio(a) => (a.source, "audio"),
-        };
-        let Some(&idx) = by_key.get(&(source, modality_name)) else {
-            return Err(Error::validation(
-                format!(
-                    "audit for source {source:?} ({modality_name}) has no matching tree after re-import",
-                ),
-                TARGET,
-            ));
-        };
-        match (&mut trees[idx], audit) {
-            (AnyTree::Text(t), AnyAudit::Text(a)) => t.root.audit = a,
-            (AnyTree::Tabular(t), AnyAudit::Tabular(a)) => t.root.audit = a,
-            (AnyTree::Image(t), AnyAudit::Image(a)) => t.root.audit = a,
-            (AnyTree::Audio(t), AnyAudit::Audio(a)) => t.root.audit = a,
-            _ => unreachable!("modality matched via key lookup"),
-        }
-    }
-    Ok(trees)
-}
-
-async fn run_one(
-    pipeline: Arc<RedactionDocumentPipeline>,
-    tree: AnyTree,
-    ctx: Arc<RedactionContext>,
-    sem: Option<Arc<Semaphore>>,
-    plan: Arc<RedactionPlan>,
-    exports: Arc<Vec<ExportFile>>,
-) -> RedactionDocumentResult {
-    let _permit = match sem {
-        Some(s) => match Arc::clone(&s).acquire_owned().await {
-            Ok(permit) => Some(permit),
-            Err(_) => {
-                return RedactionDocumentResult {
-                    tree: None,
-                    error: Some("concurrency semaphore closed".to_owned()),
-                };
-            }
-        },
-        None => None,
-    };
-
-    let mut tree = tree;
-    let phase_result = match &mut tree {
-        AnyTree::Text(t) => pipeline.run_text(&ctx, &plan, t).await,
-        AnyTree::Tabular(t) => pipeline.run_tabular(&ctx, &plan, t).await,
-        AnyTree::Image(t) => pipeline.run_image(&ctx, &plan, t).await,
-        AnyTree::Audio(t) => pipeline.run_audio(&ctx, &plan, t).await,
-    };
-    if let Err(e) = phase_result {
-        return RedactionDocumentResult {
-            tree: None,
-            error: Some(e.to_string()),
-        };
-    }
-
-    for cfg in exports.iter() {
-        if ctx.is_cancelled() {
-            return RedactionDocumentResult {
-                tree: None,
-                error: Some("redaction cancelled before export".to_owned()),
-            };
-        }
-        let exporter = Exporter::new()
-            .with_encryption(cfg.encryption.clone())
-            .with_compression(cfg.compression)
-            .with_content_ids(cfg.content_ids.clone());
-        if let Err(e) = exporter.export(&tree, ctx.shared()).await {
-            return RedactionDocumentResult {
-                tree: None,
-                error: Some(e.to_string()),
-            };
-        }
-    }
-    RedactionDocumentResult {
-        tree: Some(tree),
-        error: None,
-    }
-}
diff --git a/crates/nvisy-engine/src/redaction/override_.rs b/crates/nvisy-engine/src/redaction/override_.rs
deleted file mode 100644
index fe4a540f6..000000000
--- a/crates/nvisy-engine/src/redaction/override_.rs
+++ /dev/null
@@ -1,199 +0,0 @@
-//! Per-entity overrides applied during [`RedactionEngine::redact`].
-//!
-//! Four variants cover the human-in-the-loop review surface:
-//!
-//! - [`RedactionOverride::Accept`] — explicit approve. Implicit
-//!   for any entity not mentioned in the overrides list, so
-//!   `Accept` is rarely needed in practice but available for
-//!   completeness.
-//! - [`RedactionOverride::Reject`] — skip this entity entirely.
-//!   The audit retains the entity but `Execution` becomes
-//!   `Suppressed` instead of running an operator. The audit
-//!   records the override provenance so a reviewer can see "a
-//!   human chose to skip this entity."
-//! - [`RedactionOverride::Replace`] — substitute a different
-//!   operator for this entity, bypassing the policy chain's pick.
-//!   The audit records both the original chain pick and the
-//!   substituted operator.
-//! - [`RedactionOverride::Add`] — inject an entity the
-//!   recognisers missed. The redaction synthesises a typed
-//!   [`Entity<M>`] (UUID minted fresh) and runs the policy chain
-//!   for that entity exactly as if a recogniser had found it.
-//!   When `operator` is `Some` the override pins that operator
-//!   instead of trusting the chain.
-//!
-//! # Entity identity
-//!
-//! [`Accept`], [`Reject`], and [`Replace`] reference an entity by
-//! its `Entity::id` (the per-mention UUID). Two coreferent mentions
-//! of the same real-world entity each have distinct ids; an
-//! override targets exactly one mention. To override every mention
-//! of a coreferent entity, submit one override per mention.
-//!
-//! This is intentional: a reviewer can reject one mention while
-//! accepting another (e.g. when the recogniser linked two
-//! independent people under one coreference id).
-//!
-//! [`Accept`]: RedactionOverride::Accept
-//! [`Reject`]: RedactionOverride::Reject
-//! [`Replace`]: RedactionOverride::Replace
-//! [`RedactionEngine::redact`]: crate::redaction::RedactionEngine::redact
-//! [`Entity<M>`]: nvisy_core::entity::Entity
-
-use std::collections::HashSet;
-
-use nvisy_core::entity::EntityLabelRef;
-use nvisy_core::modality::ModalityKind;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-use crate::modality::AnyLocation;
-use crate::policy::AnyRedaction;
-
-/// A single per-entity override carried by [`RedactionInput`].
-///
-/// [`RedactionInput`]: super::RedactionInput
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum RedactionOverride {
-    /// Approve the policy chain's decision for this entity.
-    /// Equivalent to the default behaviour when an entity is not
-    /// mentioned in the overrides list; present for completeness
-    /// so callers can emit an explicit "I reviewed and approve"
-    /// record.
-    Accept {
-        /// `Entity::id` from the detection's audit. Targets a
-        /// single mention; coreferent mentions are independent.
-        entity_id: Uuid,
-    },
-    /// Skip this entity. Execution becomes `Suppressed`; no
-    /// operator runs.
-    Reject {
-        /// `Entity::id` from the detection's audit. Targets a
-        /// single mention; coreferent mentions are independent.
-        entity_id: Uuid,
-    },
-    /// Replace the operator the policy chain picked.
-    Replace {
-        /// `Entity::id` from the detection's audit. Targets a
-        /// single mention; coreferent mentions are independent.
-        entity_id: Uuid,
-        /// Operator to apply instead of the policy chain's pick.
-        /// Must match the entity's modality.
-        operator: AnyRedaction,
-    },
-    /// Inject an entity the recognisers missed. The redaction
-    /// engine synthesises a typed [`Entity<M>`] for it with a
-    /// fresh UUID and runs the policy chain as if a recogniser
-    /// had found it. The audit records this entity with
-    /// `provenance: Override::Added`.
-    ///
-    /// [`Entity<M>`]: nvisy_core::entity::Entity
-    Add(RedactionAddEntity),
-}
-
-impl RedactionOverride {
-    /// The `Entity::id` this override targets. `None` for `Add`
-    /// (the entity does not yet exist; the redaction mints a
-    /// fresh UUID).
-    #[must_use]
-    pub fn target(&self) -> Option<Uuid> {
-        match self {
-            Self::Accept { entity_id }
-            | Self::Reject { entity_id }
-            | Self::Replace { entity_id, .. } => Some(*entity_id),
-            Self::Add(_) => None,
-        }
-    }
-}
-
-/// Payload for [`RedactionOverride::Add`]. The location's modality
-/// tag carries the entity's modality — there is no parallel
-/// `modality` field, since two such fields could disagree.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct RedactionAddEntity {
-    /// Location of the entity within the document. The variant
-    /// tag (`text`/`tabular`/`image`/`audio`) determines the
-    /// entity's modality.
-    pub location: AnyLocation,
-    /// Entity label (drives policy evaluation and operator pick).
-    pub label: EntityLabelRef,
-    /// When `Some`, pins the operator the redaction will apply,
-    /// bypassing the policy chain for this entity. Must match
-    /// `location`'s modality; mismatches are rejected at
-    /// [`RedactionEngine::redact`] validation time.
-    ///
-    /// [`RedactionEngine::redact`]: crate::redaction::RedactionEngine::redact
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub operator: Option<AnyRedaction>,
-}
-
-impl RedactionAddEntity {
-    /// The modality the added entity belongs to. Derived from
-    /// the location's variant tag.
-    #[must_use]
-    pub fn modality(&self) -> ModalityKind {
-        self.location.kind()
-    }
-}
-
-/// Validate a list of overrides for internal consistency.
-///
-/// Catches the bug classes a client could submit that the engine
-/// would otherwise need to handle defensively:
-///
-/// - **Duplicate targets** — two overrides for the same
-///   `entity_id`. Order would silently determine which wins; we
-///   reject up front.
-/// - **Modality mismatch** — a `Replace` whose `operator`'s
-///   modality differs from the targeted entity (caught later by
-///   the engine since the entity-id lookup is needed) and an
-///   `Add` whose `operator`'s modality differs from the
-///   `location`'s modality (catchable here).
-///
-/// Returns the first violation encountered; callers fix and
-/// resubmit rather than iterating.
-///
-/// # Errors
-///
-/// [`ErrorKind::Validation`] with a message describing the
-/// violation and the offending override index.
-///
-/// [`ErrorKind::Validation`]: nvisy_core::ErrorKind::Validation
-pub fn validate_overrides(overrides: &[RedactionOverride]) -> Result<(), nvisy_core::Error> {
-    const COMPONENT: &str = "nvisy_engine::pipeline::redaction::override";
-
-    let mut seen: HashSet<Uuid> = HashSet::with_capacity(overrides.len());
-    for (idx, ov) in overrides.iter().enumerate() {
-        if let Some(target) = ov.target()
-            && !seen.insert(target)
-        {
-            return Err(nvisy_core::Error::validation(
-                format!("override #{idx} duplicates target entity {target}"),
-                COMPONENT,
-            ));
-        }
-        if let RedactionOverride::Add(add) = ov
-            && let Some(op) = &add.operator
-            && op.modality() != add.modality()
-        {
-            return Err(nvisy_core::Error::validation(
-                format!(
-                    "override #{idx} (add) operator modality {:?} does not match location modality {:?}",
-                    op.modality(),
-                    add.modality(),
-                ),
-                COMPONENT,
-            ));
-        }
-    }
-    Ok(())
-}
-
-// `RedactionDecision` lives in `provenance/override_decision.rs`
-// because it's a property of the audit entry, not a property of
-// the request shape. Re-exported from this module for ergonomic
-// access:
-pub use crate::document::provenance::RedactionDecision;
diff --git a/crates/nvisy-engine/src/redaction/phases/instantiate.rs b/crates/nvisy-engine/src/redaction/phases/instantiate.rs
deleted file mode 100644
index f8497d8c3..000000000
--- a/crates/nvisy-engine/src/redaction/phases/instantiate.rs
+++ /dev/null
@@ -1,105 +0,0 @@
-//! [`Instantiate<M>`]: per-modality glue that turns a redaction
-//! operator spec ([`DocumentModality::Redaction`]) into a runnable
-//! [`Anonymizer<M>`] instance.
-//!
-//! Built-in arms construct a fresh operator from the rule's params
-//! (no registry round-trip). The `Custom` arm looks up a
-//! deployment-registered operator on the toolkit-side
-//! [`RedactionRegistry<M>`].
-//!
-//! [`Anonymizer<M>`]: nvisy_toolkit::redaction::Anonymizer
-//! [`DocumentModality::Redaction`]: crate::modality::DocumentModality::Redaction
-//! [`RedactionRegistry<M>`]: nvisy_toolkit::redaction::RedactionRegistry
-
-use std::sync::Arc;
-
-use nvisy_core::modality::{Audio, Image, Modality, Text};
-use nvisy_core::{Error, Result};
-use nvisy_toolkit::redaction::anonymizer::{Hash, Keep, Mask, Redact, Replace};
-use nvisy_toolkit::redaction::{Anonymizer, RedactionRegistry};
-
-use crate::policy::redaction::{AudioRedaction, ImageRedaction, TextRedaction};
-
-const TARGET: &str = "nvisy_engine::redaction::phases::instantiate";
-
-/// Per-modality conversion from operator spec to runnable
-/// [`Anonymizer<M>`]. Implemented by the per-modality redaction
-/// enums in this module.
-///
-/// [`Anonymizer<M>`]: nvisy_toolkit::redaction::Anonymizer
-pub trait Instantiate<M: Modality> {
-    /// Resolve `self` to a runnable operator. Built-in arms
-    /// construct a fresh instance; `Custom(id)` arms look up the
-    /// operator in `registry`.
-    ///
-    /// # Errors
-    ///
-    /// Returns `Error::validation` when a `Custom` arm names an id
-    /// that is not registered in `registry`.
-    fn instantiate(&self, registry: &RedactionRegistry<M>) -> Result<Arc<dyn Anonymizer<M>>>;
-}
-
-impl Instantiate<Text> for TextRedaction {
-    fn instantiate(&self, registry: &RedactionRegistry<Text>) -> Result<Arc<dyn Anonymizer<Text>>> {
-        match self {
-            Self::Replace { template } => Ok(Arc::new(Replace::new(template.clone()))),
-            Self::Mask {
-                mask_char,
-                keep_prefix,
-                keep_suffix,
-            } => {
-                let op = Mask::new(*mask_char)
-                    .with_keep_prefix(*keep_prefix)
-                    .with_keep_suffix(*keep_suffix);
-                Ok(Arc::new(op))
-            }
-            Self::Hash { algorithm, salt } => {
-                let mut op = Hash::new(*algorithm);
-                if let Some(s) = salt {
-                    op = op.with_salt(s.clone());
-                }
-                Ok(Arc::new(op))
-            }
-            Self::Redact => Ok(Arc::new(Redact)),
-            Self::Keep => Ok(Arc::new(Keep)),
-            Self::Custom { id } => registry.resolve_id(id).cloned().ok_or_else(|| {
-                Error::validation(
-                    format!("custom text anonymizer `{id}` is not registered"),
-                    TARGET,
-                )
-            }),
-        }
-    }
-}
-
-impl Instantiate<Image> for ImageRedaction {
-    fn instantiate(
-        &self,
-        registry: &RedactionRegistry<Image>,
-    ) -> Result<Arc<dyn Anonymizer<Image>>> {
-        match self {
-            Self::Custom { id } => registry.resolve_id(id).cloned().ok_or_else(|| {
-                Error::validation(
-                    format!("custom image anonymizer `{id}` is not registered"),
-                    TARGET,
-                )
-            }),
-        }
-    }
-}
-
-impl Instantiate<Audio> for AudioRedaction {
-    fn instantiate(
-        &self,
-        registry: &RedactionRegistry<Audio>,
-    ) -> Result<Arc<dyn Anonymizer<Audio>>> {
-        match self {
-            Self::Custom { id } => registry.resolve_id(id).cloned().ok_or_else(|| {
-                Error::validation(
-                    format!("custom audio anonymizer `{id}` is not registered"),
-                    TARGET,
-                )
-            }),
-        }
-    }
-}
diff --git a/crates/nvisy-engine/src/redaction/phases/mod.rs b/crates/nvisy-engine/src/redaction/phases/mod.rs
deleted file mode 100644
index 3f41c3795..000000000
--- a/crates/nvisy-engine/src/redaction/phases/mod.rs
+++ /dev/null
@@ -1,227 +0,0 @@
-//! Redaction-side per-document phase orchestrators: the redaction
-//! phase itself plus the post-redaction validation leak check.
-//!
-//! [`RedactionPhase`] drives a two-step flow per node:
-//!
-//! 1. **Evaluate** — for each detected entity, the [`PolicyStore`]
-//!    picks `Redact` / `Suppress` / `Audit` / `Fallthrough`. The
-//!    result lands on `EntityRecord::audit` as an
-//!    [`AuditEntry<M>`] whose `decision` carries the resolved
-//!    [`ResolvedAction<M>`] — for `Redact`, the per-modality
-//!    operator picked from the rule's [`ModalityRedactions`] (or
-//!    the deployment-wide default if the rule didn't cover the
-//!    entity's modality).
-//! 2. **Apply** — pending entities have their operator spec
-//!    instantiated via [`Instantiate::instantiate`]; built-in
-//!    arms construct a fresh operator from the rule's params,
-//!    `Custom` arms look up the user-registered operator in the
-//!    [`RedactionRegistry<M>`]. The phase pulls the per-entity
-//!    source payload through [`DataAt`] and calls
-//!    [`Anonymizer::apply`]; the produced `M::Replacement` is
-//!    written back as `Execution::Applied { replacement }`.
-//!    Operator construction errors and `apply` errors both fail
-//!    the record into `Execution::Failed` with the underlying
-//!    reason.
-//!
-//! [`ValidationPhase`] runs after redaction: it streams the
-//! redacted output back through the codec, walks the audit's
-//! applied entities, and runs a per-modality [`LeakCheck`] over
-//! the redacted text to surface any value that should have been
-//! redacted but still appears in the output.
-//!
-//! Codec-side byte mutation (actually writing the replacement
-//! back at the entity's location in the handle) is still a
-//! separate follow-up — recorded as a TODO inline.
-//!
-//! [`Anonymizer::apply`]: nvisy_toolkit::redaction::Anonymizer::apply
-//! [`RedactionRegistry<M>`]: nvisy_toolkit::redaction::RedactionRegistry
-//! [`DataAt`]: nvisy_core::extraction::DataAt
-//! [`ResolvedAction<M>`]: crate::document::provenance::ResolvedAction
-//! [`ModalityRedactions`]: crate::policy::redaction::ModalityRedactions
-//! [`LeakCheck`]: nvisy_toolkit::validation::LeakCheck
-
-mod instantiate;
-pub mod phase;
-mod registries;
-pub mod validation;
-
-use nvisy_codec::content::ContentDescriptor;
-use nvisy_core::Result;
-use nvisy_core::entity::{EntityLabelCatalog, is_excluded};
-use nvisy_core::extraction::{DataAt, TextAt};
-use nvisy_core::modality::Overlap;
-use nvisy_core::primitive::ConfidenceThreshold;
-use nvisy_core::redaction::{RedactAt, Redactions};
-use nvisy_toolkit::redaction::RedactionRegistry;
-
-pub use self::instantiate::Instantiate;
-pub use self::phase::RedactionPhase;
-pub use self::registries::RedactionRegistries;
-pub use self::validation::ValidationPhase;
-use crate::core::{Decision, DocumentTree, PolicyStore};
-use crate::document::provenance::{
-    AuditEntry, Decision as AuditDecision, EntryMetadata, Execution, ResolvedAction,
-};
-use crate::modality::DocumentModality;
-use crate::policy::PolicyDecisionRef;
-use crate::policy::redaction::{ModalityRedactions, ProjectRedaction};
-
-pub(crate) const TARGET: &str = "nvisy_engine::redaction";
-
-/// Body of the redaction phase, parameterised on the resolved
-/// `default_threshold`. Crate-visible so both the phase
-/// orchestrator and the test-only path drive it through the same
-/// code.
-pub(crate) async fn run_redaction<M>(
-    default_threshold: ConfidenceThreshold,
-    tree: &mut DocumentTree<M>,
-    descriptor: &ContentDescriptor,
-    policies: &PolicyStore,
-    catalog: &EntityLabelCatalog,
-    default_operators: &ModalityRedactions,
-    registry: &RedactionRegistry<M>,
-) -> Result<()>
-where
-    M: DocumentModality + ProjectRedaction,
-    M::Location: Overlap,
-    M::Redaction: Instantiate<M>,
-    DocumentTree<M>: TextAt<M> + DataAt<M> + RedactAt<M>,
-{
-    if tree.root.audit.records.is_empty() {
-        return Ok(());
-    }
-
-    let before_filter = tree.root.audit.records.len();
-    let annotations = std::mem::take(&mut tree.root.annotations);
-    tree.root
-        .audit
-        .records
-        .retain(|record| !is_excluded(&annotations, &record.entity));
-    tree.root.annotations = annotations;
-    let dropped = before_filter - tree.root.audit.records.len();
-    if dropped > 0 {
-        tracing::debug!(
-            target: TARGET,
-            dropped,
-            "filtered entities by Assert exclusion annotations",
-        );
-    }
-
-    if tree.root.audit.records.is_empty() {
-        return Ok(());
-    }
-
-    let document_labels: Vec<&str> = tree.root.labels.iter().map(|l| l.label.as_str()).collect();
-
-    for record in &mut tree.root.audit.records {
-        if record.audit.is_some() {
-            continue;
-        }
-        if !default_threshold.admits(record.entity.confidence) {
-            continue;
-        }
-        let decision = policies.resolve::<M>(
-            &record.entity,
-            catalog,
-            default_operators,
-            &document_labels,
-            descriptor,
-        );
-        let entry = match decision {
-            Decision::Redact {
-                policy_name,
-                rule_name,
-                operator,
-            } => AuditEntry {
-                decision: AuditDecision {
-                    policy_ref: Some(PolicyDecisionRef::new(policy_name, rule_name)),
-                    action: ResolvedAction::Redact { operator },
-                },
-                execution: Execution::Pending,
-                metadata: EntryMetadata::now(),
-            },
-            Decision::Suppress {
-                policy_name,
-                rule_name,
-                reason,
-            } => AuditEntry {
-                decision: AuditDecision {
-                    policy_ref: Some(PolicyDecisionRef::new(policy_name, rule_name)),
-                    action: ResolvedAction::Suppress { reason },
-                },
-                execution: Execution::Suppressed,
-                metadata: EntryMetadata::now(),
-            },
-            Decision::Audit {
-                policy_name,
-                rule_name,
-                severity,
-            } => AuditEntry {
-                decision: AuditDecision {
-                    policy_ref: Some(PolicyDecisionRef::new(policy_name, rule_name)),
-                    action: ResolvedAction::Audit { severity },
-                },
-                execution: Execution::Suppressed,
-                metadata: EntryMetadata::now(),
-            },
-            Decision::Fallthrough => continue,
-        };
-        record.audit = Some(entry);
-    }
-
-    let mut apply_outcomes: Vec<(usize, Execution<M>)> = Vec::new();
-
-    for (idx, record) in tree.root.audit.records.iter().enumerate() {
-        let Some(audit) = record.audit.as_ref() else {
-            continue;
-        };
-        if !matches!(audit.execution, Execution::Pending) {
-            continue;
-        }
-        let ResolvedAction::Redact { operator } = &audit.decision.action else {
-            continue;
-        };
-        let outcome = match operator.instantiate(registry) {
-            Err(err) => Execution::Failed {
-                reason: err.to_string(),
-            },
-            Ok(anonymizer) => match tree.data_at(&record.entity.location).await {
-                None => Execution::Failed {
-                    reason: "source payload unavailable at entity location".into(),
-                },
-                Some(source) => match anonymizer.apply(&record.entity, &source).await {
-                    Ok(replacement) => Execution::Applied { replacement },
-                    Err(err) => Execution::Failed {
-                        reason: err.to_string(),
-                    },
-                },
-            },
-        };
-        apply_outcomes.push((idx, outcome));
-    }
-
-    let mut redactions = Redactions::<M>::new();
-    for (idx, outcome) in apply_outcomes {
-        if let Execution::Applied { replacement } = &outcome {
-            redactions.push(
-                tree.root.audit.records[idx].entity.location.clone(),
-                replacement.clone(),
-            );
-        }
-        if let Some(audit) = tree.root.audit.records[idx].audit.as_mut() {
-            audit.execution = outcome;
-        }
-    }
-
-    if !redactions.is_empty() {
-        let count = redactions.len();
-        tree.redact_at(redactions).await?;
-        tracing::debug!(
-            target: TARGET,
-            count,
-            "flushed redaction batch through codec handle",
-        );
-    }
-
-    Ok(())
-}
diff --git a/crates/nvisy-engine/src/redaction/phases/phase.rs b/crates/nvisy-engine/src/redaction/phases/phase.rs
deleted file mode 100644
index 87fa286f8..000000000
--- a/crates/nvisy-engine/src/redaction/phases/phase.rs
+++ /dev/null
@@ -1,136 +0,0 @@
-//! [`RedactionPhase`]: per-modality redaction driver.
-//!
-//! Resolves the per-run threshold from the plan (falling back to the
-//! deployment-wide [`RedactionConfig`]), then drives the per-tree
-//! `run_redaction` body in the parent module.
-//!
-//! Holds one [`RedactionRegistry<M>`] per modality, which the apply
-//! step consults when a rule's operator spec is the `Custom` arm.
-//! Built-in operators are constructed inline from the rule's spec
-//! and don't touch the registry.
-//!
-//! [`RedactionRegistry<M>`]: nvisy_toolkit::redaction::RedactionRegistry
-
-use nvisy_core::Result;
-use nvisy_core::modality::{Audio, Image, Tabular, Text};
-
-use crate::core::{DocumentTree, PhaseContext as _, RedactionContext};
-use crate::redaction::phases::registries::RedactionRegistries;
-use crate::redaction::phases::run_redaction;
-use crate::redaction::{RedactionConfig, RedactionPlan};
-
-const TARGET: &str = "nvisy_engine::redaction";
-
-/// Redaction phase orchestrator. Holds a [`RedactionConfig`] for the
-/// deployment-wide defaults and one [`RedactionRegistry<M>`] per
-/// modality for `Custom`-arm lookups at apply time.
-///
-/// [`RedactionRegistry<M>`]: nvisy_toolkit::redaction::RedactionRegistry
-pub struct RedactionPhase {
-    config: RedactionConfig,
-    registries: RedactionRegistries,
-}
-
-impl RedactionPhase {
-    pub(crate) fn new(config: RedactionConfig, registries: RedactionRegistries) -> Self {
-        Self { config, registries }
-    }
-
-    pub(crate) async fn apply_text(
-        &self,
-        ctx: &RedactionContext,
-        plan: &RedactionPlan,
-        tree: &mut DocumentTree<Text>,
-    ) -> Result<()> {
-        let threshold = plan
-            .redaction
-            .confidence_threshold
-            .unwrap_or(self.config.confidence_threshold);
-        let policies = &ctx.shared().policies;
-        let catalog = &ctx.shared().catalog;
-        let default_operators = &self.config.default_operators;
-        let descriptor = tree.descriptor.clone();
-        run_redaction(
-            threshold,
-            tree,
-            &descriptor,
-            policies,
-            catalog,
-            default_operators,
-            &self.registries.text,
-        )
-        .await
-    }
-
-    pub(crate) async fn apply_image(
-        &self,
-        ctx: &RedactionContext,
-        plan: &RedactionPlan,
-        tree: &mut DocumentTree<Image>,
-    ) -> Result<()> {
-        let threshold = plan
-            .redaction
-            .confidence_threshold
-            .unwrap_or(self.config.confidence_threshold);
-        let policies = &ctx.shared().policies;
-        let catalog = &ctx.shared().catalog;
-        let default_operators = &self.config.default_operators;
-        let descriptor = tree.descriptor.clone();
-        run_redaction(
-            threshold,
-            tree,
-            &descriptor,
-            policies,
-            catalog,
-            default_operators,
-            &self.registries.image,
-        )
-        .await
-    }
-
-    pub(crate) async fn apply_audio(
-        &self,
-        ctx: &RedactionContext,
-        plan: &RedactionPlan,
-        tree: &mut DocumentTree<Audio>,
-    ) -> Result<()> {
-        let threshold = plan
-            .redaction
-            .confidence_threshold
-            .unwrap_or(self.config.confidence_threshold);
-        let policies = &ctx.shared().policies;
-        let catalog = &ctx.shared().catalog;
-        let default_operators = &self.config.default_operators;
-        let descriptor = tree.descriptor.clone();
-        run_redaction(
-            threshold,
-            tree,
-            &descriptor,
-            policies,
-            catalog,
-            default_operators,
-            &self.registries.audio,
-        )
-        .await
-    }
-
-    /// Tabular redaction is not implemented yet — there's no
-    /// `Anonymizer<Tabular>` in the workspace. Warn if records are
-    /// pending and skip the apply step.
-    pub(crate) async fn apply_tabular(
-        &self,
-        _ctx: &RedactionContext,
-        _plan: &RedactionPlan,
-        tree: &mut DocumentTree<Tabular>,
-    ) -> Result<()> {
-        let pending = tree.root.audit.records.len();
-        if pending > 0 {
-            tracing::warn!(
-                target: TARGET,
-                pending,
-                "tabular redaction not implemented; skipping {pending} audit record(s)",
-            );
-        }
-        Ok(())
-    }
-}
diff --git a/crates/nvisy-engine/src/redaction/phases/registries.rs b/crates/nvisy-engine/src/redaction/phases/registries.rs
deleted file mode 100644
index eed4aeb52..000000000
--- a/crates/nvisy-engine/src/redaction/phases/registries.rs
+++ /dev/null
@@ -1,28 +0,0 @@
-//! [`RedactionRegistries`]: per-modality bundle of
-//! [`RedactionRegistry<M>`] instances the deployment code populates
-//! at engine startup and the redaction phase consults at apply time.
-//!
-//! Text / Image / Audio each get a slot. Tabular is omitted: it has
-//! no `Modality` impl in `nvisy-core`, so `Anonymizer<Tabular>`
-//! is not implementable today.
-//!
-//! [`RedactionRegistry<M>`]: nvisy_toolkit::redaction::RedactionRegistry
-
-use nvisy_core::modality::{Audio, Image, Text};
-use nvisy_toolkit::redaction::RedactionRegistry;
-
-/// Per-modality bundle of [`RedactionRegistry<M>`] instances. Built
-/// once at engine startup (deployment code registers custom
-/// operators) and shared with each [`RedactionPhase`].
-///
-/// [`RedactionRegistry<M>`]: nvisy_toolkit::redaction::RedactionRegistry
-/// [`RedactionPhase`]: super::phase::RedactionPhase
-#[derive(Clone, Default, Debug)]
-pub struct RedactionRegistries {
-    /// Custom text-modality operators.
-    pub text: RedactionRegistry<Text>,
-    /// Custom image-modality operators.
-    pub image: RedactionRegistry<Image>,
-    /// Custom audio-modality operators.
-    pub audio: RedactionRegistry<Audio>,
-}
diff --git a/crates/nvisy-engine/src/redaction/phases/validation.rs b/crates/nvisy-engine/src/redaction/phases/validation.rs
deleted file mode 100644
index 4baa3457d..000000000
--- a/crates/nvisy-engine/src/redaction/phases/validation.rs
+++ /dev/null
@@ -1,229 +0,0 @@
-//! [`ValidationPhase`]: per-document check-pipeline driver.
-//!
-//! Builds the canonical [`CheckPipeline`] from
-//! [`Validation`] (today: a single [`LeakCheck`] with the configured
-//! [`Severity`]) and runs it against each modality tree.
-//! Aggregates findings; any [`Severity::Fail`] finding fails the
-//! run with a validation error.
-//!
-//! [`CheckPipeline`]: nvisy_toolkit::validation::CheckPipeline
-//! [`LeakCheck`]: nvisy_toolkit::validation::LeakCheck
-//! [`Severity`]: nvisy_toolkit::validation::Severity
-//! [`Validation`]: crate::redaction::Validation
-
-use nvisy_codec::Handler;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Audio, Image, Tabular, Text};
-use nvisy_core::{Error, Result};
-use nvisy_toolkit::validation::{
-    CheckContext, CheckPipeline, Finding, FindingKind, LeakCheck, Severity,
-};
-use tracing::Instrument;
-
-use crate::core::{DocumentTree, PhaseContext as _, RedactionContext};
-use crate::document::Document;
-use crate::modality::DocumentModality;
-use crate::redaction::RedactionPlan;
-
-const TARGET: &str = "nvisy_engine::validation";
-
-/// Validation phase orchestrator.
-pub struct ValidationPhase;
-
-impl ValidationPhase {
-    /// Build the phase. Stateless.
-    pub fn new() -> Self {
-        Self
-    }
-
-    pub(crate) async fn apply_text(
-        &self,
-        ctx: &RedactionContext,
-        plan: &RedactionPlan,
-        tree: &mut DocumentTree<Text>,
-    ) -> Result<()> {
-        let span = tracing::info_span!(target: TARGET, "phase", name = "validation.text");
-        let run_id = ctx.shared().run_id;
-        let cfg = plan.validation.clone();
-        async move {
-            let redacted = stream_text(tree.handle.handler_mut()).await?;
-            let entities = applied_entities(&tree.root);
-            let pipeline: CheckPipeline<Text, DocumentTree<Text>> =
-                CheckPipeline::new().with_check(LeakCheck::new(cfg.leak_severity));
-            let mut check_ctx = CheckContext::new(&*tree).with_correlation_id(run_id);
-            if let Some(ref text) = redacted {
-                check_ctx = check_ctx.with_redacted_output(text);
-            }
-            let findings = log_findings("text", pipeline.run(&entities, &check_ctx).await);
-            finalize(&findings)
-        }
-        .instrument(span)
-        .await
-    }
-
-    pub(crate) async fn apply_tabular(
-        &self,
-        ctx: &RedactionContext,
-        plan: &RedactionPlan,
-        tree: &mut DocumentTree<Tabular>,
-    ) -> Result<()> {
-        let span = tracing::info_span!(target: TARGET, "phase", name = "validation.tabular");
-        let run_id = ctx.shared().run_id;
-        let cfg = plan.validation.clone();
-        async move {
-            let redacted = stream_tabular(tree.handle.handler_mut()).await?;
-            let entities = applied_entities(&tree.root);
-            let pipeline: CheckPipeline<Tabular, DocumentTree<Tabular>> =
-                CheckPipeline::new().with_check(LeakCheck::new(cfg.leak_severity));
-            let mut check_ctx = CheckContext::new(&*tree).with_correlation_id(run_id);
-            if let Some(ref text) = redacted {
-                check_ctx = check_ctx.with_redacted_output(text);
-            }
-            let findings = log_findings("tabular", pipeline.run(&entities, &check_ctx).await);
-            finalize(&findings)
-        }
-        .instrument(span)
-        .await
-    }
-
-    /// Image and audio have no canonical check today — succeed
-    /// immediately so the pipeline orchestrator can call this
-    /// uniformly per modality.
-    pub(crate) async fn apply_image(
-        &self,
-        _ctx: &RedactionContext,
-        _plan: &RedactionPlan,
-        _tree: &mut DocumentTree<Image>,
-    ) -> Result<()> {
-        Ok(())
-    }
-
-    pub(crate) async fn apply_audio(
-        &self,
-        _ctx: &RedactionContext,
-        _plan: &RedactionPlan,
-        _tree: &mut DocumentTree<Audio>,
-    ) -> Result<()> {
-        Ok(())
-    }
-}
-
-impl Default for ValidationPhase {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
-/// Collect the entities whose redaction the audit reports as
-/// `Applied` — the set the leak check must verify is no longer
-/// present in the post-redaction output.
-fn applied_entities<M: DocumentModality>(doc: &Document<M>) -> Vec<Entity<M>> {
-    doc.audit
-        .records
-        .iter()
-        .filter(|r| r.audit.as_ref().is_some_and(|e| e.execution.is_applied()))
-        .map(|r| r.entity.clone())
-        .collect()
-}
-
-/// Log per-tree findings at the appropriate level and return them
-/// for run-level aggregation.
-fn log_findings(modality: &'static str, findings: Vec<Finding>) -> Vec<Finding> {
-    if findings.is_empty() {
-        tracing::debug!(target: TARGET, modality, "validation passed");
-        return findings;
-    }
-    let fail_count = findings
-        .iter()
-        .filter(|f| matches!(f.severity, Severity::Fail))
-        .count();
-    if fail_count > 0 {
-        tracing::warn!(
-            target: TARGET,
-            modality,
-            findings = findings.len(),
-            failing = fail_count,
-            "validation produced failing findings",
-        );
-    } else {
-        tracing::warn!(
-            target: TARGET,
-            modality,
-            findings = findings.len(),
-            "validation produced warning findings",
-        );
-    }
-    for finding in &findings {
-        tracing::warn!(
-            target: TARGET,
-            severity = ?finding.severity,
-            message = %finding.message,
-            "validation finding",
-        );
-    }
-    findings
-}
-
-/// Convert the collected findings into a phase result. Any
-/// `Severity::Fail` finding fails the run with a validation error
-/// listing all failing findings.
-fn finalize(findings: &[Finding]) -> Result<()> {
-    let failing: Vec<&Finding> = findings
-        .iter()
-        .filter(|f| matches!(f.severity, Severity::Fail))
-        .collect();
-    if failing.is_empty() {
-        return Ok(());
-    }
-    let details = failing
-        .iter()
-        .map(|f| match &f.kind {
-            FindingKind::Leak { entity_id, value } => format!("{value}({entity_id})"),
-            FindingKind::Other => f.message.clone(),
-            _ => f.message.clone(),
-        })
-        .collect::<Vec<_>>()
-        .join(", ");
-    Err(Error::validation(
-        format!(
-            "{} validation finding(s) failed the run: {}",
-            failing.len(),
-            details,
-        ),
-        "validation",
-    ))
-}
-
-/// Drain the post-redaction text chunks into one concatenated string
-/// for substring-based leak detection. Returns `None` when the handle
-/// has no chunks at all.
-async fn stream_text(handle: &mut dyn Handler<Text>) -> Result<Option<String>> {
-    let mut buf = String::new();
-    while let Some(chunk) = handle.next_chunk().await? {
-        if !buf.is_empty() {
-            buf.push('\n');
-        }
-        buf.push_str(chunk.data.as_str());
-    }
-    Ok((!buf.is_empty()).then_some(buf))
-}
-
-/// Drain post-redaction tabular cells, separating rows with newlines.
-async fn stream_tabular(handle: &mut dyn Handler<Tabular>) -> Result<Option<String>> {
-    let mut buf = String::new();
-    let mut current_row: Option<u32> = None;
-    while let Some(chunk) = handle.next_chunk().await? {
-        let row = chunk.location.row_index;
-        if let Some(prev) = current_row
-            && prev != row
-        {
-            buf.push('\n');
-        }
-        if !buf.is_empty() && current_row == Some(row) {
-            buf.push('\t');
-        }
-        buf.push_str(chunk.data.as_str());
-        current_row = Some(row);
-    }
-    Ok((!buf.is_empty()).then_some(buf))
-}
diff --git a/crates/nvisy-engine/src/redaction/pipeline.rs b/crates/nvisy-engine/src/redaction/pipeline.rs
deleted file mode 100644
index 15c0ebd6e..000000000
--- a/crates/nvisy-engine/src/redaction/pipeline.rs
+++ /dev/null
@@ -1,195 +0,0 @@
-//! Per-pass redaction pipeline lifecycle.
-
-use std::sync::Arc;
-
-use jiff::Timestamp;
-use nvisy_codec::CodecRegistry;
-use nvisy_core::Error;
-use tokio_util::sync::CancellationToken;
-use uuid::Uuid;
-
-use super::applicator::apply_overrides;
-use super::input::RedactionInput;
-use super::orchestrator::RedactionOrchestrator;
-use super::result::RedactionResult;
-use super::state::{RedactionRecord, RedactionState};
-use super::status::RedactionStatus;
-use crate::core::ingestion::encryption::SharedKeyProvider;
-use crate::core::{RedactionContext, RedactionEngines, RuntimeConfig, SharedData};
-use crate::detection::DetectionState;
-use crate::document::provenance::AnyAudit;
-use crate::redaction::RedactionConfig;
-use crate::redaction::phases::RedactionRegistries;
-use crate::registry::Registry;
-
-const TARGET: &str = "nvisy_engine::pipeline::redaction::pipeline";
-
-pub(crate) struct RedactionEngineState {
-    pub redaction_config: Arc<RedactionConfig>,
-    pub redaction_registries: Arc<RedactionRegistries>,
-}
-
-pub(crate) struct RedactionPipeline {
-    redaction_id: Uuid,
-    registry: Registry,
-    key_provider: Option<SharedKeyProvider>,
-    detections: DetectionState,
-    redactions: RedactionState,
-    base_config: RuntimeConfig,
-    state: RedactionEngineState,
-}
-
-impl RedactionPipeline {
-    pub(crate) fn new(
-        registry: Registry,
-        key_provider: Option<SharedKeyProvider>,
-        detections: DetectionState,
-        redactions: RedactionState,
-        base_config: RuntimeConfig,
-        state: RedactionEngineState,
-    ) -> Self {
-        Self {
-            redaction_id: Uuid::now_v7(),
-            registry,
-            key_provider,
-            detections,
-            redactions,
-            base_config,
-            state,
-        }
-    }
-
-    pub(crate) fn id(&self) -> Uuid {
-        self.redaction_id
-    }
-
-    pub(crate) async fn register_pending(&self, input: &RedactionInput) {
-        self.redactions
-            .insert(
-                self.redaction_id,
-                RedactionRecord {
-                    detection_id: input.detection_id,
-                    actor_id: input.actor_id,
-                    status: RedactionStatus::Pending,
-                    created_at: Timestamp::now(),
-                    started_at: None,
-                    completed_at: None,
-                    cancel: CancellationToken::new(),
-                    audits: Vec::new(),
-                    redactions_applied: 0,
-                    error: None,
-                },
-            )
-            .await;
-    }
-
-    /// Drive the redaction to completion.
-    pub(crate) async fn execute(&self, input: RedactionInput) -> Result<(), Error> {
-        let actor_id = input.actor_id;
-
-        let detection = match self.detections.handoff(actor_id, input.detection_id).await {
-            Ok(d) => d,
-            Err(e) => {
-                self.redactions.fail(self.redaction_id, e.to_string()).await;
-                return Err(e);
-            }
-        };
-
-        if let Err(e) = super::validate_overrides(&input.overrides) {
-            self.redactions.fail(self.redaction_id, e.to_string()).await;
-            return Err(e);
-        }
-
-        let mut audits: Vec<AnyAudit> = detection.audits.clone();
-        if let Err(e) = apply_overrides(&mut audits, input.overrides.clone()) {
-            self.redactions.fail(self.redaction_id, e.to_string()).await;
-            return Err(e);
-        }
-
-        let catalog = Arc::new(detection.policies.catalog());
-
-        let mut shared_data = SharedData {
-            run_id: self.redaction_id,
-            actor_id,
-            policies: detection.policies,
-            catalog,
-            registry: self.registry.clone(),
-            codec_registry: CodecRegistry::with_builtin(),
-            key_provider: SharedKeyProvider::default(),
-        };
-        if let Some(ref kp) = self.key_provider {
-            shared_data.key_provider = kp.clone();
-        }
-
-        let cancel = CancellationToken::new();
-        let engines = RedactionEngines {
-            redaction_config: (*self.state.redaction_config).clone(),
-            redaction_registries: (*self.state.redaction_registries).clone(),
-        };
-        let concurrency = self.base_config.effective_concurrency();
-        let ctx = RedactionContext::new(cancel, Arc::new(shared_data), engines, concurrency);
-
-        self.redactions.set_started_at(self.redaction_id).await;
-
-        let orchestrator = RedactionOrchestrator::new(ctx);
-        let output = match orchestrator
-            .run(&detection.imports, &input.exports, &input.plan, audits)
-            .await
-        {
-            Ok(out) => out,
-            Err(e) => {
-                self.redactions.fail(self.redaction_id, e.to_string()).await;
-                return Err(e);
-            }
-        };
-
-        let (final_audits, redactions_applied, any_ok, any_failed) = output.into_audits();
-        let status = if !any_failed {
-            RedactionStatus::Succeeded
-        } else if any_ok {
-            RedactionStatus::PartialFailure
-        } else {
-            RedactionStatus::Failed
-        };
-
-        if matches!(
-            status,
-            RedactionStatus::Succeeded | RedactionStatus::PartialFailure
-        ) {
-            let result = RedactionResult {
-                id: self.redaction_id,
-                detection_id: input.detection_id,
-                actor_id,
-                audits: final_audits.clone(),
-                redactions_applied,
-            };
-            if let Err(e) = self
-                .registry
-                .store_redaction(actor_id, self.redaction_id, &result)
-                .await
-            {
-                tracing::error!(
-                    target: TARGET,
-                    redaction_id = %self.redaction_id,
-                    error = %e,
-                    "failed to persist redaction result; in-memory state stays valid",
-                );
-            }
-        }
-
-        self.redactions
-            .finalize(self.redaction_id, status, final_audits, redactions_applied)
-            .await;
-
-        tracing::info!(
-            target: TARGET,
-            redaction_id = %self.redaction_id,
-            detection_id = %input.detection_id,
-            ?status,
-            redactions_applied,
-            "redaction pass finalized",
-        );
-
-        Ok(())
-    }
-}
diff --git a/crates/nvisy-engine/src/redaction/plan.rs b/crates/nvisy-engine/src/redaction/plan.rs
deleted file mode 100644
index 8979b385f..000000000
--- a/crates/nvisy-engine/src/redaction/plan.rs
+++ /dev/null
@@ -1,75 +0,0 @@
-//! Per-pass plan and per-phase knobs the redaction pipeline reads
-//! against each imported document.
-//!
-//! Three concerns share this file:
-//!
-//! - [`RedactionPlan`] — top-level per-request bundle the
-//!   redaction pipeline reads once per document.
-//! - [`Redaction`] — per-plan knobs for the redaction phase
-//!   itself (confidence threshold override, metadata handling).
-//! - [`Validation`] — per-plan knobs for the post-redaction leak
-//!   check phase.
-
-use nvisy_core::primitive::ConfidenceThreshold;
-use nvisy_toolkit::validation::Severity;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Per-request bundle of redaction-side phase configs.
-///
-/// The redaction pipeline reads this once per document and routes
-/// each phase (redaction, then validation) to the matching field.
-#[derive(Debug, Clone, Default)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase", default)]
-pub struct RedactionPlan {
-    /// Redaction settings applied after policy evaluation.
-    pub redaction: Redaction,
-    /// Validation settings for the post-redaction leak check.
-    pub validation: Validation,
-}
-
-/// Per-plan knobs for the redaction phase.
-///
-/// Unset fields fall back to `[redaction]` defaults in
-/// `Nvisy.toml` ([`RedactionConfig`]); if neither is set,
-/// hard-coded defaults apply (0.5 threshold, no metadata
-/// stripping).
-///
-/// [`RedactionConfig`]: super::RedactionConfig
-#[derive(Debug, Clone, Default, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Redaction {
-    /// Minimum confidence threshold for default redaction (0.0 to
-    /// 1.0). Entities below this threshold that don't match a
-    /// policy rule are skipped. `None` falls back to
-    /// `[redaction].confidence_threshold`.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub confidence_threshold: Option<ConfidenceThreshold>,
-    /// Strip or redact document metadata (EXIF, PDF properties).
-    /// `None` falls back to `[redaction].process_metadata`.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub process_metadata: Option<bool>,
-}
-
-/// Per-plan validation settings.
-///
-/// `leak_severity` controls what the phase does when the canonical
-/// [`LeakCheck`] finds a value that should have been redacted but
-/// still appears in the output:
-///
-/// - [`Severity::Warn`] (default) — log the leak and continue.
-/// - [`Severity::Fail`] — fail the pass with a validation error
-///   listing the leaked values.
-///
-/// [`LeakCheck`]: nvisy_toolkit::validation::LeakCheck
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct Validation {
-    /// Severity stamped onto every [`Finding`] emitted by the
-    /// canonical leak check.
-    ///
-    /// [`Finding`]: nvisy_toolkit::validation::Finding
-    #[serde(default)]
-    pub leak_severity: Severity,
-}
diff --git a/crates/nvisy-engine/src/redaction/result.rs b/crates/nvisy-engine/src/redaction/result.rs
deleted file mode 100644
index c48504c18..000000000
--- a/crates/nvisy-engine/src/redaction/result.rs
+++ /dev/null
@@ -1,97 +0,0 @@
-//! [`RedactionResult`] and friends.
-
-use jiff::Timestamp;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-use super::status::RedactionStatus;
-use crate::document::provenance::AnyAudit;
-
-/// Artifact produced by one redaction pass.
-///
-/// Contains the per-document audits with `Execution` populated
-/// (applied / suppressed / failed) plus aggregate counters.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct RedactionResult {
-    /// Unique identifier for this redaction pass.
-    pub id: Uuid,
-    /// Detection pass this redaction was applied against.
-    pub detection_id: Uuid,
-    /// Identity of the actor who initiated the pass.
-    pub actor_id: Uuid,
-    /// Per-document audits with `Execution` populated.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub audits: Vec<AnyAudit>,
-    /// Total redactions actually applied across all documents.
-    pub redactions_applied: u64,
-}
-
-/// Full snapshot of a redaction pass.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct RedactionSnapshot {
-    /// Unique identifier for the redaction pass.
-    pub id: Uuid,
-    /// Detection pass this redaction was applied against.
-    pub detection_id: Uuid,
-    /// Identity of the actor who initiated the pass.
-    pub actor_id: Uuid,
-    /// Lifecycle status (queued / running / completed / failed).
-    pub status: RedactionStatus,
-    /// When the pass was created.
-    #[schemars(with = "String")]
-    pub created_at: Timestamp,
-    /// When the applicator began running, if it ever did.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub started_at: Option<Timestamp>,
-    /// When the applicator reached a terminal state, if it did.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub completed_at: Option<Timestamp>,
-    /// Result is `Some` once the pass has reached a terminal
-    /// state with at least one audit.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub result: Option<RedactionResult>,
-    /// Terminal error message, when the pass failed.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub error: Option<String>,
-}
-
-/// Lightweight summary of a redaction pass for listing endpoints.
-#[derive(Debug, Clone, Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct RedactionEntry {
-    /// Unique identifier for the redaction pass.
-    pub id: Uuid,
-    /// Detection pass this redaction was applied against.
-    pub detection_id: Uuid,
-    /// Identity of the actor who initiated the pass.
-    pub actor_id: Uuid,
-    /// Lifecycle status (queued / running / completed / failed).
-    pub status: RedactionStatus,
-    /// When the pass was created.
-    #[schemars(with = "String")]
-    pub created_at: Timestamp,
-    /// When the applicator began running, if it ever did.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub started_at: Option<Timestamp>,
-    /// When the applicator reached a terminal state, if it did.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    #[schemars(with = "Option<String>")]
-    pub completed_at: Option<Timestamp>,
-    /// Total redactions applied across all documents.
-    pub redactions_applied: u64,
-}
-
-/// Filter criteria for listing redaction passes.
-#[derive(Debug, Clone, Default)]
-pub struct RedactionFilter {
-    /// If set, only return passes with this status.
-    pub status: Option<RedactionStatus>,
-    /// If set, only return passes for this detection.
-    pub detection_id: Option<Uuid>,
-}
diff --git a/crates/nvisy-engine/src/redaction/state.rs b/crates/nvisy-engine/src/redaction/state.rs
deleted file mode 100644
index f95ea69f2..000000000
--- a/crates/nvisy-engine/src/redaction/state.rs
+++ /dev/null
@@ -1,248 +0,0 @@
-//! Volatile in-memory state for active redaction passes.
-
-use std::cmp::Reverse;
-use std::collections::HashMap;
-use std::sync::Arc;
-
-use jiff::Timestamp;
-use nvisy_core::Error;
-use tokio::sync::RwLock;
-use tokio_util::sync::CancellationToken;
-use uuid::Uuid;
-
-use super::result::{RedactionEntry, RedactionFilter, RedactionResult, RedactionSnapshot};
-use super::status::RedactionStatus;
-use crate::document::provenance::AnyAudit;
-
-const TARGET: &str = "nvisy_engine::redaction::state";
-
-#[derive(Clone, Default)]
-pub(crate) struct RedactionState {
-    inner: Arc<RwLock<HashMap<Uuid, RedactionRecord>>>,
-}
-
-pub(crate) struct RedactionRecord {
-    pub detection_id: Uuid,
-    pub actor_id: Uuid,
-    pub status: RedactionStatus,
-    pub created_at: Timestamp,
-    pub started_at: Option<Timestamp>,
-    pub completed_at: Option<Timestamp>,
-    pub cancel: CancellationToken,
-    pub audits: Vec<AnyAudit>,
-    pub redactions_applied: u64,
-    pub error: Option<String>,
-}
-
-impl RedactionRecord {
-    fn to_snapshot(&self, id: Uuid) -> RedactionSnapshot {
-        let result = match self.status {
-            RedactionStatus::Succeeded | RedactionStatus::PartialFailure => Some(RedactionResult {
-                id,
-                detection_id: self.detection_id,
-                actor_id: self.actor_id,
-                audits: self.audits.clone(),
-                redactions_applied: self.redactions_applied,
-            }),
-            _ => None,
-        };
-        RedactionSnapshot {
-            id,
-            detection_id: self.detection_id,
-            actor_id: self.actor_id,
-            status: self.status,
-            created_at: self.created_at,
-            started_at: self.started_at,
-            completed_at: self.completed_at,
-            result,
-            error: self.error.clone(),
-        }
-    }
-
-    fn to_entry(&self, id: Uuid) -> RedactionEntry {
-        RedactionEntry {
-            id,
-            detection_id: self.detection_id,
-            actor_id: self.actor_id,
-            status: self.status,
-            created_at: self.created_at,
-            started_at: self.started_at,
-            completed_at: self.completed_at,
-            redactions_applied: self.redactions_applied,
-        }
-    }
-}
-
-impl RedactionState {
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    pub async fn insert(&self, id: Uuid, record: RedactionRecord) {
-        self.inner.write().await.insert(id, record);
-    }
-
-    /// Look up a snapshot scoped by actor.
-    ///
-    /// # Errors
-    ///
-    /// [`ErrorKind::NotFound`] when the redaction does not exist
-    /// or belongs to a different actor.
-    ///
-    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
-    pub async fn snapshot(&self, actor_id: Uuid, id: Uuid) -> Result<RedactionSnapshot, Error> {
-        let guard = self.inner.read().await;
-        let Some(record) = guard.get(&id) else {
-            return Err(Error::not_found(
-                format!("redaction {id} not found"),
-                TARGET,
-            ));
-        };
-        if record.actor_id != actor_id {
-            return Err(Error::not_found(
-                format!("redaction {id} not found"),
-                TARGET,
-            ));
-        }
-        Ok(record.to_snapshot(id))
-    }
-
-    pub async fn list(&self, actor_id: Uuid, filter: RedactionFilter) -> Vec<RedactionEntry> {
-        let guard = self.inner.read().await;
-        let mut out: Vec<RedactionEntry> = guard
-            .iter()
-            .filter(|(_, r)| r.actor_id == actor_id)
-            .filter(|(_, r)| filter.status.is_none_or(|s| r.status == s))
-            .filter(|(_, r)| filter.detection_id.is_none_or(|d| r.detection_id == d))
-            .map(|(id, r)| r.to_entry(*id))
-            .collect();
-        out.sort_by_key(|e| Reverse(e.created_at));
-        out
-    }
-
-    pub async fn set_started_at(&self, id: Uuid) {
-        let mut guard = self.inner.write().await;
-        if let Some(record) = guard.get_mut(&id)
-            && record.started_at.is_none()
-        {
-            record.started_at = Some(Timestamp::now());
-            record.status = RedactionStatus::Running;
-        }
-    }
-
-    pub async fn fail(&self, id: Uuid, error: impl Into<String>) {
-        let mut guard = self.inner.write().await;
-        if let Some(record) = guard.get_mut(&id) {
-            record.status = RedactionStatus::Failed;
-            record.completed_at = Some(Timestamp::now());
-            record.error = Some(error.into());
-        }
-    }
-
-    pub async fn finalize(
-        &self,
-        id: Uuid,
-        status: RedactionStatus,
-        audits: Vec<AnyAudit>,
-        redactions_applied: u64,
-    ) {
-        let mut guard = self.inner.write().await;
-        if let Some(record) = guard.get_mut(&id) {
-            record.status = status;
-            record.completed_at = Some(Timestamp::now());
-            record.audits = audits;
-            record.redactions_applied = redactions_applied;
-        }
-    }
-
-    /// Cancel an in-progress redaction.
-    ///
-    /// # Errors
-    ///
-    /// - [`ErrorKind::NotFound`] when the redaction does not
-    ///   exist or belongs to a different actor.
-    /// - [`ErrorKind::Conflict`] when the redaction is already
-    ///   in a terminal state.
-    ///
-    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
-    /// [`ErrorKind::Conflict`]: nvisy_core::ErrorKind::Conflict
-    pub async fn cancel(&self, actor_id: Uuid, id: Uuid) -> Result<(), Error> {
-        let mut guard = self.inner.write().await;
-        let Some(record) = guard.get_mut(&id) else {
-            return Err(Error::not_found(
-                format!("redaction {id} not found"),
-                TARGET,
-            ));
-        };
-        if record.actor_id != actor_id {
-            return Err(Error::not_found(
-                format!("redaction {id} not found"),
-                TARGET,
-            ));
-        }
-        if record.status.is_terminal() {
-            return Err(Error::conflict(
-                format!(
-                    "redaction {id} already in terminal state {:?}",
-                    record.status
-                ),
-                TARGET,
-            ));
-        }
-        record.cancel.cancel();
-        record.status = RedactionStatus::Cancelled;
-        record.completed_at = Some(Timestamp::now());
-        Ok(())
-    }
-
-    /// Delete a finished redaction.
-    ///
-    /// # Errors
-    ///
-    /// - [`ErrorKind::NotFound`] when the redaction does not
-    ///   exist or belongs to a different actor.
-    /// - [`ErrorKind::Conflict`] when the redaction is still
-    ///   active.
-    ///
-    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
-    /// [`ErrorKind::Conflict`]: nvisy_core::ErrorKind::Conflict
-    pub async fn delete(&self, actor_id: Uuid, id: Uuid) -> Result<(), Error> {
-        let mut guard = self.inner.write().await;
-        let Some(record) = guard.get(&id) else {
-            return Err(Error::not_found(
-                format!("redaction {id} not found"),
-                TARGET,
-            ));
-        };
-        if record.actor_id != actor_id {
-            return Err(Error::not_found(
-                format!("redaction {id} not found"),
-                TARGET,
-            ));
-        }
-        if !record.status.is_terminal() {
-            return Err(Error::conflict(
-                format!("redaction {id} cannot be deleted while {:?}", record.status),
-                TARGET,
-            ));
-        }
-        guard.remove(&id);
-        Ok(())
-    }
-
-    /// Delete every terminal-status redaction for `actor_id`.
-    /// Returns the ids removed so the caller can cascade the
-    /// delete to other stores (registry persistence).
-    pub async fn delete_all(&self, actor_id: Uuid) -> Vec<Uuid> {
-        let mut guard = self.inner.write().await;
-        let to_delete: Vec<Uuid> = guard
-            .iter()
-            .filter(|(_, r)| r.actor_id == actor_id && r.status.is_terminal())
-            .map(|(id, _)| *id)
-            .collect();
-        for id in &to_delete {
-            guard.remove(id);
-        }
-        to_delete
-    }
-}
diff --git a/crates/nvisy-engine/src/redaction/status.rs b/crates/nvisy-engine/src/redaction/status.rs
deleted file mode 100644
index 6f5bd6946..000000000
--- a/crates/nvisy-engine/src/redaction/status.rs
+++ /dev/null
@@ -1,37 +0,0 @@
-//! [`RedactionStatus`]: lifecycle states for a redaction pass.
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Lifecycle states for a single redaction pass.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum RedactionStatus {
-    /// Submitted, not yet started.
-    Pending,
-    /// Apply / validate / export running.
-    Running,
-    /// All documents redacted, validated, and exported.
-    Succeeded,
-    /// Some documents redacted while others failed.
-    PartialFailure,
-    /// Every document errored or the pass failed before any
-    /// document was processed.
-    Failed,
-    /// The caller cancelled the redaction pass before it finished.
-    Cancelled,
-}
-
-impl RedactionStatus {
-    /// `true` for [`Self::Succeeded`] / [`Self::PartialFailure`] /
-    /// [`Self::Failed`] / [`Self::Cancelled`] — anything past the
-    /// running phase.
-    #[must_use]
-    pub fn is_terminal(self) -> bool {
-        matches!(
-            self,
-            Self::Succeeded | Self::PartialFailure | Self::Failed | Self::Cancelled
-        )
-    }
-}
diff --git a/crates/nvisy-engine/src/registry/composite_key.rs b/crates/nvisy-engine/src/registry/composite_key.rs
deleted file mode 100644
index 2163abda6..000000000
--- a/crates/nvisy-engine/src/registry/composite_key.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-//! Actor-scoped composite key for registry storage.
-
-use derive_more::{AsRef, Deref, From};
-use uuid::Uuid;
-
-/// 32-byte actor-scoped key: `[actor_id: 16][resource_id: 16]`.
-///
-/// Every registry entry is scoped to an actor. This type encodes
-/// that invariant and provides ergonomic construction from two UUIDs.
-#[derive(Clone, Copy, Deref, AsRef, From)]
-#[as_ref(forward)]
-pub(crate) struct CompositeKey(#[deref] [u8; 32]);
-
-impl CompositeKey {
-    /// Build a key from an actor UUID and a resource UUID.
-    pub fn new(actor_id: Uuid, resource_id: Uuid) -> Self {
-        let mut key = [0u8; 32];
-        key[..16].copy_from_slice(actor_id.as_bytes());
-        key[16..].copy_from_slice(resource_id.as_bytes());
-        Self(key)
-    }
-
-    /// Extract the actor UUID from the leading 16 bytes.
-    pub fn actor_id(&self) -> Uuid {
-        let mut bytes = [0u8; 16];
-        bytes.copy_from_slice(&self.0[..16]);
-        Uuid::from_bytes(bytes)
-    }
-
-    /// Extract the resource UUID from the trailing 16 bytes.
-    pub fn resource_id(&self) -> Uuid {
-        let mut bytes = [0u8; 16];
-        bytes.copy_from_slice(&self.0[16..]);
-        Uuid::from_bytes(bytes)
-    }
-}
diff --git a/crates/nvisy-engine/src/registry/content_handle.rs b/crates/nvisy-engine/src/registry/content_handle.rs
deleted file mode 100644
index a5336c797..000000000
--- a/crates/nvisy-engine/src/registry/content_handle.rs
+++ /dev/null
@@ -1,104 +0,0 @@
-//! [`ContentHandle`]: async handle to stored content data and record.
-
-use std::fmt;
-
-use bytes::Bytes;
-use fjall::Keyspace;
-use nvisy_codec::content::{
-    Content, ContentData, ContentDescriptor, ContentDigest, ContentRecord, ContentSource,
-};
-use nvisy_core::Result;
-use uuid::Uuid;
-
-use super::composite_key::CompositeKey;
-use super::fjall_ext::{FjallKeyspaceExt, blocking, not_found};
-
-/// Lightweight handle to a content entry stored in the registry.
-///
-/// Holds references to the fjall keyspaces so it can read content
-/// data and its persisted [`ContentRecord`] on demand.
-#[derive(Clone)]
-pub struct ContentHandle {
-    actor_id: Uuid,
-    content_source: ContentSource,
-    content_ks: Keyspace,
-    content_meta_ks: Keyspace,
-}
-
-impl fmt::Debug for ContentHandle {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("ContentHandle")
-            .field("actor_id", &self.actor_id)
-            .field("content_source", &self.content_source)
-            .finish_non_exhaustive()
-    }
-}
-
-impl ContentHandle {
-    pub(crate) fn new(
-        actor_id: Uuid,
-        content_source: ContentSource,
-        content_ks: Keyspace,
-        content_meta_ks: Keyspace,
-    ) -> Self {
-        Self {
-            actor_id,
-            content_source,
-            content_ks,
-            content_meta_ks,
-        }
-    }
-
-    /// Returns the content source identifier.
-    #[must_use]
-    pub fn content_source(&self) -> ContentSource {
-        self.content_source
-    }
-
-    /// Reads the content bytes from the store.
-    pub async fn content_data(&self) -> Result<ContentData> {
-        let key = CompositeKey::new(self.actor_id, self.content_source.as_uuid());
-        let source = self.content_source;
-        let ks = self.content_ks.clone();
-
-        blocking(move || {
-            let bytes = ks
-                .get_bytes(key)?
-                .ok_or_else(|| not_found("content", Uuid::nil(), source.as_uuid()))?;
-            Ok(ContentData::new(source, Bytes::from(bytes)))
-        })
-        .await
-    }
-
-    /// Reads the persisted [`ContentRecord`] (descriptor + digest).
-    ///
-    /// Returns a default record (empty descriptor, zero digest) when
-    /// no record was stored — historically the metadata keyspace can
-    /// hold entries written before the record format existed.
-    pub async fn record(&self) -> Result<ContentRecord> {
-        let key = CompositeKey::new(self.actor_id, self.content_source.as_uuid());
-        let ks = self.content_meta_ks.clone();
-
-        blocking(move || match ks.get_bytes(key)? {
-            Some(bytes) => Ok(serde_json::from_slice(&bytes)?),
-            None => Ok(ContentRecord {
-                descriptor: ContentDescriptor::default(),
-                digest: ContentDigest {
-                    size: 0,
-                    sha256: String::new(),
-                    detected_content_type: None,
-                },
-            }),
-        })
-        .await
-    }
-
-    /// Reads bytes + descriptor, returning a [`Content`] suitable for
-    /// re-importing through the codec pipeline. The digest half of
-    /// the record is dropped — codec resolution only needs the
-    /// caller-supplied bits (extension, MIME hint).
-    pub async fn content(&self) -> Result<Content> {
-        let (data, record) = tokio::try_join!(self.content_data(), self.record())?;
-        Ok(Content::with_descriptor(data, record.descriptor))
-    }
-}
diff --git a/crates/nvisy-engine/src/registry/fjall_ext.rs b/crates/nvisy-engine/src/registry/fjall_ext.rs
index 5efddabcb..2cb03979b 100644
--- a/crates/nvisy-engine/src/registry/fjall_ext.rs
+++ b/crates/nvisy-engine/src/registry/fjall_ext.rs
@@ -1,8 +1,7 @@
 //! Fjall extension traits and async helpers for the registry module.
 //!
-//! [`FjallKeyspaceExt`] and [`FjallDatabaseExt`] wrap raw fjall
-//! operations with [`CompositeKey`] support and consistent error
-//! mapping to [`Error`].
+//! [`FjallDatabaseExt`] wraps raw fjall database operations with
+//! consistent error mapping to [`Error`].
 //!
 //! [`Error`]: nvisy_core::Error
 
@@ -13,76 +12,8 @@ use fjall::{Database, Keyspace, KeyspaceCreateOptions, KvSeparationOptions};
 use nvisy_core::{Error, ErrorKind, Result};
 use uuid::Uuid;
 
-use super::composite_key::CompositeKey;
-
 const COMPONENT: &str = "registry";
 
-/// Extension trait for fjall [`Keyspace`] with consistent error mapping.
-///
-/// Wraps raw fjall operations to accept [`CompositeKey`] and return
-/// [`Result`] with standardized error context.
-///
-/// [`Result`]: nvisy_core::Result
-pub(crate) trait FjallKeyspaceExt {
-    /// Insert a value at the given composite key.
-    fn put(&self, key: CompositeKey, value: &[u8]) -> Result<()>;
-    /// Read the value at a key, returning `None` if absent.
-    fn get_bytes(&self, key: CompositeKey) -> Result<Option<Vec<u8>>>;
-    /// Remove the entry at a key.
-    fn delete(&self, key: CompositeKey) -> Result<()>;
-    /// Check whether a key exists.
-    fn exists(&self, key: CompositeKey) -> Result<bool>;
-    /// List all resource UUIDs for the given actor (sorted).
-    fn resource_ids(&self, actor_id: Uuid) -> Result<Vec<Uuid>>;
-    /// Collect all composite keys sharing the given byte prefix.
-    fn prefix_keys(&self, prefix: &[u8]) -> Result<Vec<CompositeKey>>;
-}
-
-impl FjallKeyspaceExt for Keyspace {
-    fn put(&self, key: CompositeKey, value: &[u8]) -> Result<()> {
-        self.insert(*key, value)
-            .map_err(|err| fjall_err("failed to write entry", err))
-    }
-
-    fn get_bytes(&self, key: CompositeKey) -> Result<Option<Vec<u8>>> {
-        self.get(*key)
-            .map(|opt| opt.map(|guard| guard.to_vec()))
-            .map_err(|err| fjall_err("failed to read entry", err))
-    }
-
-    fn delete(&self, key: CompositeKey) -> Result<()> {
-        self.remove(*key)
-            .map_err(|err| fjall_err("failed to remove entry", err))
-    }
-
-    fn exists(&self, key: CompositeKey) -> Result<bool> {
-        self.contains_key(*key)
-            .map_err(|err| fjall_err("failed to check key", err))
-    }
-
-    fn resource_ids(&self, actor_id: Uuid) -> Result<Vec<Uuid>> {
-        let keys = self.prefix_keys(actor_id.as_bytes())?;
-        let mut ids: Vec<Uuid> = keys.iter().map(|k| k.resource_id()).collect();
-        ids.sort();
-        Ok(ids)
-    }
-
-    fn prefix_keys(&self, prefix: &[u8]) -> Result<Vec<CompositeKey>> {
-        self.prefix(prefix)
-            .map(|guard| {
-                let key = guard
-                    .key()
-                    .map_err(|err| fjall_err("failed to iterate keyspace", err))?;
-                let bytes: [u8; 32] = key.as_ref().try_into().map_err(|_| {
-                    Error::new(ErrorKind::Internal, "unexpected key length")
-                        .with_component(COMPONENT)
-                })?;
-                Ok(CompositeKey::from(bytes))
-            })
-            .collect()
-    }
-}
-
 /// Extension trait for fjall [`Database`] with consistent error mapping.
 ///
 /// Provides ergonomic database and keyspace creation methods that
diff --git a/crates/nvisy-engine/src/registry/handle.rs b/crates/nvisy-engine/src/registry/handle.rs
new file mode 100644
index 000000000..3854da160
--- /dev/null
+++ b/crates/nvisy-engine/src/registry/handle.rs
@@ -0,0 +1,125 @@
+//! [`RegistryHandle`]: owns the fjall [`Database`] + every
+//! keyspace the engine resource modules read and write.
+//!
+//! Constructed once at server startup with
+//! [`RegistryHandle::open`]; cheaply cloneable (the fjall types
+//! are `Arc`-backed) so it can be passed to every resource API as
+//! a borrowed handle.
+
+use std::path::Path;
+use std::sync::Arc;
+
+use fjall::{Database, Keyspace};
+use nvisy_core::Result;
+
+use super::fjall_ext::FjallDatabaseExt;
+
+/// Keyspace name constants. Each is one logical resource class.
+const KS_POLICIES: &str = "policies";
+const KS_CONTEXTS: &str = "contexts";
+const KS_FILES_METADATA: &str = "files_metadata";
+const KS_FILES_CONTENT: &str = "files_content";
+const KS_RUN_HEADERS: &str = "run_headers";
+const KS_RUN_DOCS: &str = "run_docs";
+
+/// Cheaply-cloneable handle over the engine's fjall keyspaces.
+#[derive(Clone)]
+pub struct RegistryHandle {
+    inner: Arc<RegistryInner>,
+}
+
+struct RegistryInner {
+    db: Database,
+    policies: Keyspace,
+    contexts: Keyspace,
+    files_metadata: Keyspace,
+    files_content: Keyspace,
+    run_headers: Keyspace,
+    run_docs: Keyspace,
+}
+
+impl RegistryHandle {
+    /// Open (or create) the engine database at `path` and pre-open
+    /// every keyspace. Idempotent — re-opening an existing database
+    /// at the same path returns a handle to the same data.
+    pub fn open(path: &Path) -> Result<Self> {
+        let db = Database::open_at(path)?;
+        let policies = db.open_blob_keyspace(KS_POLICIES)?;
+        let contexts = db.open_blob_keyspace(KS_CONTEXTS)?;
+        let files_metadata = db.open_keyspace(KS_FILES_METADATA)?;
+        let files_content = db.open_blob_keyspace(KS_FILES_CONTENT)?;
+        let run_headers = db.open_keyspace(KS_RUN_HEADERS)?;
+        let run_docs = db.open_blob_keyspace(KS_RUN_DOCS)?;
+        Ok(Self {
+            inner: Arc::new(RegistryInner {
+                db,
+                policies,
+                contexts,
+                files_metadata,
+                files_content,
+                run_headers,
+                run_docs,
+            }),
+        })
+    }
+
+    /// Keyspace for persisted `Policy` resources. Blob-separated:
+    /// policy bodies range from small to large (rule lists +
+    /// catalogs).
+    pub(crate) fn policies(&self) -> &Keyspace {
+        &self.inner.policies
+    }
+
+    /// Keyspace for persisted `Context` resources. Blob-separated.
+    pub(crate) fn contexts(&self) -> &Keyspace {
+        &self.inner.contexts
+    }
+
+    /// Keyspace for [`FileMetadata`] blobs — small JSON
+    /// descriptors keyed by `(actor_id, file_id)`. Plain
+    /// keyspace (no blob separation) because the bodies are
+    /// tiny; range-scan for `list_files` is cheap.
+    ///
+    /// [`FileMetadata`]: nvisy_core::FileMetadata
+    pub(crate) fn files_metadata(&self) -> &Keyspace {
+        &self.inner.files_metadata
+    }
+
+    /// Keyspace for raw file bytes keyed by `(actor_id,
+    /// file_id)`. Blob-separated — files routinely run from
+    /// kilobytes to many megabytes and the bytes are only
+    /// loaded when the caller asks for `GET /files/{id}/content`.
+    pub(crate) fn files_content(&self) -> &Keyspace {
+        &self.inner.files_content
+    }
+
+    /// Keyspace for [`Run`] headers — short metadata blobs holding
+    /// run state + per-policy / per-context refs.
+    ///
+    /// [`Run`]: super::super::runs::Run
+    pub(crate) fn run_headers(&self) -> &Keyspace {
+        &self.inner.run_headers
+    }
+
+    /// Keyspace for per-document run bodies — one entry per
+    /// `(actor_id, run_id, doc_id)` carrying that document's
+    /// recognized entities + reviewer overrides. Bytes (both
+    /// input and post-apply redacted output) live in the
+    /// [`files_metadata`](Self::files_metadata) /
+    /// [`files_content`](Self::files_content) keyspaces; the
+    /// row's [`RunDocument::input_file_id`] and
+    /// [`output_file_id`](RunDocument::output_file_id) point at
+    /// them.
+    ///
+    /// [`RunDocument::input_file_id`]: super::super::runs::RunDocument::input_file_id
+    /// [`output_file_id`]: super::super::runs::RunDocument::output_file_id
+    pub(crate) fn run_docs(&self) -> &Keyspace {
+        &self.inner.run_docs
+    }
+
+    /// Flush all pending writes to disk. The engine HTTP layer
+    /// calls this on graceful shutdown.
+    pub fn sync(&self) -> Result<()> {
+        self.inner.db.sync()
+    }
+}
diff --git a/crates/nvisy-engine/src/registry/key.rs b/crates/nvisy-engine/src/registry/key.rs
new file mode 100644
index 000000000..ae9a05cf1
--- /dev/null
+++ b/crates/nvisy-engine/src/registry/key.rs
@@ -0,0 +1,138 @@
+//! Composite keys for the fjall keyspaces.
+//!
+//! Three key shapes, all big-endian byte-encoded so fjall's lex
+//! ordering matches the natural prefix order:
+//!
+//! - [`CompositeKey`] — `(actor_id, resource_id)`, 32 bytes. Existing
+//!   2-component scoping inherited from the salvaged registry.
+//! - [`VersionedKey`] — `(actor_id, resource_id, version_serialised)`,
+//!   variable length. For per-version resource storage (policies,
+//!   contexts). Prefix scan by `(actor, id)` returns every version of
+//!   the same logical resource in lex order.
+//! - [`TripleKey`] — `(actor_id, run_id, doc_id)`, 48 bytes. For
+//!   per-document run bodies. Prefix scan by `(actor, run)` returns
+//!   every document in a run.
+
+use derive_more::{AsRef, Deref, From};
+use semver::Version;
+use uuid::Uuid;
+
+/// 32-byte actor-scoped key: `[actor_id: 16][resource_id: 16]`.
+///
+/// Identical to the salvaged shape; one entry per resource.
+#[derive(Clone, Copy, Deref, AsRef, From)]
+#[as_ref(forward)]
+pub(crate) struct CompositeKey(#[deref] [u8; 32]);
+
+impl CompositeKey {
+    /// Build from an actor + resource UUID.
+    pub fn new(actor_id: Uuid, resource_id: Uuid) -> Self {
+        let mut key = [0u8; 32];
+        key[..16].copy_from_slice(actor_id.as_bytes());
+        key[16..].copy_from_slice(resource_id.as_bytes());
+        Self(key)
+    }
+
+    /// Prefix bytes for "every resource of `actor`": 16 bytes.
+    /// Use with `Keyspace::prefix` to range over every entry an
+    /// actor owns in a 2-component keyspace.
+    pub fn actor_prefix(actor_id: Uuid) -> [u8; 16] {
+        *actor_id.as_bytes()
+    }
+}
+
+/// 48-byte three-component key: `[a: 16][b: 16][c: 16]`. Used for
+/// per-document run-body entries — `(actor_id, run_id, doc_id)`.
+#[derive(Clone, Copy, Deref, AsRef, From)]
+#[as_ref(forward)]
+pub(crate) struct TripleKey(#[deref] [u8; 48]);
+
+impl TripleKey {
+    /// Build from three UUIDs.
+    pub fn new(a: Uuid, b: Uuid, c: Uuid) -> Self {
+        let mut key = [0u8; 48];
+        key[..16].copy_from_slice(a.as_bytes());
+        key[16..32].copy_from_slice(b.as_bytes());
+        key[32..].copy_from_slice(c.as_bytes());
+        Self(key)
+    }
+
+    /// Prefix bytes for "every C of `(a, b)`": 32 bytes. Use with
+    /// `Keyspace::prefix` to range over every per-doc entry of one
+    /// run (e.g. cascade-delete across `run_docs` /
+    /// `run_artifacts` / `run_inputs`).
+    pub fn ab_prefix(a: Uuid, b: Uuid) -> [u8; 32] {
+        let mut prefix = [0u8; 32];
+        prefix[..16].copy_from_slice(a.as_bytes());
+        prefix[16..].copy_from_slice(b.as_bytes());
+        prefix
+    }
+}
+
+/// Variable-length per-version key:
+/// `[actor: 16][resource: 16][version_be_serialised: N]`.
+///
+/// The trailing version bytes use a stable big-endian encoding of
+/// `(major, minor, patch)` u64s — `[major: 8][minor: 8][patch: 8]`
+/// — so lex order on the byte sequence matches semver order. Pre-
+/// release / build metadata are dropped from the key (they ride
+/// along in the value); two `(id, version)` pairs are considered
+/// distinct if and only if their `(major, minor, patch)` differs.
+pub(crate) struct VersionedKey(Vec<u8>);
+
+impl VersionedKey {
+    /// Build from `(actor, id, version)`.
+    pub fn new(actor_id: Uuid, resource_id: Uuid, version: &Version) -> Self {
+        let mut buf = Vec::with_capacity(16 + 16 + 24);
+        buf.extend_from_slice(actor_id.as_bytes());
+        buf.extend_from_slice(resource_id.as_bytes());
+        buf.extend_from_slice(&version.major.to_be_bytes());
+        buf.extend_from_slice(&version.minor.to_be_bytes());
+        buf.extend_from_slice(&version.patch.to_be_bytes());
+        Self(buf)
+    }
+
+    /// Borrow the underlying bytes for fjall.
+    pub fn as_bytes(&self) -> &[u8] {
+        &self.0
+    }
+
+    /// Prefix bytes for "every version of `(actor, id)`": 32 bytes,
+    /// just `[actor][id]`. Use with `Keyspace::prefix` to range over
+    /// all versions of one logical resource.
+    pub fn prefix(actor_id: Uuid, resource_id: Uuid) -> [u8; 32] {
+        let mut prefix = [0u8; 32];
+        prefix[..16].copy_from_slice(actor_id.as_bytes());
+        prefix[16..].copy_from_slice(resource_id.as_bytes());
+        prefix
+    }
+
+    /// Prefix bytes for "every resource of `actor`": 16 bytes.
+    pub fn actor_prefix(actor_id: Uuid) -> [u8; 16] {
+        *actor_id.as_bytes()
+    }
+
+    /// Read the version triple back from a full versioned-key byte
+    /// slice (last 24 bytes).
+    pub fn version_from_bytes(bytes: &[u8]) -> Option<(u64, u64, u64)> {
+        if bytes.len() < 32 + 24 {
+            return None;
+        }
+        let tail = &bytes[32..32 + 24];
+        let major = u64::from_be_bytes(tail[..8].try_into().ok()?);
+        let minor = u64::from_be_bytes(tail[8..16].try_into().ok()?);
+        let patch = u64::from_be_bytes(tail[16..24].try_into().ok()?);
+        Some((major, minor, patch))
+    }
+
+    /// Read the resource id (bytes 16..32) from a full versioned-key
+    /// byte slice.
+    pub fn resource_id_from_bytes(bytes: &[u8]) -> Option<Uuid> {
+        if bytes.len() < 32 {
+            return None;
+        }
+        let mut b = [0u8; 16];
+        b.copy_from_slice(&bytes[16..32]);
+        Some(Uuid::from_bytes(b))
+    }
+}
diff --git a/crates/nvisy-engine/src/registry/mod.rs b/crates/nvisy-engine/src/registry/mod.rs
index 2edf075de..83fc628eb 100644
--- a/crates/nvisy-engine/src/registry/mod.rs
+++ b/crates/nvisy-engine/src/registry/mod.rs
@@ -1,17 +1,27 @@
-//! Actor-scoped content and audit storage backed by fjall.
+//! Multi-tenant registry over fjall.
 //!
-//! The [`Registry`] stores content blobs, content metadata,
-//! annotations, audits, and detection / redaction results. All
-//! entries are actor-scoped via a composite key. Policies are NOT
-//! persisted — they are submitted inline on every detection input
-//! and snapshotted onto the in-memory detection record.
+//! Three keyspace shapes, all actor-scoped:
+//!
+//! - **policies** + **contexts** keyspaces hold [`Policy`] and
+//!   [`Context`] resources versioned in place. Lookup by
+//!   `(actor_id, resource_id, version)`; latest-version lookup
+//!   range-scans by `(actor_id, resource_id)`.
+//! - **run_headers** holds short metadata blobs for each run (state,
+//!   refs to policies/contexts, timestamps). Lookup by
+//!   `(actor_id, run_id)`.
+//! - **run_docs** holds the per-document body for each run
+//!   (recognized entities + reviewer overrides + post-apply
+//!   redacted bytes/audit). Lookup by `(actor_id, run_id, doc_id)`.
+//!
+//! [`Policy`]: nvisy_core::policy::Policy
+//! [`Context`]: nvisy_core::context::Context
 
-mod composite_key;
-mod content_handle;
 mod fjall_ext;
+mod handle;
+mod key;
 mod paged;
-mod registry_store;
 
-pub use self::content_handle::ContentHandle;
+pub(crate) use self::fjall_ext::{blocking, not_found};
+pub use self::handle::RegistryHandle;
+pub(crate) use self::key::{CompositeKey, TripleKey, VersionedKey};
 pub use self::paged::PagedResult;
-pub use self::registry_store::Registry;
diff --git a/crates/nvisy-engine/src/registry/registry_store.rs b/crates/nvisy-engine/src/registry/registry_store.rs
deleted file mode 100644
index d1ad571ac..000000000
--- a/crates/nvisy-engine/src/registry/registry_store.rs
+++ /dev/null
@@ -1,626 +0,0 @@
-//! [`Registry`]: actor-scoped content and policy store.
-
-use std::borrow::Cow;
-use std::fmt;
-use std::path::{Path, PathBuf};
-use std::sync::Arc;
-
-use fjall::{Database, Keyspace};
-use nvisy_codec::content::{
-    Content, ContentDescriptor, ContentDigest, ContentRecord, ContentSource,
-};
-use nvisy_core::Result;
-use nvisy_core::health::{Healthcheck, ServiceStatus};
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-use super::composite_key::CompositeKey;
-use super::content_handle::ContentHandle;
-use super::fjall_ext::{FjallDatabaseExt, FjallKeyspaceExt, blocking, not_found};
-use super::paged::PagedResult;
-use crate::document::AnyAnnotations;
-use crate::document::provenance::AnyAudit;
-
-const TARGET: &str = "nvisy_engine::ingestion::registry";
-
-/// Actor-scoped content and policy store backed by fjall.
-///
-/// Cheaply cloneable (`Arc` internally).
-#[derive(Clone)]
-pub struct Registry {
-    inner: Arc<RegistryInner>,
-}
-
-struct RegistryInner {
-    base_dir: PathBuf,
-    db: Database,
-    content_ks: Keyspace,
-    content_meta_ks: Keyspace,
-    annotations_ks: Keyspace,
-    audits_ks: Keyspace,
-    /// Persisted [`DetectionResult`]s, keyed by
-    /// `(actor_id, detection_id)`. JSON-encoded; one entry per
-    /// terminal-state detection pass.
-    ///
-    /// [`DetectionResult`]: crate::detection::DetectionResult
-    detections_ks: Keyspace,
-    /// Persisted [`RedactionResult`]s, keyed by
-    /// `(actor_id, redaction_id)`. JSON-encoded; one entry per
-    /// terminal-state redaction pass.
-    ///
-    /// [`RedactionResult`]: crate::redaction::RedactionResult
-    redactions_ks: Keyspace,
-}
-
-impl fmt::Debug for Registry {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("Registry")
-            .field("base_dir", &self.inner.base_dir)
-            .finish_non_exhaustive()
-    }
-}
-
-impl Registry {
-    /// Opens (or creates) the database at `path`.
-    #[tracing::instrument(target = TARGET, name = "registry.open", fields(path = %path.as_ref().display()))]
-    pub fn open(path: impl AsRef<Path>) -> Result<Self> {
-        let base_dir = path.as_ref().to_path_buf();
-        let db = Database::open_at(&base_dir)?;
-        let content_ks = db.open_blob_keyspace("content")?;
-        let content_meta_ks = db.open_keyspace("content_meta")?;
-        let annotations_ks = db.open_keyspace("annotations")?;
-        let audits_ks = db.open_keyspace("audits")?;
-        let detections_ks = db.open_keyspace("detections")?;
-        let redactions_ks = db.open_keyspace("redactions")?;
-
-        tracing::debug!(target: TARGET, "registry opened");
-        Ok(Self {
-            inner: Arc::new(RegistryInner {
-                base_dir,
-                db,
-                content_ks,
-                content_meta_ks,
-                annotations_ks,
-                audits_ks,
-                detections_ks,
-                redactions_ks,
-            }),
-        })
-    }
-
-    /// Returns the base directory path.
-    #[must_use]
-    pub fn base_dir(&self) -> &Path {
-        &self.inner.base_dir
-    }
-
-    /// Registers content: stores the raw bytes, builds the
-    /// [`ContentRecord`] (descriptor + freshly computed digest), and
-    /// optionally attaches user-supplied annotations — all in one
-    /// durable write.
-    ///
-    /// All three keyspaces (content, content_meta, annotations) are
-    /// updated inside one `blocking` closure and committed by a
-    /// single `db.sync()`, so a crash mid-call leaves either every
-    /// write or none — never an orphan content blob with missing
-    /// metadata or annotations.
-    #[tracing::instrument(target = TARGET, name = "registry.register_content", skip(self, content, annotations), fields(%actor_id))]
-    pub async fn register_content(
-        &self,
-        actor_id: Uuid,
-        content: Content,
-        annotations: Option<&AnyAnnotations>,
-    ) -> Result<ContentHandle> {
-        let content_source = content.content_source();
-        let key = CompositeKey::new(actor_id, content_source.as_uuid());
-        let data = content.as_bytes().to_vec();
-
-        let (content_data, descriptor) = content.into_parts();
-        let descriptor = descriptor.unwrap_or_default();
-        let digest = ContentDigest {
-            size: content_data.size() as u64,
-            sha256: content_data.sha256_hex(),
-            detected_content_type: content_data.detect_mime(),
-        };
-        let record = ContentRecord { descriptor, digest };
-
-        let record_bytes = serde_json::to_vec(&record)?;
-        let annotations_bytes = match annotations {
-            Some(a) if !a.is_empty() => Some(serde_json::to_vec(a)?),
-            _ => None,
-        };
-
-        let content_ks = self.inner.content_ks.clone();
-        let meta_ks = self.inner.content_meta_ks.clone();
-        let annotations_ks = self.inner.annotations_ks.clone();
-        let db = self.inner.db.clone();
-
-        blocking(move || {
-            content_ks.put(key, &data)?;
-            meta_ks.put(key, &record_bytes)?;
-            if let Some(bytes) = annotations_bytes {
-                annotations_ks.put(key, &bytes)?;
-            }
-            db.sync()
-        })
-        .await?;
-
-        tracing::trace!(target: TARGET, source_id = %content_source.as_uuid(), "content registered");
-        Ok(ContentHandle::new(
-            actor_id,
-            content_source,
-            self.inner.content_ks.clone(),
-            self.inner.content_meta_ks.clone(),
-        ))
-    }
-
-    /// Reads previously registered content.
-    #[tracing::instrument(target = TARGET, name = "registry.read_content", skip(self), fields(%actor_id, %content_id))]
-    pub async fn read_content(&self, actor_id: Uuid, content_id: Uuid) -> Result<ContentHandle> {
-        let key = CompositeKey::new(actor_id, content_id);
-        let ks = self.inner.content_ks.clone();
-        let exists = blocking(move || ks.exists(key)).await?;
-        if !exists {
-            return Err(not_found("content", actor_id, content_id));
-        }
-        Ok(ContentHandle::new(
-            actor_id,
-            ContentSource::from(content_id),
-            self.inner.content_ks.clone(),
-            self.inner.content_meta_ks.clone(),
-        ))
-    }
-
-    /// Removes a single content entry (data + metadata).
-    /// Removes a content entry plus its metadata. Returns
-    /// `NotFound` if no entry exists for the given actor/content
-    /// pair.
-    ///
-    /// The check-then-delete sequence is not transactionally
-    /// atomic: two concurrent calls for the same key race on
-    /// `exists`, one wins, the other returns `NotFound`. That's
-    /// the intended outcome (the second caller's view is "the key
-    /// is gone"), so no synchronisation is needed beyond fjall's
-    /// per-key linearisability.
-    #[tracing::instrument(target = TARGET, name = "registry.unregister_content", skip(self), fields(%actor_id, %content_id))]
-    pub async fn unregister_content(&self, actor_id: Uuid, content_id: Uuid) -> Result<()> {
-        let key = CompositeKey::new(actor_id, content_id);
-        let content_ks = self.inner.content_ks.clone();
-        let meta_ks = self.inner.content_meta_ks.clone();
-        let annotations_ks = self.inner.annotations_ks.clone();
-        let db = self.inner.db.clone();
-
-        blocking(move || {
-            if !content_ks.exists(key)? {
-                return Err(not_found("content", actor_id, content_id));
-            }
-            content_ks.delete(key)?;
-            meta_ks.delete(key)?;
-            annotations_ks.delete(key)?;
-            db.sync()
-        })
-        .await
-    }
-
-    /// Removes all content for an actor. Returns the number removed.
-    #[tracing::instrument(target = TARGET, name = "registry.unregister_all_content", skip(self), fields(%actor_id))]
-    pub async fn unregister_all_content(&self, actor_id: Uuid) -> Result<usize> {
-        let content_ks = self.inner.content_ks.clone();
-        let meta_ks = self.inner.content_meta_ks.clone();
-        let annotations_ks = self.inner.annotations_ks.clone();
-        let db = self.inner.db.clone();
-
-        blocking(move || {
-            let keys = content_ks.prefix_keys(actor_id.as_bytes())?;
-            let count = keys.len();
-            for key in &keys {
-                content_ks.delete(*key)?;
-                meta_ks.delete(*key)?;
-                annotations_ks.delete(*key)?;
-            }
-            if count > 0 {
-                db.sync()?;
-            }
-            Ok(count)
-        })
-        .await
-    }
-
-    /// Lists all content IDs for the given actor.
-    #[tracing::instrument(target = TARGET, name = "registry.list_content", skip(self), fields(%actor_id))]
-    pub async fn list_content(&self, actor_id: Uuid) -> Result<Vec<Uuid>> {
-        let ks = self.inner.content_ks.clone();
-        blocking(move || ks.resource_ids(actor_id)).await
-    }
-
-    /// Lists a window of stored records for the given actor.
-    ///
-    /// Iterates the content keyspace prefix to count total items
-    /// (cheap — keys only, no value reads), then deserialises only
-    /// the `[offset, offset + limit)` slice. Returns the windowed
-    /// records plus the total count via [`PagedResult`].
-    #[tracing::instrument(target = TARGET, name = "registry.list_content_with_record", skip(self), fields(%actor_id, offset, limit))]
-    pub async fn list_content_with_record(
-        &self,
-        actor_id: Uuid,
-        offset: usize,
-        limit: usize,
-    ) -> Result<PagedResult<(Uuid, ContentRecord)>> {
-        let content_ks = self.inner.content_ks.clone();
-        let meta_ks = self.inner.content_meta_ks.clone();
-
-        blocking(move || {
-            let ids = content_ks.resource_ids(actor_id)?;
-            let total = ids.len();
-            let window: Vec<Uuid> = ids.into_iter().skip(offset).take(limit).collect();
-            let mut items = Vec::with_capacity(window.len());
-            for id in window {
-                let key = CompositeKey::new(actor_id, id);
-                let record = match meta_ks.get_bytes(key)? {
-                    Some(bytes) => serde_json::from_slice(&bytes)?,
-                    None => ContentRecord {
-                        descriptor: ContentDescriptor::default(),
-                        digest: ContentDigest {
-                            size: 0,
-                            sha256: String::new(),
-                            detected_content_type: None,
-                        },
-                    },
-                };
-                items.push((id, record));
-            }
-            Ok(PagedResult { items, total })
-        })
-        .await
-    }
-
-    /// Stores a JSON-serializable value in a keyspace.
-    async fn store_json<T: Serialize + Send + 'static>(
-        &self,
-        ks: &Keyspace,
-        key: CompositeKey,
-        value: &T,
-    ) -> Result<()> {
-        let bytes = serde_json::to_vec(value)?;
-        let ks = ks.clone();
-        let db = self.inner.db.clone();
-        blocking(move || {
-            ks.put(key, &bytes)?;
-            db.sync()
-        })
-        .await
-    }
-
-    /// Loads a JSON-deserializable value from a keyspace.
-    async fn load_json<T: for<'de> Deserialize<'de> + Send + 'static>(
-        &self,
-        ks: &Keyspace,
-        key: CompositeKey,
-        kind: &'static str,
-    ) -> Result<T> {
-        let actor_id = key.actor_id();
-        let resource_id = key.resource_id();
-        let ks = ks.clone();
-        blocking(move || {
-            let bytes = ks
-                .get_bytes(key)?
-                .ok_or_else(|| not_found(kind, actor_id, resource_id))?;
-            Ok(serde_json::from_slice(&bytes)?)
-        })
-        .await
-    }
-
-    /// Removes a single entry from a keyspace.
-    async fn remove_entry(
-        &self,
-        ks: &Keyspace,
-        key: CompositeKey,
-        kind: &'static str,
-    ) -> Result<()> {
-        let actor_id = key.actor_id();
-        let resource_id = key.resource_id();
-        let ks = ks.clone();
-        let db = self.inner.db.clone();
-        blocking(move || {
-            if !ks.exists(key)? {
-                return Err(not_found(kind, actor_id, resource_id));
-            }
-            ks.delete(key)?;
-            db.sync()
-        })
-        .await
-    }
-
-    /// Persist user-supplied annotations for a piece of content.
-    ///
-    /// Annotations live per-`ContentSource`, written at upload time
-    /// and read at import time. Calling this overwrites any prior
-    /// annotations for the same content.
-    #[tracing::instrument(target = TARGET, name = "registry.store_annotations", skip(self, annotations), fields(%actor_id, %content_id))]
-    pub async fn store_annotations(
-        &self,
-        actor_id: Uuid,
-        content_id: Uuid,
-        annotations: &AnyAnnotations,
-    ) -> Result<()> {
-        let key = CompositeKey::new(actor_id, content_id);
-        self.store_json(&self.inner.annotations_ks, key, annotations)
-            .await
-    }
-
-    /// Load annotations for a piece of content. Returns
-    /// [`AnyAnnotations::default`] when none were stored — annotation
-    /// absence is the common case, not an error.
-    #[tracing::instrument(target = TARGET, name = "registry.load_annotations", skip(self), fields(%actor_id, %content_id))]
-    pub async fn load_annotations(
-        &self,
-        actor_id: Uuid,
-        content_id: Uuid,
-    ) -> Result<AnyAnnotations> {
-        let key = CompositeKey::new(actor_id, content_id);
-        let ks = self.inner.annotations_ks.clone();
-        blocking(move || match ks.get_bytes(key)? {
-            Some(bytes) => Ok(serde_json::from_slice(&bytes)?),
-            None => Ok(AnyAnnotations::default()),
-        })
-        .await
-    }
-
-    /// Persist audit trails for a completed pipeline run.
-    #[tracing::instrument(target = TARGET, name = "registry.store_audits", skip(self, audits), fields(%actor_id, %run_id))]
-    pub async fn store_audits(
-        &self,
-        actor_id: Uuid,
-        run_id: Uuid,
-        audits: Vec<AnyAudit>,
-    ) -> Result<()> {
-        let key = CompositeKey::new(actor_id, run_id);
-        let count = audits.len();
-        self.store_json(&self.inner.audits_ks, key, &audits).await?;
-        tracing::trace!(target: TARGET, count, "audits stored");
-        Ok(())
-    }
-
-    /// Load persisted audit trails for a pipeline run.
-    #[tracing::instrument(target = TARGET, name = "registry.load_audits", skip(self), fields(%actor_id, %run_id))]
-    pub async fn load_audits(&self, actor_id: Uuid, run_id: Uuid) -> Result<Vec<AnyAudit>> {
-        let key = CompositeKey::new(actor_id, run_id);
-        self.load_json(&self.inner.audits_ks, key, "audits").await
-    }
-
-    /// Remove persisted audit trails for a pipeline run.
-    #[tracing::instrument(target = TARGET, name = "registry.unregister_audits", skip(self), fields(%actor_id, %run_id))]
-    pub async fn unregister_audits(&self, actor_id: Uuid, run_id: Uuid) -> Result<()> {
-        let key = CompositeKey::new(actor_id, run_id);
-        self.remove_entry(&self.inner.audits_ks, key, "audits")
-            .await
-    }
-
-    /// Persist a completed [`DetectionResult`].
-    ///
-    /// The write is atomic per-key (fjall blob-store semantics);
-    /// a crash mid-write leaves either the prior value or no
-    /// value, never a partial one. Callers must finalize the
-    /// in-memory state only after the persisted write completes
-    /// so a restart can resume from disk.
-    ///
-    /// [`DetectionResult`]: crate::detection::DetectionResult
-    #[tracing::instrument(target = TARGET, name = "registry.store_detection", skip(self, detection), fields(%actor_id, %detection_id))]
-    pub async fn store_detection(
-        &self,
-        actor_id: Uuid,
-        detection_id: Uuid,
-        detection: &crate::detection::DetectionResult,
-    ) -> Result<()> {
-        let key = CompositeKey::new(actor_id, detection_id);
-        self.store_json(&self.inner.detections_ks, key, detection)
-            .await?;
-        tracing::trace!(target: TARGET, "detection stored");
-        Ok(())
-    }
-
-    /// Load a persisted [`DetectionResult`].
-    ///
-    /// Returns the typed error from `load_json` — `NotFound`
-    /// for an absent key, `Serialization` for a corrupted blob.
-    ///
-    /// [`DetectionResult`]: crate::detection::DetectionResult
-    #[tracing::instrument(target = TARGET, name = "registry.load_detection", skip(self), fields(%actor_id, %detection_id))]
-    pub async fn load_detection(
-        &self,
-        actor_id: Uuid,
-        detection_id: Uuid,
-    ) -> Result<crate::detection::DetectionResult> {
-        let key = CompositeKey::new(actor_id, detection_id);
-        self.load_json(&self.inner.detections_ks, key, "detection")
-            .await
-    }
-
-    /// Remove a persisted detection.
-    #[tracing::instrument(target = TARGET, name = "registry.unregister_detection", skip(self), fields(%actor_id, %detection_id))]
-    pub async fn unregister_detection(&self, actor_id: Uuid, detection_id: Uuid) -> Result<()> {
-        let key = CompositeKey::new(actor_id, detection_id);
-        self.remove_entry(&self.inner.detections_ks, key, "detection")
-            .await
-    }
-
-    /// Persist a completed [`RedactionResult`].
-    ///
-    /// [`RedactionResult`]: crate::redaction::RedactionResult
-    #[tracing::instrument(target = TARGET, name = "registry.store_redaction", skip(self, redaction), fields(%actor_id, %redaction_id))]
-    pub async fn store_redaction(
-        &self,
-        actor_id: Uuid,
-        redaction_id: Uuid,
-        redaction: &crate::redaction::RedactionResult,
-    ) -> Result<()> {
-        let key = CompositeKey::new(actor_id, redaction_id);
-        self.store_json(&self.inner.redactions_ks, key, redaction)
-            .await?;
-        tracing::trace!(target: TARGET, "redaction stored");
-        Ok(())
-    }
-
-    /// Load a persisted [`RedactionResult`].
-    ///
-    /// [`RedactionResult`]: crate::redaction::RedactionResult
-    #[tracing::instrument(target = TARGET, name = "registry.load_redaction", skip(self), fields(%actor_id, %redaction_id))]
-    pub async fn load_redaction(
-        &self,
-        actor_id: Uuid,
-        redaction_id: Uuid,
-    ) -> Result<crate::redaction::RedactionResult> {
-        let key = CompositeKey::new(actor_id, redaction_id);
-        self.load_json(&self.inner.redactions_ks, key, "redaction")
-            .await
-    }
-
-    /// Remove a persisted redaction.
-    #[tracing::instrument(target = TARGET, name = "registry.unregister_redaction", skip(self), fields(%actor_id, %redaction_id))]
-    pub async fn unregister_redaction(&self, actor_id: Uuid, redaction_id: Uuid) -> Result<()> {
-        let key = CompositeKey::new(actor_id, redaction_id);
-        self.remove_entry(&self.inner.redactions_ks, key, "redaction")
-            .await
-    }
-}
-
-#[async_trait::async_trait]
-impl Healthcheck for Registry {
-    fn name(&self) -> Cow<'static, str> {
-        Cow::Borrowed("registry")
-    }
-
-    /// Probes the underlying fjall database by listing the
-    /// content keyspace prefix (cheap — keys only, no value
-    /// reads). Failure to read maps to [`ServiceStatus::Unhealthy`].
-    #[tracing::instrument(target = TARGET, name = "registry.healthcheck", skip(self))]
-    async fn healthcheck(&self) -> ServiceStatus {
-        let ks = self.inner.content_ks.clone();
-        match blocking(move || ks.prefix_keys(&[]).map(|_| ())).await {
-            Ok(()) => ServiceStatus::Healthy,
-            Err(_) => ServiceStatus::Unhealthy,
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_codec::content::{Content, ContentData};
-    use nvisy_core::ErrorKind;
-
-    use super::*;
-
-    fn temp_registry() -> anyhow::Result<(tempfile::TempDir, Registry)> {
-        let temp = tempfile::tempdir()?;
-        let registry = Registry::open(temp.path().join("data"))?;
-        Ok((temp, registry))
-    }
-
-    #[tokio::test]
-    async fn register_and_read_content() -> anyhow::Result<()> {
-        let (_temp, registry) = temp_registry()?;
-        let actor = Uuid::now_v7();
-        let handle = registry
-            .register_content(
-                actor,
-                Content::new(ContentData::from("Hello, world!")),
-                None,
-            )
-            .await?;
-        let data = handle.content_data().await?;
-        assert_eq!(data.as_str().unwrap(), "Hello, world!");
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn content_scoped_by_actor() -> anyhow::Result<()> {
-        let (_temp, registry) = temp_registry()?;
-        let actor_a = Uuid::now_v7();
-        let actor_b = Uuid::now_v7();
-        let handle = registry
-            .register_content(
-                actor_a,
-                Content::new(ContentData::from("actor A only")),
-                None,
-            )
-            .await?;
-        let id = handle.content_source().as_uuid();
-        assert_eq!(
-            registry.read_content(actor_b, id).await.unwrap_err().kind(),
-            ErrorKind::NotFound
-        );
-        registry.read_content(actor_a, id).await?;
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn list_content_per_actor() -> anyhow::Result<()> {
-        let (_temp, registry) = temp_registry()?;
-        let actor_a = Uuid::now_v7();
-        let actor_b = Uuid::now_v7();
-        registry
-            .register_content(actor_a, Content::new(ContentData::from("a1")), None)
-            .await?;
-        registry
-            .register_content(actor_a, Content::new(ContentData::from("a2")), None)
-            .await?;
-        registry
-            .register_content(actor_b, Content::new(ContentData::from("b1")), None)
-            .await?;
-        assert_eq!(registry.list_content(actor_a).await?.len(), 2);
-        assert_eq!(registry.list_content(actor_b).await?.len(), 1);
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn unregister_content() -> anyhow::Result<()> {
-        let (_temp, registry) = temp_registry()?;
-        let actor = Uuid::now_v7();
-        let content = Content::new(ContentData::from("delete me"));
-        let id = content.content_source().as_uuid();
-        registry.register_content(actor, content, None).await?;
-        registry.unregister_content(actor, id).await?;
-        assert_eq!(
-            registry.read_content(actor, id).await.unwrap_err().kind(),
-            ErrorKind::NotFound
-        );
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn unregister_all_content() -> anyhow::Result<()> {
-        let (_temp, registry) = temp_registry()?;
-        let actor = Uuid::now_v7();
-        registry
-            .register_content(actor, Content::new(ContentData::from("first")), None)
-            .await?;
-        registry
-            .register_content(actor, Content::new(ContentData::from("second")), None)
-            .await?;
-        assert_eq!(registry.unregister_all_content(actor).await?, 2);
-        assert!(registry.list_content(actor).await?.is_empty());
-        Ok(())
-    }
-
-    #[tokio::test]
-    async fn data_persists_across_reopen() -> anyhow::Result<()> {
-        let temp = tempfile::tempdir()?;
-        let path = temp.path().join("persist");
-        let actor = Uuid::now_v7();
-        let content = Content::new(ContentData::from("persistent"));
-        let id = content.content_source().as_uuid();
-
-        let registry = Registry::open(&path)?;
-        registry.register_content(actor, content, None).await?;
-        drop(registry);
-
-        let registry = Registry::open(&path)?;
-        let data = registry
-            .read_content(actor, id)
-            .await?
-            .content_data()
-            .await?;
-        assert_eq!(data.as_str().unwrap(), "persistent");
-        Ok(())
-    }
-}
diff --git a/crates/nvisy-engine/src/runs/filter.rs b/crates/nvisy-engine/src/runs/filter.rs
new file mode 100644
index 000000000..831aaf148
--- /dev/null
+++ b/crates/nvisy-engine/src/runs/filter.rs
@@ -0,0 +1,66 @@
+//! Evaluate a [`Policy::applies_when`]
+//! [`DocumentPredicate`] against the merged per-doc fact set.
+//!
+//! Document facts come from two sources, merged at evaluation
+//! time: the doc's caller-supplied descriptor (labels +
+//! metadata), and the per-request metadata from
+//! [`StartBatch::metadata`]. Per-request keys override descriptor
+//! keys on conflict (per the design decision: "explicit-per-request
+//! wins").
+//!
+//! [`Policy::applies_when`]: nvisy_core::policy::Policy::applies_when
+//! [`DocumentPredicate`]: nvisy_core::policy::DocumentPredicate
+//! [`StartBatch::metadata`]: super::input::StartBatch::metadata
+
+use std::collections::HashMap;
+
+use nvisy_core::policy::{DocumentPredicate, Policy};
+
+/// The merged per-document fact set [`DocumentPredicate`]
+/// evaluates against.
+pub(crate) struct DocumentFacts<'a> {
+    /// Labels the descriptor carries (doc-level tags assigned at
+    /// upload time).
+    pub labels: &'a [String],
+    /// Metadata: descriptor metadata + per-request metadata
+    /// already merged with per-request keys winning conflicts.
+    pub metadata: &'a HashMap<String, String>,
+}
+
+/// `true` when `predicate` holds against `facts`. Returns `true`
+/// unconditionally when `predicate` is `None` (no gate).
+pub(crate) fn policy_applies(policy: &Policy, facts: &DocumentFacts<'_>) -> bool {
+    match &policy.applies_when {
+        None => true,
+        Some(predicate) => eval(predicate, facts),
+    }
+}
+
+fn eval(predicate: &DocumentPredicate, facts: &DocumentFacts<'_>) -> bool {
+    match predicate {
+        DocumentPredicate::HasLabel { label } => facts.labels.iter().any(|l| l == label),
+        DocumentPredicate::HasMetadata { key, value } => match facts.metadata.get(key) {
+            Some(actual) => match value {
+                Some(expected) => actual == expected,
+                None => true,
+            },
+            None => false,
+        },
+        DocumentPredicate::All { all } => all.iter().all(|p| eval(p, facts)),
+        DocumentPredicate::Any { any } => any.iter().any(|p| eval(p, facts)),
+        DocumentPredicate::Not { not } => !eval(not, facts),
+    }
+}
+
+/// Merge per-doc descriptor metadata with per-request metadata.
+/// Per-request keys win on conflict.
+pub(crate) fn merge_metadata(
+    descriptor: &HashMap<String, String>,
+    request: &HashMap<String, String>,
+) -> HashMap<String, String> {
+    let mut merged = descriptor.clone();
+    for (k, v) in request {
+        merged.insert(k.clone(), v.clone());
+    }
+    merged
+}
diff --git a/crates/nvisy-engine/src/runs/input.rs b/crates/nvisy-engine/src/runs/input.rs
new file mode 100644
index 000000000..e16a1f303
--- /dev/null
+++ b/crates/nvisy-engine/src/runs/input.rs
@@ -0,0 +1,62 @@
+//! Inputs for [`crate::runs::start`].
+
+use std::collections::HashMap;
+
+use nvisy_core::plan::AnalyzerParams;
+use serde::{Deserialize, Serialize};
+use uuid::Uuid;
+
+use super::state::ResourceRef;
+
+/// Per-call input to [`crate::runs::start`].
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct StartBatch {
+    /// Policies to apply, by `(id, version)`. Engine resolves
+    /// each against [`crate::PolicyRegistry::get_policy`]; missing
+    /// refs fail the start call with [`ErrorKind::NotFound`].
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    pub policy_refs: Vec<ResourceRef>,
+    /// Contexts to apply, by `(id, version)`. Same resolution path
+    /// as [`policy_refs`](Self::policy_refs).
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub context_refs: Vec<ResourceRef>,
+    /// Input documents — each is a file id previously uploaded
+    /// via [`crate::FileRegistry::put_file`]. Engine resolves
+    /// every id at start time; missing files fail the call with
+    /// [`ErrorKind::NotFound`].
+    ///
+    /// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+    pub documents: Vec<DocumentInput>,
+    /// Per-request metadata merged with each document's
+    /// descriptor at [`DocumentPredicate::HasMetadata`] evaluation
+    /// time. Per-request keys override descriptor keys on
+    /// conflict.
+    ///
+    /// [`DocumentPredicate::HasMetadata`]: nvisy_core::policy::DocumentPredicate::HasMetadata
+    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
+    pub metadata: HashMap<String, String>,
+    /// Recognition plan — which recognizers + enrichers run, the
+    /// dedup pipeline, the request scope. Engine compiles this
+    /// into an [`elide::detection::Analyzer`] per modality at start time.
+    pub analyzer: AnalyzerParams,
+    /// Cap on per-doc analyses + applies running concurrently.
+    /// `None` falls back to a sensible default at start time.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub concurrency: Option<usize>,
+}
+
+/// One input document handed to [`crate::runs::start`]. Files
+/// are uploaded once via [`crate::FileRegistry`]; runs reference
+/// them by id and inherit their extension + descriptor labels +
+/// metadata.
+///
+/// [`crate::FileRegistry`]: crate::FileRegistry
+#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct DocumentInput {
+    /// File the run analyses + redacts. Must exist under
+    /// `(actor_id, file_id)` in the engine's files keyspace.
+    pub file_id: Uuid,
+}
diff --git a/crates/nvisy-engine/src/runs/mod.rs b/crates/nvisy-engine/src/runs/mod.rs
new file mode 100644
index 000000000..82d018145
--- /dev/null
+++ b/crates/nvisy-engine/src/runs/mod.rs
@@ -0,0 +1,39 @@
+//! Run orchestrator: drives the two-phase analyze + apply
+//! lifecycle over the fjall registry.
+//!
+//! Surface, all keyed by `(actor_id, run_id)`:
+//!
+//! - [`start`] — submit a batch, mints a run id, fans the
+//!   analyzer out, lands in [`RunState::AwaitingReview`].
+//! - [`get`] / [`get_doc`] — read the run header / a per-doc body.
+//! - [`list`] — list every run for an actor.
+//! - [`override_entity`] — reviewer flips a per-entity decision
+//!   before apply.
+//! - [`apply`] — fan the anonymizer out, lands in
+//!   [`RunState::Applied`] or [`RunState::PartiallyApplied`].
+//! - [`cancel`] — mark an in-flight or awaiting-review run
+//!   [`RunState::Failed`] with `reason = "cancelled"`.
+//! - [`delete`] — cascade-remove a run across all four
+//!   keyspaces.
+//!
+//! Symmetric with [`crate::keyspace`]: all
+//! engine state lives in fjall keyspaces on the shared
+//! [`RegistryHandle`]; entry points read/write through the
+//! [`EngineHandle`] that wraps it.
+//!
+//! [`EngineHandle`]: crate::EngineHandle
+//! [`RegistryHandle`]: crate::registry::RegistryHandle
+
+mod filter;
+mod input;
+mod orchestrate;
+mod persist;
+mod pipeline;
+mod state;
+
+pub use self::input::{DocumentInput, StartBatch};
+pub use self::orchestrate::{apply, cancel, delete, get, get_doc, list, override_entity, start};
+pub use self::state::{
+    DocBody, EntityRecord, FailureReason, ModalityKind, ResourceRef, Run, RunDocState, RunDocument,
+    RunState,
+};
diff --git a/crates/nvisy-engine/src/runs/orchestrate.rs b/crates/nvisy-engine/src/runs/orchestrate.rs
new file mode 100644
index 000000000..808bb8618
--- /dev/null
+++ b/crates/nvisy-engine/src/runs/orchestrate.rs
@@ -0,0 +1,591 @@
+//! Run lifecycle orchestration: [`start`] resolves input files,
+//! writes Queued per-doc rows, fans the analyzer out, and flips
+//! the run into [`RunState::AwaitingReview`]. [`apply`] runs the
+//! per-doc anonymizer pass, writes the redacted bytes back to
+//! the [`FileRegistry`] as new files (stamped with
+//! [`FileLineage::RedactedFrom`]), records each output's id on
+//! the per-doc row, and flips the run into [`RunState::Applied`]
+//! or [`RunState::PartiallyApplied`].
+//!
+//! [`FileRegistry`]: crate::FileRegistry
+//! [`FileLineage::RedactedFrom`]: nvisy_core::FileLineage::RedactedFrom
+//!
+//! Fan-out shape (shared by analyze and apply):
+//!
+//! - [`futures::stream::iter`] over the per-doc workload,
+//!   capped at [`Run::concurrency`] in-flight by
+//!   [`StreamExt::buffer_unordered`].
+//! - Each per-doc future is wrapped in [`tokio::time::timeout`]
+//!   so a single stuck recognizer never stalls the run.
+//! - Per-doc failures (validation, timeout) are recorded in the
+//!   per-doc [`RunDocState`]; they do not fail the run as a
+//!   whole — apply will only run for docs that reached
+//!   [`RunDocState::AwaitingReview`].
+//!
+//! Stream-based concurrency (not [`tokio::task::JoinSet`]) is
+//! deliberate: each per-doc future borrows the [`EngineHandle`]
+//! and the analyzer spec, and `JoinSet::spawn` would force them
+//! to `'static`. The stream combinator keeps the borrows live
+//! for free.
+//!
+//! [`EngineHandle`]: crate::EngineHandle
+
+use std::time::Duration;
+
+use futures::StreamExt;
+use jiff::Timestamp;
+use nvisy_core::file::FileMetadata;
+use nvisy_core::{FileLineage, Result};
+use uuid::Uuid;
+
+use super::filter::{DocumentFacts, merge_metadata};
+use super::input::StartBatch;
+use super::persist::RunRegistry;
+use super::pipeline::{analyze_document, apply_document};
+use super::state::{DocBody, ModalityKind, Run, RunDocState, RunDocument, RunState};
+use crate::keyspace::FileDescriptor;
+use crate::{EngineHandle, FileRegistry, PolicyRegistry};
+
+/// Default per-run concurrency cap when the caller's
+/// [`StartBatch::concurrency`] is `None`.
+const DEFAULT_CONCURRENCY: usize = 4;
+
+/// Per-doc analyze hard timeout. Recognizers that exceed this
+/// land in [`RunDocState::TimedOut`]; the rest of the run
+/// continues.
+const PER_DOC_TIMEOUT: Duration = Duration::from_secs(120);
+
+/// Start a run.
+///
+/// Mints a UUIDv7 run id and per-doc ids, persists the input
+/// bytes together with a Queued per-doc row for every input,
+/// writes the run header in [`RunState::Analyzing`], then fans
+/// the analyzer out across docs. Each task rewrites its per-doc
+/// row with the recognized entities and new state. When the
+/// fan-out finishes, the header flips to
+/// [`RunState::AwaitingReview`].
+///
+/// Returns the new run id; the per-doc results are queryable via
+/// [`super::get_doc`].
+pub async fn start(engine: &EngineHandle, actor_id: Uuid, batch: StartBatch) -> Result<Uuid> {
+    let run_id = Uuid::now_v7();
+    let now = Timestamp::now();
+    let concurrency = batch.concurrency.unwrap_or(DEFAULT_CONCURRENCY).max(1);
+
+    // Mint a doc id per input and write a Queued per-doc row up
+    // front so the run is fully queryable the moment the header
+    // lands. `modality` defaults to Text; the analyzer pass
+    // overwrites it once the codec resolves. `body` stays empty
+    // until analyze finishes.
+    let registry = engine.registry();
+    let mut document_ids = Vec::with_capacity(batch.documents.len());
+    for doc in &batch.documents {
+        let doc_id = Uuid::now_v7();
+        document_ids.push(doc_id);
+
+        let modality = ModalityKind::Text;
+        let document = RunDocument {
+            id: doc_id,
+            input_file_id: doc.file_id,
+            output_file_id: None,
+            state: RunDocState::Queued,
+            modality,
+            body: DocBody::empty(modality),
+        };
+        registry.put_run_doc(actor_id, run_id, &document).await?;
+    }
+
+    let run = Run {
+        id: run_id,
+        state: RunState::Analyzing,
+        started_at: now,
+        updated_at: now,
+        policy_refs: batch.policy_refs,
+        context_refs: batch.context_refs,
+        metadata: batch.metadata,
+        document_ids: document_ids.clone(),
+        analyzer: batch.analyzer,
+        concurrency,
+    };
+    registry.put_run(actor_id, &run).await?;
+
+    // Materialise (doc_id, file_id) pairs as owned `Copy`
+    // tuples; closures clone the analyzer spec per item so the
+    // resulting futures hold no borrows on this stack frame —
+    // important when this future is consumed from an axum
+    // handler where the outer future must be `Send + 'static`.
+    let analyze_inputs: Vec<(Uuid, Uuid)> = document_ids
+        .iter()
+        .zip(batch.documents.iter())
+        .map(|(doc_id, doc)| (*doc_id, doc.file_id))
+        .collect();
+    let analyzer_spec = run.analyzer.clone();
+    let work = analyze_inputs.into_iter().map(|(doc_id, file_id)| {
+        analyze_one_doc(
+            engine,
+            actor_id,
+            run_id,
+            doc_id,
+            file_id,
+            analyzer_spec.clone(),
+        )
+    });
+    futures::stream::iter(work)
+        .buffer_unordered(concurrency)
+        .for_each(|()| async {})
+        .await;
+
+    // All per-doc rows have settled in their terminal analyze
+    // state (AwaitingReview / Failed / TimedOut); flip the header
+    // so callers see the run is ready for review.
+    let mut header = run;
+    header.state = RunState::AwaitingReview;
+    header.updated_at = Timestamp::now();
+    registry.put_run(actor_id, &header).await?;
+
+    Ok(run_id)
+}
+
+/// One per-doc unit of the analyze fan-out. Owns the lifecycle
+/// transitions for one row; never returns an error (failures
+/// land in the row's state).
+///
+/// `spec` is owned (not borrowed) so the resulting future
+/// captures no references on its caller's stack — keeps the
+/// outer `runs::start` future composable with handler-level
+/// `Send + 'static` bounds.
+async fn analyze_one_doc(
+    engine: &EngineHandle,
+    actor_id: Uuid,
+    run_id: Uuid,
+    doc_id: Uuid,
+    input_file_id: Uuid,
+    spec: nvisy_core::plan::AnalyzerParams,
+) {
+    let registry = engine.registry();
+    mark_analyzing(registry, actor_id, run_id, doc_id).await;
+
+    let outcome = match load_input(registry, actor_id, input_file_id).await {
+        Ok((file, bytes)) => {
+            let analyze = analyze_document(engine.formats(), bytes, file.extension.as_str(), &spec);
+            match tokio::time::timeout(PER_DOC_TIMEOUT, analyze).await {
+                Ok(Ok(outcome)) => Ok(outcome),
+                Ok(Err(err)) => Err(DocFailure::Failed(err.to_string())),
+                Err(_) => Err(DocFailure::TimedOut),
+            }
+        }
+        Err(err) => Err(DocFailure::Failed(err.to_string())),
+    };
+
+    write_outcome(registry, actor_id, run_id, doc_id, outcome).await;
+}
+
+/// Fetch the file metadata + bytes for one input file. Two
+/// reads (metadata then content) because the file API splits
+/// them across keyspaces.
+async fn load_input(
+    registry: &crate::registry::RegistryHandle,
+    actor_id: Uuid,
+    file_id: Uuid,
+) -> Result<(FileMetadata, bytes::Bytes)> {
+    let file = registry.get_file(actor_id, file_id).await?;
+    let bytes = registry.get_file_bytes(actor_id, file_id).await?;
+    Ok((file, bytes))
+}
+
+enum DocFailure {
+    Failed(String),
+    TimedOut,
+}
+
+async fn mark_analyzing(
+    registry: &crate::registry::RegistryHandle,
+    actor_id: Uuid,
+    run_id: Uuid,
+    doc_id: Uuid,
+) {
+    let Ok(mut doc) = registry.get_run_doc(actor_id, run_id, doc_id).await else {
+        return;
+    };
+    doc.state = RunDocState::Analyzing;
+    let _ = registry.put_run_doc(actor_id, run_id, &doc).await;
+}
+
+async fn write_outcome(
+    registry: &crate::registry::RegistryHandle,
+    actor_id: Uuid,
+    run_id: Uuid,
+    doc_id: Uuid,
+    outcome: std::result::Result<super::pipeline::AnalyzeOutcome, DocFailure>,
+) {
+    let Ok(mut doc) = registry.get_run_doc(actor_id, run_id, doc_id).await else {
+        return;
+    };
+    match outcome {
+        Ok(out) => {
+            doc.modality = out.modality;
+            doc.body = out.body;
+            doc.state = RunDocState::AwaitingReview;
+        }
+        Err(DocFailure::Failed(reason)) => {
+            doc.state = RunDocState::Failed { reason };
+        }
+        Err(DocFailure::TimedOut) => {
+            doc.state = RunDocState::TimedOut;
+        }
+    }
+    let _ = registry.put_run_doc(actor_id, run_id, &doc).await;
+}
+
+/// Apply a run.
+///
+/// Loads the run header, verifies it is in
+/// [`RunState::AwaitingReview`], resolves every referenced
+/// policy, fans the per-doc anonymizer pass out under the run's
+/// concurrency cap, persists the redacted bytes as artifacts,
+/// and rewrites the header to [`RunState::Applied`] (every doc
+/// applied) or [`RunState::PartiallyApplied`] (at least one
+/// failed).
+pub async fn apply(engine: &EngineHandle, actor_id: Uuid, run_id: Uuid) -> Result<()> {
+    let registry = engine.registry();
+    let mut run = registry.get_run(actor_id, run_id).await?;
+    if !matches!(run.state, RunState::AwaitingReview) {
+        return Err(nvisy_core::Error::conflict(
+            format!(
+                "run {run_id} is in state {:?}; apply only valid from AwaitingReview",
+                run.state,
+            ),
+            "runs::apply",
+        ));
+    }
+
+    // Resolve every referenced policy up-front. Apply needs the
+    // full set so each per-doc task can filter by
+    // `Policy::applies_when` against that doc's facts.
+    let policies =
+        std::sync::Arc::new(resolve_policies(registry, actor_id, &run.policy_refs).await?);
+
+    // Materialise doc ids as owned `Copy` values; clone the
+    // shared inputs (analyzer spec + metadata + Arc<policies>)
+    // into each closure so the stream futures hold no borrows
+    // on this stack frame — keeps the outer apply future
+    // composable with handler-level `Send + 'static` bounds.
+    let doc_ids: Vec<Uuid> = run.document_ids.clone();
+    let request_metadata = run.metadata.clone();
+    let analyzer_spec = run.analyzer.clone();
+    let concurrency = run.concurrency;
+    let work = doc_ids.into_iter().map(|doc_id| {
+        apply_one_doc(
+            engine,
+            actor_id,
+            run_id,
+            doc_id,
+            request_metadata.clone(),
+            analyzer_spec.clone(),
+            std::sync::Arc::clone(&policies),
+        )
+    });
+    let outcomes: Vec<bool> = futures::stream::iter(work)
+        .buffer_unordered(concurrency)
+        .collect()
+        .await;
+
+    let all_applied = outcomes.iter().all(|ok| *ok);
+    run.state = if all_applied {
+        RunState::Applied
+    } else {
+        RunState::PartiallyApplied
+    };
+    run.updated_at = Timestamp::now();
+    registry.put_run(actor_id, &run).await?;
+    Ok(())
+}
+
+/// Read the run header at `(actor_id, run_id)`. Public read-side
+/// counterpart to the `pub(crate)` `RunRegistry::get_run`;
+/// external API consumers reach run state through this rather than
+/// the persistence trait.
+pub async fn get(engine: &EngineHandle, actor_id: Uuid, run_id: Uuid) -> Result<Run> {
+    engine.registry().get_run(actor_id, run_id).await
+}
+
+/// Read a per-doc body at `(actor_id, run_id, doc_id)`.
+pub async fn get_doc(
+    engine: &EngineHandle,
+    actor_id: Uuid,
+    run_id: Uuid,
+    doc_id: Uuid,
+) -> Result<RunDocument> {
+    engine
+        .registry()
+        .get_run_doc(actor_id, run_id, doc_id)
+        .await
+}
+
+/// List every run for `actor_id`. Returns full headers; callers
+/// filter by [`RunState`] (e.g. the HTTP layer's `/detections`
+/// view shows non-applied runs, `/redactions` shows applied
+/// runs).
+pub async fn list(engine: &EngineHandle, actor_id: Uuid) -> Vec<Run> {
+    engine
+        .registry()
+        .list_runs(actor_id)
+        .await
+        .unwrap_or_default()
+}
+
+/// Cancel a run.
+///
+/// Marks the header [`RunState::Failed`] with `reason = "cancelled"`.
+/// Cancel is only valid from [`RunState::Analyzing`] or
+/// [`RunState::AwaitingReview`]; calling it on a terminal run
+/// ([`RunState::Applied`] / [`RunState::PartiallyApplied`] /
+/// [`RunState::Failed`]) returns [`ErrorKind::Conflict`].
+///
+/// Today's cancel is a *header* operation: in-flight per-doc
+/// fan-out tasks are not interrupted; they finish their current
+/// step and write their outcome into a doc row no one will read.
+/// Cooperative interruption needs a [`tokio_util::sync::CancellationToken`]
+/// threaded through the pipeline and lands as a follow-up slice.
+///
+/// [`ErrorKind::Conflict`]: nvisy_core::ErrorKind::Conflict
+/// [`tokio_util::sync::CancellationToken`]: https://docs.rs/tokio-util/latest/tokio_util/sync/struct.CancellationToken.html
+pub async fn cancel(engine: &EngineHandle, actor_id: Uuid, run_id: Uuid) -> Result<()> {
+    let registry = engine.registry();
+    let mut run = registry.get_run(actor_id, run_id).await?;
+    match run.state {
+        RunState::Analyzing | RunState::AwaitingReview => {}
+        ref other => {
+            return Err(nvisy_core::Error::conflict(
+                format!(
+                    "run {run_id} is in state {other:?}; cancel only valid from Analyzing or AwaitingReview"
+                ),
+                "runs::cancel",
+            ));
+        }
+    }
+    run.state = RunState::Failed {
+        reason: "cancelled".to_owned(),
+    };
+    run.updated_at = Timestamp::now();
+    registry.put_run(actor_id, &run).await
+}
+
+/// Delete a run, cascading across all four run keyspaces.
+///
+/// Removes the header (`run_headers`) plus every per-doc body
+/// (`run_docs`), artifact (`run_artifacts`), and input
+/// (`run_inputs`) belonging to the run. Returns
+/// [`ErrorKind::NotFound`] when the header is missing.
+///
+/// In-flight cancellation is **not** part of delete; callers
+/// should [`cancel`] first when the run is still active, then
+/// delete. Calling `delete` on an active run is allowed —
+/// per-doc tasks that complete after the delete write into
+/// keyspaces that no longer have a header, and are reaped on the
+/// next list scan.
+///
+/// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+pub async fn delete(engine: &EngineHandle, actor_id: Uuid, run_id: Uuid) -> Result<()> {
+    let registry = engine.registry();
+    // Surface a clean NotFound when the run doesn't exist instead
+    // of silently succeeding.
+    let _header = registry.get_run(actor_id, run_id).await?;
+    registry.delete_run_bodies(actor_id, run_id).await?;
+    registry.delete_run(actor_id, run_id).await
+}
+
+/// Resolve every `(policy_id, version)` ref in the run header to
+/// its persisted [`Policy`] blob. Missing refs fail the apply
+/// call (the per-doc filter operates on the full set; a missing
+/// policy can't be silently dropped).
+async fn resolve_policies(
+    registry: &crate::registry::RegistryHandle,
+    actor_id: Uuid,
+    refs: &[super::state::ResourceRef],
+) -> Result<Vec<nvisy_core::policy::Policy>> {
+    let mut out = Vec::with_capacity(refs.len());
+    for r in refs {
+        let policy = registry
+            .get_policy(actor_id, r.id, r.version.clone())
+            .await?;
+        out.push(policy);
+    }
+    Ok(out)
+}
+
+/// One per-doc unit of the apply fan-out. Returns `true` on
+/// success, `false` when any step failed (failures land in the
+/// row's state; the boolean only drives the header transition).
+///
+/// Inputs are owned (analyzer spec, metadata, Arc-shared
+/// policy set) so the resulting future captures no references
+/// on its caller's stack — keeps the outer `runs::apply`
+/// future composable with handler-level `Send + 'static`
+/// bounds.
+async fn apply_one_doc(
+    engine: &EngineHandle,
+    actor_id: Uuid,
+    run_id: Uuid,
+    doc_id: Uuid,
+    request_metadata: std::collections::HashMap<String, String>,
+    spec: nvisy_core::plan::AnalyzerParams,
+    policies: std::sync::Arc<Vec<nvisy_core::policy::Policy>>,
+) -> bool {
+    let registry = engine.registry();
+    let Ok(mut doc) = registry.get_run_doc(actor_id, run_id, doc_id).await else {
+        return false;
+    };
+    // Only apply rows that finished analyze. Failed / TimedOut /
+    // already-Applied rows pass through untouched.
+    if !matches!(doc.state, RunDocState::AwaitingReview) {
+        return matches!(doc.state, RunDocState::Applied);
+    }
+
+    let (input_file, input_bytes) = match load_input(registry, actor_id, doc.input_file_id).await {
+        Ok(loaded) => loaded,
+        Err(err) => {
+            doc.state = RunDocState::Failed {
+                reason: format!("input file unavailable: {err}"),
+            };
+            let _ = registry.put_run_doc(actor_id, run_id, &doc).await;
+            return false;
+        }
+    };
+
+    let merged = merge_metadata(&input_file.descriptor_metadata, &request_metadata);
+    let facts = DocumentFacts {
+        labels: &input_file.descriptor_labels,
+        metadata: &merged,
+    };
+
+    let outcome = apply_document(
+        engine.formats(),
+        input_bytes,
+        input_file.extension.as_str(),
+        &spec,
+        policies.as_slice(),
+        &facts,
+        &doc.body,
+    )
+    .await;
+
+    match outcome {
+        Ok(out) => {
+            let descriptor = redacted_descriptor(&input_file, run_id);
+            match registry.put_file(actor_id, descriptor, out.bytes).await {
+                Ok(output_file) => {
+                    doc.output_file_id = Some(output_file.id);
+                    doc.state = RunDocState::Applied;
+                    let _ = registry.put_run_doc(actor_id, run_id, &doc).await;
+                    true
+                }
+                Err(err) => {
+                    doc.state = RunDocState::Failed {
+                        reason: format!("writing redacted output file failed: {err}"),
+                    };
+                    let _ = registry.put_run_doc(actor_id, run_id, &doc).await;
+                    false
+                }
+            }
+        }
+        Err(err) => {
+            doc.state = RunDocState::Failed {
+                reason: err.to_string(),
+            };
+            let _ = registry.put_run_doc(actor_id, run_id, &doc).await;
+            false
+        }
+    }
+}
+
+/// Build the descriptor for a redacted output file. Inherits
+/// extension + content_type + descriptor labels/metadata from
+/// the input file, appends `-redacted` to the filename, stamps
+/// the lineage so audits + clients can trace the output back to
+/// its source and the run that produced it.
+fn redacted_descriptor(input: &FileMetadata, run_id: Uuid) -> FileDescriptor {
+    let filename = input
+        .filename
+        .as_ref()
+        .map(|name| redacted_filename(name.as_str()).into());
+    FileDescriptor {
+        filename,
+        content_type: input.content_type.clone(),
+        extension: input.extension.clone(),
+        lineage: Some(FileLineage::RedactedFrom {
+            run_id,
+            source_file_id: input.id,
+        }),
+        descriptor_labels: input.descriptor_labels.clone(),
+        descriptor_metadata: input.descriptor_metadata.clone(),
+    }
+}
+
+/// Insert `-redacted` before the file's final extension —
+/// `report.pdf` → `report-redacted.pdf`, `archive.tar.gz` →
+/// `archive.tar-redacted.gz`. Pure naming convenience for the
+/// download UX; identity is the UUID.
+fn redacted_filename(name: &str) -> String {
+    match name.rfind('.') {
+        Some(idx) if idx > 0 => format!("{}-redacted{}", &name[..idx], &name[idx..]),
+        _ => format!("{name}-redacted"),
+    }
+}
+
+/// Apply a reviewer override to one entity.
+///
+/// Loads the doc body, finds the entity by id, sets the override,
+/// writes the body back. Idempotent — overriding the same entity
+/// twice is well-defined (second write wins).
+///
+/// Returns [`ErrorKind::NotFound`] when the run, doc, or entity
+/// id is unknown.
+///
+/// [`ErrorKind::NotFound`]: nvisy_core::ErrorKind::NotFound
+pub async fn override_entity(
+    engine: &EngineHandle,
+    actor_id: Uuid,
+    run_id: Uuid,
+    doc_id: Uuid,
+    entity_id: Uuid,
+    action: nvisy_core::policy::RuleAction,
+) -> Result<()> {
+    let registry = engine.registry();
+    let mut doc = registry.get_run_doc(actor_id, run_id, doc_id).await?;
+    let found = patch_override(&mut doc.body, entity_id, action);
+    if !found {
+        return Err(nvisy_core::Error::not_found(
+            format!("entity {entity_id} not found in run {run_id} doc {doc_id}"),
+            "runs::override_entity",
+        ));
+    }
+    registry.put_run_doc(actor_id, run_id, &doc).await
+}
+
+fn patch_override(
+    body: &mut DocBody,
+    entity_id: Uuid,
+    action: nvisy_core::policy::RuleAction,
+) -> bool {
+    match body {
+        DocBody::Text { entities } => patch_each(entities, entity_id, action),
+        DocBody::Tabular { entities } => patch_each(entities, entity_id, action),
+        DocBody::Image { entities } => patch_each(entities, entity_id, action),
+        DocBody::Audio { entities } => patch_each(entities, entity_id, action),
+    }
+}
+
+fn patch_each<M: elide_core::modality::Modality>(
+    entities: &mut [super::state::EntityRecord<M>],
+    entity_id: Uuid,
+    action: nvisy_core::policy::RuleAction,
+) -> bool {
+    for record in entities.iter_mut() {
+        if record.entity.id == entity_id {
+            record.r#override = Some(action);
+            return true;
+        }
+    }
+    false
+}
diff --git a/crates/nvisy-engine/src/runs/persist.rs b/crates/nvisy-engine/src/runs/persist.rs
new file mode 100644
index 000000000..8de780fff
--- /dev/null
+++ b/crates/nvisy-engine/src/runs/persist.rs
@@ -0,0 +1,188 @@
+//! fjall read/write for run state.
+//!
+//! Surfaced as the crate-private [`RunRegistry`] extension trait
+//! on [`RegistryHandle`] so the orchestrator reaches storage the
+//! same way the public resource APIs do — `registry.put_run(...)`,
+//! `registry.get_run_doc(...)`, etc. Trait is `pub(crate)`: a run
+//! is the orchestrator's invariant, so external code should never
+//! manipulate it directly; the public surface is
+//! [`crate::runs::start`] / [`apply`](crate::runs::apply) and
+//! their siblings.
+//!
+//! Two keyspaces, both keyed by `(actor_id, run_id, …)`:
+//!
+//! - **run_headers** — [`CompositeKey(actor, run)`] holding the
+//!   [`Run`] metadata blob (state, refs, timestamps).
+//! - **run_docs** — [`TripleKey(actor, run, doc)`] holding the
+//!   per-document body (entities + overrides + `input_file_id`
+//!   + `output_file_id`).
+//!
+//! Bytes (inputs + redacted outputs) live in the
+//! [`crate::FileRegistry`], not in any run keyspace. The
+//! per-doc row points at them by id.
+//!
+//! [`CompositeKey(actor, run)`]: crate::registry::CompositeKey
+//! [`TripleKey(actor, run, doc)`]: crate::registry::TripleKey
+
+use nvisy_core::{Error, Result};
+use uuid::Uuid;
+
+use super::state::{Run, RunDocument};
+use crate::registry::{CompositeKey, RegistryHandle, TripleKey, blocking, not_found};
+
+const COMPONENT: &str = "runs::persist";
+const KIND_RUN: &str = "run";
+const KIND_DOC: &str = "run_document";
+
+/// Crate-private extension trait adding run-lifecycle storage to
+/// [`RegistryHandle`]. Consumed exclusively by the orchestrator
+/// and the public reader functions in [`crate::runs`].
+pub(crate) trait RunRegistry {
+    /// Write the run header at `(actor_id, run.id)`. Overwrites
+    /// any existing header — state transitions go through this.
+    fn put_run(&self, actor_id: Uuid, run: &Run) -> impl Future<Output = Result<()>> + Send;
+
+    /// Read the run header at `(actor_id, run_id)`.
+    fn get_run(&self, actor_id: Uuid, run_id: Uuid) -> impl Future<Output = Result<Run>> + Send;
+
+    /// List every run header for `actor_id`. Returns full
+    /// headers; callers filter by [`super::state::RunState`].
+    fn list_runs(&self, actor_id: Uuid) -> impl Future<Output = Result<Vec<Run>>> + Send;
+
+    /// Delete the run header at `(actor_id, run_id)`. Caller
+    /// drives the per-doc cascade via
+    /// [`delete_run_bodies`](Self::delete_run_bodies) first.
+    fn delete_run(&self, actor_id: Uuid, run_id: Uuid) -> impl Future<Output = Result<()>> + Send;
+
+    /// Write a per-doc body at `(actor_id, run_id, doc.id)`.
+    /// Overwrites; reviewer overrides and per-doc state
+    /// transitions flow through this.
+    fn put_run_doc(
+        &self,
+        actor_id: Uuid,
+        run_id: Uuid,
+        doc: &RunDocument,
+    ) -> impl Future<Output = Result<()>> + Send;
+
+    /// Read a per-doc body at `(actor_id, run_id, doc_id)`.
+    fn get_run_doc(
+        &self,
+        actor_id: Uuid,
+        run_id: Uuid,
+        doc_id: Uuid,
+    ) -> impl Future<Output = Result<RunDocument>> + Send;
+
+    /// Remove every per-doc body under `(actor_id, run_id)` in
+    /// `run_docs`. Returns the number of rows removed. Does not
+    /// touch the input or output files in [`FileRegistry`] —
+    /// per the no-cascade-delete contract, file lifecycle is
+    /// independent of run lifecycle.
+    ///
+    /// [`FileRegistry`]: crate::FileRegistry
+    fn delete_run_bodies(
+        &self,
+        actor_id: Uuid,
+        run_id: Uuid,
+    ) -> impl Future<Output = Result<usize>> + Send;
+}
+
+impl RunRegistry for RegistryHandle {
+    async fn put_run(&self, actor_id: Uuid, run: &Run) -> Result<()> {
+        let key = CompositeKey::new(actor_id, run.id);
+        let value = serde_json::to_vec(run)?;
+        let headers = self.run_headers().clone();
+        blocking(move || {
+            headers.insert(*key, value).map_err(fjall_err)?;
+            Ok(())
+        })
+        .await
+    }
+
+    async fn get_run(&self, actor_id: Uuid, run_id: Uuid) -> Result<Run> {
+        let key = CompositeKey::new(actor_id, run_id);
+        let headers = self.run_headers().clone();
+        blocking(move || {
+            let value = headers
+                .get(*key)
+                .map_err(fjall_err)?
+                .ok_or_else(|| not_found(KIND_RUN, actor_id, run_id))?;
+            serde_json::from_slice(&value).map_err(Into::into)
+        })
+        .await
+    }
+
+    async fn list_runs(&self, actor_id: Uuid) -> Result<Vec<Run>> {
+        let headers = self.run_headers().clone();
+        blocking(move || {
+            let prefix = CompositeKey::actor_prefix(actor_id);
+            let mut out: Vec<Run> = Vec::new();
+            for guard in headers.prefix(prefix) {
+                let (_, value) = guard.into_inner().map_err(fjall_err)?;
+                let run: Run = serde_json::from_slice(&value)?;
+                out.push(run);
+            }
+            Ok(out)
+        })
+        .await
+    }
+
+    async fn delete_run(&self, actor_id: Uuid, run_id: Uuid) -> Result<()> {
+        let key = CompositeKey::new(actor_id, run_id);
+        let headers = self.run_headers().clone();
+        blocking(move || {
+            headers.remove(*key).map_err(fjall_err)?;
+            Ok(())
+        })
+        .await
+    }
+
+    async fn put_run_doc(&self, actor_id: Uuid, run_id: Uuid, doc: &RunDocument) -> Result<()> {
+        let key = TripleKey::new(actor_id, run_id, doc.id);
+        let value = serde_json::to_vec(doc)?;
+        let docs = self.run_docs().clone();
+        blocking(move || {
+            docs.insert(*key, value).map_err(fjall_err)?;
+            Ok(())
+        })
+        .await
+    }
+
+    async fn get_run_doc(&self, actor_id: Uuid, run_id: Uuid, doc_id: Uuid) -> Result<RunDocument> {
+        let key = TripleKey::new(actor_id, run_id, doc_id);
+        let docs = self.run_docs().clone();
+        blocking(move || {
+            let value = docs
+                .get(*key)
+                .map_err(fjall_err)?
+                .ok_or_else(|| not_found(KIND_DOC, actor_id, doc_id))?;
+            serde_json::from_slice(&value).map_err(Into::into)
+        })
+        .await
+    }
+
+    async fn delete_run_bodies(&self, actor_id: Uuid, run_id: Uuid) -> Result<usize> {
+        let docs = self.run_docs().clone();
+        blocking(move || {
+            let prefix = TripleKey::ab_prefix(actor_id, run_id);
+            // Collect first, then remove — iterating while
+            // mutating a fjall keyspace is unsupported.
+            let keys: Vec<Vec<u8>> = docs
+                .prefix(prefix)
+                .map(|guard| {
+                    let key = guard.key().map_err(fjall_err)?;
+                    Ok(key.as_ref().to_vec())
+                })
+                .collect::<Result<Vec<_>>>()?;
+            let removed = keys.len();
+            for key in keys {
+                docs.remove(&key).map_err(fjall_err)?;
+            }
+            Ok(removed)
+        })
+        .await
+    }
+}
+
+fn fjall_err(err: impl std::error::Error + Send + Sync + 'static) -> Error {
+    Error::internal("fjall operation failed", COMPONENT).with_source(err)
+}
diff --git a/crates/nvisy-engine/src/runs/pipeline.rs b/crates/nvisy-engine/src/runs/pipeline.rs
new file mode 100644
index 000000000..7e0e91978
--- /dev/null
+++ b/crates/nvisy-engine/src/runs/pipeline.rs
@@ -0,0 +1,324 @@
+//! Per-document analyze + apply pipelines.
+//!
+//! - [`analyze_document`] decodes bytes, resolves modality,
+//!   compiles the per-modality analyzer from the
+//!   [`AnalyzerParams`], recognizes entities, and wraps them in a
+//!   [`DocBody`].
+//! - [`apply_document`] decodes bytes again, layers reviewer
+//!   overrides on top of the policy-driven anonymizer, applies
+//!   it to the persisted entities, and returns the redacted
+//!   bytes.
+//!
+//! Pure functions over the elide toolkit; no fjall I/O. The
+//! orchestrator in [`super::orchestrate`] drives them as a
+//! bounded-concurrency stream and persists the resulting body /
+//! artifact.
+
+use bytes::Bytes;
+use elide::codec::{DocumentHandle, FormatRegistry};
+use elide::detection::Analyzer;
+use elide::redaction::Anonymizer;
+use elide_core::entity::Entity;
+use elide_core::modality::Modality;
+use elide_core::modality::audio::Audio;
+use elide_core::modality::image::Image;
+use elide_core::modality::tabular::Tabular;
+use elide_core::modality::text::Text;
+use elide_core::recognition::Scope;
+use nvisy_core::plan::AnalyzerParams;
+use nvisy_core::policy::Policy;
+use nvisy_core::{Error, Result};
+
+use super::filter::{DocumentFacts, policy_applies};
+use super::state::{DocBody, EntityRecord, ModalityKind};
+use crate::analyzer::{build_catalog, compile_audio, compile_image, compile_tabular, compile_text};
+use crate::anonymizer::{
+    attach_override_audio, attach_override_image, attach_override_tabular, attach_override_text,
+    attach_policies_audio, attach_policies_image, attach_policies_tabular, attach_policies_text,
+};
+
+const COMPONENT: &str = "runs::pipeline";
+
+/// Outcome of analyzing one document end-to-end.
+pub(super) struct AnalyzeOutcome {
+    /// Modality elide's codec resolved the bytes to.
+    pub modality: ModalityKind,
+    /// Recognized entities, wrapped in [`EntityRecord`] for
+    /// persistence (no overrides set yet — those flow through the
+    /// reviewer surface).
+    pub body: DocBody,
+}
+
+/// Decode `bytes` via `registry`, resolve the modality, compile
+/// the per-modality analyzer from `spec`, recognize entities,
+/// return them in the modality-specific [`DocBody`] variant.
+///
+/// `extension` is the codec discriminator (case-insensitive, no
+/// leading dot) — e.g. `"txt"`, `"csv"`, `"png"`, `"wav"`.
+pub(super) async fn analyze_document(
+    registry: &FormatRegistry,
+    bytes: Bytes,
+    extension: &str,
+    spec: &AnalyzerParams,
+) -> Result<AnalyzeOutcome> {
+    let handle = registry.decode(bytes, extension).await.map_err(|err| {
+        Error::validation(
+            format!("codec decode failed for extension {extension:?}"),
+            COMPONENT,
+        )
+        .with_source(err)
+    })?;
+
+    if handle.is::<Text>() {
+        let typed = handle
+            .into::<Text>()
+            .map_err(|_| modality_mismatch_err("Text"))?;
+        let (analyzer, scope) = compile_text(spec).map_err(compile_err)?;
+        let entities = recognize::<Text>(analyzer, scope, typed).await?;
+        let body = DocBody::Text {
+            entities: entities.into_iter().map(EntityRecord::new).collect(),
+        };
+        return Ok(AnalyzeOutcome {
+            modality: ModalityKind::Text,
+            body,
+        });
+    }
+
+    if handle.is::<Tabular>() {
+        let typed = handle
+            .into::<Tabular>()
+            .map_err(|_| modality_mismatch_err("Tabular"))?;
+        let (analyzer, scope) = compile_tabular(spec).map_err(compile_err)?;
+        let entities = recognize::<Tabular>(analyzer, scope, typed).await?;
+        let body = DocBody::Tabular {
+            entities: entities.into_iter().map(EntityRecord::new).collect(),
+        };
+        return Ok(AnalyzeOutcome {
+            modality: ModalityKind::Tabular,
+            body,
+        });
+    }
+
+    if handle.is::<Image>() {
+        let typed = handle
+            .into::<Image>()
+            .map_err(|_| modality_mismatch_err("Image"))?;
+        let (analyzer, scope) = compile_image(spec).map_err(compile_err)?;
+        let entities = recognize::<Image>(analyzer, scope, typed).await?;
+        let body = DocBody::Image {
+            entities: entities.into_iter().map(EntityRecord::new).collect(),
+        };
+        return Ok(AnalyzeOutcome {
+            modality: ModalityKind::Image,
+            body,
+        });
+    }
+
+    if handle.is::<Audio>() {
+        let typed = handle
+            .into::<Audio>()
+            .map_err(|_| modality_mismatch_err("Audio"))?;
+        let (analyzer, scope) = compile_audio(spec).map_err(compile_err)?;
+        let entities = recognize::<Audio>(analyzer, scope, typed).await?;
+        let body = DocBody::Audio {
+            entities: entities.into_iter().map(EntityRecord::new).collect(),
+        };
+        return Ok(AnalyzeOutcome {
+            modality: ModalityKind::Audio,
+            body,
+        });
+    }
+
+    Err(Error::validation(
+        format!("codec resolved {extension:?} to an unsupported modality"),
+        COMPONENT,
+    ))
+}
+
+async fn recognize<M: Modality>(
+    analyzer: Analyzer<M>,
+    scope: Scope<M>,
+    mut handle: DocumentHandle<M>,
+) -> Result<Vec<Entity<M>>> {
+    analyzer
+        .analyze_stream(&mut handle, &scope)
+        .await
+        .map_err(|err| Error::internal("analyzer fanout failed", COMPONENT).with_source(err))
+}
+
+fn modality_mismatch_err(expected: &'static str) -> Error {
+    Error::internal(
+        format!(
+            "codec advertised modality {expected} but downcast failed — \
+             elide handle::is/into mismatch"
+        ),
+        COMPONENT,
+    )
+}
+
+/// Translate an `elide::Error` from a `compile_*` call into the
+/// runtime's error type. Compile failures are caller-driven
+/// (e.g. unsupported recognizer for the modality), so they map
+/// to `Validation`.
+fn compile_err(err: elide::Error) -> Error {
+    Error::validation(format!("analyzer compile failed: {err}"), COMPONENT).with_source(err)
+}
+
+/// Outcome of applying redactions to one document.
+pub(super) struct ApplyOutcome {
+    /// Encoded bytes of the redacted document, ready to persist
+    /// in the `run_artifacts` keyspace.
+    pub bytes: Bytes,
+}
+
+/// Re-decode `bytes`, layer reviewer overrides + applicable
+/// policies into a per-modality anonymizer, apply it, re-encode,
+/// return the redacted bytes.
+///
+/// `entities` is the persisted recognition output (one
+/// [`EntityRecord`] per entity, possibly carrying a reviewer
+/// override). `policies` is the full resolved policy set; the
+/// pipeline filters to those whose [`Policy::applies_when`]
+/// holds against `facts` (the merged descriptor + per-request
+/// metadata). `spec` is the same [`AnalyzerParams`] that drove
+/// analyze — needed for the label catalog.
+pub(super) async fn apply_document(
+    registry: &FormatRegistry,
+    bytes: Bytes,
+    extension: &str,
+    spec: &AnalyzerParams,
+    policies: &[Policy],
+    facts: &DocumentFacts<'_>,
+    body: &DocBody,
+) -> Result<ApplyOutcome> {
+    let handle = registry.decode(bytes, extension).await.map_err(|err| {
+        Error::validation(
+            format!("codec decode failed for extension {extension:?}"),
+            COMPONENT,
+        )
+        .with_source(err)
+    })?;
+
+    let scoped = || policies.iter().filter(|p| policy_applies(p, facts));
+
+    match body {
+        DocBody::Text { entities } => {
+            let typed = handle
+                .into::<Text>()
+                .map_err(|_| modality_mismatch_err("Text"))?;
+            let anonymizer = build_text_anonymizer(spec, scoped(), entities)?;
+            let bytes = run_anonymize::<Text>(anonymizer, typed, entities).await?;
+            Ok(ApplyOutcome { bytes })
+        }
+        DocBody::Tabular { entities } => {
+            let typed = handle
+                .into::<Tabular>()
+                .map_err(|_| modality_mismatch_err("Tabular"))?;
+            let anonymizer = build_tabular_anonymizer(spec, scoped(), entities)?;
+            let bytes = run_anonymize::<Tabular>(anonymizer, typed, entities).await?;
+            Ok(ApplyOutcome { bytes })
+        }
+        DocBody::Image { entities } => {
+            let typed = handle
+                .into::<Image>()
+                .map_err(|_| modality_mismatch_err("Image"))?;
+            let anonymizer = build_image_anonymizer(spec, scoped(), entities);
+            let bytes = run_anonymize::<Image>(anonymizer, typed, entities).await?;
+            Ok(ApplyOutcome { bytes })
+        }
+        DocBody::Audio { entities } => {
+            let typed = handle
+                .into::<Audio>()
+                .map_err(|_| modality_mismatch_err("Audio"))?;
+            let anonymizer = build_audio_anonymizer(spec, scoped(), entities);
+            let bytes = run_anonymize::<Audio>(anonymizer, typed, entities).await?;
+            Ok(ApplyOutcome { bytes })
+        }
+    }
+}
+
+fn build_text_anonymizer<'a>(
+    spec: &AnalyzerParams,
+    policies: impl Iterator<Item = &'a Policy>,
+    entities: &[EntityRecord<Text>],
+) -> Result<Anonymizer<Text>> {
+    let catalog = build_catalog(spec);
+    let mut anonymizer = Anonymizer::<Text>::new().with_catalog(catalog);
+    for record in entities {
+        if let Some(action) = &record.r#override {
+            anonymizer =
+                attach_override_text(anonymizer, record.entity.id, action).map_err(compile_err)?;
+        }
+    }
+    attach_policies_text(anonymizer, policies).map_err(compile_err)
+}
+
+fn build_tabular_anonymizer<'a>(
+    spec: &AnalyzerParams,
+    policies: impl Iterator<Item = &'a Policy>,
+    entities: &[EntityRecord<Tabular>],
+) -> Result<Anonymizer<Tabular>> {
+    let catalog = build_catalog(spec);
+    let mut anonymizer = Anonymizer::<Tabular>::new().with_catalog(catalog);
+    for record in entities {
+        if let Some(action) = &record.r#override {
+            anonymizer = attach_override_tabular(anonymizer, record.entity.id, action)
+                .map_err(compile_err)?;
+        }
+    }
+    attach_policies_tabular(anonymizer, policies).map_err(compile_err)
+}
+
+fn build_image_anonymizer<'a>(
+    spec: &AnalyzerParams,
+    policies: impl Iterator<Item = &'a Policy>,
+    entities: &[EntityRecord<Image>],
+) -> Anonymizer<Image> {
+    let catalog = build_catalog(spec);
+    let mut anonymizer = Anonymizer::<Image>::new().with_catalog(catalog);
+    for record in entities {
+        if let Some(action) = &record.r#override {
+            anonymizer = attach_override_image(anonymizer, record.entity.id, action);
+        }
+    }
+    attach_policies_image(anonymizer, policies)
+}
+
+fn build_audio_anonymizer<'a>(
+    spec: &AnalyzerParams,
+    policies: impl Iterator<Item = &'a Policy>,
+    entities: &[EntityRecord<Audio>],
+) -> Anonymizer<Audio> {
+    let catalog = build_catalog(spec);
+    let mut anonymizer = Anonymizer::<Audio>::new().with_catalog(catalog);
+    for record in entities {
+        if let Some(action) = &record.r#override {
+            anonymizer = attach_override_audio(anonymizer, record.entity.id, action);
+        }
+    }
+    attach_policies_audio(anonymizer, policies)
+}
+
+async fn run_anonymize<M>(
+    anonymizer: Anonymizer<M>,
+    mut handle: DocumentHandle<M>,
+    entities: &[EntityRecord<M>],
+) -> Result<Bytes>
+where
+    M: Modality + Clone,
+{
+    // elide's `anonymize` takes `&mut [Entity<M>]` and updates
+    // provenance in place; we clone the entities out of the
+    // persisted records since the persisted body is the
+    // source of truth for the next call (idempotent apply).
+    let mut working: Vec<Entity<M>> = entities.iter().map(|r| r.entity.clone()).collect();
+    anonymizer
+        .anonymize(&mut handle, &mut working)
+        .await
+        .map_err(|err| Error::internal("anonymize failed", COMPONENT).with_source(err))?;
+
+    let content = handle.encode().map_err(|err| {
+        Error::internal("post-anonymize encode failed", COMPONENT).with_source(err)
+    })?;
+    Ok(content.into_bytes())
+}
diff --git a/crates/nvisy-engine/src/runs/state.rs b/crates/nvisy-engine/src/runs/state.rs
new file mode 100644
index 000000000..cb7cd07b6
--- /dev/null
+++ b/crates/nvisy-engine/src/runs/state.rs
@@ -0,0 +1,263 @@
+//! Persisted run state: the [`Run`] header + per-document
+//! [`RunDocument`] bodies.
+//!
+//! Two-layer persistence in fjall:
+//!
+//! - [`Run`] (a metadata header) lives in the `run_headers`
+//!   keyspace under [`CompositeKey(actor_id, run_id)`].
+//! - [`RunDocument`] (one per input doc) lives in the `run_docs`
+//!   keyspace under [`TripleKey(actor_id, run_id, doc_id)`]. Its
+//!   [`body`](RunDocument::body) carries the recognized entities +
+//!   reviewer overrides.
+//!
+//! Bytes (input + redacted output) are **not** persisted in run
+//! keyspaces. Inputs live in the [`crate::FileRegistry`] before
+//! the run starts; redacted outputs land back in the same files
+//! keyspace via [`FileRegistry::put_file`] stamped with a
+//! [`FileLineage::RedactedFrom`] so they're traceable back to
+//! the run + source file. A [`RunDocument`] tracks both id sides
+//! via [`input_file_id`](RunDocument::input_file_id) +
+//! [`output_file_id`](RunDocument::output_file_id).
+//!
+//! [`CompositeKey(actor_id, run_id)`]: crate::registry::CompositeKey
+//! [`TripleKey(actor_id, run_id, doc_id)`]: crate::registry::TripleKey
+//! [`FileRegistry`]: crate::FileRegistry
+//! [`FileRegistry::put_file`]: crate::FileRegistry::put_file
+//! [`FileLineage::RedactedFrom`]: nvisy_core::FileLineage::RedactedFrom
+
+use std::collections::HashMap;
+
+use elide_core::entity::Entity;
+use elide_core::modality::audio::Audio;
+use elide_core::modality::image::Image;
+use elide_core::modality::tabular::Tabular;
+use elide_core::modality::text::Text;
+use hipstr::HipStr;
+use jiff::Timestamp;
+use nvisy_core::plan::AnalyzerParams;
+use nvisy_core::policy::RuleAction;
+use semver::Version;
+use serde::{Deserialize, Serialize};
+use uuid::Uuid;
+
+/// Top-level state of one run.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "state", rename_all = "snake_case")]
+pub enum RunState {
+    /// Analyze phase is in flight (or queued).
+    Analyzing,
+    /// All per-doc analyses finished; reports are ready for review
+    /// and override.
+    AwaitingReview,
+    /// Apply ran; every document succeeded.
+    Applied,
+    /// Apply ran; some documents succeeded, others failed. Per-doc
+    /// [`RunDocState`] carries the detail.
+    PartiallyApplied,
+    /// The whole run failed before any per-doc work completed.
+    Failed {
+        /// Human-readable reason (e.g. "couldn't load policy X").
+        reason: String,
+    },
+}
+
+/// State of one document inside a run.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "state", rename_all = "snake_case")]
+pub enum RunDocState {
+    /// Awaiting its turn on the analyze semaphore.
+    Queued,
+    /// Analyze is running.
+    Analyzing,
+    /// Analyze succeeded; entities + overrides recorded; awaiting
+    /// reviewer / apply.
+    AwaitingReview,
+    /// Apply ran; redacted bytes + audit recorded.
+    Applied,
+    /// Analyze or apply errored; reason in `reason`.
+    Failed {
+        /// Human-readable failure reason.
+        reason: String,
+    },
+    /// Per-doc timeout fired before completion.
+    TimedOut,
+}
+
+/// Run header — short metadata blob persisted under
+/// `(actor_id, run_id)`.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct Run {
+    /// Stable identifier. Engine-minted UUIDv7 at start time so
+    /// the value carries its own creation order.
+    pub id: Uuid,
+    /// Run state.
+    pub state: RunState,
+    /// UUIDv7 timestamp the run was started.
+    pub started_at: Timestamp,
+    /// UUIDv7 timestamp of the most recent state transition.
+    pub updated_at: Timestamp,
+    /// Policies the caller submitted, as resource refs. Loaded
+    /// from [`crate::keyspace::policy`] at start time; stable for the
+    /// lifetime of the run.
+    pub policy_refs: Vec<ResourceRef>,
+    /// Contexts the caller submitted, as resource refs.
+    pub context_refs: Vec<ResourceRef>,
+    /// Per-request metadata the caller attached (merged with each
+    /// document's descriptor at policy-evaluation time).
+    pub metadata: HashMap<String, String>,
+    /// One entry per input document; the body lives in `run_docs`
+    /// under a `(actor_id, run_id, doc_id)` triple key.
+    pub document_ids: Vec<Uuid>,
+    /// Recognition plan the caller supplied — engine compiles it
+    /// per modality at analyze time. Persisted on the header so
+    /// apply (run separately, potentially after a process restart)
+    /// can re-resolve the same recognition shape.
+    pub analyzer: AnalyzerParams,
+    /// Cap on how many per-doc analyses run concurrently. Honoured
+    /// by both analyze and apply.
+    pub concurrency: usize,
+}
+
+/// Reference to a stored policy / context resource.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct ResourceRef {
+    /// Resource UUID.
+    pub id: Uuid,
+    /// Resource version. Two `(id, version)` pairs reference
+    /// distinct stored blobs.
+    pub version: Version,
+}
+
+/// One document inside a run — the per-doc body persisted under
+/// `(actor_id, run_id, doc_id)`.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct RunDocument {
+    /// Engine-minted UUIDv7 at start time.
+    pub id: Uuid,
+    /// Id of the input file this doc analyses. Persisted on the
+    /// row so apply (run separately, potentially after a process
+    /// restart) can re-resolve the bytes via
+    /// [`FileRegistry::get_file_bytes`] without re-deriving from
+    /// the run header's `document_ids`.
+    ///
+    /// [`FileRegistry::get_file_bytes`]: crate::FileRegistry::get_file_bytes
+    pub input_file_id: Uuid,
+    /// Id of the redacted output file, when apply succeeded for
+    /// this doc. `None` pre-apply and on per-doc failure. Lookup
+    /// via [`FileRegistry::get_file_bytes`]; the metadata blob's
+    /// [`lineage`](nvisy_core::FileMetadata::lineage) field
+    /// points back at this run.
+    ///
+    /// [`FileRegistry::get_file_bytes`]: crate::FileRegistry::get_file_bytes
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub output_file_id: Option<Uuid>,
+    /// State of the per-doc lifecycle.
+    pub state: RunDocState,
+    /// The modality elide's codec resolved this doc to; pins the
+    /// active variant of [`body`](Self::body).
+    pub modality: ModalityKind,
+    /// Per-modality recognized entities + reviewer overrides.
+    pub body: DocBody,
+}
+
+/// Discriminator for the modality the codec resolved a document
+/// to. Pinned at decode time inside [`RunDocument::modality`].
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum ModalityKind {
+    /// `Text` modality.
+    Text,
+    /// `Tabular` modality.
+    Tabular,
+    /// `Image` modality.
+    Image,
+    /// `Audio` modality.
+    Audio,
+}
+
+/// Per-modality body: the entities recognized in this doc plus
+/// any reviewer overrides. Tagged by `modality` so the read side
+/// can pick the right variant from the JSON.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(tag = "modality", rename_all = "snake_case")]
+pub enum DocBody {
+    /// Text entities.
+    Text {
+        /// Recognized entities, in source-coordinate order.
+        entities: Vec<EntityRecord<Text>>,
+    },
+    /// Tabular entities.
+    Tabular {
+        /// Recognized entities, in source-coordinate order.
+        entities: Vec<EntityRecord<Tabular>>,
+    },
+    /// Image entities.
+    Image {
+        /// Recognized entities, in source-coordinate order.
+        entities: Vec<EntityRecord<Image>>,
+    },
+    /// Audio entities.
+    Audio {
+        /// Recognized entities, in source-coordinate order.
+        entities: Vec<EntityRecord<Audio>>,
+    },
+}
+
+impl DocBody {
+    /// Empty body for the given modality. Engine constructs this
+    /// before recognition runs so the doc state is queryable from
+    /// the moment the run starts.
+    pub fn empty(modality: ModalityKind) -> Self {
+        match modality {
+            ModalityKind::Text => Self::Text {
+                entities: Vec::new(),
+            },
+            ModalityKind::Tabular => Self::Tabular {
+                entities: Vec::new(),
+            },
+            ModalityKind::Image => Self::Image {
+                entities: Vec::new(),
+            },
+            ModalityKind::Audio => Self::Audio {
+                entities: Vec::new(),
+            },
+        }
+    }
+}
+
+/// One recognized entity plus the optional reviewer override.
+///
+/// The serde bound mirrors elide's [`Entity<M>`]: serialization
+/// needs `M::Location` and `M::Data` (de)serializable, which all
+/// four modalities elide ships satisfy under the `serde` feature.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+#[serde(bound = "M::Location: Serialize + for<'a> Deserialize<'a>, \
+                  M::Data: Serialize + for<'a> Deserialize<'a>")]
+pub struct EntityRecord<M: elide_core::modality::Modality> {
+    /// The elide entity, as recognition produced it.
+    pub entity: Entity<M>,
+    /// Reviewer-supplied override. `None` means "use the policy's
+    /// decision"; `Some(action)` overrides it for this specific
+    /// entity at apply time.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub r#override: Option<RuleAction>,
+}
+
+impl<M: elide_core::modality::Modality> EntityRecord<M> {
+    /// New record over `entity`, no override.
+    pub fn new(entity: Entity<M>) -> Self {
+        Self {
+            entity,
+            r#override: None,
+        }
+    }
+}
+
+/// Free-form reason carrier reused by `Failed`/`TimedOut` states
+/// at the per-doc API surface, when callers want a typed handle.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct FailureReason(pub HipStr<'static>);
diff --git a/crates/nvisy-engine/tests/configuration.rs b/crates/nvisy-engine/tests/configuration.rs
deleted file mode 100644
index 7c5ccb350..000000000
--- a/crates/nvisy-engine/tests/configuration.rs
+++ /dev/null
@@ -1,43 +0,0 @@
-//! Configuration parsing, validation, and example TOML tests.
-
-use nvisy_engine::core::{EngineConfig, RuntimeConfig};
-use validator::Validate;
-
-#[test]
-fn example_toml_parses() {
-    let contents = include_str!("../../../Nvisy.example.toml");
-
-    // Parse into a loose Value first to verify TOML syntax is valid,
-    // then extract the engine section which is always feature-independent.
-    // Subsystem provider sections (llm, stt) use cfg-gated enum
-    // variants that may not be available in the test binary.
-    let table: toml::Table =
-        toml::from_str(contents).expect("Nvisy.example.toml should be valid TOML");
-
-    assert!(table.contains_key("version"), "example should have version");
-    assert!(table.contains_key("engine"), "example should have [engine]");
-    assert!(table.contains_key("server"), "example should have [server]");
-
-    // Verify the engine section parses into our struct.
-    let engine_toml = toml::to_string(table.get("engine").unwrap()).unwrap();
-    let engine: EngineConfig =
-        toml::from_str(&engine_toml).expect("[engine] section should parse into EngineConfig");
-    engine
-        .validate()
-        .expect("[engine] section should pass validation");
-}
-
-#[test]
-fn empty_toml_uses_defaults() {
-    let config: RuntimeConfig = toml::from_str("").unwrap();
-    assert!(config.engine.is_none());
-    assert!(config.extraction.is_none());
-    assert!(config.detection.is_none());
-    assert!(config.validate().is_ok());
-}
-
-#[test]
-fn validation_accepts_defaults() {
-    let config = RuntimeConfig::default();
-    assert!(config.validate().is_ok());
-}
diff --git a/crates/nvisy-engine/tests/deduplication.rs b/crates/nvisy-engine/tests/deduplication.rs
deleted file mode 100644
index b041f35a4..000000000
--- a/crates/nvisy-engine/tests/deduplication.rs
+++ /dev/null
@@ -1,264 +0,0 @@
-//! Deduplication layer integration tests.
-//!
-//! Exercises the toolkit's `LayerPipeline` against real
-//! `DocumentTree<Text>` values constructed from in-memory text via
-//! the codec registry. These tests live here rather than in
-//! `nvisy-toolkit` because they cross the toolkit ↔ document
-//! boundary (they need `DocumentTree`, `TextMetadata`, and a real
-//! decoded codec handle).
-
-use nvisy_codec::content::{ContentData, ContentDescriptor, ContentSource};
-use nvisy_codec::{CodecRegistry, UntypedDocumentHandle};
-use nvisy_core::entity::{Entity, ModelProvenance, TrailProvenance, TrailStep, TrailStepKind};
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::Text;
-use nvisy_core::primitive::{Confidence, ConfidenceThreshold};
-use nvisy_engine::core::DocumentTree;
-use nvisy_engine::document::Document;
-use nvisy_engine::modality::{TextExtraction, TextMetadata};
-use nvisy_toolkit::deduplication::fuse::{
-    DeduplicationStrategy, FuseLayer, FusionWeights, GroupingCriteria,
-};
-use nvisy_toolkit::deduplication::{Layer, LayerContext, LayerParams, LayerPipeline};
-use uuid::Uuid;
-
-fn conf(v: f64) -> Confidence {
-    Confidence::new(v).expect("confidence in [0,1]")
-}
-
-fn ner_step(confidence: Confidence) -> TrailStep {
-    TrailStep::recognition(
-        "ner",
-        confidence,
-        TrailProvenance::Model(ModelProvenance::new("test")),
-        "",
-    )
-}
-
-const TEXT: &str = "John Smith";
-
-async fn tree_from(text: &str) -> DocumentTree<Text> {
-    let registry = CodecRegistry::with_builtin();
-    let format = registry.by_extension("txt").expect("txt codec registered");
-    let data = ContentData::new(ContentSource::new(), text.as_bytes().to_vec().into());
-    let untyped = format.decode(data).await.expect("decode");
-    let handle = match untyped {
-        UntypedDocumentHandle::Text(h) => h,
-        _ => panic!("txt loader must produce Text"),
-    };
-    let source = handle.handler().source();
-    let doc = Document::<Text>::new(
-        TextMetadata {
-            extraction: TextExtraction::Native,
-            languages: Vec::new(),
-        },
-        source,
-    );
-    let descriptor = ContentDescriptor::new().with_content_type("text/plain");
-    DocumentTree::new(doc, handle, descriptor)
-}
-
-async fn fuse_with(
-    strategy: DeduplicationStrategy,
-    criteria: GroupingCriteria,
-    tree: &DocumentTree<Text>,
-    entities: &mut Vec<Entity<Text>>,
-) {
-    let ctx = LayerContext::new(tree);
-    let layer = FuseLayer::new(strategy, criteria);
-    let dropped = layer.apply(entities, &ctx).await;
-    assert!(dropped.is_empty(), "fuse never drops");
-}
-
-#[tokio::test]
-async fn confidence_threshold_filters() {
-    let tree = tree_from("John......Jane").await;
-    let params = LayerParams {
-        confidence_threshold: Some(ConfidenceThreshold::clamped(0.85)),
-        ..Default::default()
-    };
-    let entities: Vec<Entity<Text>> = vec![
-        Entity::test_builder(0, 4).test_build(),
-        Entity::test_builder(10, 14)
-            .with_confidence(conf(0.5))
-            .test_build(),
-    ];
-    let pipeline: LayerPipeline<Text, _> =
-        LayerPipeline::from_params(&params).expect("pipeline builds");
-    let ctx = LayerContext::new(&tree).with_correlation_id(Uuid::nil());
-    let result = pipeline.run(entities, &ctx).await;
-    assert_eq!(result.len(), 1);
-    let value = tree.text_at(&result[0].location).await;
-    assert_eq!(value.as_deref(), Some("John"));
-}
-
-#[tokio::test]
-async fn full_pipeline() {
-    let tree = tree_from(TEXT).await;
-    let entities: Vec<Entity<Text>> = vec![
-        Entity::test_builder(0, 4)
-            .with_confidence(conf(0.7))
-            .test_build(),
-        Entity::test_builder(0, 4)
-            .with_confidence(conf(0.8))
-            .test_build(),
-        Entity::test_builder(0, 4)
-            .with_trail(vec![TrailStep::recognition(
-                "ner",
-                conf(0.85),
-                TrailProvenance::Model(ModelProvenance::new("test")),
-                "",
-            )])
-            .with_confidence(conf(0.85))
-            .test_build(),
-    ];
-    let pipeline: LayerPipeline<Text, _> =
-        LayerPipeline::from_params(&LayerParams::default()).expect("pipeline builds");
-    let ctx = LayerContext::new(&tree).with_correlation_id(Uuid::nil());
-    let result = pipeline.run(entities, &ctx).await;
-    assert_eq!(result.len(), 1);
-    assert!((result[0].confidence.get() - 0.85).abs() < f64::EPSILON);
-}
-
-#[tokio::test]
-async fn strict_grouping_fuses_identical_spans_with_max_confidence() {
-    let tree = tree_from(TEXT).await;
-    let mut entities: Vec<_> = vec![
-        Entity::test_builder(0, 4)
-            .with_confidence(conf(0.8))
-            .test_build(),
-        Entity::test_builder(0, 4).test_build(),
-    ];
-    fuse_with(
-        DeduplicationStrategy::MaxConfidence,
-        GroupingCriteria::Strict,
-        &tree,
-        &mut entities,
-    )
-    .await;
-    assert_eq!(entities.len(), 1);
-    assert!((entities[0].confidence.get() - 0.9).abs() < f64::EPSILON);
-}
-
-#[tokio::test]
-async fn narrowing_groups_substring_with_overlap() {
-    let tree = tree_from(TEXT).await;
-    let mut entities: Vec<_> = vec![
-        Entity::test_builder(0, 4)
-            .with_confidence(conf(0.8))
-            .test_build(),
-        Entity::test_builder(0, 10)
-            .with_trail(vec![ner_step(conf(0.9))])
-            .test_build(),
-    ];
-    fuse_with(
-        DeduplicationStrategy::MaxConfidence,
-        GroupingCriteria::Narrowing,
-        &tree,
-        &mut entities,
-    )
-    .await;
-    assert_eq!(entities.len(), 1);
-    let value = tree.text_at(&entities[0].location).await;
-    assert_eq!(value.as_deref(), Some("John Smith"));
-}
-
-#[tokio::test]
-async fn widening_groups_across_non_overlapping_locations() {
-    let text = format!("{:<100}John Smith", TEXT);
-    let tree = tree_from(&text).await;
-    let mut entities: Vec<_> = vec![
-        Entity::test_builder(0, 4).test_build(),
-        Entity::test_builder(100, 110)
-            .with_trail(vec![ner_step(conf(0.9))])
-            .test_build(),
-    ];
-    fuse_with(
-        DeduplicationStrategy::MaxConfidence,
-        GroupingCriteria::Widening,
-        &tree,
-        &mut entities,
-    )
-    .await;
-    assert_eq!(entities.len(), 1);
-}
-
-#[tokio::test]
-async fn noisy_or_strategy() {
-    let tree = tree_from(TEXT).await;
-    let mut entities: Vec<_> = vec![
-        Entity::test_builder(0, 4)
-            .with_confidence(conf(0.7))
-            .test_build(),
-        Entity::test_builder(0, 4)
-            .with_trail(vec![ner_step(conf(0.9))])
-            .with_confidence(conf(0.8))
-            .test_build(),
-    ];
-    fuse_with(
-        DeduplicationStrategy::NoisyOr {
-            weights: Default::default(),
-        },
-        GroupingCriteria::default(),
-        &tree,
-        &mut entities,
-    )
-    .await;
-    assert_eq!(entities.len(), 1);
-    // 1 - (1 - 0.7)(1 - 0.8) = 1 - 0.06 = 0.94
-    assert!((entities[0].confidence.get() - 0.94).abs() < 0.001);
-}
-
-#[tokio::test]
-async fn weighted_average_strategy() {
-    let tree = tree_from(TEXT).await;
-    let mut weights = FusionWeights::new();
-    weights.insert("pattern", 1.0);
-    weights.insert("ner", 2.0);
-
-    let mut entities: Vec<_> = vec![
-        Entity::test_builder(0, 4)
-            .with_confidence(conf(0.6))
-            .test_build(),
-        Entity::test_builder(0, 4)
-            .with_trail(vec![ner_step(conf(0.9))])
-            .test_build(),
-    ];
-    fuse_with(
-        DeduplicationStrategy::WeightedAverage { weights },
-        GroupingCriteria::default(),
-        &tree,
-        &mut entities,
-    )
-    .await;
-    assert_eq!(entities.len(), 1);
-    // (0.6 * 1.0 + 0.9 * 2.0) / 3.0 = 0.8
-    assert!((entities[0].confidence.get() - 0.8).abs() < 0.001);
-}
-
-#[tokio::test]
-async fn different_detector_tagged_as_ensemble_fusion() {
-    let tree = tree_from(TEXT).await;
-    let mut entities: Vec<_> = vec![
-        Entity::test_builder(0, 4)
-            .with_confidence(conf(0.8))
-            .test_build(),
-        Entity::test_builder(0, 4)
-            .with_trail(vec![ner_step(conf(0.9))])
-            .test_build(),
-    ];
-    fuse_with(
-        DeduplicationStrategy::MaxConfidence,
-        GroupingCriteria::default(),
-        &tree,
-        &mut entities,
-    )
-    .await;
-    assert_eq!(entities.len(), 1);
-    assert!(
-        entities[0]
-            .trail
-            .iter()
-            .any(|s| matches!(s.kind, TrailStepKind::Fusion) && s.reason.contains("ensemble"))
-    );
-}
diff --git a/crates/nvisy-engine/tests/detection_engine.rs b/crates/nvisy-engine/tests/detection_engine.rs
deleted file mode 100644
index 7913ba9d6..000000000
--- a/crates/nvisy-engine/tests/detection_engine.rs
+++ /dev/null
@@ -1,41 +0,0 @@
-//! Detection-engine construction and basic lifecycle tests.
-
-use std::time::Duration;
-
-use nvisy_engine::core::{EngineConfig, ResourceLimits, RuntimeConfig};
-use nvisy_engine::detection::DetectionEngine;
-
-#[tokio::test]
-async fn temp_engine_points_to_temp_dir() {
-    let (engine, dir) = DetectionEngine::temp().await;
-    assert_eq!(engine.data_dir(), dir.path());
-}
-
-#[tokio::test]
-async fn open_with_default_config() -> anyhow::Result<()> {
-    let dir = tempfile::tempdir()?;
-    let engine = DetectionEngine::open(dir.path(), RuntimeConfig::default()).await?;
-    assert_eq!(engine.data_dir(), dir.path());
-    Ok(())
-}
-
-#[tokio::test]
-async fn open_with_custom_limits() -> anyhow::Result<()> {
-    let dir = tempfile::tempdir()?;
-    let config = RuntimeConfig {
-        engine: Some(EngineConfig {
-            limits: ResourceLimits {
-                concurrency: None,
-                run_timeout: Some(Duration::from_secs(5)),
-            },
-            ..Default::default()
-        }),
-        ..Default::default()
-    };
-    let engine = DetectionEngine::open(dir.path(), config).await?;
-    assert_eq!(
-        engine.config().engine.as_ref().unwrap().limits.run_timeout,
-        Some(Duration::from_secs(5))
-    );
-    Ok(())
-}
diff --git a/crates/nvisy-engine/tests/redaction_policy.rs b/crates/nvisy-engine/tests/redaction_policy.rs
deleted file mode 100644
index 1e72ab7e8..000000000
--- a/crates/nvisy-engine/tests/redaction_policy.rs
+++ /dev/null
@@ -1,221 +0,0 @@
-//! End-to-end test for the policy → audit redaction wire shape.
-//!
-//! Loads each [`TextRedaction`] variant from TOML, asserts the
-//! deserialised operator matches the spec, then confirms a full
-//! `Action::Redact(ModalityRedactions)` round-trips through the JSON
-//! wire format the audit record uses.
-//!
-//! This exercises the user-visible surface end-to-end without
-//! touching the codec handle — the apply-time bridge from operator
-//! spec to runnable instance is covered separately by
-//! `policy::redaction::Instantiate` impls.
-
-use nvisy_core::modality::Text;
-use nvisy_engine::policy::redaction::{HashAlgorithm, ModalityRedactions, TextRedaction};
-use nvisy_engine::policy::{Action, Policy, PolicyRule};
-use nvisy_toolkit::redaction::AnonymizerId;
-
-/// Each built-in arm deserialises with its declared params.
-#[test]
-fn text_redaction_each_variant_round_trips() {
-    let toml = r#"
-        [[arms]]
-        kind = "replace"
-        template = "[EMAIL]"
-
-        [[arms]]
-        kind = "mask"
-        mask_char = "*"
-        keep_prefix = 4
-        keep_suffix = 4
-
-        [[arms]]
-        kind = "hash"
-        algorithm = "sha256"
-        salt = "pepper"
-
-        [[arms]]
-        kind = "redact"
-
-        [[arms]]
-        kind = "keep"
-
-        [[arms]]
-        kind = "custom"
-        id = "kms_encrypt"
-    "#;
-
-    #[derive(serde::Deserialize)]
-    struct Wrapper {
-        arms: Vec<TextRedaction>,
-    }
-    let parsed: Wrapper = toml::from_str(toml).expect("parse arms");
-    assert_eq!(parsed.arms.len(), 6);
-
-    match &parsed.arms[0] {
-        TextRedaction::Replace { template } => assert_eq!(template, "[EMAIL]"),
-        other => panic!("expected Replace, got {other:?}"),
-    }
-    match &parsed.arms[1] {
-        TextRedaction::Mask {
-            mask_char,
-            keep_prefix,
-            keep_suffix,
-        } => {
-            assert_eq!(*mask_char, '*');
-            assert_eq!(*keep_prefix, 4);
-            assert_eq!(*keep_suffix, 4);
-        }
-        other => panic!("expected Mask, got {other:?}"),
-    }
-    match &parsed.arms[2] {
-        TextRedaction::Hash { algorithm, salt } => {
-            assert_eq!(*algorithm, HashAlgorithm::Sha256);
-            assert_eq!(salt.as_deref(), Some("pepper"));
-        }
-        other => panic!("expected Hash, got {other:?}"),
-    }
-    assert!(matches!(parsed.arms[3], TextRedaction::Redact));
-    assert!(matches!(parsed.arms[4], TextRedaction::Keep));
-    match &parsed.arms[5] {
-        TextRedaction::Custom { id } => assert_eq!(id.as_str(), "kms_encrypt"),
-        other => panic!("expected Custom, got {other:?}"),
-    }
-}
-
-/// Defaulted fields (template, mask_char) come back with their
-/// declared defaults when the TOML omits them.
-#[test]
-fn text_redaction_defaults_fill_in() {
-    let parsed: TextRedaction = toml::from_str(r#"kind = "replace""#).expect("parse");
-    match parsed {
-        TextRedaction::Replace { template } => assert_eq!(template, "[{entity_kind}]"),
-        other => panic!("expected Replace, got {other:?}"),
-    }
-    let parsed: TextRedaction = toml::from_str(r#"kind = "mask""#).expect("parse");
-    match parsed {
-        TextRedaction::Mask {
-            mask_char,
-            keep_prefix,
-            keep_suffix,
-        } => {
-            assert_eq!(mask_char, '*');
-            assert_eq!(keep_prefix, 0);
-            assert_eq!(keep_suffix, 0);
-        }
-        other => panic!("expected Mask, got {other:?}"),
-    }
-}
-
-/// A full `Action::Redact(ModalityRedactions)` round-trips through
-/// JSON — the wire shape audit records use.
-#[test]
-fn action_redact_round_trips_through_json() {
-    let action = Action::Redact(ModalityRedactions {
-        text: Some(TextRedaction::Replace {
-            template: "[EMAIL]".into(),
-        }),
-        ..Default::default()
-    });
-    let json = serde_json::to_string(&action).expect("serialize");
-    assert!(
-        json.contains("\"redact\""),
-        "expected redact tag, got {json}"
-    );
-    assert!(
-        json.contains("\"text\""),
-        "expected text operator, got {json}"
-    );
-
-    let parsed: Action = serde_json::from_str(&json).expect("deserialize");
-    match parsed {
-        Action::Redact(operators) => match operators.text {
-            Some(TextRedaction::Replace { template }) => assert_eq!(template, "[EMAIL]"),
-            other => panic!("expected nested Replace, got {other:?}"),
-        },
-        other => panic!("expected Redact, got {other:?}"),
-    }
-}
-
-/// A `Custom`-arm action survives a JSON round-trip with its
-/// [`AnonymizerId<Text>`] intact.
-#[test]
-fn action_redact_custom_id_round_trips() {
-    let id = AnonymizerId::<Text>::from_static("kms_encrypt");
-    let action = Action::Redact(ModalityRedactions {
-        text: Some(TextRedaction::Custom { id: id.clone() }),
-        ..Default::default()
-    });
-    let json = serde_json::to_string(&action).expect("serialize");
-    let parsed: Action = serde_json::from_str(&json).expect("deserialize");
-    match parsed {
-        Action::Redact(operators) => match operators.text {
-            Some(TextRedaction::Custom { id: round_tripped }) => assert_eq!(round_tripped, id),
-            other => panic!("expected Custom redact, got {other:?}"),
-        },
-        other => panic!("expected Redact, got {other:?}"),
-    }
-}
-
-/// A `Policy` with mixed rules deserialises from TOML and the
-/// `defaultAction` carries its operator spec verbatim.
-#[test]
-fn policy_with_redact_rules_round_trips_from_toml() {
-    let toml = r##"
-        name = "email-and-card"
-        version = "1.0.0"
-        description = "redact emails, mask cards, drop everything else by default"
-
-        [[rules]]
-        name = "redact-email"
-        match = { labels = ["email_address"] }
-        [rules.redact]
-        text = { kind = "replace", template = "[EMAIL]" }
-
-        [[rules]]
-        name = "mask-card"
-        match = { labels = ["payment_card"] }
-        [rules.redact]
-        text = { kind = "mask", mask_char = "#", keep_suffix = 4 }
-
-        [defaultAction.redact]
-        text = { kind = "redact" }
-    "##;
-
-    let policy: Policy = toml::from_str(toml).expect("parse policy");
-    assert_eq!(policy.name, "email-and-card");
-    assert_eq!(policy.rules.len(), 2);
-
-    let PolicyRule { action, .. } = &policy.rules[0];
-    match action {
-        Action::Redact(operators) => match &operators.text {
-            Some(TextRedaction::Replace { template }) => assert_eq!(template, "[EMAIL]"),
-            other => panic!("expected Replace, got {other:?}"),
-        },
-        other => panic!("expected Redact, got {other:?}"),
-    }
-
-    let PolicyRule { action, .. } = &policy.rules[1];
-    match action {
-        Action::Redact(operators) => match &operators.text {
-            Some(TextRedaction::Mask {
-                mask_char,
-                keep_suffix,
-                ..
-            }) => {
-                assert_eq!(*mask_char, '#');
-                assert_eq!(*keep_suffix, 4);
-            }
-            other => panic!("expected Mask, got {other:?}"),
-        },
-        other => panic!("expected Redact, got {other:?}"),
-    }
-
-    match policy.default_action.as_ref() {
-        Some(Action::Redact(operators)) => match &operators.text {
-            Some(TextRedaction::Redact) => {}
-            other => panic!("expected default Redact, got {other:?}"),
-        },
-        other => panic!("expected default Redact action, got {other:?}"),
-    }
-}
diff --git a/crates/nvisy-fake/Cargo.toml b/crates/nvisy-fake/Cargo.toml
deleted file mode 100644
index d2bdba916..000000000
--- a/crates/nvisy-fake/Cargo.toml
+++ /dev/null
@@ -1,41 +0,0 @@
-# https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[package]
-name = "nvisy-fake"
-description = "Locale-aware fake-data Anonymizer for Nvisy entities"
-keywords = ["nvisy", "fake", "anonymizer", "pii"]
-categories = ["text-processing"]
-readme = "README.md"
-
-version = { workspace = true }
-rust-version = { workspace = true }
-edition = { workspace = true }
-license = { workspace = true }
-publish = true
-
-authors = { workspace = true }
-repository = { workspace = true }
-homepage = { workspace = true }
-documentation = { workspace = true }
-
-[package.metadata.docs.rs]
-all-features = true
-rustdoc-args = ["--cfg", "docsrs"]
-
-[dependencies]
-# Internal crates
-nvisy-core = { workspace = true, features = [] }
-
-# Primitive datatypes (UUIDv4 for fake `DeviceId`)
-uuid = { workspace = true, features = ["v4"] }
-
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-
-# Utilities
-fake = { workspace = true, features = [] }
-
-[dev-dependencies]
-nvisy-core = { workspace = true, features = ["test-utils"] }
-nvisy-toolkit = { workspace = true, features = [] }
-tokio = { workspace = true, features = ["rt", "macros"] }
diff --git a/crates/nvisy-fake/README.md b/crates/nvisy-fake/README.md
deleted file mode 100644
index 39293a19e..000000000
--- a/crates/nvisy-fake/README.md
+++ /dev/null
@@ -1,42 +0,0 @@
-# nvisy-fake
-
-[![Build](https://img.shields.io/github/actions/workflow/status/nvisycom/runtime/build.yml?branch=main&label=build%20%26%20test&style=flat-square)](https://github.com/nvisycom/runtime/actions/workflows/build.yml)
-
-Locale-aware fake-data anonymizer for the Nvisy runtime.
-
-## Overview
-
-`Fake` is a text-modality `Anonymizer` that swaps a detected entity
-for a plausible fake value drawn from the [`fake`](https://docs.rs/fake)
-crate's locale tables. The locale is picked per-entity from the
-entity's BCP-47 `language` field, falling back to the
-`default_language` passed at construction when no tag is present.
-
-Construct with `Fake::new(language_tag)`; tune behaviour with
-`.with_seed(u64)`, `.length_preserving()`, and `.format_preserving()`.
-RNG state is derived from `entity_id` (or the UUID when no
-coreference id is present) so coreferent mentions collapse to the
-same fake value within a run.
-
-Entity kinds outside the core PII set (`PersonName`, `EmailAddress`,
-`PhoneNumber`, `Address`, `PostalCode`, `Url`, `DateOfBirth`, `Age`,
-`PaymentCard`, `Iban`, `BankAccount`, `Currency`) fall through to a
-`[{entity_kind}]` placeholder.
-
-## Documentation
-
-See [`docs/`](../../docs/) for architecture, security, and API documentation.
-
-## Changelog
-
-See [CHANGELOG.md](../../CHANGELOG.md) for release notes and version history.
-
-## License
-
-Apache 2.0 License, see [LICENSE.txt](../../LICENSE.txt)
-
-## Support
-
-- **Documentation**: [docs.nvisy.com](https://docs.nvisy.com)
-- **Issues**: [GitHub Issues](https://github.com/nvisycom/runtime/issues)
-- **Email**: [support@nvisy.com](mailto:support@nvisy.com)
diff --git a/crates/nvisy-fake/src/anonymizer/mod.rs b/crates/nvisy-fake/src/anonymizer/mod.rs
deleted file mode 100644
index e73c63631..000000000
--- a/crates/nvisy-fake/src/anonymizer/mod.rs
+++ /dev/null
@@ -1,382 +0,0 @@
-//! [`Fake`]: text-modality [`Anonymizer`] that swaps detected
-//! entities for plausible fake values.
-
-use std::fmt;
-use std::hash::{DefaultHasher, Hash, Hasher};
-use std::str::FromStr;
-use std::sync::Arc;
-
-use fake::rand::SeedableRng;
-use fake::rand::rngs::SmallRng;
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Modality, Tabular, Text, TextData, TextLocation};
-use nvisy_core::primitive::LanguageTag;
-use nvisy_core::redaction::{Anonymizer, LeakProfile, TabularReplacement, TextReplacement};
-
-use crate::generator;
-use crate::locale::Locale;
-
-/// Text-modality fake-data anonymizer.
-///
-/// Picks a locale from the entity's BCP-47 `language` field, falling
-/// back to the `default_language` (English unless overridden) when
-/// the entity carries no tag. RNG state is derived per-call from
-/// the entity's coreference id (or its UUID when there is none),
-/// so coreferent mentions of the same real-world entity collapse to
-/// the same fake value within a run.
-///
-/// Entity kinds outside the core PII catalogue delegate to the
-/// `fallback` anonymizer passed at construction. There is no
-/// implicit default: the caller must say what should happen for
-/// unsupported kinds.
-///
-/// Structured kinds (IBAN, payment card, postal code, phone,
-/// date-of-birth, etc.) always pattern-preserve the original — the
-/// output's length and character-class layout matches the input,
-/// only the digits and letters are randomised. Free-form kinds
-/// (names, addresses, organisations) emit a fresh locale-aware
-/// fake whose length doesn't need to match.
-#[derive(Clone)]
-pub struct Fake {
-    fallback: Arc<dyn Anonymizer<Text>>,
-    default_language: LanguageTag,
-    seed: u64,
-}
-
-impl fmt::Debug for Fake {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("Fake")
-            .field("default_language", &self.default_language)
-            .field("seed", &self.seed)
-            .field("fallback", &"Arc<dyn Anonymizer<Text>>")
-            .finish()
-    }
-}
-
-impl Fake {
-    /// Build a `Fake` operator with `fallback` as the anonymizer
-    /// used for entity kinds outside the core PII catalogue.
-    pub fn new<A>(fallback: A) -> Self
-    where
-        A: Anonymizer<Text> + 'static,
-    {
-        Self {
-            fallback: Arc::new(fallback),
-            default_language: LanguageTag::from_str("en").expect("en is BCP-47"),
-            seed: 0,
-        }
-    }
-
-    /// Override the default language used when the entity carries
-    /// no `language` tag. Initial value is `"en"`.
-    #[must_use]
-    pub fn with_default_language(mut self, language: LanguageTag) -> Self {
-        self.default_language = language;
-        self
-    }
-
-    /// Salt the per-call RNG with `seed`. Two operators with the
-    /// same seed produce the same fake value for the same entity.
-    #[must_use]
-    pub fn with_seed(mut self, seed: u64) -> Self {
-        self.seed = seed;
-        self
-    }
-
-    fn locale_for(&self, language: Option<&LanguageTag>) -> Locale {
-        Locale::from_tag(language.unwrap_or(&self.default_language))
-    }
-
-    fn rng_for(&self, identity: Identity<'_>) -> SmallRng {
-        let mut hasher = DefaultHasher::new();
-        self.seed.hash(&mut hasher);
-        identity.hash(&mut hasher);
-        SmallRng::seed_from_u64(hasher.finish())
-    }
-
-    /// Try the generator for `label`; return `None` if it has no
-    /// entry, so the caller can delegate to the fallback.
-    fn try_generate(
-        &self,
-        locale: Locale,
-        label: &str,
-        identity: Identity<'_>,
-        source: &str,
-    ) -> Option<String> {
-        let mut rng = self.rng_for(identity);
-        generator::Context::new(locale, label, source).generate(&mut rng)
-    }
-}
-
-#[async_trait::async_trait]
-impl Anonymizer<Text> for Fake {
-    fn leak_profile(&self) -> LeakProfile {
-        // The original value is gone; only the entity's position
-        // and approximate shape (length differs from the original)
-        // leak.
-        LeakProfile::Partial
-    }
-
-    async fn apply(&self, entity: &Entity<Text>, source: &TextData) -> Result<TextReplacement> {
-        let locale = self.locale_for(entity.language.as_ref());
-        match self.try_generate(
-            locale,
-            entity.label.as_str(),
-            Identity::from(entity),
-            source.text.as_str(),
-        ) {
-            Some(value) => Ok(TextReplacement::substituted(value)),
-            None => self.fallback.apply(entity, source).await,
-        }
-    }
-}
-
-#[async_trait::async_trait]
-impl Anonymizer<Tabular> for Fake {
-    fn leak_profile(&self) -> LeakProfile {
-        LeakProfile::Partial
-    }
-
-    async fn apply(
-        &self,
-        entity: &Entity<Tabular>,
-        source: &TextData,
-    ) -> Result<TabularReplacement> {
-        let locale = self.locale_for(entity.language.as_ref());
-        if let Some(value) = self.try_generate(
-            locale,
-            entity.label.as_str(),
-            Identity::from(entity),
-            source.text.as_str(),
-        ) {
-            return Ok(TabularReplacement::substituted(value));
-        }
-        // No Tabular fallback — borrow the Text fallback by
-        // adapting the synthetic Text entity. The replacement
-        // values are both string-typed, so the adaptation is just
-        // a variant rewrap.
-        let text_entity = adapt_to_text(entity);
-        match self.fallback.apply(&text_entity, source).await? {
-            TextReplacement::Substituted { value } => Ok(TabularReplacement::substituted(value)),
-            TextReplacement::Removed => Ok(TabularReplacement::ColumnDropped),
-        }
-    }
-}
-
-/// Adapt a `Entity<Tabular>` into an `Entity<Text>` for delegating
-/// to a Text-shaped fallback. Only the identity-relevant fields
-/// (id, entity_id, entity_kind, language) need to round-trip; the
-/// location is a synthetic full-span TextLocation since the
-/// fallback's `apply` body reads source directly.
-fn adapt_to_text(entity: &Entity<Tabular>) -> Entity<Text> {
-    let mut builder = Entity::<Text>::builder()
-        .with_id(entity.id)
-        .with_label(entity.label.clone())
-        .with_location(TextLocation::new(0, 0))
-        .with_confidence(entity.confidence)
-        .with_trail(entity.trail.clone());
-    if let Some(id) = entity.entity_id.clone() {
-        builder = builder.with_entity_id(id);
-    }
-    if let Some(lang) = entity.language.clone() {
-        builder = builder.with_language(lang);
-    }
-    builder.build().expect("text entity adaptation")
-}
-
-/// Identity key used to seed the rng. Prefers the coreference id
-/// (shared across mentions of the same real-world entity) and falls
-/// back to the entity's UUID bytes.
-enum Identity<'a> {
-    Coref(&'a str),
-    Uuid([u8; 16]),
-}
-
-impl<'a, M> From<&'a Entity<M>> for Identity<'a>
-where
-    M: Modality,
-{
-    fn from(entity: &'a Entity<M>) -> Self {
-        match entity.entity_id.as_deref() {
-            Some(id) => Identity::Coref(id),
-            None => Identity::Uuid(*entity.id.as_bytes()),
-        }
-    }
-}
-
-impl Hash for Identity<'_> {
-    fn hash<H: Hasher>(&self, state: &mut H) {
-        match self {
-            Identity::Coref(s) => {
-                0u8.hash(state);
-                s.hash(state);
-            }
-            Identity::Uuid(bytes) => {
-                1u8.hash(state);
-                bytes.hash(state);
-            }
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::collections::HashSet;
-
-    use nvisy_core::entity::{EntityLabelRef, builtins};
-    use nvisy_core::modality::TabularLocation;
-    use nvisy_core::primitive::Confidence;
-    use nvisy_core::redaction::Anonymizer as _;
-    use nvisy_toolkit::redaction::anonymizer::{Mask, Replace};
-
-    use super::*;
-
-    fn fake() -> Fake {
-        Fake::new(Mask::stars())
-    }
-
-    /// Build an entity whose location spans the entire `source`
-    /// string. Tests don't care about offsets — making them match
-    /// the source makes them self-documenting.
-    fn entity_over(label: EntityLabelRef, source: &str) -> Entity<Text> {
-        Entity::test_builder(0, source.len())
-            .with_label(label)
-            .test_build()
-    }
-
-    fn coref_entity(label: EntityLabelRef, source: &str, coref_id: &str) -> Entity<Text> {
-        Entity::test_builder(0, source.len())
-            .with_label(label)
-            .with_entity_id(coref_id.to_string())
-            .test_build()
-    }
-
-    #[tokio::test]
-    async fn unsupported_kind_delegates_to_fallback() {
-        // Diagnosis isn't faked — sensitive clinical kinds are
-        // intentionally excluded — so it falls through to the
-        // fallback anonymizer.
-        let op = Fake::new(Replace::new("[redacted]"));
-        let source = TextData::new("hypertension");
-        let entity = entity_over(builtins::DIAGNOSIS.label_ref(), source.text.as_str());
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(out, TextReplacement::substituted("[redacted]"));
-    }
-
-    #[tokio::test]
-    async fn fallback_can_be_mask() {
-        let op = Fake::new(Mask::stars());
-        let source = TextData::new("hypertension");
-        let entity = entity_over(builtins::DIAGNOSIS.label_ref(), source.text.as_str());
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(out, TextReplacement::substituted("************"));
-    }
-
-    #[tokio::test]
-    async fn same_seed_and_entity_id_produces_same_output() {
-        let op = fake().with_seed(42);
-        let source = TextData::new("alice");
-        let entity = entity_over(builtins::PERSON_NAME.label_ref(), source.text.as_str());
-        let a = op.apply(&entity, &source).await.unwrap();
-        let b = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(a, b);
-    }
-
-    #[tokio::test]
-    async fn entity_language_overrides_default() {
-        let op = fake();
-        let source = TextData::new("名前");
-        let mut entity = entity_over(builtins::PERSON_NAME.label_ref(), source.text.as_str());
-        entity.language = Some("ja".parse().unwrap());
-        let out = op.apply(&entity, &source).await.unwrap();
-        let TextReplacement::Substituted { value } = out else {
-            panic!("expected Substituted variant");
-        };
-        assert!(!value.is_empty(), "ja name should not be empty");
-    }
-
-    #[tokio::test]
-    async fn default_language_applies_when_entity_unlanguaged() {
-        let op = fake().with_default_language("ja".parse().unwrap());
-        let source = TextData::new("name");
-        let entity = entity_over(builtins::PERSON_NAME.label_ref(), source.text.as_str());
-        let out = op.apply(&entity, &source).await.unwrap();
-        let TextReplacement::Substituted { value } = out else {
-            panic!("expected Substituted variant");
-        };
-        assert!(!value.is_empty());
-    }
-
-    #[tokio::test]
-    async fn coreferent_entities_collapse_to_same_fake() {
-        let op = fake();
-        let source = TextData::new("alice");
-        let a = coref_entity(
-            builtins::PERSON_NAME.label_ref(),
-            source.text.as_str(),
-            "ENTITY_42",
-        );
-        let b = coref_entity(
-            builtins::PERSON_NAME.label_ref(),
-            source.text.as_str(),
-            "ENTITY_42",
-        );
-        let out_a = op.apply(&a, &source).await.unwrap();
-        let out_b = op.apply(&b, &source).await.unwrap();
-        assert_eq!(out_a, out_b);
-        let c = coref_entity(
-            builtins::PERSON_NAME.label_ref(),
-            source.text.as_str(),
-            "ENTITY_99",
-        );
-        let out_c = op.apply(&c, &source).await.unwrap();
-        assert_ne!(
-            out_a, out_c,
-            "different coref id should yield different fake"
-        );
-    }
-
-    #[tokio::test]
-    async fn distinct_entities_get_distinct_fakes() {
-        let op = fake();
-        let source = TextData::new("seed");
-        let mut outputs: HashSet<String> = HashSet::new();
-        for _ in 0..32 {
-            let entity = entity_over(builtins::PERSON_NAME.label_ref(), source.text.as_str());
-            let out = op.apply(&entity, &source).await.unwrap();
-            let TextReplacement::Substituted { value } = out else {
-                panic!("expected Substituted");
-            };
-            outputs.insert(value);
-        }
-        assert_eq!(
-            outputs.len(),
-            32,
-            "expected 32 distinct fakes across 32 fresh entity ids"
-        );
-    }
-
-    #[tokio::test]
-    async fn tabular_impl_emits_substituted() {
-        let op = fake();
-        let entity = Entity::<Tabular>::builder()
-            .with_label(builtins::PERSON_NAME.label_ref())
-            .with_location(TabularLocation {
-                row_index: 0u32,
-                column_index: 0u32,
-                start_offset: None,
-                end_offset: None,
-                column_name: None,
-                sheet_name: None,
-            })
-            .with_confidence(Confidence::clamped(1.0))
-            .build()
-            .unwrap();
-        let out = op.apply(&entity, &TextData::new("alice")).await.unwrap();
-        let TabularReplacement::Substituted { value } = out else {
-            panic!("expected Substituted");
-        };
-        assert!(!value.is_empty());
-    }
-}
diff --git a/crates/nvisy-fake/src/generator/case_id.rs b/crates/nvisy-fake/src/generator/case_id.rs
deleted file mode 100644
index 394334515..000000000
--- a/crates/nvisy-fake/src/generator/case_id.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-//! Opaque numeric identifiers: internal IDs, case numbers.
-
-use fake::rand::RngExt;
-
-use super::digits;
-
-/// 10-digit numeric identifier shared by `InternalId` and
-/// `CaseNumber` — both are opaque IDs without a globally
-/// standardised format.
-pub(super) fn internal_id<R: RngExt + ?Sized>(rng: &mut R) -> String {
-    digits(10, rng)
-}
diff --git a/crates/nvisy-fake/src/generator/contact.rs b/crates/nvisy-fake/src/generator/contact.rs
deleted file mode 100644
index 34f2d6045..000000000
--- a/crates/nvisy-fake/src/generator/contact.rs
+++ /dev/null
@@ -1,67 +0,0 @@
-//! Free-form contact generators: street address composition, URL
-//! synthesis. The structured kinds (EmailAddress, PhoneNumber,
-//! PostalCode, Coordinates, LicensePlate) pattern-preserve their
-//! original and don't go through this module.
-
-use fake::faker::address::raw as address;
-use fake::faker::internet::raw as internet;
-use fake::rand::RngExt;
-
-use super::dispatch::fan_locale;
-use crate::locale::Locale;
-
-pub(super) fn street_address<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    let building: String = fan_locale!(locale, rng, address::BuildingNumber);
-    let street: String = fan_locale!(locale, rng, address::StreetName);
-    let city: String = fan_locale!(locale, rng, address::CityName);
-    match locale {
-        // CJK addresses go big-to-small (prefecture → ward → block →
-        // building) and don't concatenate street + building the way
-        // Latin-script ones do. This is "less wrong than English
-        // ordering," not a faithful rendering.
-        Locale::JaJp | Locale::ZhCn | Locale::ZhTw => format!("{city}{street}{building}"),
-        _ => format!("{building} {street}, {city}"),
-    }
-}
-
-pub(super) fn url<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    let user: String = fan_locale!(locale, rng, internet::Username);
-    let domain: String = fan_locale!(locale, rng, internet::DomainSuffix);
-    let host = sanitise_hostname_label(&user);
-    let host = if host.is_empty() {
-        "site"
-    } else {
-        host.as_str()
-    };
-    format!("https://www.{host}.{domain}")
-}
-
-/// Strip characters that aren't valid in a DNS label
-/// (RFC 1035: ASCII letters, digits, and hyphens), and trim leading
-/// or trailing hyphens. Returns lowercase output.
-fn sanitise_hostname_label(input: &str) -> String {
-    let mut out = String::with_capacity(input.len());
-    for ch in input.chars() {
-        if ch.is_ascii_alphanumeric() || ch == '-' {
-            out.push(ch.to_ascii_lowercase());
-        }
-    }
-    out.trim_matches('-').to_owned()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn strips_non_dns_characters() {
-        assert_eq!(sanitise_hostname_label("Ali_ce"), "alice");
-        assert_eq!(sanitise_hostname_label("Bob.Smith"), "bobsmith");
-        assert_eq!(sanitise_hostname_label("-mid-"), "mid");
-    }
-
-    #[test]
-    fn handles_empty_after_strip() {
-        assert!(sanitise_hostname_label("___").is_empty());
-    }
-}
diff --git a/crates/nvisy-fake/src/generator/device.rs b/crates/nvisy-fake/src/generator/device.rs
deleted file mode 100644
index 07a0fecf6..000000000
--- a/crates/nvisy-fake/src/generator/device.rs
+++ /dev/null
@@ -1,69 +0,0 @@
-//! Free-form device/credential generators: passwords, API tokens,
-//! device UUIDs. Structured kinds (IpAddress, MacAddress)
-//! pattern-preserve their original and don't go through this
-//! module.
-
-use fake::Fake;
-use fake::faker::internet::raw as internet;
-use fake::locales::{
-    AR_SA, CY_GB, DE_DE, EN, FA_IR, FR_FR, IT_IT, JA_JP, NL_NL, PT_BR, PT_PT, TR_TR, ZH_CN, ZH_TW,
-};
-use fake::rand::RngExt;
-use uuid::Uuid;
-
-use crate::locale::Locale;
-
-/// `internet::Password` takes both a locale and a length range, so
-/// it can't go through the `fan_locale!` macro (which assumes a
-/// one-arg faker constructor). Manual per-locale dispatch instead.
-pub(super) fn password<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    let range = 12..24;
-    match locale {
-        Locale::En => internet::Password(EN, range).fake_with_rng(rng),
-        Locale::FrFr => internet::Password(FR_FR, range).fake_with_rng(rng),
-        Locale::JaJp => internet::Password(JA_JP, range).fake_with_rng(rng),
-        Locale::ZhCn => internet::Password(ZH_CN, range).fake_with_rng(rng),
-        Locale::ZhTw => internet::Password(ZH_TW, range).fake_with_rng(rng),
-        Locale::DeDe => internet::Password(DE_DE, range).fake_with_rng(rng),
-        Locale::ItIt => internet::Password(IT_IT, range).fake_with_rng(rng),
-        Locale::PtBr => internet::Password(PT_BR, range).fake_with_rng(rng),
-        Locale::PtPt => internet::Password(PT_PT, range).fake_with_rng(rng),
-        Locale::NlNl => internet::Password(NL_NL, range).fake_with_rng(rng),
-        Locale::TrTr => internet::Password(TR_TR, range).fake_with_rng(rng),
-        Locale::ArSa => internet::Password(AR_SA, range).fake_with_rng(rng),
-        Locale::FaIr => internet::Password(FA_IR, range).fake_with_rng(rng),
-        Locale::CyGb => internet::Password(CY_GB, range).fake_with_rng(rng),
-    }
-}
-
-/// Random 32-char lowercase hex — looks like an API token without
-/// claiming any particular provider's shape.
-pub(super) fn api_key<R: RngExt + ?Sized>(rng: &mut R) -> String {
-    hex_chars(32, rng)
-}
-
-/// Random 48-char lowercase hex — slightly longer than `api_key`
-/// so the two are visually distinguishable.
-pub(super) fn auth_token<R: RngExt + ?Sized>(rng: &mut R) -> String {
-    hex_chars(48, rng)
-}
-
-/// UUIDv4 in canonical hex-with-hyphens form.
-pub(super) fn device_id<R: RngExt + ?Sized>(rng: &mut R) -> String {
-    let mut bytes = [0u8; 16];
-    for b in &mut bytes {
-        let n: u32 = (0..256u32).fake_with_rng(rng);
-        *b = n as u8;
-    }
-    Uuid::from_bytes(bytes).to_string()
-}
-
-fn hex_chars<R: RngExt + ?Sized>(len: usize, rng: &mut R) -> String {
-    const HEX: &[u8; 16] = b"0123456789abcdef";
-    let mut out = String::with_capacity(len);
-    for _ in 0..len {
-        let n: u32 = (0..16u32).fake_with_rng(rng);
-        out.push(HEX[n as usize] as char);
-    }
-    out
-}
diff --git a/crates/nvisy-fake/src/generator/dispatch.rs b/crates/nvisy-fake/src/generator/dispatch.rs
deleted file mode 100644
index c3f458e15..000000000
--- a/crates/nvisy-fake/src/generator/dispatch.rs
+++ /dev/null
@@ -1,45 +0,0 @@
-//! Locale-dispatch macro for fake-rs fakers.
-//!
-//! Each `fake` locale (`EN`, `FR_FR`, …) implements a different set
-//! of `*GenFn` trait gates per faker, so a generic function bounded
-//! on [`fake::locales::Data`] can't reach every faker. The
-//! [`fan_locale!`] macro fans a single faker constructor across all
-//! 14 [`Locale`] variants at the call site.
-//!
-//! The macro is self-contained — it imports `fake::Fake` and every
-//! locale constant it needs into the expansion's scope, so callers
-//! don't have to.
-//!
-//! [`Locale`]: crate::locale::Locale
-
-/// Invoke `$faker(locale)` for each [`Locale`] variant and call
-/// `.fake_with_rng($rng)`.
-///
-/// [`Locale`]: crate::locale::Locale
-macro_rules! fan_locale {
-    ($locale:expr, $rng:expr, $faker:expr) => {{
-        use ::fake::Fake as _;
-        use ::fake::locales::{
-            AR_SA, CY_GB, DE_DE, EN, FA_IR, FR_FR, IT_IT, JA_JP, NL_NL, PT_BR, PT_PT, TR_TR, ZH_CN,
-            ZH_TW,
-        };
-        match $locale {
-            $crate::locale::Locale::En => $faker(EN).fake_with_rng($rng),
-            $crate::locale::Locale::FrFr => $faker(FR_FR).fake_with_rng($rng),
-            $crate::locale::Locale::JaJp => $faker(JA_JP).fake_with_rng($rng),
-            $crate::locale::Locale::ZhCn => $faker(ZH_CN).fake_with_rng($rng),
-            $crate::locale::Locale::ZhTw => $faker(ZH_TW).fake_with_rng($rng),
-            $crate::locale::Locale::DeDe => $faker(DE_DE).fake_with_rng($rng),
-            $crate::locale::Locale::ItIt => $faker(IT_IT).fake_with_rng($rng),
-            $crate::locale::Locale::PtBr => $faker(PT_BR).fake_with_rng($rng),
-            $crate::locale::Locale::PtPt => $faker(PT_PT).fake_with_rng($rng),
-            $crate::locale::Locale::NlNl => $faker(NL_NL).fake_with_rng($rng),
-            $crate::locale::Locale::TrTr => $faker(TR_TR).fake_with_rng($rng),
-            $crate::locale::Locale::ArSa => $faker(AR_SA).fake_with_rng($rng),
-            $crate::locale::Locale::FaIr => $faker(FA_IR).fake_with_rng($rng),
-            $crate::locale::Locale::CyGb => $faker(CY_GB).fake_with_rng($rng),
-        }
-    }};
-}
-
-pub(crate) use fan_locale;
diff --git a/crates/nvisy-fake/src/generator/finance.rs b/crates/nvisy-fake/src/generator/finance.rs
deleted file mode 100644
index d13ad89ba..000000000
--- a/crates/nvisy-fake/src/generator/finance.rs
+++ /dev/null
@@ -1,57 +0,0 @@
-//! Free-form financial generators: currency code, monetary
-//! amounts, quantities. The structured kinds (IBAN, PaymentCard,
-//! BankAccount, BankRouting, SwiftCode, CardSecurityCode,
-//! CardExpiry) all pattern-preserve their original and don't go
-//! through this module.
-
-use fake::Fake;
-use fake::faker::currency::raw as currency;
-use fake::rand::RngExt;
-
-use super::dispatch::fan_locale;
-use crate::locale::Locale;
-
-pub(super) fn currency_code<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    fan_locale!(locale, rng, currency::CurrencyCode)
-}
-
-/// Locale-aware monetary amount. SI/UK locales use `1234.56`;
-/// continental EU locales use `1234,56`.
-pub(super) fn amount<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    let cents: u32 = (0..10_000_000u32).fake_with_rng(rng);
-    let whole = cents / 100;
-    let frac = cents % 100;
-    let sep = decimal_separator(locale);
-    format!("{whole}{sep}{frac:02}")
-}
-
-/// Plain integer in `0..=10_000`.
-pub(super) fn quantity<R: RngExt + ?Sized>(rng: &mut R) -> String {
-    let n: u32 = (0..=10_000u32).fake_with_rng(rng);
-    n.to_string()
-}
-
-fn decimal_separator(locale: Locale) -> char {
-    match locale {
-        Locale::DeDe | Locale::NlNl | Locale::FrFr | Locale::ItIt | Locale::PtPt | Locale::PtBr => {
-            ','
-        }
-        _ => '.',
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use fake::rand::SeedableRng;
-    use fake::rand::rngs::SmallRng;
-
-    use super::*;
-
-    #[test]
-    fn amount_uses_locale_decimal_separator() {
-        let mut rng = SmallRng::seed_from_u64(1);
-        assert!(amount(Locale::DeDe, &mut rng).contains(','));
-        let mut rng = SmallRng::seed_from_u64(1);
-        assert!(amount(Locale::En, &mut rng).contains('.'));
-    }
-}
diff --git a/crates/nvisy-fake/src/generator/identity.rs b/crates/nvisy-fake/src/generator/identity.rs
deleted file mode 100644
index 377dbe3b4..000000000
--- a/crates/nvisy-fake/src/generator/identity.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-//! People and categorical identifiers: names, organisations, jobs,
-//! and short categorical pick-from-list values (gender, language,
-//! nationality, citizenship).
-
-use fake::Fake;
-use fake::faker::address::raw as address;
-use fake::faker::company::raw as company;
-use fake::faker::internet::raw as internet;
-use fake::faker::job::raw as job;
-use fake::faker::name::raw as name;
-use fake::rand::RngExt;
-
-use super::dispatch::fan_locale;
-use crate::locale::Locale;
-
-pub(super) fn person_name<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    fan_locale!(locale, rng, name::Name)
-}
-
-pub(super) fn organization_name<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    fan_locale!(locale, rng, company::CompanyName)
-}
-
-pub(super) fn occupation<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    fan_locale!(locale, rng, job::Position)
-}
-
-pub(super) fn username<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    fan_locale!(locale, rng, internet::Username)
-}
-
-pub(super) fn gender<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    let options = gender_options(locale);
-    pick(options, rng).to_owned()
-}
-
-pub(super) fn language<R: RngExt + ?Sized>(rng: &mut R) -> String {
-    // BCP-47 tags are locale-invariant identifiers.
-    const TAGS: &[&str] = &[
-        "en", "fr", "de", "ja", "zh", "es", "it", "pt", "ar", "ru", "nl", "tr", "ko",
-    ];
-    pick(TAGS, rng).to_owned()
-}
-
-pub(super) fn nationality<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    fan_locale!(locale, rng, address::CountryName)
-}
-
-pub(super) fn citizenship<R: RngExt + ?Sized>(locale: Locale, rng: &mut R) -> String {
-    fan_locale!(locale, rng, address::CountryName)
-}
-
-/// Per-locale gender label sets. English is the fallback for
-/// locales without an explicit list.
-fn gender_options(locale: Locale) -> &'static [&'static str] {
-    match locale {
-        Locale::FrFr => &["féminin", "masculin", "non-binaire", "autre"],
-        Locale::DeDe => &["weiblich", "männlich", "nicht-binär", "andere"],
-        Locale::ItIt => &["femminile", "maschile", "non-binario", "altro"],
-        Locale::PtBr | Locale::PtPt => &["feminino", "masculino", "não-binário", "outro"],
-        Locale::NlNl => &["vrouwelijk", "mannelijk", "non-binair", "anders"],
-        Locale::JaJp => &["女性", "男性", "その他"],
-        Locale::ZhCn | Locale::ZhTw => &["女性", "男性", "其他"],
-        _ => &["female", "male", "non-binary", "other", "prefer not to say"],
-    }
-}
-
-fn pick<'a, R: RngExt + ?Sized>(options: &'a [&'a str], rng: &mut R) -> &'a str {
-    let i: usize = (0..options.len()).fake_with_rng(rng);
-    options[i]
-}
diff --git a/crates/nvisy-fake/src/generator/mod.rs b/crates/nvisy-fake/src/generator/mod.rs
deleted file mode 100644
index 25948e503..000000000
--- a/crates/nvisy-fake/src/generator/mod.rs
+++ /dev/null
@@ -1,187 +0,0 @@
-//! Per-label fake-value generation, dispatched by [`Locale`].
-//!
-//! [`Context::generate`] returns `Some(string)` for every label the
-//! catalogue covers, or `None` for labels the fake-data layer
-//! doesn't support — the caller delegates to its fallback
-//! anonymizer in that case.
-//!
-//! Two paths:
-//!
-//! - **Structured labels** (IBAN, payment cards, dates, IPs, …)
-//!   pattern-preserve the original string: same length, same
-//!   character-class layout, randomised digits and letters.
-//!   See [`pattern::pattern_preserve`].
-//! - **Free-form labels** (names, addresses, organisations, …)
-//!   emit a fresh locale-aware fake whose length doesn't need to
-//!   match. These go through per-domain submodules.
-
-mod case_id;
-mod contact;
-mod device;
-mod dispatch;
-mod finance;
-mod identity;
-mod pattern;
-mod temporal;
-
-use fake::Fake;
-use fake::faker::number::raw as number;
-use fake::locales::EN;
-use fake::rand::RngExt;
-
-use crate::locale::Locale;
-
-/// Per-call options threaded through to each label generator.
-pub(crate) struct Context<'a> {
-    locale: Locale,
-    label: &'a str,
-    original: &'a str,
-}
-
-impl<'a> Context<'a> {
-    /// Build a generation request.
-    pub(crate) fn new(locale: Locale, label: &'a str, original: &'a str) -> Self {
-        Self {
-            locale,
-            label,
-            original,
-        }
-    }
-
-    /// Generate a fake replacement string for this context, using
-    /// `rng` as the entropy source. Returns `None` when the entity
-    /// label isn't covered.
-    ///
-    /// Two paths:
-    /// - **Structured** labels reshape the original string in place
-    ///   via [`pattern::pattern_preserve`]; they return `None` when
-    ///   `original` is empty since there's no pattern to copy.
-    /// - **Free-form** labels emit a fresh locale-aware fake whose
-    ///   length doesn't need to match `original`.
-    pub(crate) fn generate<R: RngExt + ?Sized>(self, rng: &mut R) -> Option<String> {
-        let l = self.locale;
-        let preserve = |rng: &mut R| {
-            (!self.original.is_empty()).then(|| pattern::pattern_preserve(self.original, rng))
-        };
-        let value = match self.label {
-            // identity (free-form)
-            "person_name" => identity::person_name(l, rng),
-            "organization_name" => identity::organization_name(l, rng),
-            "occupation" => identity::occupation(l, rng),
-            "username" => identity::username(l, rng),
-            "gender" => identity::gender(l, rng),
-            "language" => identity::language(rng),
-            "nationality" => identity::nationality(l, rng),
-            "citizenship" => identity::citizenship(l, rng),
-
-            // contact
-            "address" => contact::street_address(l, rng),
-            "url" => contact::url(l, rng),
-            "email_address" | "phone_number" | "postal_code" => return preserve(rng),
-
-            // temporal
-            "age" => temporal::age(rng),
-            "date_of_birth" | "date_time" => return preserve(rng),
-
-            // finance (free-form subset)
-            "currency" => finance::currency_code(l, rng),
-            "amount" => finance::amount(l, rng),
-            "quantity" => finance::quantity(rng),
-
-            // finance (structured)
-            "iban" | "payment_card" | "card_security_code" | "card_expiry" | "bank_account"
-            | "bank_routing" | "swift_code" | "crypto_address" => return preserve(rng),
-
-            // device (free-form tokens)
-            "password" => device::password(l, rng),
-            "api_key" => device::api_key(rng),
-            "auth_token" => device::auth_token(rng),
-            "device_id" => device::device_id(rng),
-
-            // device (structured)
-            "ip_address" | "mac_address" | "coordinates" => return preserve(rng),
-
-            // case ids (free-form)
-            "internal_id" | "case_number" => case_id::internal_id(rng),
-
-            // ids (structured)
-            "government_id"
-            | "tax_id"
-            | "drivers_license"
-            | "passport_number"
-            | "national_insurance_number"
-            | "vehicle_id"
-            | "license_plate"
-            | "medical_id"
-            | "insurance_id"
-            | "prescription_id" => return preserve(rng),
-
-            _ => return None,
-        };
-        Some(value)
-    }
-}
-
-/// Shared helper for labels that synthesise digit groups outside
-/// the fake-rs locale tables (bank account, IDs).
-pub(crate) fn digits<R: RngExt + ?Sized>(len: usize, rng: &mut R) -> String {
-    let fmt = "#".repeat(len);
-    number::NumberWithFormat(EN, fmt.as_str()).fake_with_rng(rng)
-}
-
-#[cfg(test)]
-mod tests {
-    use fake::rand::SeedableRng;
-    use fake::rand::rngs::SmallRng;
-
-    use super::*;
-
-    fn rng() -> SmallRng {
-        SmallRng::seed_from_u64(7)
-    }
-
-    fn ctx<'a>(locale: Locale, label: &'a str, original: &'a str) -> Context<'a> {
-        Context::new(locale, label, original)
-    }
-
-    #[test]
-    fn unsupported_labels_return_none() {
-        let mut rng = rng();
-        for label in ["fingerprint", "face", "religion", "diagnosis"] {
-            assert!(
-                ctx(Locale::En, label, "").generate(&mut rng).is_none(),
-                "{label} should be None"
-            );
-        }
-    }
-
-    #[test]
-    fn structured_label_with_empty_source_returns_none() {
-        let mut rng = rng();
-        // No pattern to copy → can't pattern-preserve.
-        assert!(ctx(Locale::En, "iban", "").generate(&mut rng).is_none());
-    }
-
-    #[test]
-    fn structured_labels_preserve_original_shape() {
-        let cases: &[(&str, &str)] = &[
-            ("iban", "GB82WEST12345698765432"),
-            ("payment_card", "4111-1111-1111-1111"),
-            ("phone_number", "+1-555-123-4567"),
-            ("date_of_birth", "1985-03-12"),
-            ("ip_address", "192.168.1.1"),
-            ("postal_code", "SW1A 1AA"),
-        ];
-        for &(label, original) in cases {
-            let mut rng = rng();
-            let out = ctx(Locale::En, label, original).generate(&mut rng).unwrap();
-            assert_eq!(out.len(), original.len(), "{label}: length mismatch");
-            // Separator positions match.
-            for (i, (a, b)) in out.chars().zip(original.chars()).enumerate() {
-                if !a.is_ascii_alphanumeric() {
-                    assert_eq!(a, b, "{label}: separator mismatch at {i} ({a:?} vs {b:?})");
-                }
-            }
-        }
-    }
-}
diff --git a/crates/nvisy-fake/src/generator/pattern.rs b/crates/nvisy-fake/src/generator/pattern.rs
deleted file mode 100644
index 7d773d79f..000000000
--- a/crates/nvisy-fake/src/generator/pattern.rs
+++ /dev/null
@@ -1,98 +0,0 @@
-//! Pattern-preserving reshape: emit a string whose length and
-//! character-class layout matches `original` exactly. Digits in
-//! `original` map to fresh random digits in the output; letters
-//! map to fresh random letters of matching case; everything else
-//! is copied verbatim.
-//!
-//! Used by structured kinds (IBAN, payment cards, postal codes,
-//! phone numbers, dates) where the shape carries meaning. Free-form
-//! kinds (names, addresses) don't go through this — their output
-//! comes straight from the locale-aware generator.
-
-use fake::Fake;
-use fake::rand::RngExt;
-
-/// Walk `original` left-to-right and emit a same-length output:
-/// - digit `→` random ASCII digit `0..=9`
-/// - uppercase ASCII letter `→` random uppercase `A..=Z`
-/// - lowercase ASCII letter `→` random lowercase `a..=z`
-/// - other code-point (separators, multibyte, punctuation) `→`
-///   copied unchanged
-pub(crate) fn pattern_preserve<R: RngExt + ?Sized>(original: &str, rng: &mut R) -> String {
-    const DIGITS: &[u8; 10] = b"0123456789";
-    const UPPER: &[u8; 26] = b"ABCDEFGHIJKLMNOPQRSTUVWXYZ";
-    const LOWER: &[u8; 26] = b"abcdefghijklmnopqrstuvwxyz";
-
-    let mut out = String::with_capacity(original.len());
-    for ch in original.chars() {
-        if ch.is_ascii_digit() {
-            let i: usize = (0..DIGITS.len()).fake_with_rng(rng);
-            out.push(DIGITS[i] as char);
-        } else if ch.is_ascii_uppercase() {
-            let i: usize = (0..UPPER.len()).fake_with_rng(rng);
-            out.push(UPPER[i] as char);
-        } else if ch.is_ascii_lowercase() {
-            let i: usize = (0..LOWER.len()).fake_with_rng(rng);
-            out.push(LOWER[i] as char);
-        } else {
-            out.push(ch);
-        }
-    }
-    out
-}
-
-#[cfg(test)]
-mod tests {
-    use fake::rand::SeedableRng;
-    use fake::rand::rngs::SmallRng;
-
-    use super::*;
-
-    fn rng() -> SmallRng {
-        SmallRng::seed_from_u64(7)
-    }
-
-    #[test]
-    fn preserves_iban_shape() {
-        let out = pattern_preserve("GB82WEST12345698765432", &mut rng());
-        assert_eq!(out.len(), 22);
-        // Char-class layout: 2 upper + 2 digit + 4 upper + 14 digit
-        assert!(out.as_bytes()[0].is_ascii_uppercase());
-        assert!(out.as_bytes()[1].is_ascii_uppercase());
-        assert!(out.as_bytes()[2].is_ascii_digit());
-        assert!(out.as_bytes()[3].is_ascii_digit());
-        for i in 4..8 {
-            assert!(out.as_bytes()[i].is_ascii_uppercase(), "pos {i}");
-        }
-        for i in 8..22 {
-            assert!(out.as_bytes()[i].is_ascii_digit(), "pos {i}");
-        }
-    }
-
-    #[test]
-    fn preserves_phone_separators() {
-        let out = pattern_preserve("555-123-4567", &mut rng());
-        assert_eq!(out.len(), 12);
-        assert_eq!(&out[3..4], "-");
-        assert_eq!(&out[7..8], "-");
-    }
-
-    #[test]
-    fn preserves_cjk_date_separators() {
-        let out = pattern_preserve("1985年3月12日", &mut rng());
-        assert!(out.contains('年') && out.contains('月') && out.contains('日'));
-    }
-
-    #[test]
-    fn preserves_us_date_format() {
-        let out = pattern_preserve("03/14/2024", &mut rng());
-        assert_eq!(out.len(), 10);
-        assert_eq!(&out[2..3], "/");
-        assert_eq!(&out[5..6], "/");
-    }
-
-    #[test]
-    fn empty_input_yields_empty_output() {
-        assert_eq!(pattern_preserve("", &mut rng()), "");
-    }
-}
diff --git a/crates/nvisy-fake/src/generator/temporal.rs b/crates/nvisy-fake/src/generator/temporal.rs
deleted file mode 100644
index 985a4867e..000000000
--- a/crates/nvisy-fake/src/generator/temporal.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-//! Age generator. Date-of-birth and date-time pattern-preserve
-//! their original and don't go through this module.
-
-use fake::Fake;
-use fake::rand::RngExt;
-
-pub(super) fn age<R: RngExt + ?Sized>(rng: &mut R) -> String {
-    let years: u8 = (1..=99u8).fake_with_rng(rng);
-    years.to_string()
-}
diff --git a/crates/nvisy-fake/src/lib.rs b/crates/nvisy-fake/src/lib.rs
deleted file mode 100644
index 0b4465ae1..000000000
--- a/crates/nvisy-fake/src/lib.rs
+++ /dev/null
@@ -1,9 +0,0 @@
-#![forbid(unsafe_code)]
-#![cfg_attr(docsrs, feature(doc_cfg))]
-#![doc = include_str!("../README.md")]
-
-mod anonymizer;
-mod generator;
-mod locale;
-
-pub use self::anonymizer::Fake;
diff --git a/crates/nvisy-fake/src/locale/mod.rs b/crates/nvisy-fake/src/locale/mod.rs
deleted file mode 100644
index d4e488026..000000000
--- a/crates/nvisy-fake/src/locale/mod.rs
+++ /dev/null
@@ -1,96 +0,0 @@
-//! Internal locale picker: maps BCP-47 tags to `fake` crate locale
-//! variants.
-
-use nvisy_core::primitive::LanguageTag;
-
-/// Locale selector for the fake-data generator. Maps to one of the
-/// `fake` crate's locale modules. Internal — callers express locale
-/// via [`LanguageTag`].
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default)]
-pub(crate) enum Locale {
-    #[default]
-    En,
-    FrFr,
-    JaJp,
-    ZhCn,
-    ZhTw,
-    DeDe,
-    ItIt,
-    PtBr,
-    PtPt,
-    NlNl,
-    TrTr,
-    ArSa,
-    FaIr,
-    CyGb,
-}
-
-impl Locale {
-    /// Map a BCP-47 language tag to the closest supported locale.
-    /// Unknown primary languages fall back to [`Locale::En`].
-    pub(crate) fn from_tag(tag: &LanguageTag) -> Self {
-        let primary = tag.primary_language().to_ascii_lowercase();
-        let region = region_subtag(tag.as_str()).map(|s| s.to_ascii_uppercase());
-
-        match (primary.as_str(), region.as_deref()) {
-            ("en", _) => Self::En,
-            ("fr", _) => Self::FrFr,
-            ("ja", _) => Self::JaJp,
-            ("zh", Some("TW")) => Self::ZhTw,
-            ("zh", _) => Self::ZhCn,
-            ("de", _) => Self::DeDe,
-            ("it", _) => Self::ItIt,
-            ("pt", Some("BR")) => Self::PtBr,
-            ("pt", _) => Self::PtPt,
-            ("nl", _) => Self::NlNl,
-            ("tr", _) => Self::TrTr,
-            ("ar", _) => Self::ArSa,
-            ("fa", _) => Self::FaIr,
-            ("cy", _) => Self::CyGb,
-            _ => Self::En,
-        }
-    }
-}
-
-/// Pull the BCP-47 region subtag out of a tag like `zh-TW` or
-/// `pt-BR`. Returns `None` if the tag has no region.
-fn region_subtag(tag: &str) -> Option<&str> {
-    tag.split('-')
-        .nth(1)
-        .filter(|part| part.len() == 2 && part.chars().all(|c| c.is_ascii_alphabetic()))
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn tag(s: &str) -> LanguageTag {
-        s.parse().expect("valid bcp-47")
-    }
-
-    #[test]
-    fn english_variants_map_to_en() {
-        assert_eq!(Locale::from_tag(&tag("en")), Locale::En);
-        assert_eq!(Locale::from_tag(&tag("en-US")), Locale::En);
-        assert_eq!(Locale::from_tag(&tag("en-GB")), Locale::En);
-    }
-
-    #[test]
-    fn chinese_region_splits_simplified_traditional() {
-        assert_eq!(Locale::from_tag(&tag("zh-CN")), Locale::ZhCn);
-        assert_eq!(Locale::from_tag(&tag("zh-TW")), Locale::ZhTw);
-        assert_eq!(Locale::from_tag(&tag("zh")), Locale::ZhCn);
-    }
-
-    #[test]
-    fn portuguese_region_splits_brazil_portugal() {
-        assert_eq!(Locale::from_tag(&tag("pt-BR")), Locale::PtBr);
-        assert_eq!(Locale::from_tag(&tag("pt-PT")), Locale::PtPt);
-        assert_eq!(Locale::from_tag(&tag("pt")), Locale::PtPt);
-    }
-
-    #[test]
-    fn unknown_falls_back_to_english() {
-        assert_eq!(Locale::from_tag(&tag("kl")), Locale::En);
-    }
-}
diff --git a/crates/nvisy-llm/Cargo.toml b/crates/nvisy-llm/Cargo.toml
deleted file mode 100644
index 6e12f04f3..000000000
--- a/crates/nvisy-llm/Cargo.toml
+++ /dev/null
@@ -1,76 +0,0 @@
-# https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[package]
-name = "nvisy-llm"
-description = "LLM-mediated entity recognizer (text NER + image VLM) for the Nvisy runtime"
-keywords = ["nvisy", "rig", "llm", "ner"]
-categories = ["text-processing", "api-bindings"]
-readme = "README.md"
-
-version = { workspace = true }
-rust-version = { workspace = true }
-edition = { workspace = true }
-license = { workspace = true }
-publish = { workspace = true }
-
-authors = { workspace = true }
-repository = { workspace = true }
-homepage = { workspace = true }
-documentation = { workspace = true }
-
-[features]
-default = []
-
-## Enable OpenAI GPT completion provider.
-openai-gpt = []
-## Enable Anthropic Claude completion provider.
-anthropic-claude = []
-## Enable Google Gemini completion provider.
-google-gemini = []
-
-[package.metadata.docs.rs]
-all-features = true
-rustdoc-args = ["--cfg", "docsrs"]
-
-[dependencies]
-# Internal crates
-nvisy-core = { workspace = true, features = [] }
-
-# Serialization
-serde = { workspace = true, features = [] }
-serde_json = { workspace = true, features = [] }
-schemars = { workspace = true, features = [] }
-toml = { workspace = true, features = ["parse"] }
-minijinja = { workspace = true, features = [] }
-humantime-serde = { workspace = true, features = [] }
-
-# Derive macros and error handling
-derive_builder = { workspace = true, features = [] }
-derive_more = { workspace = true, features = ["add_assign"] }
-thiserror = { workspace = true, features = [] }
-
-# Encoding and hashing
-base64 = { workspace = true, features = [] }
-
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-
-# Observability
-tracing = { workspace = true, features = [] }
-
-# Text processing
-unicode-normalization = { workspace = true, features = [] }
-
-# AI / LLM frameworks
-rig = { workspace = true, features = ["derive", "reqwest-middleware"] }
-
-# HTTP client and middleware (shared retry/tracing layers)
-reqwest-middleware = { workspace = true, features = [] }
-reqwest-retry = { workspace = true, features = [] }
-reqwest-tracing = { workspace = true, features = [] }
-
-[dev-dependencies]
-tokio = { workspace = true, features = ["macros", "rt"] }
-
-[package.metadata.cargo-machete]
-ignored = ["humantime-serde"]
diff --git a/crates/nvisy-llm/README.md b/crates/nvisy-llm/README.md
deleted file mode 100644
index b51ba8ed2..000000000
--- a/crates/nvisy-llm/README.md
+++ /dev/null
@@ -1,41 +0,0 @@
-# nvisy-llm
-
-[![Build](https://img.shields.io/github/actions/workflow/status/nvisycom/runtime/build.yml?branch=main&label=build%20%26%20test&style=flat-square)](https://github.com/nvisycom/runtime/actions/workflows/build.yml)
-
-LLM-mediated entity recognition for the Nvisy runtime, built on the
-[`rig`](https://github.com/0xPlaygrounds/rig) framework.
-
-## Overview
-
-`LlmRecognizer` implements `EntityRecognizer<M>` against a
-caller-supplied `Prompt` (the `FilePrompt` helper loads templates
-from disk); pair it with any provider in `provider::*` and plug it
-into a `RecognizerRegistry`. The crate doesn't impose a fixed agent
-shape — the prompt + provider combination decides what the LLM is
-asked to do (NER, VLM, verification).
-
-Provider features (`openai-gpt`, `anthropic-claude`, `google-gemini`)
-are independently selectable; none are on by default — the CLI/server
-entry points opt in.
-
-Speech-to-text lives in [`nvisy-stt`](../nvisy-stt) — it ships with
-its own backend trait and segment-shaped output type, decoupled from
-the LLM family entirely.
-
-## Documentation
-
-See [`docs/`](../../docs/) for architecture, security, and API documentation.
-
-## Changelog
-
-See [CHANGELOG.md](../../CHANGELOG.md) for release notes and version history.
-
-## License
-
-Apache 2.0 License, see [LICENSE.txt](../../LICENSE.txt)
-
-## Support
-
-- **Documentation**: [docs.nvisy.com](https://docs.nvisy.com)
-- **Issues**: [GitHub Issues](https://github.com/nvisycom/runtime/issues)
-- **Email**: [support@nvisy.com](mailto:support@nvisy.com)
diff --git a/crates/nvisy-llm/src/backend/http/config.rs b/crates/nvisy-llm/src/backend/http/config.rs
deleted file mode 100644
index 5ca728ce6..000000000
--- a/crates/nvisy-llm/src/backend/http/config.rs
+++ /dev/null
@@ -1,72 +0,0 @@
-//! HTTP client configuration.
-
-use std::time::Duration;
-
-use serde::{Deserialize, Serialize};
-
-/// Default request timeout (2 minutes).
-const DEFAULT_TIMEOUT: Duration = Duration::from_secs(120);
-/// Default connection timeout (10 seconds).
-const DEFAULT_CONNECT_TIMEOUT: Duration = Duration::from_secs(10);
-/// Default maximum number of retries.
-const DEFAULT_MAX_RETRIES: u32 = 3;
-/// Default keep-alive idle timeout (90 seconds).
-const DEFAULT_IDLE_TIMEOUT: Duration = Duration::from_secs(90);
-
-/// Configuration for the shared HTTP client.
-///
-/// Durations accept any human-readable form (`"120s"`, `"2min"`,
-/// `"500ms"`) via [`humantime_serde`].
-#[derive(Debug, Clone, Serialize, Deserialize)]
-#[serde(rename_all = "camelCase")]
-pub struct HttpConfig {
-    /// Maximum number of retries for transient failures (default: 3).
-    #[serde(default = "default_max_retries", alias = "max_retries")]
-    pub max_retries: u32,
-    /// Per-request timeout (default: 2m).
-    #[serde(
-        default = "default_timeout",
-        with = "humantime_serde",
-        alias = "timeout"
-    )]
-    pub timeout: Duration,
-    /// TCP connection timeout (default: 10s).
-    #[serde(
-        default = "default_connect_timeout",
-        with = "humantime_serde",
-        alias = "connect_timeout"
-    )]
-    pub connect_timeout: Duration,
-    /// Keep-alive pool idle timeout (default: 90s).
-    #[serde(
-        default = "default_idle_timeout",
-        with = "humantime_serde",
-        alias = "idle_timeout",
-        alias = "pool_idle_timeout"
-    )]
-    pub idle_timeout: Duration,
-}
-
-fn default_max_retries() -> u32 {
-    DEFAULT_MAX_RETRIES
-}
-fn default_timeout() -> Duration {
-    DEFAULT_TIMEOUT
-}
-fn default_connect_timeout() -> Duration {
-    DEFAULT_CONNECT_TIMEOUT
-}
-fn default_idle_timeout() -> Duration {
-    DEFAULT_IDLE_TIMEOUT
-}
-
-impl Default for HttpConfig {
-    fn default() -> Self {
-        Self {
-            max_retries: DEFAULT_MAX_RETRIES,
-            timeout: DEFAULT_TIMEOUT,
-            connect_timeout: DEFAULT_CONNECT_TIMEOUT,
-            idle_timeout: DEFAULT_IDLE_TIMEOUT,
-        }
-    }
-}
diff --git a/crates/nvisy-llm/src/backend/http/middleware.rs b/crates/nvisy-llm/src/backend/http/middleware.rs
deleted file mode 100644
index e981fd5db..000000000
--- a/crates/nvisy-llm/src/backend/http/middleware.rs
+++ /dev/null
@@ -1,23 +0,0 @@
-//! HTTP middleware layers: exponential-backoff retry +
-//! OpenTelemetry tracing.
-
-use reqwest_retry::RetryTransientMiddleware;
-use reqwest_retry::policies::ExponentialBackoff;
-use reqwest_tracing::{DefaultSpanBackend, TracingMiddleware};
-
-/// Build an exponential-backoff policy with the given maximum retries.
-pub(super) fn backoff_policy(max_retries: u32) -> ExponentialBackoff {
-    ExponentialBackoff::builder().build_with_max_retries(max_retries)
-}
-
-/// Wrap a backoff policy into the retry middleware layer.
-pub(super) fn retry_layer(
-    policy: ExponentialBackoff,
-) -> RetryTransientMiddleware<ExponentialBackoff> {
-    RetryTransientMiddleware::new_with_policy(policy)
-}
-
-/// Create the default tracing middleware layer.
-pub(super) fn tracing_layer() -> TracingMiddleware<DefaultSpanBackend> {
-    TracingMiddleware::default()
-}
diff --git a/crates/nvisy-llm/src/backend/http/mod.rs b/crates/nvisy-llm/src/backend/http/mod.rs
deleted file mode 100644
index 3be58cc9e..000000000
--- a/crates/nvisy-llm/src/backend/http/mod.rs
+++ /dev/null
@@ -1,60 +0,0 @@
-//! Shared HTTP client factory (reqwest + retry + tracing middleware).
-//!
-//! [`build_http_client`] builds a [`ClientWithMiddleware`] from an
-//! [`HttpConfig`] with exponential-backoff retry and OpenTelemetry
-//! tracing layers pre-installed. Callers hand the returned client to
-//! whatever HTTP-aware library they're driving (rig provider, STT
-//! provider, raw `reqwest_middleware` requests). [`HttpConfig`]
-//! deserialises durations via [`humantime_serde`] so config files
-//! accept `"120s"`, `"2min"`, etc.
-//!
-//! [`ClientWithMiddleware`]: reqwest_middleware::ClientWithMiddleware
-//! [`humantime_serde`]: https://crates.io/crates/humantime-serde
-
-mod config;
-mod middleware;
-
-use nvisy_core::{Error, Result};
-use reqwest_middleware::reqwest::Client;
-use reqwest_middleware::{ClientBuilder, ClientWithMiddleware};
-
-pub use self::config::HttpConfig;
-use self::middleware::{backoff_policy, retry_layer, tracing_layer};
-
-const TARGET: &str = "nvisy_llm::http";
-
-/// Build a [`ClientWithMiddleware`] from the given configuration.
-///
-/// The returned client has exponential-backoff retry and OpenTelemetry
-/// tracing middleware pre-installed; callers wire it into rig
-/// providers, bento clients, or raw `reqwest_middleware` request
-/// builders.
-///
-/// # Errors
-///
-/// Returns an error if the underlying `reqwest::Client` cannot be
-/// built (e.g. TLS backend initialisation failure).
-pub fn build_http_client(config: &HttpConfig) -> Result<ClientWithMiddleware> {
-    tracing::debug!(
-        target: TARGET,
-        max_retries = config.max_retries,
-        timeout = ?config.timeout,
-        connect_timeout = ?config.connect_timeout,
-        idle_timeout = ?config.idle_timeout,
-        "building HTTP client"
-    );
-
-    let policy = backoff_policy(config.max_retries);
-
-    let client = Client::builder()
-        .timeout(config.timeout)
-        .connect_timeout(config.connect_timeout)
-        .pool_idle_timeout(config.idle_timeout)
-        .build()
-        .map_err(|e| Error::runtime(format!("failed to build HTTP client: {e}"), "http", false))?;
-
-    Ok(ClientBuilder::new(client)
-        .with(tracing_layer())
-        .with(retry_layer(policy))
-        .build())
-}
diff --git a/crates/nvisy-llm/src/backend/mod.rs b/crates/nvisy-llm/src/backend/mod.rs
deleted file mode 100644
index 327a5ddbc..000000000
--- a/crates/nvisy-llm/src/backend/mod.rs
+++ /dev/null
@@ -1,42 +0,0 @@
-//! Backend layer: the modality-agnostic [`LlmBackend`] trait that
-//! turns a prompt + schema into the model's reply, plus its shipped
-//! impls.
-//!
-//! Modality-specific work (prompt construction, response → entity
-//! lifting) lives in [`crate::LlmRecognizer`]; backends only handle
-//! provider dispatch, structured-output, retries, and usage
-//! tracking.
-
-pub mod http;
-mod request;
-mod response;
-pub mod rig;
-
-use nvisy_core::Result;
-
-pub use self::request::LlmRequest;
-pub use self::response::LlmResponse;
-
-/// Per-call LLM backend.
-///
-/// Implemented by everything that turns a `(prompt, schema)` pair
-/// into the model's text reply — rig-backed providers (OpenAI,
-/// Anthropic, Gemini, Ollama), externalised inference gateways, the
-/// in-process no-op test stub.
-///
-/// Object-safe: recognizers hold `Arc<dyn LlmBackend>` and dispatch
-/// per call.
-#[async_trait::async_trait]
-pub trait LlmBackend: Send + Sync + 'static {
-    /// Send `request` to the model and return its reply.
-    ///
-    /// # Errors
-    ///
-    /// Returns the underlying transport / provider / parse error.
-    async fn predict(&self, request: LlmRequest<'_>) -> Result<LlmResponse>;
-
-    /// Model name the backend is configured to call. Recognizers
-    /// stamp this into entity trail provenance so post-hoc analysis
-    /// can attribute scores to a specific model.
-    fn model(&self) -> &str;
-}
diff --git a/crates/nvisy-llm/src/backend/request.rs b/crates/nvisy-llm/src/backend/request.rs
deleted file mode 100644
index 4d98ea712..000000000
--- a/crates/nvisy-llm/src/backend/request.rs
+++ /dev/null
@@ -1,20 +0,0 @@
-//! [`LlmRequest`]: per-call input to an [`LlmBackend`].
-//!
-//! [`LlmBackend`]: super::LlmBackend
-
-/// One per-call LLM request handed to a [`LlmBackend`].
-///
-/// [`LlmBackend`]: super::LlmBackend
-#[derive(Debug, Clone)]
-pub struct LlmRequest<'a> {
-    /// Fully-rendered user prompt. The recognizer is responsible for
-    /// folding the source text, hints, labels, and any base64-encoded
-    /// binary payloads (images, audio) into this string.
-    pub prompt: &'a str,
-    /// Optional JSON schema the backend asks the model to constrain
-    /// output against. Backends that support structured output
-    /// (rig's `output_schema`) use it; backends that don't, ignore
-    /// it. `None` means the recognizer's prompt is responsible for
-    /// describing the expected output shape inline.
-    pub schema: Option<&'a schemars::Schema>,
-}
diff --git a/crates/nvisy-llm/src/backend/response.rs b/crates/nvisy-llm/src/backend/response.rs
deleted file mode 100644
index 1c62bbc3b..000000000
--- a/crates/nvisy-llm/src/backend/response.rs
+++ /dev/null
@@ -1,26 +0,0 @@
-//! [`LlmResponse`]: per-call output from an [`LlmBackend`].
-//!
-//! [`LlmBackend`]: super::LlmBackend
-
-/// One per-call LLM response from a [`LlmBackend`].
-///
-/// Wraps the model's reply text verbatim. The recognizer
-/// deserialises it (and applies whatever markdown-fence /
-/// sentinel-text forgiveness its [`Prompt<M>`] needs) on the way
-/// out.
-///
-/// [`LlmBackend`]: super::LlmBackend
-/// [`Prompt<M>`]: crate::Prompt
-#[derive(Debug, Clone, Default)]
-pub struct LlmResponse {
-    /// Model's reply text, verbatim.
-    pub text: String,
-}
-
-impl LlmResponse {
-    /// Construct a response from raw text.
-    #[must_use]
-    pub fn new(text: impl Into<String>) -> Self {
-        Self { text: text.into() }
-    }
-}
diff --git a/crates/nvisy-llm/src/backend/rig/config.rs b/crates/nvisy-llm/src/backend/rig/config.rs
deleted file mode 100644
index 52d44458e..000000000
--- a/crates/nvisy-llm/src/backend/rig/config.rs
+++ /dev/null
@@ -1,54 +0,0 @@
-//! [`LlmConfig`]: sampling, retry, context-window, and preamble
-//! settings for a [`RigBackend`].
-//!
-//! [`RigBackend`]: super::RigBackend
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::context::ContextWindow;
-
-/// Sampling, retry, context-window, and preamble settings for a
-/// [`RigBackend`].
-///
-/// [`RigBackend`]: super::RigBackend
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-pub struct LlmConfig {
-    /// Sampling temperature (default: 0.1).
-    pub temperature: f64,
-    /// Maximum output tokens (default: 4096).
-    pub max_tokens: u64,
-    /// Maximum retries for transient HTTP errors (default: 3).
-    pub max_retries: u32,
-    /// Context window for chunking large inputs. When set, prompts
-    /// exceeding the input budget are summarised via an extra LLM
-    /// call before the real call. See [`compact`].
-    ///
-    /// [`compact`]: Self::compact
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub context_window: Option<ContextWindow>,
-    /// Whether to silently compact over-budget prompts. Only fires
-    /// when `context_window` is also set. Defaults to `true`.
-    #[serde(default = "default_true")]
-    pub compact: bool,
-    /// System prompt prepended to every request.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub preamble: Option<String>,
-}
-
-impl Default for LlmConfig {
-    fn default() -> Self {
-        Self {
-            temperature: 0.1,
-            max_tokens: 4096,
-            max_retries: 3,
-            context_window: None,
-            compact: true,
-            preamble: None,
-        }
-    }
-}
-
-fn default_true() -> bool {
-    true
-}
diff --git a/crates/nvisy-llm/src/backend/rig/context.rs b/crates/nvisy-llm/src/backend/rig/context.rs
deleted file mode 100644
index ae5dc69ad..000000000
--- a/crates/nvisy-llm/src/backend/rig/context.rs
+++ /dev/null
@@ -1,74 +0,0 @@
-//! Token budget estimation and prompt compaction trigger.
-//!
-//! [`ContextWindow`] provides a script-aware heuristic (~4 chars/token
-//! for Latin, ~1 token per CJK/emoji character) to decide whether
-//! text fits within a model's input budget. When it doesn't, the
-//! [`RigBackend`] sends an extra LLM call to summarise the prompt
-//! before the real call.
-//!
-//! [`RigBackend`]: super::RigBackend
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Token budget manager for a single model context window.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-pub struct ContextWindow {
-    /// Maximum tokens the model supports.
-    max_tokens: usize,
-    /// Tokens reserved for the output/completion.
-    reserve_output: usize,
-}
-
-impl ContextWindow {
-    /// Build a context window from the model's maximum token budget
-    /// and the share reserved for the completion.
-    pub fn new(max_tokens: usize, reserve_output: usize) -> Self {
-        Self {
-            max_tokens,
-            reserve_output,
-        }
-    }
-
-    /// Rough token count.
-    ///
-    /// Uses character count (not byte length) so that CJK, Arabic,
-    /// emoji, and other multi-byte scripts are estimated more
-    /// accurately. The heuristic assumes ~4 characters per token for
-    /// Latin text; CJK and other ideographic characters are counted
-    /// as ~1 token each.
-    pub fn estimate_tokens(text: &str) -> usize {
-        let mut latin_chars = 0usize;
-        let mut wide_chars = 0usize;
-        for ch in text.chars() {
-            if ch.is_ascii() || ch.len_utf8() <= 2 {
-                latin_chars += 1;
-            } else {
-                wide_chars += 1;
-            }
-        }
-        latin_chars.div_ceil(4) + wide_chars
-    }
-
-    /// Input token budget (`max_tokens − reserve_output`).
-    pub(crate) fn input_budget(&self) -> usize {
-        self.max_tokens.saturating_sub(self.reserve_output)
-    }
-
-    /// Whether `text` fits within the input budget.
-    pub fn fits(&self, text: &str) -> bool {
-        Self::estimate_tokens(text) <= self.input_budget()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn fits_within_budget() {
-        let cw = ContextWindow::new(100, 20);
-        assert!(cw.fits(&"a".repeat(300)));
-        assert!(!cw.fits(&"a".repeat(400)));
-    }
-}
diff --git a/crates/nvisy-llm/src/backend/rig/inner.rs b/crates/nvisy-llm/src/backend/rig/inner.rs
deleted file mode 100644
index 537c1a181..000000000
--- a/crates/nvisy-llm/src/backend/rig/inner.rs
+++ /dev/null
@@ -1,41 +0,0 @@
-//! Private rig-agent dispatch enum + macro.
-//!
-//! Wraps the four provider-specific `rig::Agent<M>` instances behind
-//! one enum so the rest of [`super`] can call them uniformly without
-//! caring which provider produced the agent.
-
-use reqwest_middleware::ClientWithMiddleware;
-use rig::agent::Agent;
-#[cfg(feature = "anthropic-claude")]
-use rig::providers::anthropic::completion::CompletionModel as AnthropicCompletionModel;
-#[cfg(feature = "google-gemini")]
-use rig::providers::gemini::completion::CompletionModel as GeminiCompletionModel;
-use rig::providers::ollama::CompletionModel as OllamaCompletionModel;
-#[cfg(feature = "openai-gpt")]
-use rig::providers::openai::completion::CompletionModel as OpenAiCompletionModel;
-
-pub(super) enum RigInner {
-    #[cfg(feature = "openai-gpt")]
-    OpenAi(Agent<OpenAiCompletionModel<ClientWithMiddleware>>),
-    #[cfg(feature = "anthropic-claude")]
-    Anthropic(Agent<AnthropicCompletionModel<ClientWithMiddleware>>),
-    #[cfg(feature = "google-gemini")]
-    Gemini(Agent<GeminiCompletionModel<ClientWithMiddleware>>),
-    Ollama(Agent<OllamaCompletionModel<ClientWithMiddleware>>),
-}
-
-macro_rules! dispatch {
-    ($inner:expr, |$agent:ident| $body:expr) => {
-        match $inner {
-            #[cfg(feature = "openai-gpt")]
-            $crate::backend::rig::inner::RigInner::OpenAi($agent) => $body,
-            #[cfg(feature = "anthropic-claude")]
-            $crate::backend::rig::inner::RigInner::Anthropic($agent) => $body,
-            #[cfg(feature = "google-gemini")]
-            $crate::backend::rig::inner::RigInner::Gemini($agent) => $body,
-            $crate::backend::rig::inner::RigInner::Ollama($agent) => $body,
-        }
-    };
-}
-
-pub(super) use dispatch;
diff --git a/crates/nvisy-llm/src/backend/rig/mod.rs b/crates/nvisy-llm/src/backend/rig/mod.rs
deleted file mode 100644
index b33036219..000000000
--- a/crates/nvisy-llm/src/backend/rig/mod.rs
+++ /dev/null
@@ -1,243 +0,0 @@
-//! [`RigBackend`]: rig-backed [`LlmBackend`].
-//!
-//! Wraps one of the four supported rig providers (OpenAI, Anthropic,
-//! Gemini, Ollama) behind the modality-agnostic [`LlmBackend`]
-//! surface. Owns its [`UsageTracker`] and [`LlmConfig`].
-//!
-//! [`LlmBackend`]: crate::backend::LlmBackend
-
-mod config;
-mod context;
-mod inner;
-mod usage;
-
-use std::borrow::Cow;
-
-use nvisy_core::{Error as CoreError, Result};
-use rig::agent::{Agent, AgentBuilder};
-use rig::client::CompletionClient;
-use rig::completion::{AssistantContent, Completion, CompletionModel, Message};
-
-pub use self::config::LlmConfig;
-pub use self::context::ContextWindow;
-use self::inner::{RigInner, dispatch};
-pub use self::usage::{UsageStats, UsageTracker};
-use super::http::{HttpConfig, build_http_client};
-use super::{LlmBackend, LlmRequest, LlmResponse};
-use crate::error::Error;
-use crate::provider::LlmProvider;
-
-const TARGET: &str = "nvisy_llm::backend::rig";
-
-/// Rig-backed LLM backend.
-///
-/// Construct via [`builder`]. Owns the provider-specific rig agent
-/// (created at build time), a [`UsageTracker`] that accumulates
-/// per-call token usage, and the [`LlmConfig`] driving sampling +
-/// compaction policy.
-///
-/// [`builder`]: Self::builder
-pub struct RigBackend {
-    agent: RigInner,
-    config: LlmConfig,
-    tracker: UsageTracker,
-    model_name: String,
-}
-
-impl RigBackend {
-    /// Start the chainable builder.
-    #[must_use]
-    pub fn builder() -> RigBackendBuilder {
-        RigBackendBuilder::default()
-    }
-
-    /// Snapshot the cumulative token-usage counters.
-    #[must_use]
-    pub fn usage(&self) -> UsageStats {
-        self.tracker.snapshot()
-    }
-
-    /// Reset the usage counters to zero.
-    pub fn reset_usage(&self) {
-        self.tracker.reset();
-    }
-
-    /// If `compact` is enabled, a context window is configured, and
-    /// `prompt` exceeds the input budget, summarise it via an extra
-    /// LLM call so it fits. Otherwise return the prompt unchanged.
-    async fn maybe_compact<'a>(&self, prompt: &'a str) -> Result<Cow<'a, str>> {
-        let Some(cw) = self.config.context_window.as_ref() else {
-            return Ok(Cow::Borrowed(prompt));
-        };
-        if !self.config.compact || cw.fits(prompt) {
-            return Ok(Cow::Borrowed(prompt));
-        }
-
-        let budget = cw.input_budget();
-        tracing::info!(
-            target: TARGET,
-            prompt_len = prompt.len(),
-            budget,
-            "prompt exceeds input budget, compacting"
-        );
-
-        let compact_prompt = format!(
-            "Summarize the following text so it fits within {budget} tokens. \
-             Preserve all key facts and details.\n\n{prompt}"
-        );
-        let text = self.complete_text(&compact_prompt, None).await?;
-        Ok(Cow::Owned(text))
-    }
-
-    /// Send a completion request and return the model's text reply.
-    /// `schema`, when `Some`, is passed through to rig's
-    /// `output_schema` so providers constrain the response.
-    async fn complete_text(
-        &self,
-        prompt: &str,
-        schema: Option<&schemars::Schema>,
-    ) -> Result<String> {
-        let (text, usage) = dispatch!(&self.agent, |agent| {
-            let mut builder = agent
-                .completion(prompt, Vec::<Message>::new())
-                .await
-                .map_err(Error::from)
-                .map_err(crate::error::convert)?;
-            if let Some(schema) = schema {
-                builder = builder.output_schema(schema.clone());
-            }
-            let response = builder
-                .send()
-                .await
-                .map_err(Error::from)
-                .map_err(crate::error::convert)?;
-            let text = extract_text(response.choice.iter())?;
-            Ok::<_, Error>((text, response.usage))
-        })?;
-        self.tracker.record(&usage, 0);
-        Ok(text)
-    }
-}
-
-#[async_trait::async_trait]
-impl LlmBackend for RigBackend {
-    #[tracing::instrument(target = TARGET, skip_all, fields(model = %self.model_name))]
-    async fn predict(&self, request: LlmRequest<'_>) -> Result<LlmResponse> {
-        let prompt = self.maybe_compact(request.prompt).await?;
-        let text = self.complete_text(&prompt, request.schema).await?;
-        Ok(LlmResponse::new(text))
-    }
-
-    fn model(&self) -> &str {
-        &self.model_name
-    }
-}
-
-/// Builder for [`RigBackend`].
-#[derive(Debug, Default)]
-pub struct RigBackendBuilder {
-    provider: Option<LlmProvider>,
-    config: Option<LlmConfig>,
-}
-
-impl RigBackendBuilder {
-    /// Set the LLM provider. Required.
-    #[must_use]
-    pub fn with_provider(mut self, provider: LlmProvider) -> Self {
-        self.provider = Some(provider);
-        self
-    }
-
-    /// Set the agent config. Defaults to [`LlmConfig::default`].
-    #[must_use]
-    pub fn with_config(mut self, config: LlmConfig) -> Self {
-        self.config = Some(config);
-        self
-    }
-
-    /// Build the backend.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when `provider` is unset, and the
-    /// underlying rig / HTTP error when client construction fails.
-    pub fn build(self) -> Result<RigBackend> {
-        let provider = self
-            .provider
-            .ok_or_else(|| CoreError::validation("RigBackendBuilder requires a provider", "rig"))?;
-        let config = self.config.unwrap_or_default();
-
-        let http = build_http_client(&HttpConfig {
-            max_retries: config.max_retries,
-            ..HttpConfig::default()
-        })
-        .map_err(|e| Error::Request(e.to_string()))
-        .map_err(crate::error::convert)?;
-
-        let preamble = config.preamble.as_deref();
-        let agent = match &provider {
-            #[cfg(feature = "openai-gpt")]
-            LlmProvider::OpenAi(p) => {
-                let client = p.openai_client(http).map_err(crate::error::convert)?;
-                let model = client.completions_api().completion_model(p.model.as_str());
-                RigInner::OpenAi(build_agent(model, &config, preamble))
-            }
-            #[cfg(feature = "anthropic-claude")]
-            LlmProvider::Anthropic(p) => {
-                let client = p.anthropic_client(http).map_err(crate::error::convert)?;
-                let model = client.completion_model(p.model.as_str());
-                RigInner::Anthropic(build_agent(model, &config, preamble))
-            }
-            #[cfg(feature = "google-gemini")]
-            LlmProvider::Gemini(p) => {
-                let client = p.gemini_client(http).map_err(crate::error::convert)?;
-                let model = client.completion_model(p.model.as_str());
-                RigInner::Gemini(build_agent(model, &config, preamble))
-            }
-            LlmProvider::Ollama(p) => {
-                let client = p.ollama_client(http).map_err(crate::error::convert)?;
-                let model = client.completion_model(p.model.as_str());
-                RigInner::Ollama(build_agent(model, &config, preamble))
-            }
-        };
-
-        Ok(RigBackend {
-            agent,
-            config,
-            tracker: UsageTracker::new(),
-            model_name: provider.model().to_owned(),
-        })
-    }
-}
-
-fn build_agent<M: CompletionModel>(
-    model: M,
-    config: &LlmConfig,
-    preamble: Option<&str>,
-) -> Agent<M> {
-    let mut b = AgentBuilder::new(model)
-        .temperature(config.temperature)
-        .max_tokens(config.max_tokens);
-    if let Some(p) = preamble {
-        b = b.preamble(p);
-    }
-    b.build()
-}
-
-fn extract_text<'a>(choices: impl Iterator<Item = &'a AssistantContent>) -> Result<String> {
-    let texts: Vec<&str> = choices
-        .filter_map(|c| match c {
-            AssistantContent::Text(t) => Some(t.text.as_str()),
-            _ => None,
-        })
-        .collect();
-
-    if texts.is_empty() {
-        return Err(CoreError::runtime(
-            "LLM response contained no text content",
-            "rig",
-            false,
-        ));
-    }
-    Ok(texts.join("\n"))
-}
diff --git a/crates/nvisy-llm/src/backend/rig/usage.rs b/crates/nvisy-llm/src/backend/rig/usage.rs
deleted file mode 100644
index cc4ff321b..000000000
--- a/crates/nvisy-llm/src/backend/rig/usage.rs
+++ /dev/null
@@ -1,83 +0,0 @@
-//! Cumulative token-usage tracking across LLM requests.
-
-use std::sync::atomic::{AtomicU64, Ordering};
-
-use derive_more::AddAssign;
-use rig::completion::Usage;
-
-/// Thread-safe accumulator for LLM token usage.
-///
-/// Uses lock-free atomics instead of a mutex — all operations are
-/// wait-free counter increments/loads. Each [`RigBackend`] owns one
-/// tracker; callers snapshot it to inspect costs.
-///
-/// [`RigBackend`]: super::RigBackend
-pub struct UsageTracker {
-    input_tokens: AtomicU64,
-    output_tokens: AtomicU64,
-    requests: AtomicU64,
-    retries: AtomicU64,
-}
-
-/// Point-in-time snapshot of accumulated usage counters.
-///
-/// Field-wise `AddAssign` is derived so pipelines holding multiple
-/// backends can sum their snapshots into a single aggregate.
-#[derive(Debug, Default, Clone, AddAssign)]
-pub struct UsageStats {
-    /// Cumulative input (prompt) tokens billed across every request.
-    pub total_input_tokens: u64,
-    /// Cumulative output (completion) tokens billed across every request.
-    pub total_output_tokens: u64,
-    /// Number of completed LLM requests.
-    pub total_requests: u64,
-    /// Number of retry attempts triggered by transient failures.
-    pub total_retries: u64,
-}
-
-impl UsageTracker {
-    /// Build a tracker with every counter at zero.
-    pub fn new() -> Self {
-        Self {
-            input_tokens: AtomicU64::new(0),
-            output_tokens: AtomicU64::new(0),
-            requests: AtomicU64::new(0),
-            retries: AtomicU64::new(0),
-        }
-    }
-
-    /// Record a single LLM request's token usage and retry count.
-    pub fn record(&self, usage: &Usage, retries: u32) {
-        self.input_tokens
-            .fetch_add(usage.input_tokens, Ordering::Relaxed);
-        self.output_tokens
-            .fetch_add(usage.output_tokens, Ordering::Relaxed);
-        self.requests.fetch_add(1, Ordering::Relaxed);
-        self.retries
-            .fetch_add(u64::from(retries), Ordering::Relaxed);
-    }
-
-    /// Snapshot the current counters without resetting them.
-    pub fn snapshot(&self) -> UsageStats {
-        UsageStats {
-            total_input_tokens: self.input_tokens.load(Ordering::Relaxed),
-            total_output_tokens: self.output_tokens.load(Ordering::Relaxed),
-            total_requests: self.requests.load(Ordering::Relaxed),
-            total_retries: self.retries.load(Ordering::Relaxed),
-        }
-    }
-
-    /// Reset all counters to zero.
-    pub fn reset(&self) {
-        self.input_tokens.store(0, Ordering::Relaxed);
-        self.output_tokens.store(0, Ordering::Relaxed);
-        self.requests.store(0, Ordering::Relaxed);
-        self.retries.store(0, Ordering::Relaxed);
-    }
-}
-
-impl Default for UsageTracker {
-    fn default() -> Self {
-        Self::new()
-    }
-}
diff --git a/crates/nvisy-llm/src/error.rs b/crates/nvisy-llm/src/error.rs
deleted file mode 100644
index 636b5fda5..000000000
--- a/crates/nvisy-llm/src/error.rs
+++ /dev/null
@@ -1,175 +0,0 @@
-//! Unified error type covering LLM provider, serialization, and tool failures.
-
-use nvisy_core::{Error as CoreError, ErrorKind as CoreErrorKind};
-use rig::completion::{CompletionError, PromptError, StructuredOutputError};
-use rig::http_client::Error as HttpClientError;
-
-/// Internal error type for LLM provider interactions.
-///
-/// Converted to [`CoreError`] at public API boundaries via the
-/// [`convert`] helper.
-///
-/// [`CoreError`]: nvisy_core::Error
-#[derive(Debug, thiserror::Error)]
-pub(crate) enum Error {
-    /// An HTTP / network error from the LLM provider.
-    #[error("HTTP error: {0}")]
-    Http(HttpClientError),
-
-    /// A JSON (de)serialization error.
-    #[error("JSON error: {0}")]
-    Json(#[from] serde_json::Error),
-
-    /// The LLM provider returned an error response.
-    #[error("Provider error: {0}")]
-    Provider(String),
-
-    /// The LLM response was malformed or unexpected.
-    #[error("Response error: {0}")]
-    Response(String),
-
-    /// A request construction or validation error.
-    #[error("Request error: {0}")]
-    Request(String),
-
-    /// A runtime error (tool failure, agent limits, generation errors, etc.).
-    #[error("{0}")]
-    Runtime(String),
-}
-
-impl From<CompletionError> for Error {
-    fn from(err: CompletionError) -> Self {
-        match err {
-            CompletionError::HttpError(e) => Self::Http(e),
-            CompletionError::JsonError(e) => Self::Json(e),
-            CompletionError::ProviderError(msg) => Self::Provider(msg),
-            CompletionError::ResponseError(msg) => Self::Response(msg),
-            CompletionError::RequestError(e) => Self::Request(e.to_string()),
-            CompletionError::UrlError(e) => Self::Request(format!("URL: {e}")),
-        }
-    }
-}
-
-impl From<PromptError> for Error {
-    fn from(err: PromptError) -> Self {
-        match err {
-            PromptError::CompletionError(e) => Self::from(e),
-            PromptError::ToolError(e) => Self::Runtime(format!("tool: {e}")),
-            PromptError::ToolServerError(e) => Self::Runtime(format!("tool server: {e}")),
-            PromptError::MaxTurnsError { max_turns, .. } => {
-                Self::Runtime(format!("agent exceeded max turn limit ({max_turns})"))
-            }
-            PromptError::PromptCancelled { reason, .. } => {
-                Self::Runtime(format!("prompt cancelled: {reason}"))
-            }
-            PromptError::UnknownToolCall { tool_name, .. } => {
-                Self::Runtime(format!("agent called unknown tool: {tool_name}"))
-            }
-        }
-    }
-}
-
-impl From<StructuredOutputError> for Error {
-    fn from(err: StructuredOutputError) -> Self {
-        match err {
-            StructuredOutputError::PromptError(e) => Self::from(*e),
-            StructuredOutputError::DeserializationError(e) => {
-                Self::Response(format!("structured output: {e}"))
-            }
-            StructuredOutputError::EmptyResponse => {
-                Self::Response("model returned no content".to_string())
-            }
-        }
-    }
-}
-
-/// Convert any error that can be turned into a provider [`Error`] into a
-/// [`CoreError`]. Intended for `.map_err(crate::error::convert)` at
-/// public API boundaries.
-///
-/// [`CoreError`]: nvisy_core::Error
-pub(crate) fn convert<E: Into<Error>>(e: E) -> CoreError {
-    CoreError::from(e.into())
-}
-
-impl From<Error> for CoreError {
-    fn from(err: Error) -> Self {
-        match &err {
-            Error::Http(_) => CoreError::connection(err.to_string(), "rig", true),
-            Error::Json(_) => {
-                CoreError::new(CoreErrorKind::Serialization, err.to_string()).with_component("rig")
-            }
-            Error::Provider(msg) => {
-                let retryable = is_retryable_provider_error(msg);
-                CoreError::connection(err.to_string(), "rig", retryable)
-            }
-            Error::Response(_) => CoreError::runtime(err.to_string(), "rig", false),
-            Error::Request(_) => CoreError::validation(err.to_string(), "rig"),
-            Error::Runtime(_) => CoreError::runtime(err.to_string(), "rig", false),
-        }
-    }
-}
-
-/// Check if a provider error message indicates a retryable condition.
-///
-/// This uses substring matching against known error patterns from OpenAI,
-/// Anthropic, and Google. If upstream providers change their error
-/// message format this may need updating — the test suite below
-/// validates all known patterns.
-fn is_retryable_provider_error(msg: &str) -> bool {
-    let lower = msg.to_lowercase();
-    lower.contains("rate_limit")
-        || lower.contains("rate limit")
-        || lower.contains("overloaded")
-        || lower.contains("timeout")
-        || lower.contains("429")
-        || lower.contains("503")
-        || lower.contains("529")
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn retryable_patterns() {
-        // OpenAI
-        assert!(is_retryable_provider_error("Rate limit reached for gpt-4o"));
-        assert!(is_retryable_provider_error(
-            "Error code: 429 - You exceeded your current quota"
-        ));
-
-        // Anthropic
-        assert!(is_retryable_provider_error("overloaded_error: Overloaded"));
-        assert!(is_retryable_provider_error(
-            "rate_limit_error: Rate limited"
-        ));
-        assert!(is_retryable_provider_error("Error code: 529 - Overloaded"));
-
-        // Google
-        assert!(is_retryable_provider_error("503 Service Unavailable"));
-        assert!(is_retryable_provider_error("Request timeout"));
-    }
-
-    #[test]
-    fn non_retryable_patterns() {
-        assert!(!is_retryable_provider_error("invalid_api_key"));
-        assert!(!is_retryable_provider_error("model not found"));
-        assert!(!is_retryable_provider_error("content policy violation"));
-        assert!(!is_retryable_provider_error(""));
-    }
-
-    #[test]
-    fn provider_error_conversion_is_retryable() {
-        let err = Error::Provider("Rate limit exceeded".to_string());
-        let core_err: CoreError = err.into();
-        assert!(core_err.is_retryable());
-    }
-
-    #[test]
-    fn provider_error_conversion_is_not_retryable() {
-        let err = Error::Provider("invalid model".to_string());
-        let core_err: CoreError = err.into();
-        assert!(!core_err.is_retryable());
-    }
-}
diff --git a/crates/nvisy-llm/src/lib.rs b/crates/nvisy-llm/src/lib.rs
deleted file mode 100644
index 0c9d7018f..000000000
--- a/crates/nvisy-llm/src/lib.rs
+++ /dev/null
@@ -1,13 +0,0 @@
-#![forbid(unsafe_code)]
-#![cfg_attr(docsrs, feature(doc_cfg))]
-#![doc = include_str!("../README.md")]
-
-pub mod backend;
-pub(crate) mod error;
-pub mod provider;
-mod recognition;
-
-pub use self::provider::{AuthenticatedProvider, LlmProvider, UnauthenticatedProvider};
-pub use self::recognition::{
-    DefaultPrompt, FilePrompt, LlmRecognizer, LlmRecognizerBuilder, Prompt,
-};
diff --git a/crates/nvisy-llm/src/provider/authenticated.rs b/crates/nvisy-llm/src/provider/authenticated.rs
deleted file mode 100644
index d3d41d436..000000000
--- a/crates/nvisy-llm/src/provider/authenticated.rs
+++ /dev/null
@@ -1,99 +0,0 @@
-//! LLM providers that require an API key.
-
-use std::fmt;
-
-#[cfg(any(
-    feature = "openai-gpt",
-    feature = "anthropic-claude",
-    feature = "google-gemini"
-))]
-use reqwest_middleware::ClientWithMiddleware;
-#[cfg(feature = "anthropic-claude")]
-use rig::providers::anthropic;
-#[cfg(feature = "google-gemini")]
-use rig::providers::gemini;
-#[cfg(feature = "openai-gpt")]
-use rig::providers::openai;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-#[cfg(any(
-    feature = "openai-gpt",
-    feature = "anthropic-claude",
-    feature = "google-gemini"
-))]
-use crate::error::Error;
-
-/// Provider that requires an API key (OpenAI, Anthropic, Gemini).
-///
-/// Connection parameters used by the completion-side
-/// [`crate::backend::rig`] path.
-#[derive(Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-pub struct AuthenticatedProvider {
-    /// API key for the provider.
-    pub api_key: String,
-    /// Model name (e.g. `gpt-4o`, `claude-3-5-sonnet-20241022`).
-    pub model: String,
-    /// Optional base URL override. `None` uses the provider's default.
-    pub base_url: Option<String>,
-}
-
-impl fmt::Debug for AuthenticatedProvider {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("AuthenticatedProvider")
-            .field("api_key", &"***")
-            .field("model", &self.model)
-            .field("base_url", &self.base_url)
-            .finish()
-    }
-}
-
-impl AuthenticatedProvider {
-    /// Build an OpenAI rig-core client.
-    #[cfg(feature = "openai-gpt")]
-    #[cfg_attr(docsrs, doc(cfg(feature = "openai-gpt")))]
-    pub(crate) fn openai_client(
-        &self,
-        http: ClientWithMiddleware,
-    ) -> Result<openai::Client<ClientWithMiddleware>, Error> {
-        let mut b = openai::Client::builder()
-            .api_key(&self.api_key)
-            .http_client(http);
-        if let Some(url) = &self.base_url {
-            b = b.base_url(url);
-        }
-        b.build().map_err(|e| Error::Request(e.to_string()))
-    }
-
-    /// Build a Gemini rig-core client.
-    #[cfg(feature = "google-gemini")]
-    #[cfg_attr(docsrs, doc(cfg(feature = "google-gemini")))]
-    pub(crate) fn gemini_client(
-        &self,
-        http: ClientWithMiddleware,
-    ) -> Result<gemini::Client<ClientWithMiddleware>, Error> {
-        let mut b = gemini::Client::builder()
-            .api_key(&self.api_key)
-            .http_client(http);
-        if let Some(url) = &self.base_url {
-            b = b.base_url(url);
-        }
-        b.build().map_err(|e| Error::Request(e.to_string()))
-    }
-
-    /// Build an Anthropic rig-core client.
-    #[cfg(feature = "anthropic-claude")]
-    #[cfg_attr(docsrs, doc(cfg(feature = "anthropic-claude")))]
-    pub(crate) fn anthropic_client(
-        &self,
-        http: ClientWithMiddleware,
-    ) -> Result<anthropic::Client<ClientWithMiddleware>, Error> {
-        let mut b = anthropic::Client::builder()
-            .api_key(&self.api_key)
-            .http_client(http);
-        if let Some(url) = &self.base_url {
-            b = b.base_url(url);
-        }
-        b.build().map_err(|e| Error::Request(e.to_string()))
-    }
-}
diff --git a/crates/nvisy-llm/src/provider/llm.rs b/crates/nvisy-llm/src/provider/llm.rs
deleted file mode 100644
index d0802fb9d..000000000
--- a/crates/nvisy-llm/src/provider/llm.rs
+++ /dev/null
@@ -1,103 +0,0 @@
-//! [`LlmProvider`]: tagged union of LLM provider connection
-//! parameters used by [`RigBackend`].
-//!
-//! Holds connection parameters and the model name only; the actual
-//! rig-core client is constructed lazily when a [`RigBackend`] is
-//! built.
-//!
-//! [`RigBackend`]: crate::backend::RigBackend
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-#[cfg(any(
-    feature = "openai-gpt",
-    feature = "anthropic-claude",
-    feature = "google-gemini"
-))]
-use super::AuthenticatedProvider;
-use super::UnauthenticatedProvider;
-
-/// Supported LLM providers for agent-based tasks.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "kebab-case")]
-pub enum LlmProvider {
-    /// OpenAI (GPT-4o, GPT-4, etc.).
-    #[cfg(feature = "openai-gpt")]
-    #[cfg_attr(docsrs, doc(cfg(feature = "openai-gpt")))]
-    OpenAi(AuthenticatedProvider),
-    /// Anthropic (Claude).
-    #[cfg(feature = "anthropic-claude")]
-    #[cfg_attr(docsrs, doc(cfg(feature = "anthropic-claude")))]
-    Anthropic(AuthenticatedProvider),
-    /// Google Gemini.
-    #[cfg(feature = "google-gemini")]
-    #[cfg_attr(docsrs, doc(cfg(feature = "google-gemini")))]
-    Gemini(AuthenticatedProvider),
-    /// Ollama (local models).
-    Ollama(UnauthenticatedProvider),
-}
-
-impl LlmProvider {
-    /// Create an OpenAI provider.
-    #[cfg(feature = "openai-gpt")]
-    #[cfg_attr(docsrs, doc(cfg(feature = "openai-gpt")))]
-    pub fn openai(api_key: &str, model: &str) -> Self {
-        Self::OpenAi(AuthenticatedProvider {
-            api_key: api_key.to_owned(),
-            model: model.to_owned(),
-            base_url: None,
-        })
-    }
-
-    /// Create an Anthropic provider.
-    #[cfg(feature = "anthropic-claude")]
-    #[cfg_attr(docsrs, doc(cfg(feature = "anthropic-claude")))]
-    pub fn anthropic(api_key: &str, model: &str) -> Self {
-        Self::Anthropic(AuthenticatedProvider {
-            api_key: api_key.to_owned(),
-            model: model.to_owned(),
-            base_url: None,
-        })
-    }
-
-    /// Create a Google Gemini provider.
-    #[cfg(feature = "google-gemini")]
-    #[cfg_attr(docsrs, doc(cfg(feature = "google-gemini")))]
-    pub fn gemini(api_key: &str, model: &str) -> Self {
-        Self::Gemini(AuthenticatedProvider {
-            api_key: api_key.to_owned(),
-            model: model.to_owned(),
-            base_url: None,
-        })
-    }
-
-    /// Create an Ollama provider using the default local URL.
-    pub fn ollama(model: &str) -> Self {
-        Self::Ollama(UnauthenticatedProvider {
-            model: model.to_owned(),
-            base_url: None,
-        })
-    }
-
-    /// Create an Ollama provider with a custom base URL.
-    pub fn ollama_with_url(model: &str, url: &str) -> Self {
-        Self::Ollama(UnauthenticatedProvider {
-            model: model.to_owned(),
-            base_url: Some(url.to_owned()),
-        })
-    }
-
-    /// The model name for this provider.
-    pub fn model(&self) -> &str {
-        match self {
-            #[cfg(feature = "openai-gpt")]
-            Self::OpenAi(p) => &p.model,
-            #[cfg(feature = "anthropic-claude")]
-            Self::Anthropic(p) => &p.model,
-            #[cfg(feature = "google-gemini")]
-            Self::Gemini(p) => &p.model,
-            Self::Ollama(p) => &p.model,
-        }
-    }
-}
diff --git a/crates/nvisy-llm/src/provider/mod.rs b/crates/nvisy-llm/src/provider/mod.rs
deleted file mode 100644
index ddd807a26..000000000
--- a/crates/nvisy-llm/src/provider/mod.rs
+++ /dev/null
@@ -1,18 +0,0 @@
-//! Shared LLM provider connection parameters.
-//!
-//! The rig completion backend ([`backend::rig`]) consumes the
-//! [`LlmProvider`] enum to dispatch to OpenAI / Anthropic / Google /
-//! Ollama clients. Connection parameters (API key, model, optional base
-//! URL) are split into [`AuthenticatedProvider`] +
-//! [`UnauthenticatedProvider`] so providers that don't require a key
-//! (Ollama) stay distinct in the type system.
-//!
-//! [`backend::rig`]: crate::backend
-
-mod authenticated;
-mod llm;
-mod unauthenticated;
-
-pub use self::authenticated::AuthenticatedProvider;
-pub use self::llm::LlmProvider;
-pub use self::unauthenticated::UnauthenticatedProvider;
diff --git a/crates/nvisy-llm/src/provider/unauthenticated.rs b/crates/nvisy-llm/src/provider/unauthenticated.rs
deleted file mode 100644
index b89e22ca5..000000000
--- a/crates/nvisy-llm/src/provider/unauthenticated.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-//! LLM providers that do not require an API key.
-
-use reqwest_middleware::ClientWithMiddleware;
-use rig::client::Nothing;
-use rig::providers::ollama;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use crate::error::Error;
-
-/// Provider that does not require an API key (Ollama, future local
-/// STT backends).
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-pub struct UnauthenticatedProvider {
-    /// Model name (e.g. `llama3.1:8b`).
-    pub model: String,
-    /// Optional base URL override. `None` uses the provider's default
-    /// (Ollama: `http://localhost:11434`).
-    pub base_url: Option<String>,
-}
-
-impl UnauthenticatedProvider {
-    /// Build an Ollama rig-core client.
-    pub(crate) fn ollama_client(
-        &self,
-        http: ClientWithMiddleware,
-    ) -> Result<ollama::Client<ClientWithMiddleware>, Error> {
-        let mut b = ollama::Client::builder().api_key(Nothing).http_client(http);
-        if let Some(url) = &self.base_url {
-            b = b.base_url(url);
-        }
-        b.build().map_err(|e| Error::Request(e.to_string()))
-    }
-}
diff --git a/crates/nvisy-llm/src/recognition/candidates.rs b/crates/nvisy-llm/src/recognition/candidates.rs
deleted file mode 100644
index 195c74040..000000000
--- a/crates/nvisy-llm/src/recognition/candidates.rs
+++ /dev/null
@@ -1,67 +0,0 @@
-//! Structured-output candidate types — the typed schemas the model
-//! is asked to produce.
-
-use nvisy_core::primitive::NormalizedBoundingBox;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Serde wrapper matching the model's `{"entities": [...]}`
-/// response for the [`Text`] modality.
-///
-/// [`Text`]: nvisy_core::modality::Text
-#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
-pub(super) struct TextCandidates {
-    /// Detected candidates.
-    pub entities: Vec<TextCandidate>,
-}
-
-/// One entity candidate produced by the model for the text
-/// modality. Carries the surface form (`value`) plus a surrounding
-/// `context` snippet the recognizer uses to localize the value back
-/// into a byte range in the source text.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-pub(super) struct TextCandidate {
-    /// Model-assigned identifier for the underlying real-world
-    /// entity. Stable across coreferent mentions within one call.
-    #[serde(default)]
-    pub entity_id: Option<String>,
-    /// Label name. Missing (`None`) means the model declined to
-    /// type the candidate; the recognizer drops these.
-    pub entity_type: Option<String>,
-    /// The matched text value — the literal surface form to flag.
-    pub value: String,
-    /// Model-asserted confidence in `[0.0, 1.0]`.
-    #[serde(default)]
-    pub confidence: Option<f64>,
-    /// Short surrounding text intended to uniquely locate `value`.
-    /// Missing or non-unique `context` causes the candidate to be
-    /// dropped per the recognizer's policy.
-    #[serde(default)]
-    pub context: Option<String>,
-    /// Brief description of the real-world entity (advisory).
-    #[serde(default)]
-    pub description: Option<String>,
-}
-
-/// Serde wrapper matching the model's `{"entities": [...]}`
-/// response for the [`Image`] modality.
-///
-/// [`Image`]: nvisy_core::modality::Image
-#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
-pub(super) struct VlmCandidates {
-    pub entities: Vec<VlmCandidate>,
-}
-
-/// One image entity discovered by the VLM. Bounding box is
-/// normalised (`[0, 1]`); the recognizer scales to pixel
-/// coordinates using the source image's dimensions.
-#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
-pub(super) struct VlmCandidate {
-    pub label: String,
-    #[serde(flatten)]
-    pub bbox: NormalizedBoundingBox,
-    #[serde(default)]
-    pub confidence: Option<f64>,
-    #[serde(default)]
-    pub description: Option<String>,
-}
diff --git a/crates/nvisy-llm/src/recognition/default_prompt.rs b/crates/nvisy-llm/src/recognition/default_prompt.rs
deleted file mode 100644
index 1268304f7..000000000
--- a/crates/nvisy-llm/src/recognition/default_prompt.rs
+++ /dev/null
@@ -1,74 +0,0 @@
-//! [`DefaultPrompt`]: the shipped [`Prompt`] impl, covering both
-//! [`Text`] and [`Image`].
-//!
-//! Both impls follow the same pattern: build a structured-output
-//! prompt with shared system instructions, ask the model to return
-//! `{ "entities": [...] }`, deserialise into a candidate vec, and
-//! lift each candidate into an `Entity<M>`. For text, the
-//! candidate's `context` field is used to localize the value back
-//! into a byte range; for image, the bounding box arrives in
-//! normalised `[0, 1]` coordinates and is scaled to pixel space.
-//!
-//! No label-map or labels-to-ignore filtering is applied here — the
-//! model's kinds pass through verbatim. Use [`FilePrompt`] when you
-//! need to remap raw model labels.
-//!
-//! [`FilePrompt`]: super::file_prompt::FilePrompt
-
-use base64::Engine;
-use base64::engine::general_purpose::STANDARD;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Image, Text};
-use nvisy_core::recognition::{LabelMap, RecognizerInput};
-use schemars::Schema;
-
-use super::candidates::{TextCandidates, VlmCandidates};
-use super::lift::{lift_image, lift_text};
-use super::prompt::Prompt;
-use super::response_parser::parse_json;
-use super::schemas::{text_schema, vlm_schema};
-use super::text_prompt::TextPromptBuilder;
-use super::vlm_prompt::VlmPromptBuilder;
-use crate::backend::LlmResponse;
-
-/// Shipped [`Prompt`] impl covering both [`Text`] and [`Image`].
-///
-/// Stateless zero-sized type. Customise behaviour by writing your
-/// own [`Prompt<M>`] impl rather than tweaking this one.
-#[derive(Debug, Default, Clone, Copy)]
-pub struct DefaultPrompt;
-
-impl Prompt<Text> for DefaultPrompt {
-    fn build(&self, input: &RecognizerInput<Text>) -> String {
-        TextPromptBuilder::new(input.data.text.as_str(), &input.hints, &input.labels).build()
-    }
-
-    fn schema(&self) -> Option<&Schema> {
-        Some(text_schema())
-    }
-
-    fn lift(&self, response: &LlmResponse, input: &RecognizerInput<Text>) -> Vec<Entity<Text>> {
-        let Ok(parsed): Result<TextCandidates, _> = parse_json(&response.text) else {
-            return Vec::new();
-        };
-        lift_text(input, parsed.entities, &LabelMap::new(), &[])
-    }
-}
-
-impl Prompt<Image> for DefaultPrompt {
-    fn build(&self, input: &RecognizerInput<Image>) -> String {
-        let image_b64 = STANDARD.encode(input.data.bytes.as_ref());
-        VlmPromptBuilder::new(&input.hints, &input.labels).build(&image_b64)
-    }
-
-    fn schema(&self) -> Option<&Schema> {
-        Some(vlm_schema())
-    }
-
-    fn lift(&self, response: &LlmResponse, input: &RecognizerInput<Image>) -> Vec<Entity<Image>> {
-        let Ok(parsed): Result<VlmCandidates, _> = parse_json(&response.text) else {
-            return Vec::new();
-        };
-        lift_image(input, parsed.entities, &LabelMap::new(), &[])
-    }
-}
diff --git a/crates/nvisy-llm/src/recognition/file_prompt.rs b/crates/nvisy-llm/src/recognition/file_prompt.rs
deleted file mode 100644
index bd78f7096..000000000
--- a/crates/nvisy-llm/src/recognition/file_prompt.rs
+++ /dev/null
@@ -1,324 +0,0 @@
-//! [`FilePrompt`]: load a [`Prompt`] from a TOML file.
-//!
-//! Prompt-as-data shape: the user-prompt template plus the label
-//! map plus the labels-to-ignore set all live in a single TOML
-//! file. Users swap behaviour by editing the file, not by writing
-//! Rust. Templates use Jinja2 syntax via `minijinja`.
-//!
-//! # TOML schema
-//!
-//! ```toml
-//! schema_version = 1
-//!
-//! [meta]
-//! name = "ner-default"
-//! modality = "text"   # or "image"
-//!
-//! # Optional. Maps model-emitted labels to canonical entity
-//! # labels. Use snake_case label names on the right-hand side.
-//! [label_map]
-//! person = "person_name"
-//! email = "email_address"
-//!
-//! # Optional. Drop spans the model returned with any of these
-//! # labels (case-sensitive).
-//! labels_to_ignore = ["MISC", "O"]
-//!
-//! # Jinja2-syntax template. Available variables:
-//! #   - text: source text (text modality only)
-//! #   - image_b64: base64-encoded image bytes (image modality only)
-//! #   - hints: list of { name, kind, value, snippet } (text) or
-//! #            { name, kind, bbox: { x, y, width, height } } (image)
-//! #   - labels: list of document context labels
-//! template = """
-//! Detect every sensitive entity in:
-//! ---
-//! {{ text }}
-//! ---
-//! {% if labels %}Labels: {{ labels | join(", ") }}{% endif %}
-//! Return JSON: {"entities": [...]}.
-//! """
-//! ```
-
-use std::collections::HashMap;
-use std::fs;
-use std::marker::PhantomData;
-use std::path::Path;
-
-use base64::Engine;
-use base64::engine::general_purpose::STANDARD;
-use minijinja::{Environment, context};
-use nvisy_core::entity::{Entity, EntityLabelRef};
-use nvisy_core::modality::{Image, Text};
-use nvisy_core::recognition::{LabelMap, RecognizerInput};
-use nvisy_core::{Error, Result};
-use schemars::Schema;
-use serde::Deserialize;
-
-use super::candidates::{TextCandidates, VlmCandidates};
-use super::lift::{lift_image, lift_text};
-use super::prompt::Prompt;
-use super::response_parser::parse_json;
-use super::schemas::{text_schema, vlm_schema};
-use crate::backend::LlmResponse;
-
-/// Half-width of the snippet rendered around a hint's location.
-const HINT_SNIPPET_HALF_WIDTH: usize = 80;
-
-/// File-driven [`Prompt`] impl.
-///
-/// Construct via [`from_toml_file`] or [`from_toml`]; the modality
-/// (`M`) is checked against `meta.modality` at parse time.
-///
-/// [`from_toml_file`]: Self::from_toml_file
-/// [`from_toml`]: Self::from_toml
-pub struct FilePrompt<M> {
-    template: String,
-    label_map: LabelMap,
-    labels_to_ignore: Vec<String>,
-    env: Environment<'static>,
-    _modality: PhantomData<fn() -> M>,
-}
-
-#[derive(Debug, Deserialize)]
-struct PromptFile {
-    #[allow(dead_code)]
-    schema_version: Option<u32>,
-    meta: PromptMeta,
-    #[serde(default)]
-    label_map: Option<HashMap<String, String>>,
-    #[serde(default)]
-    labels_to_ignore: Vec<String>,
-    template: String,
-}
-
-#[derive(Debug, Deserialize)]
-struct PromptMeta {
-    #[allow(dead_code)]
-    name: Option<String>,
-    modality: String,
-}
-
-impl<M> FilePrompt<M> {
-    fn from_parsed(parsed: PromptFile, expected_modality: &str) -> Result<Self> {
-        if parsed.meta.modality != expected_modality {
-            return Err(Error::validation(
-                format!(
-                    "prompt file modality is {:?}, expected {:?}",
-                    parsed.meta.modality, expected_modality
-                ),
-                "file-prompt",
-            ));
-        }
-
-        let mut label_map = LabelMap::new();
-        if let Some(entries) = parsed.label_map {
-            for (model_label, entity_label) in entries {
-                label_map = label_map.with_entry(model_label, EntityLabelRef::from(entity_label));
-            }
-        }
-
-        let mut env = Environment::new();
-        env.add_template_owned("prompt", parsed.template.clone())
-            .map_err(|e| {
-                Error::validation(format!("template compile error: {e}"), "file-prompt")
-            })?;
-
-        Ok(Self {
-            template: parsed.template,
-            label_map,
-            labels_to_ignore: parsed.labels_to_ignore,
-            env,
-            _modality: PhantomData,
-        })
-    }
-
-    /// Borrow the loaded template source.
-    #[must_use]
-    pub fn template(&self) -> &str {
-        &self.template
-    }
-
-    /// Borrow the configured label map.
-    #[must_use]
-    pub fn label_map(&self) -> &LabelMap {
-        &self.label_map
-    }
-}
-
-impl FilePrompt<Text> {
-    /// Load a text-modality prompt from a TOML file.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when the file is missing, malformed,
-    /// declares a non-`text` modality, or contains an invalid Jinja2
-    /// template.
-    pub fn from_toml_file(path: impl AsRef<Path>) -> Result<Self> {
-        let raw = fs::read_to_string(path.as_ref())
-            .map_err(|e| Error::validation(format!("reading prompt file: {e}"), "file-prompt"))?;
-        Self::from_toml(&raw)
-    }
-
-    /// Load a text-modality prompt from a TOML string.
-    ///
-    /// # Errors
-    ///
-    /// See [`from_toml_file`].
-    ///
-    /// [`from_toml_file`]: Self::from_toml_file
-    pub fn from_toml(raw: &str) -> Result<Self> {
-        let parsed: PromptFile = toml::from_str(raw)
-            .map_err(|e| Error::validation(format!("parsing prompt TOML: {e}"), "file-prompt"))?;
-        Self::from_parsed(parsed, "text")
-    }
-}
-
-impl FilePrompt<Image> {
-    /// Load an image-modality prompt from a TOML file.
-    ///
-    /// # Errors
-    ///
-    /// Same as the text-modality loader.
-    pub fn from_toml_file(path: impl AsRef<Path>) -> Result<Self> {
-        let raw = fs::read_to_string(path.as_ref())
-            .map_err(|e| Error::validation(format!("reading prompt file: {e}"), "file-prompt"))?;
-        Self::from_toml(&raw)
-    }
-
-    /// Load an image-modality prompt from a TOML string.
-    ///
-    /// # Errors
-    ///
-    /// Same as the text-modality loader.
-    pub fn from_toml(raw: &str) -> Result<Self> {
-        let parsed: PromptFile = toml::from_str(raw)
-            .map_err(|e| Error::validation(format!("parsing prompt TOML: {e}"), "file-prompt"))?;
-        Self::from_parsed(parsed, "image")
-    }
-}
-
-impl Prompt<Text> for FilePrompt<Text> {
-    fn build(&self, input: &RecognizerInput<Text>) -> String {
-        let text = input.data.text.as_str();
-        let hints: Vec<_> = input
-            .hints
-            .iter()
-            .map(|h| {
-                let value = value_at(text, h.location.start, h.location.end);
-                let snippet = snippet_around(text, h.location.start, h.location.end);
-                context! {
-                    name => h.name.as_deref().unwrap_or(""),
-                    kind => h.label.as_ref().map(|l| l.to_string()).unwrap_or_else(|| "unknown".to_owned()),
-                    value => value,
-                    snippet => snippet,
-                }
-            })
-            .collect();
-        let ctx = context! {
-            text => text,
-            hints => hints,
-            labels => input.labels.clone(),
-        };
-        self.env
-            .get_template("prompt")
-            .and_then(|t| t.render(ctx))
-            .unwrap_or_default()
-    }
-
-    fn schema(&self) -> Option<&Schema> {
-        Some(text_schema())
-    }
-
-    fn lift(&self, response: &LlmResponse, input: &RecognizerInput<Text>) -> Vec<Entity<Text>> {
-        let Ok(parsed): Result<TextCandidates, _> = parse_json(&response.text) else {
-            return Vec::new();
-        };
-        lift_text(
-            input,
-            parsed.entities,
-            &self.label_map,
-            &self.labels_to_ignore,
-        )
-    }
-}
-
-impl Prompt<Image> for FilePrompt<Image> {
-    fn build(&self, input: &RecognizerInput<Image>) -> String {
-        let image_b64 = STANDARD.encode(input.data.bytes.as_ref());
-        let hints: Vec<_> = input
-            .hints
-            .iter()
-            .map(|h| {
-                let bbox = &h.location.bounding_box;
-                context! {
-                    name => h.name.as_deref().unwrap_or(""),
-                    kind => h.label.as_ref().map(|l| l.to_string()).unwrap_or_else(|| "unknown".to_owned()),
-                    bbox => context! {
-                        x => bbox.x,
-                        y => bbox.y,
-                        width => bbox.width,
-                        height => bbox.height,
-                    },
-                }
-            })
-            .collect();
-        let ctx = context! {
-            image_b64 => image_b64,
-            hints => hints,
-            labels => input.labels.clone(),
-        };
-        self.env
-            .get_template("prompt")
-            .and_then(|t| t.render(ctx))
-            .unwrap_or_default()
-    }
-
-    fn schema(&self) -> Option<&Schema> {
-        Some(vlm_schema())
-    }
-
-    fn lift(&self, response: &LlmResponse, input: &RecognizerInput<Image>) -> Vec<Entity<Image>> {
-        let Ok(parsed): Result<VlmCandidates, _> = parse_json(&response.text) else {
-            return Vec::new();
-        };
-        lift_image(
-            input,
-            parsed.entities,
-            &self.label_map,
-            &self.labels_to_ignore,
-        )
-    }
-}
-
-fn value_at(text: &str, start: usize, end: usize) -> &str {
-    if start < end
-        && end <= text.len()
-        && text.is_char_boundary(start)
-        && text.is_char_boundary(end)
-    {
-        &text[start..end]
-    } else {
-        ""
-    }
-}
-
-fn snippet_around(text: &str, start: usize, end: usize) -> &str {
-    let lo = floor_char_boundary(text, start.saturating_sub(HINT_SNIPPET_HALF_WIDTH));
-    let hi = ceil_char_boundary(text, (end + HINT_SNIPPET_HALF_WIDTH).min(text.len()));
-    &text[lo..hi]
-}
-
-fn floor_char_boundary(s: &str, mut pos: usize) -> usize {
-    while pos > 0 && !s.is_char_boundary(pos) {
-        pos -= 1;
-    }
-    pos
-}
-
-fn ceil_char_boundary(s: &str, mut pos: usize) -> usize {
-    while pos < s.len() && !s.is_char_boundary(pos) {
-        pos += 1;
-    }
-    pos
-}
diff --git a/crates/nvisy-llm/src/recognition/lift.rs b/crates/nvisy-llm/src/recognition/lift.rs
deleted file mode 100644
index 10f39e566..000000000
--- a/crates/nvisy-llm/src/recognition/lift.rs
+++ /dev/null
@@ -1,147 +0,0 @@
-//! Candidate → [`Entity`] lifting shared by [`DefaultPrompt`] and
-//! [`FilePrompt`].
-//!
-//! Both prompts produce the same `{TextCandidates,VlmCandidates}`
-//! JSON shape and walk it the same way to emit entities. The only
-//! axis they differ on is whether they apply a [`LabelMap`] +
-//! `labels_to_ignore` filter on the model's emitted label string —
-//! [`DefaultPrompt`] passes an empty map + empty slice (no
-//! filtering); [`FilePrompt`] passes its loaded config.
-//!
-//! [`DefaultPrompt`]: super::default_prompt::DefaultPrompt
-//! [`FilePrompt`]: super::file_prompt::FilePrompt
-
-use nvisy_core::entity::{Entity, EntityLabelRef, ModelProvenance, TrailProvenance, TrailStep};
-use nvisy_core::modality::{Image, ImageLocation, Text, TextLocation};
-use nvisy_core::primitive::Confidence;
-use nvisy_core::recognition::{LabelMap, RecognizerInput};
-
-use super::candidates::{TextCandidate, VlmCandidate};
-use super::localize::{UnresolvedCandidatePolicy, localize_all};
-
-/// Default confidence assigned to a candidate when the LLM didn't
-/// score it.
-const DEFAULT_CONFIDENCE: f64 = 0.5;
-
-/// Lift a parsed text-candidate batch into `Entity<Text>` values.
-///
-/// `label_map` and `labels_to_ignore` together implement the
-/// model-label → canonical-name translation. Pass an empty map + an
-/// empty slice for no filtering.
-pub(super) fn lift_text(
-    input: &RecognizerInput<Text>,
-    candidates: Vec<TextCandidate>,
-    label_map: &LabelMap,
-    labels_to_ignore: &[String],
-) -> Vec<Entity<Text>> {
-    let text = input.data.text.as_str();
-    let localized = localize_all(text, candidates, UnresolvedCandidatePolicy::default());
-    let model = ModelProvenance::new("llm".to_owned());
-
-    let mut out = Vec::with_capacity(localized.len());
-    for l in localized {
-        let Some(label) = resolve_text_label(
-            l.candidate.entity_type.as_deref(),
-            l.candidate.value.as_str(),
-            label_map,
-            labels_to_ignore,
-        ) else {
-            continue;
-        };
-        let raw = l.candidate.confidence.unwrap_or(DEFAULT_CONFIDENCE);
-        let Some(confidence) = Confidence::new(raw.clamp(0.0, 1.0)) else {
-            continue;
-        };
-        let location = TextLocation::new(l.start_offset, l.end_offset);
-        let reason = format!("llm identified {label}");
-        let step = TrailStep::recognition(
-            "llm-ner",
-            confidence,
-            TrailProvenance::Model(model.clone()),
-            reason,
-        );
-
-        let mut b = Entity::builder()
-            .with_label(label)
-            .with_trail(vec![step])
-            .with_confidence(confidence)
-            .with_location(location);
-        if let Some(id) = l.candidate.entity_id {
-            b = b.with_entity_id(id);
-        }
-        out.push(b.build().expect("required fields provided"));
-    }
-    out
-}
-
-/// Lift a parsed VLM-candidate batch into `Entity<Image>` values.
-pub(super) fn lift_image(
-    input: &RecognizerInput<Image>,
-    candidates: Vec<VlmCandidate>,
-    label_map: &LabelMap,
-    labels_to_ignore: &[String],
-) -> Vec<Entity<Image>> {
-    let dims = input.data.dims;
-    let model = ModelProvenance::new("llm".to_owned());
-
-    let mut out = Vec::with_capacity(candidates.len());
-    for d in candidates {
-        if labels_to_ignore.iter().any(|l| l == &d.label) {
-            continue;
-        }
-        let label = label_map
-            .lookup(&d.label)
-            .cloned()
-            .unwrap_or_else(|| EntityLabelRef::from(d.label.clone()));
-        let raw = d.confidence.unwrap_or(DEFAULT_CONFIDENCE);
-        let Some(confidence) = Confidence::new(raw.clamp(0.0, 1.0)) else {
-            continue;
-        };
-        let bbox = d.bbox.to_pixel(dims);
-        let location = ImageLocation::new(bbox);
-        let reason = format!("vlm identified {label}");
-        let step = TrailStep::recognition(
-            "llm-vlm",
-            confidence,
-            TrailProvenance::Model(model.clone()),
-            reason,
-        );
-        let entity = Entity::builder()
-            .with_label(label)
-            .with_trail(vec![step])
-            .with_confidence(confidence)
-            .with_location(location)
-            .build()
-            .expect("required fields provided");
-        out.push(entity);
-    }
-    out
-}
-
-/// Pick the canonical label name for a text candidate.
-///
-/// Priority order: (1) the model's typed label, after label-map +
-/// ignore-list filtering; (2) literal-value lookup in the label
-/// map (covers raw-string-label backends); (3) drop.
-fn resolve_text_label(
-    typed: Option<&str>,
-    value: &str,
-    label_map: &LabelMap,
-    labels_to_ignore: &[String],
-) -> Option<EntityLabelRef> {
-    if let Some(model_label) = typed {
-        if labels_to_ignore.iter().any(|l| l == model_label) {
-            return None;
-        }
-        return Some(
-            label_map
-                .lookup(model_label)
-                .cloned()
-                .unwrap_or_else(|| EntityLabelRef::from(model_label.to_owned())),
-        );
-    }
-    if labels_to_ignore.iter().any(|l| l == value) {
-        return None;
-    }
-    label_map.lookup(value).cloned()
-}
diff --git a/crates/nvisy-llm/src/recognition/llm_recognizer.rs b/crates/nvisy-llm/src/recognition/llm_recognizer.rs
deleted file mode 100644
index 9ad80a6aa..000000000
--- a/crates/nvisy-llm/src/recognition/llm_recognizer.rs
+++ /dev/null
@@ -1,117 +0,0 @@
-//! [`LlmRecognizer`]: LLM-driven recognizer.
-//!
-//! Generic over [`Modality`] so one type drives text and image
-//! detection through the same surface. Holds an
-//! `Arc<dyn LlmBackend>` for the swappable LLM plumbing plus an
-//! `Arc<dyn Prompt<M>>` for the swappable modality-specific
-//! prompt-build + response-lift. The recognizer body is
-//! modality-agnostic: build prompt, send to backend, lift reply.
-
-use std::sync::Arc;
-
-use derive_builder::Builder;
-use nvisy_core::modality::Modality;
-use nvisy_core::recognition::{EntityRecognizer, RecognizerInput, RecognizerOutput};
-use nvisy_core::{Error, Result};
-
-use super::prompt::Prompt;
-use crate::backend::{LlmBackend, LlmRequest};
-
-/// LLM-driven recognizer.
-#[derive(Clone, Builder)]
-#[builder(
-    name = "LlmRecognizerBuilder",
-    pattern = "owned",
-    setter(into, prefix = "with"),
-    build_fn(error = "Error", name = "try_build", private)
-)]
-pub struct LlmRecognizer<M: Modality> {
-    /// Recognizer name. Surfaced in trail provenance and used as the
-    /// registry key.
-    name: String,
-    /// Backend that sends the prompt to the model and returns its
-    /// reply. Required. Set via [`with_backend`], which accepts any
-    /// concrete [`LlmBackend`] impl by value and wraps it in `Arc`
-    /// internally.
-    ///
-    /// [`with_backend`]: LlmRecognizerBuilder::with_backend
-    #[builder(setter(custom))]
-    backend: Arc<dyn LlmBackend>,
-    /// Modality-specific prompt builder + response lifter. Required.
-    /// Set via [`with_prompt`], which accepts any concrete
-    /// [`Prompt<M>`] impl by value and wraps it in `Arc` internally.
-    ///
-    /// [`with_prompt`]: LlmRecognizerBuilder::with_prompt
-    #[builder(setter(custom))]
-    prompt: Arc<dyn Prompt<M>>,
-}
-
-impl<M: Modality> LlmRecognizer<M> {
-    /// Start the chainable builder. `name`, `backend`, and `prompt`
-    /// are required — calling [`build`] without them returns a
-    /// validation error.
-    ///
-    /// [`build`]: LlmRecognizerBuilder::build
-    #[must_use]
-    pub fn builder() -> LlmRecognizerBuilder<M> {
-        LlmRecognizerBuilder::default()
-    }
-
-    /// Recognizer name.
-    #[must_use]
-    pub fn name(&self) -> &str {
-        &self.name
-    }
-
-    /// Borrow the configured backend.
-    #[must_use]
-    pub fn backend(&self) -> &Arc<dyn LlmBackend> {
-        &self.backend
-    }
-
-    /// Borrow the configured prompt.
-    #[must_use]
-    pub fn prompt(&self) -> &Arc<dyn Prompt<M>> {
-        &self.prompt
-    }
-}
-
-impl<M: Modality> LlmRecognizerBuilder<M> {
-    /// Set the [`LlmBackend`] that powers this recognizer. Accepts
-    /// any concrete impl by value and wraps it in `Arc`. Required —
-    /// `build` errors when this hasn't been called.
-    #[must_use]
-    pub fn with_backend<B: LlmBackend>(mut self, backend: B) -> Self {
-        self.backend = Some(Arc::new(backend));
-        self
-    }
-
-    /// Set the modality-specific [`Prompt`] strategy. Accepts any
-    /// concrete impl by value and wraps it in `Arc`. Required —
-    /// `build` errors when this hasn't been called.
-    #[must_use]
-    pub fn with_prompt<P: Prompt<M>>(mut self, prompt: P) -> Self {
-        self.prompt = Some(Arc::new(prompt));
-        self
-    }
-
-    /// Finish the builder. Errors when `name`, `backend`, or
-    /// `prompt` is unset.
-    pub fn build(self) -> Result<LlmRecognizer<M>> {
-        self.try_build()
-    }
-}
-
-#[async_trait::async_trait]
-impl<M: Modality> EntityRecognizer<M> for LlmRecognizer<M> {
-    async fn recognize(&self, input: &RecognizerInput<M>) -> Result<RecognizerOutput<M>> {
-        let prompt = self.prompt.build(input);
-        let request = LlmRequest {
-            prompt: &prompt,
-            schema: self.prompt.schema(),
-        };
-        let response = self.backend.predict(request).await?;
-        let entities = self.prompt.lift(&response, input);
-        Ok(RecognizerOutput::new(entities))
-    }
-}
diff --git a/crates/nvisy-llm/src/recognition/localize.rs b/crates/nvisy-llm/src/recognition/localize.rs
deleted file mode 100644
index 4efdb1d8c..000000000
--- a/crates/nvisy-llm/src/recognition/localize.rs
+++ /dev/null
@@ -1,172 +0,0 @@
-//! Localization: map a [`TextCandidate`] to a byte range in the
-//! source text using its `context` hint.
-//!
-//! Both context and source are normalized (NFC + whitespace
-//! collapse) before searching to absorb LLM whitespace drift. Byte
-//! offsets returned are in the *original*, un-normalized text.
-
-use unicode_normalization::UnicodeNormalization;
-
-use super::candidates::TextCandidate;
-
-const TARGET: &str = "nvisy_llm::recognition::localize";
-
-/// A candidate that's been resolved to a byte range in the source.
-#[derive(Debug, Clone)]
-pub(super) struct LocalizedCandidate {
-    pub candidate: TextCandidate,
-    pub start_offset: usize,
-    pub end_offset: usize,
-}
-
-/// What to do with candidates that can't be uniquely localized.
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
-#[derive(serde::Serialize, serde::Deserialize, schemars::JsonSchema)]
-#[serde(rename_all = "kebab-case")]
-pub(super) enum UnresolvedCandidatePolicy {
-    /// Drop ambiguous and missing candidates. Default.
-    #[default]
-    Drop,
-    /// Pick the first match for ambiguous candidates; drop only
-    /// when there are zero matches.
-    FirstMatch,
-}
-
-/// Localize every candidate against the source text.
-pub(super) fn localize_all(
-    text: &str,
-    candidates: Vec<TextCandidate>,
-    policy: UnresolvedCandidatePolicy,
-) -> Vec<LocalizedCandidate> {
-    let (normalized_text, index_maps) = normalize_with_index_map(text);
-
-    let mut out = Vec::with_capacity(candidates.len());
-    for c in candidates {
-        if let Some(localized) = localize_one(&normalized_text, &index_maps, &c, policy) {
-            out.push(localized);
-        }
-    }
-    out
-}
-
-fn localize_one(
-    normalized_text: &str,
-    orig_index: &(Vec<usize>, Vec<usize>),
-    candidate: &TextCandidate,
-    policy: UnresolvedCandidatePolicy,
-) -> Option<LocalizedCandidate> {
-    let context = match candidate.context.as_deref() {
-        Some(c) => c,
-        None => {
-            warn_dropped(candidate, "no context");
-            return None;
-        }
-    };
-    let (normalized_context, _) = normalize_with_index_map(context);
-    let (normalized_value, _) = normalize_with_index_map(&candidate.value);
-
-    let context_matches: Vec<usize> = normalized_text
-        .match_indices(&normalized_context)
-        .map(|(i, _)| i)
-        .collect();
-
-    let context_start = match context_matches.len() {
-        0 => {
-            warn_dropped(candidate, "context not found");
-            return None;
-        }
-        1 => context_matches[0],
-        _ => match policy {
-            UnresolvedCandidatePolicy::FirstMatch => context_matches[0],
-            _ => {
-                warn_dropped(candidate, "context ambiguous");
-                return None;
-            }
-        },
-    };
-
-    let context_end = context_start + normalized_context.len();
-    let window = &normalized_text[context_start..context_end];
-    let value_matches: Vec<usize> = window
-        .match_indices(&normalized_value)
-        .map(|(i, _)| i)
-        .collect();
-    let value_offset = match value_matches.len() {
-        0 => {
-            warn_dropped(candidate, "value not found in context");
-            return None;
-        }
-        1 => value_matches[0],
-        _ => match policy {
-            UnresolvedCandidatePolicy::FirstMatch => value_matches[0],
-            _ => {
-                warn_dropped(candidate, "value ambiguous within context");
-                return None;
-            }
-        },
-    };
-
-    let norm_start = context_start + value_offset;
-    let norm_end = norm_start + normalized_value.len();
-
-    let (start_index, end_index) = orig_index;
-    let start_offset = *start_index.get(norm_start)?;
-    let end_offset = if norm_end == 0 {
-        start_offset
-    } else {
-        *end_index.get(norm_end - 1)?
-    };
-
-    Some(LocalizedCandidate {
-        candidate: candidate.clone(),
-        start_offset,
-        end_offset,
-    })
-}
-
-fn warn_dropped(c: &TextCandidate, reason: &str) {
-    tracing::warn!(
-        target: TARGET,
-        entity_id = ?c.entity_id,
-        value = %c.value,
-        reason,
-        "dropping unresolvable text candidate"
-    );
-}
-
-/// Normalize text (NFC + whitespace collapse) and return parallel
-/// maps from normalized byte positions to original byte positions.
-fn normalize_with_index_map(text: &str) -> (String, (Vec<usize>, Vec<usize>)) {
-    let mut out = String::with_capacity(text.len());
-    let mut start_index: Vec<usize> = Vec::with_capacity(text.len());
-    let mut end_index: Vec<usize> = Vec::with_capacity(text.len());
-    let mut last_was_space = false;
-    let mut orig_pos = 0usize;
-
-    for orig_ch in text.chars() {
-        let orig_ch_len = orig_ch.len_utf8();
-        let orig_end = orig_pos + orig_ch_len;
-
-        if orig_ch.is_whitespace() {
-            if !last_was_space {
-                start_index.push(orig_pos);
-                end_index.push(orig_end);
-                out.push(' ');
-                last_was_space = true;
-            }
-        } else {
-            for nfc_ch in orig_ch.to_string().nfc() {
-                for _ in 0..nfc_ch.len_utf8() {
-                    start_index.push(orig_pos);
-                    end_index.push(orig_end);
-                }
-                out.push(nfc_ch);
-            }
-            last_was_space = false;
-        }
-        orig_pos += orig_ch_len;
-    }
-    start_index.push(orig_pos);
-    end_index.push(orig_pos);
-    (out, (start_index, end_index))
-}
diff --git a/crates/nvisy-llm/src/recognition/mod.rs b/crates/nvisy-llm/src/recognition/mod.rs
deleted file mode 100644
index b0541996e..000000000
--- a/crates/nvisy-llm/src/recognition/mod.rs
+++ /dev/null
@@ -1,26 +0,0 @@
-//! Recognizer layer: the LLM-driven [`LlmRecognizer`].
-//!
-//! `LlmRecognizer<M>` composes a modality-agnostic [`LlmBackend`]
-//! with a modality-specific [`Prompt<M>`]. [`DefaultPrompt`] is the
-//! shipped impl, covering both text and image; users wanting
-//! different prompt wording or different response handling
-//! implement [`Prompt<M>`] and pass their impl through the builder.
-//!
-//! [`LlmBackend`]: crate::backend::LlmBackend
-
-mod candidates;
-mod default_prompt;
-mod file_prompt;
-mod lift;
-mod llm_recognizer;
-mod localize;
-mod prompt;
-mod response_parser;
-mod schemas;
-mod text_prompt;
-mod vlm_prompt;
-
-pub use self::default_prompt::DefaultPrompt;
-pub use self::file_prompt::FilePrompt;
-pub use self::llm_recognizer::{LlmRecognizer, LlmRecognizerBuilder};
-pub use self::prompt::Prompt;
diff --git a/crates/nvisy-llm/src/recognition/prompt.rs b/crates/nvisy-llm/src/recognition/prompt.rs
deleted file mode 100644
index 3a5ac1751..000000000
--- a/crates/nvisy-llm/src/recognition/prompt.rs
+++ /dev/null
@@ -1,52 +0,0 @@
-//! [`Prompt`]: pluggable per-modality prompt builder + response
-//! lifter consumed by [`LlmRecognizer`].
-//!
-//! One trait per modality (`Prompt<Text>`, `Prompt<Image>`) — the
-//! recognizer holds `Arc<dyn Prompt<M>>` and dispatches through it.
-//! [`DefaultPrompt`] is the shipped impl, covering both modalities;
-//! users wanting different wording, different response shapes, or
-//! different localization policy implement [`Prompt<M>`] themselves
-//! and pass their impl to
-//! [`LlmRecognizerBuilder::with_prompt`].
-//!
-//! [`LlmRecognizer`]: super::LlmRecognizer
-//! [`LlmRecognizerBuilder::with_prompt`]: super::LlmRecognizerBuilder::with_prompt
-
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::Modality;
-use nvisy_core::recognition::RecognizerInput;
-
-use crate::backend::LlmResponse;
-
-/// Pluggable prompt builder + response lifter for one modality.
-///
-/// Implementors own both halves of the modality-specific work: turn
-/// a [`RecognizerInput<M>`] into the prompt string the backend
-/// receives, then turn the backend's reply into entities. Keeping
-/// both halves on one trait means the lifter has access to whatever
-/// state the builder stamped into the prompt (label maps, hint
-/// indices, etc.) by construction.
-pub trait Prompt<M>: Send + Sync + 'static
-where
-    M: Modality,
-{
-    /// Render the user prompt for `input`. Fold in source data,
-    /// hints, labels, and any base64-encoded binary payloads
-    /// (images, audio) the model needs to see.
-    fn build(&self, input: &RecognizerInput<M>) -> String;
-
-    /// Optional JSON schema describing the expected response shape.
-    /// When `Some`, the backend asks the model to constrain its
-    /// reply (via rig's `output_schema` or whatever the backend
-    /// uses). When `None`, the prompt itself is responsible for
-    /// describing the expected shape inline.
-    fn schema(&self) -> Option<&schemars::Schema> {
-        None
-    }
-
-    /// Parse the response text into entities. The recognizer wraps
-    /// these into a [`RecognizerOutput`].
-    ///
-    /// [`RecognizerOutput`]: nvisy_core::recognition::RecognizerOutput
-    fn lift(&self, response: &LlmResponse, input: &RecognizerInput<M>) -> Vec<Entity<M>>;
-}
diff --git a/crates/nvisy-llm/src/recognition/response_parser.rs b/crates/nvisy-llm/src/recognition/response_parser.rs
deleted file mode 100644
index f5f46a5f5..000000000
--- a/crates/nvisy-llm/src/recognition/response_parser.rs
+++ /dev/null
@@ -1,46 +0,0 @@
-//! Forgiving JSON parser for LLM responses.
-//!
-//! Strips markdown fences when present. Returns `T::default()` for
-//! empty / `"none"` / `"no entities"` responses.
-
-use serde::de::DeserializeOwned;
-
-/// Parse `text` as JSON into `T`, with markdown-fence + sentinel
-/// fallback handling.
-pub(super) fn parse_json<T: DeserializeOwned + Default>(
-    text: &str,
-) -> Result<T, serde_json::Error> {
-    let trimmed = text.trim();
-
-    if trimmed.is_empty()
-        || trimmed.eq_ignore_ascii_case("none")
-        || trimmed.eq_ignore_ascii_case("no entities")
-    {
-        return Ok(T::default());
-    }
-
-    let json_str = extract_fenced_json(trimmed).unwrap_or(trimmed);
-    serde_json::from_str::<T>(json_str)
-}
-
-/// Extract JSON content from markdown fences (```` ```json ... ``` ````).
-fn extract_fenced_json(text: &str) -> Option<&str> {
-    let start_marker = if let Some(pos) = text.find("```json") {
-        pos + "```json".len()
-    } else if let Some(pos) = text.find("```") {
-        pos + "```".len()
-    } else {
-        return None;
-    };
-
-    let rest = &text[start_marker..];
-    let rest = rest.strip_prefix('\n').unwrap_or(rest);
-    let end = rest.find("```")?;
-    let content = rest[..end].trim();
-
-    if content.is_empty() {
-        None
-    } else {
-        Some(content)
-    }
-}
diff --git a/crates/nvisy-llm/src/recognition/schemas.rs b/crates/nvisy-llm/src/recognition/schemas.rs
deleted file mode 100644
index 2057486af..000000000
--- a/crates/nvisy-llm/src/recognition/schemas.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-//! Cached JSON schemas for the two candidate-response shapes
-//! ([`TextCandidates`] and [`VlmCandidates`]).
-//!
-//! Both [`DefaultPrompt`] and [`FilePrompt`] return these from
-//! [`Prompt::schema`]; caching them in a `OnceLock` so the
-//! `schemars::schema_for!` macro only runs once per process keeps
-//! per-call cost negligible.
-//!
-//! [`DefaultPrompt`]: super::default_prompt::DefaultPrompt
-//! [`FilePrompt`]: super::file_prompt::FilePrompt
-//! [`Prompt::schema`]: super::prompt::Prompt::schema
-
-use std::sync::OnceLock;
-
-use schemars::Schema;
-
-use super::candidates::{TextCandidates, VlmCandidates};
-
-/// Cached JSON schema for [`TextCandidates`].
-pub(super) fn text_schema() -> &'static Schema {
-    static CACHE: OnceLock<Schema> = OnceLock::new();
-    CACHE.get_or_init(|| schemars::schema_for!(TextCandidates))
-}
-
-/// Cached JSON schema for [`VlmCandidates`].
-pub(super) fn vlm_schema() -> &'static Schema {
-    static CACHE: OnceLock<Schema> = OnceLock::new();
-    CACHE.get_or_init(|| schemars::schema_for!(VlmCandidates))
-}
diff --git a/crates/nvisy-llm/src/recognition/text_prompt.rs b/crates/nvisy-llm/src/recognition/text_prompt.rs
deleted file mode 100644
index 591fac412..000000000
--- a/crates/nvisy-llm/src/recognition/text_prompt.rs
+++ /dev/null
@@ -1,105 +0,0 @@
-//! Text prompt builder used by [`DefaultPrompt`]'s [`Prompt<Text>`]
-//! impl.
-//!
-//! [`DefaultPrompt`]: super::DefaultPrompt
-//! [`Prompt<Text>`]: super::Prompt
-
-use nvisy_core::modality::Text;
-use nvisy_core::recognition::Hint;
-
-/// Snippet window (in bytes) emitted on each side of a hint's
-/// range so the LLM has surrounding context for judgement.
-const SNIPPET_HALF_WIDTH: usize = 80;
-
-/// Builds user prompts for the text detect pass.
-pub(super) struct TextPromptBuilder<'a> {
-    text: &'a str,
-    hints: &'a [Hint<Text>],
-    labels: &'a [String],
-}
-
-impl<'a> TextPromptBuilder<'a> {
-    pub fn new(text: &'a str, hints: &'a [Hint<Text>], labels: &'a [String]) -> Self {
-        Self {
-            text,
-            hints,
-            labels,
-        }
-    }
-
-    pub fn build(&self) -> String {
-        let mut prompt = String::new();
-        prompt.push_str(
-            "Detect every sensitive entity in the following text. \
-             Return a JSON object with an \"entities\" key whose value is an array of \
-             candidates. Each candidate has keys: entity_id, entity_type, \
-             value, confidence, context, description.",
-        );
-        prompt.push_str("\n\n---\n");
-        prompt.push_str(self.text);
-        prompt.push_str("\n---");
-
-        if !self.labels.is_empty() {
-            let labels = self.labels.join(", ");
-            prompt.push_str(&format!(
-                "\n\nDocument context labels (adjust sensitivity to \
-                 domain-specific terms accordingly): {labels}."
-            ));
-        }
-
-        if !self.hints.is_empty() {
-            prompt.push_str(
-                "\n\nThe uploader marked these regions as likely sensitive. \
-                 Use them as priors when scoring candidates. Hints:",
-            );
-            for (i, h) in self.hints.iter().enumerate() {
-                let value = value_at(self.text, h.location.start, h.location.end);
-                let snippet = snippet_around(self.text, h.location.start, h.location.end);
-                let name = h.name.as_deref().unwrap_or("");
-                let kind = h
-                    .label
-                    .as_ref()
-                    .map(|l| l.to_string())
-                    .unwrap_or_else(|| "unknown".to_string());
-                prompt.push_str(&format!(
-                    "\n[hint {i}] name=\"{name}\", kind={kind}, \
-                     value=\"{value}\"\n  snippet: \"{snippet}\""
-                ));
-            }
-        }
-
-        prompt
-    }
-}
-
-fn snippet_around(text: &str, start: usize, end: usize) -> &str {
-    let lo = floor_char_boundary(text, start.saturating_sub(SNIPPET_HALF_WIDTH));
-    let hi = ceil_char_boundary(text, (end + SNIPPET_HALF_WIDTH).min(text.len()));
-    &text[lo..hi]
-}
-
-fn value_at(text: &str, start: usize, end: usize) -> &str {
-    if start < end
-        && end <= text.len()
-        && text.is_char_boundary(start)
-        && text.is_char_boundary(end)
-    {
-        &text[start..end]
-    } else {
-        ""
-    }
-}
-
-fn floor_char_boundary(s: &str, mut pos: usize) -> usize {
-    while pos > 0 && !s.is_char_boundary(pos) {
-        pos -= 1;
-    }
-    pos
-}
-
-fn ceil_char_boundary(s: &str, mut pos: usize) -> usize {
-    while pos < s.len() && !s.is_char_boundary(pos) {
-        pos += 1;
-    }
-    pos
-}
diff --git a/crates/nvisy-llm/src/recognition/vlm_prompt.rs b/crates/nvisy-llm/src/recognition/vlm_prompt.rs
deleted file mode 100644
index 7e562bdc1..000000000
--- a/crates/nvisy-llm/src/recognition/vlm_prompt.rs
+++ /dev/null
@@ -1,66 +0,0 @@
-//! Image (VLM) prompt builder used by [`DefaultPrompt`]'s
-//! [`Prompt<Image>`] impl.
-//!
-//! [`DefaultPrompt`]: super::DefaultPrompt
-//! [`Prompt<Image>`]: super::Prompt
-
-use nvisy_core::modality::Image;
-use nvisy_core::recognition::Hint;
-
-/// Builds user prompts for the VLM detect pass.
-pub(super) struct VlmPromptBuilder<'a> {
-    hints: &'a [Hint<Image>],
-    labels: &'a [String],
-}
-
-impl<'a> VlmPromptBuilder<'a> {
-    pub fn new(hints: &'a [Hint<Image>], labels: &'a [String]) -> Self {
-        Self { hints, labels }
-    }
-
-    pub fn build(&self, image_b64: &str) -> String {
-        let mut prompt = String::from(
-            "Find every sensitive entity visible in the image below. \
-             Draw a tight bounding box around each. Return a JSON object \
-             with an \"entities\" key whose value is an array of detections.",
-        );
-
-        if !self.labels.is_empty() {
-            let labels = self.labels.join(", ");
-            prompt.push_str(&format!(
-                "\n\nDocument context labels (adjust sensitivity to domain-specific \
-                 visual content accordingly): {labels}."
-            ));
-        }
-
-        if !self.hints.is_empty() {
-            prompt.push_str(
-                "\n\nThe uploader marked these regions as likely sensitive. \
-                 Confirm or relocate each via your detections; ignore those you \
-                 disagree with. Hints:",
-            );
-            for (i, h) in self.hints.iter().enumerate() {
-                let bbox = &h.location.bounding_box;
-                let kind = h
-                    .label
-                    .as_ref()
-                    .map(|l| l.to_string())
-                    .unwrap_or_else(|| "unknown".to_string());
-                let name = h.name.as_deref().unwrap_or("");
-                prompt.push_str(&format!(
-                    "\n[hint {i}] name=\"{name}\", kind={kind}, \
-                     bbox=({x}, {y}, {w}, {h})",
-                    x = bbox.x,
-                    y = bbox.y,
-                    w = bbox.width,
-                    h = bbox.height,
-                ));
-            }
-        }
-
-        prompt.push_str("\n\nImage (base64):\n");
-        prompt.push_str(image_b64);
-
-        prompt
-    }
-}
diff --git a/crates/nvisy-llm/testdata/prompts/ner.toml b/crates/nvisy-llm/testdata/prompts/ner.toml
deleted file mode 100644
index 25c452627..000000000
--- a/crates/nvisy-llm/testdata/prompts/ner.toml
+++ /dev/null
@@ -1,33 +0,0 @@
-schema_version = 1
-
-# Drop spans the model returned with these EntityKinds. Hoisted
-# above [label_map] because TOML tables continue until the next
-# header.
-labels_to_ignore = ["diagnosis"]
-
-# Jinja2 template exercising: variable interpolation, `{% for %}` over
-# hints, `{% if %}` gating on labels, the `| join` filter, and hint
-# snippet/value rendering.
-template = """
-You are a PII detector.
-Document text:
----
-{{ text }}
----
-{% if labels %}Document labels: {{ labels | join(", ") }}{% endif %}
-{% if hints %}Uploader hints:
-{% for h in hints %}- {{ h.name }} ({{ h.kind }}): value={{ h.value }} | snippet="{{ h.snippet }}"
-{% endfor %}{% endif %}
-Return JSON: {"entities": [...]}.
-"""
-
-[meta]
-name = "ner-e2e"
-modality = "text"
-
-# Remap one canonical EntityKind into another so the test can prove
-# `label_map` is invoked on the lift side. `TextCandidate.entity_type`
-# is strictly typed, so the model emits snake_case kinds; the map runs
-# over the rendered kind string.
-[label_map]
-person_name = "date_of_birth"
diff --git a/crates/nvisy-llm/testdata/prompts/vlm.toml b/crates/nvisy-llm/testdata/prompts/vlm.toml
deleted file mode 100644
index 1a94598f3..000000000
--- a/crates/nvisy-llm/testdata/prompts/vlm.toml
+++ /dev/null
@@ -1,25 +0,0 @@
-schema_version = 1
-
-# Drop spans the VLM emits with these EntityKinds. Hoisted above
-# [label_map] so they don't get swallowed into that table.
-labels_to_ignore = ["url"]
-
-# Jinja2 template exercising: variable interpolation, `{% for %}` over
-# hints, the `{% if %}` gate on labels, and nested bbox access.
-template = """
-Inspect this image (base64): {{ image_b64 }}
-{% if labels %}Tags: {{ labels | join(", ") }}{% endif %}
-{% if hints %}Uploader regions:
-{% for h in hints %}- {{ h.name }} ({{ h.kind }}): bbox=[{{ h.bbox.x }},{{ h.bbox.y }},{{ h.bbox.width }}x{{ h.bbox.height }}]
-{% endfor %}{% endif %}
-Return JSON: {"entities": [...]}.
-"""
-
-[meta]
-name = "vlm-e2e"
-modality = "image"
-
-# Remap one canonical kind into another so the test can prove
-# `label_map` is invoked on the lift side.
-[label_map]
-person_name = "date_of_birth"
diff --git a/crates/nvisy-llm/tests/file_prompt.rs b/crates/nvisy-llm/tests/file_prompt.rs
deleted file mode 100644
index 4710071bf..000000000
--- a/crates/nvisy-llm/tests/file_prompt.rs
+++ /dev/null
@@ -1,185 +0,0 @@
-//! End-to-end: load [`FilePrompt`] from a TOML fixture (Jinja2
-//! template + label map + ignore list), then exercise both halves of
-//! the [`Prompt`] trait — render against a populated
-//! [`RecognizerInput`] and lift a stub [`LlmResponse`] back into
-//! entities.
-//!
-//! Covers the minijinja code path end-to-end (variable interpolation,
-//! `{% for %}`, `{% if %}`, the `| join` filter, hint snippet
-//! rendering for text, bbox access for image), plus the
-//! `label_map` / `labels_to_ignore` policy on the lift side.
-
-use nvisy_core::entity::{EntityLabelRef, builtins};
-use nvisy_core::modality::{Image, ImageData, ImageLocation, Text, TextData, TextLocation};
-use nvisy_core::primitive::{BoundingBox, Dimensions};
-use nvisy_core::recognition::{Hint, RecognizerInput};
-use nvisy_llm::backend::LlmResponse;
-use nvisy_llm::{FilePrompt, Prompt};
-const NER_TOML: &str = include_str!("../testdata/prompts/ner.toml");
-const VLM_TOML: &str = include_str!("../testdata/prompts/vlm.toml");
-
-#[test]
-fn text_prompt_renders_template_and_lifts_entities() {
-    let prompt = FilePrompt::<Text>::from_toml(NER_TOML).expect("ner.toml parses");
-
-    // Realistic doc text. Hint coordinates pick out "Alice Carter".
-    let body = "From: Alice Carter <alice.carter@acme.test>\nSubject: hello";
-    let alice_start = body.find("Alice Carter").expect("alice substring");
-    let alice_end = alice_start + "Alice Carter".len();
-
-    let hint = Hint::<Text>::new(TextLocation::new(alice_start, alice_end))
-        .with_name("uploader-alice")
-        .with_label(builtins::PERSON_NAME.label_ref());
-
-    let input = RecognizerInput::new(TextData::new(body))
-        .with_hints(vec![hint])
-        .with_labels(vec!["medical".to_owned(), "gdpr-request".to_owned()]);
-
-    // -- build()
-    let rendered = prompt.build(&input);
-    assert!(rendered.contains(body), "source text missing: {rendered}");
-    assert!(
-        rendered.contains("Document labels: medical, gdpr-request"),
-        "labels `| join` filter not applied: {rendered}",
-    );
-    assert!(
-        rendered.contains("Uploader hints:"),
-        "{{% if hints %}} branch not taken: {rendered}",
-    );
-    assert!(
-        rendered.contains("uploader-alice (person_name): value=Alice Carter"),
-        "{{% for %}} over hints didn't render name/kind/value: {rendered}",
-    );
-    assert!(
-        rendered
-            .contains("snippet=\"From: Alice Carter <alice.carter@acme.test>\nSubject: hello\""),
-        "hint snippet not rendered (full body is within ±80 chars): {rendered}",
-    );
-
-    // -- lift(): the TOML maps `person_name → date_of_birth` and
-    // ignores `diagnosis`. The model emits snake_case label names
-    // (TextCandidate.entity_type is `Option<String>`); we expect
-    // person_name → date_of_birth via the map, email_address
-    // untouched, and diagnosis dropped by the ignore list.
-    let response = LlmResponse::new(
-        r#"{"entities":[
-            {"entity_type":"person_name","value":"Alice Carter","context":"From: Alice Carter <","confidence":0.9},
-            {"entity_type":"email_address","value":"alice.carter@acme.test","context":"<alice.carter@acme.test>","confidence":0.95},
-            {"entity_type":"diagnosis","value":"hello","context":"Subject: hello","confidence":0.1}
-        ]}"#,
-    );
-    let entities = prompt.lift(&response, &input);
-
-    let kinds: Vec<EntityLabelRef> = entities.iter().map(|e| e.label.clone()).collect();
-    assert!(
-        kinds.contains(&builtins::DATE_OF_BIRTH.label_ref()),
-        "person_name should have been remapped to DateOfBirth via label_map: {kinds:?}",
-    );
-    assert!(
-        kinds.contains(&builtins::EMAIL_ADDRESS.label_ref()),
-        "email_address (no map entry) should pass through: {kinds:?}",
-    );
-    assert!(
-        !kinds.contains(&builtins::DIAGNOSIS.label_ref()),
-        "diagnosis was in labels_to_ignore but appeared: {kinds:?}",
-    );
-    assert_eq!(
-        entities.len(),
-        2,
-        "expected 2 entities (diagnosis dropped), got {}: {kinds:?}",
-        entities.len(),
-    );
-}
-
-#[test]
-fn image_prompt_renders_template_and_lifts_entities() {
-    let prompt = FilePrompt::<Image>::from_toml(VLM_TOML).expect("vlm.toml parses");
-
-    // Tiny PNG-shaped payload — the prompt only base64-encodes it.
-    let bytes = b"\x89PNG\r\n\x1a\nfake-image-bytes".to_vec();
-    let dims = Dimensions::new(640, 480);
-
-    let hint = Hint::<Image>::new(ImageLocation::new(BoundingBox::new(
-        10.0, 20.0, 100.0, 50.0,
-    )))
-    .with_name("uploader-face")
-    .with_label(builtins::PERSON_NAME.label_ref());
-
-    let input = RecognizerInput::new(ImageData::new(bytes.clone(), dims))
-        .with_hints(vec![hint])
-        .with_labels(vec!["badge".to_owned()]);
-
-    // -- build()
-    let rendered = prompt.build(&input);
-    let expected_b64 = base64_encode(&bytes);
-    assert!(
-        rendered.contains(&expected_b64),
-        "image_b64 missing from render: {rendered}",
-    );
-    assert!(
-        rendered.contains("Tags: badge"),
-        "labels `| join` not rendered: {rendered}",
-    );
-    assert!(
-        rendered.contains("uploader-face (person_name): bbox=[10.0,20.0,100.0x50.0]"),
-        "{{% for %}} over image hints with nested bbox access broken: {rendered}",
-    );
-
-    // -- lift(): label_map maps `person_name → date_of_birth`, and
-    // labels_to_ignore drops `url`. The VLM emits typed snake_case
-    // kinds; assert remap + ignore both fire.
-    let response = LlmResponse::new(
-        r#"{"entities":[
-            {"label":"person_name","x":0.1,"y":0.1,"width":0.2,"height":0.2,"confidence":0.85},
-            {"label":"license_plate","x":0.5,"y":0.5,"width":0.1,"height":0.05,"confidence":0.7},
-            {"label":"url","x":0.0,"y":0.0,"width":0.05,"height":0.05,"confidence":0.9}
-        ]}"#,
-    );
-    let entities = prompt.lift(&response, &input);
-
-    let kinds: Vec<EntityLabelRef> = entities.iter().map(|e| e.label.clone()).collect();
-    assert!(
-        kinds.contains(&builtins::DATE_OF_BIRTH.label_ref()),
-        "person_name should have been remapped to DateOfBirth via label_map: {kinds:?}",
-    );
-    assert!(
-        kinds.contains(&builtins::LICENSE_PLATE.label_ref()),
-        "license_plate (no map entry) should pass through: {kinds:?}",
-    );
-    assert!(
-        !kinds.contains(&builtins::URL.label_ref()),
-        "url was in labels_to_ignore but appeared: {kinds:?}",
-    );
-    assert_eq!(
-        entities.len(),
-        2,
-        "expected 2 entities (url dropped), got {}: {kinds:?}",
-        entities.len(),
-    );
-}
-
-/// Local base64 encoder so the test asserts on the exact string the
-/// prompt rendered, without pulling in `base64` as a dev-dep.
-fn base64_encode(bytes: &[u8]) -> String {
-    const TABLE: &[u8; 64] = b"ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
-    let mut out = String::with_capacity(bytes.len().div_ceil(3) * 4);
-    for chunk in bytes.chunks(3) {
-        let b0 = chunk[0] as u32;
-        let b1 = chunk.get(1).copied().unwrap_or(0) as u32;
-        let b2 = chunk.get(2).copied().unwrap_or(0) as u32;
-        let n = (b0 << 16) | (b1 << 8) | b2;
-        out.push(TABLE[((n >> 18) & 63) as usize] as char);
-        out.push(TABLE[((n >> 12) & 63) as usize] as char);
-        if chunk.len() >= 2 {
-            out.push(TABLE[((n >> 6) & 63) as usize] as char);
-        } else {
-            out.push('=');
-        }
-        if chunk.len() == 3 {
-            out.push(TABLE[(n & 63) as usize] as char);
-        } else {
-            out.push('=');
-        }
-    }
-    out
-}
diff --git a/crates/nvisy-ner/Cargo.toml b/crates/nvisy-ner/Cargo.toml
deleted file mode 100644
index a6e3739e7..000000000
--- a/crates/nvisy-ner/Cargo.toml
+++ /dev/null
@@ -1,60 +0,0 @@
-# https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[package]
-name = "nvisy-ner"
-description = "NER traits and language detection for the Nvisy runtime"
-keywords = ["nvisy", "ner", "nlp", "pii"]
-categories = ["text-processing"]
-readme = "README.md"
-
-version = { workspace = true }
-rust-version = { workspace = true }
-edition = { workspace = true }
-license = { workspace = true }
-publish = { workspace = true }
-
-authors = { workspace = true }
-repository = { workspace = true }
-homepage = { workspace = true }
-documentation = { workspace = true }
-
-[features]
-default = []
-
-## Enables [`BentoBackend`]: zero-shot NER over an externalized
-## HTTP inference service (the `inference-gliner` Bento in
-## https://github.com/nvisycom/inference).
-bento = ["dep:bentoml"]
-
-[package.metadata.docs.rs]
-all-features = true
-rustdoc-args = ["--cfg", "docsrs"]
-
-[dependencies]
-# Internal crates
-nvisy-core = { workspace = true, features = [] }
-
-# Serialization
-serde = { workspace = true, features = ["derive"] }
-
-# Derive macros and error handling
-derive_builder = { workspace = true, features = [] }
-
-# Primitive datatypes
-uuid = { workspace = true, features = ["v7"] }
-type-map = { workspace = true, features = [] }
-
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-
-# Observability
-tracing = { workspace = true, features = [] }
-
-# Text processing (language detection)
-lingua = { workspace = true, features = [] }
-
-# AI / LLM frameworks (feature-gated)
-bentoml = { workspace = true, optional = true }
-
-[dev-dependencies]
-tokio = { workspace = true, features = ["macros", "rt"] }
diff --git a/crates/nvisy-ner/README.md b/crates/nvisy-ner/README.md
deleted file mode 100644
index 7971b2a68..000000000
--- a/crates/nvisy-ner/README.md
+++ /dev/null
@@ -1,40 +0,0 @@
-# nvisy-ner
-
-[![Build](https://img.shields.io/github/actions/workflow/status/nvisycom/runtime/build.yml?branch=main&label=build%20%26%20test&style=flat-square)](https://github.com/nvisycom/runtime/actions/workflows/build.yml)
-
-NER recognizer + pluggable inference `Backend` trait for the Nvisy
-runtime.
-
-## Overview
-
-`NerRecognizer` turns model-produced spans into typed `Entity<Text>`
-values via any implementation of the crate's `Backend` trait.
-Shipped backends: `NoopBackend` (test stub) and the feature-gated
-`BentoBackend`, an HTTP client into the externalised
-`inference-gliner` Bento. Model-bearing inference is intentionally
-out-of-process — see [`nvisycom/inference`].
-
-`LabelMap` projects raw backend labels onto the canonical
-`EntityKind` set so consumers reason about one fixed taxonomy
-regardless of the upstream model. LLM-mediated NER lives in
-`nvisy-llm`.
-
-[`nvisycom/inference`]: https://github.com/nvisycom/inference
-
-## Documentation
-
-See [`docs/`](../../docs/) for architecture, security, and API documentation.
-
-## Changelog
-
-See [CHANGELOG.md](../../CHANGELOG.md) for release notes and version history.
-
-## License
-
-Apache 2.0 License, see [LICENSE.txt](../../LICENSE.txt)
-
-## Support
-
-- **Documentation**: [docs.nvisy.com](https://docs.nvisy.com)
-- **Issues**: [GitHub Issues](https://github.com/nvisycom/runtime/issues)
-- **Email**: [support@nvisy.com](mailto:support@nvisy.com)
diff --git a/crates/nvisy-ner/src/backend/bento_backend.rs b/crates/nvisy-ner/src/backend/bento_backend.rs
deleted file mode 100644
index e84014af4..000000000
--- a/crates/nvisy-ner/src/backend/bento_backend.rs
+++ /dev/null
@@ -1,123 +0,0 @@
-//! [`BentoBackend`]: NER backend over the externalised
-//! `inference-gliner` Bento.
-//!
-//! Implements [`NerBackend`]. The wire shape mirrors
-//! `nvisy_core.ner.v1` from [`nvisycom/inference`]; per-request
-//! `correlation_id` propagation rides on the `x-request-id`
-//! header. Today the service is zero-shot — it requires a per-call
-//! `labels` list — so the backend errors out when called with
-//! `labels = None`.
-//!
-//! Wire compatibility note: the inference service returns the
-//! service-side label string verbatim; the recognizer's
-//! [`LabelMap`] re-canonicalises it into a workspace label name.
-//!
-//! [`nvisycom/inference`]: https://github.com/nvisycom/inference
-//! [`LabelMap`]: nvisy_core::recognition::LabelMap
-
-use bentoml::prelude::*;
-use nvisy_core::entity::ModelProvenance;
-use nvisy_core::{Error, Result};
-use uuid::Uuid;
-
-use super::bento_types::{WireBatch, WireRequest, WireResponse};
-use super::ner_backend::{NerBackend, NerRequest, NerResponse};
-use super::ner_span::RawNerSpan;
-
-const COMPONENT: &str = "ner-bento";
-const RECOGNIZE_ROUTE: &str = "recognize";
-
-/// Construction parameters for [`BentoBackend`].
-#[derive(Debug, Clone)]
-pub struct BentoParams {
-    /// Base URL of the `inference-gliner` Bento (e.g.
-    /// `http://localhost:3000`).
-    pub base_url: String,
-}
-
-impl BentoParams {
-    /// Construct with the given service URL.
-    pub fn new(base_url: impl Into<String>) -> Self {
-        Self {
-            base_url: base_url.into(),
-        }
-    }
-}
-
-/// [`NerBackend`] backed by an externalised BentoML service.
-#[derive(Debug)]
-pub struct BentoBackend {
-    endpoint: Endpoint,
-}
-
-impl BentoBackend {
-    /// Build a backend against the given parameters.
-    ///
-    /// # Errors
-    ///
-    /// Returns an error if the underlying HTTP client cannot be
-    /// constructed (invalid `base_url`).
-    pub fn new(params: BentoParams) -> Result<Self> {
-        let client = Client::builder()
-            .with_base_url(&params.base_url)
-            .build()
-            .map_err(|e| Error::runtime(format!("bentoml client init: {e}"), COMPONENT, false))?;
-        Ok(Self {
-            endpoint: client.endpoint(RECOGNIZE_ROUTE),
-        })
-    }
-}
-
-#[async_trait::async_trait]
-impl NerBackend for BentoBackend {
-    fn provenance(&self) -> ModelProvenance {
-        ModelProvenance::new("bento-ner")
-    }
-
-    #[tracing::instrument(skip_all)]
-    async fn recognize(&self, request: NerRequest<'_>) -> Result<NerResponse> {
-        let Some(labels) = request.labels else {
-            return Err(Error::validation(
-                "BentoBackend requires per-call labels (the inference-gliner service is zero-shot)",
-                COMPONENT,
-            ));
-        };
-        if labels.is_empty() {
-            return Ok(NerResponse::default());
-        }
-
-        let language = request.language.map(|l| l.as_str().to_owned());
-        let wire_request = WireRequest {
-            text: request.text.to_owned(),
-            labels: labels.iter().map(|s| (*s).to_owned()).collect(),
-            threshold: 0.0,
-            language,
-        };
-        let request_id = request
-            .correlation_id
-            .unwrap_or_else(Uuid::now_v7)
-            .to_string();
-
-        let responses: Vec<WireResponse> = self
-            .endpoint
-            .clone()
-            .with_request_id(&request_id)
-            .invoke(&WireBatch {
-                requests: vec![wire_request],
-            })
-            .await
-            .map_err(|e| Error::runtime(format!("bento ner call: {e}"), COMPONENT, true))?;
-
-        let mut spans = Vec::new();
-        for response in &responses {
-            for entity in &response.entities {
-                spans.push(RawNerSpan::new(
-                    entity.label.clone(),
-                    entity.score,
-                    entity.start..entity.end,
-                ));
-            }
-        }
-        Ok(NerResponse::new(spans))
-    }
-}
diff --git a/crates/nvisy-ner/src/backend/bento_types.rs b/crates/nvisy-ner/src/backend/bento_types.rs
deleted file mode 100644
index 0aa1a2edc..000000000
--- a/crates/nvisy-ner/src/backend/bento_types.rs
+++ /dev/null
@@ -1,78 +0,0 @@
-//! Wire types for the externalised `inference-gliner` Bento.
-//!
-//! Mirror of `nvisy_core.ner.v1` from [`nvisycom/inference`].
-//! Schema version v1. Field names are camelCase on the wire to
-//! match the Python service.
-//!
-//! [`BentoBackend`]: super::BentoBackend
-//! [`nvisycom/inference`]: https://github.com/nvisycom/inference
-
-use serde::{Deserialize, Serialize};
-
-/// Outer batch wrapper. Single- and multi-text calls share the
-/// same shape — a single recognition is just a batch of one.
-#[derive(Serialize)]
-pub(super) struct WireBatch {
-    /// One recognition request per text. Response array has the
-    /// same length and ordering.
-    pub requests: Vec<WireRequest>,
-}
-
-/// One recognition request: a text plus the per-call inference
-/// knobs the GLiNER model exposes.
-#[derive(Serialize)]
-#[serde(rename_all = "camelCase")]
-pub(super) struct WireRequest {
-    /// The text to recognise entities in.
-    pub text: String,
-    /// Entity label names the caller is interested in. GLiNER is
-    /// zero-shot — sending an empty list is meaningless and the
-    /// runtime short-circuits before making the call.
-    pub labels: Vec<String>,
-    /// Lower bound on per-entity score. The runtime keeps this at
-    /// `0.0` and post-filters locally so threshold decisions stay
-    /// in one place (the engine-side detection driver).
-    pub threshold: f64,
-    /// Optional BCP-47 language hint. Multilingual GLiNER variants
-    /// may ignore it; monolingual variants may validate it.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub language: Option<String>,
-}
-
-/// One per-request recognition result.
-#[derive(Deserialize)]
-#[serde(rename_all = "camelCase")]
-pub(super) struct WireResponse {
-    /// Identity of the underlying model that produced these
-    /// entities — surfaced onto the entity's trail step for
-    /// audit attribution. Echoed per response so a service hosting
-    /// multiple models can report different identities within one
-    /// batch.
-    #[allow(dead_code)]
-    pub model: String,
-    /// Recognised entities. Each entity's label is the
-    /// service-side classification, returned as a string so the
-    /// recognizer's [`LabelMap`] can re-canonicalise it.
-    ///
-    /// [`LabelMap`]: nvisy_core::recognition::LabelMap
-    #[serde(default)]
-    pub entities: Vec<WireEntity>,
-}
-
-/// One recognised entity span on the wire.
-#[derive(Deserialize)]
-#[serde(rename_all = "camelCase")]
-pub(super) struct WireEntity {
-    /// Service-side classification of this span. Translated to
-    /// the workspace label vocabulary by the recognizer's
-    /// `LabelMap`.
-    pub label: String,
-    /// Raw model score in `[0.0, 1.0]`.
-    pub score: f64,
-    /// Byte offset of the entity's start within
-    /// [`WireRequest::text`].
-    pub start: usize,
-    /// Byte offset one past the entity's end within
-    /// [`WireRequest::text`] (half-open `start..end`).
-    pub end: usize,
-}
diff --git a/crates/nvisy-ner/src/backend/mod.rs b/crates/nvisy-ner/src/backend/mod.rs
deleted file mode 100644
index d4e7612df..000000000
--- a/crates/nvisy-ner/src/backend/mod.rs
+++ /dev/null
@@ -1,24 +0,0 @@
-//! Backend layer: the [`NerBackend`] trait and its shipped impls.
-//!
-//! One trait covers zero-shot backends (per-call labels via
-//! [`NerRequest::labels`] = `Some(...)`) and fixed-label backends
-//! (labels baked into the model, `labels = None`). Built-in
-//! [`NoopBackend`] (returns no spans; test stub) and feature-gated
-//! [`BentoBackend`] (HTTP call into the externalised
-//! `inference-gliner` service).
-
-mod ner_backend;
-mod ner_span;
-mod noop_backend;
-
-#[cfg(feature = "bento")]
-mod bento_backend;
-#[cfg(feature = "bento")]
-mod bento_types;
-
-#[cfg(feature = "bento")]
-#[cfg_attr(docsrs, doc(cfg(feature = "bento")))]
-pub use self::bento_backend::{BentoBackend, BentoParams};
-pub use self::ner_backend::{NerBackend, NerRequest, NerResponse};
-pub use self::ner_span::RawNerSpan;
-pub use self::noop_backend::NoopBackend;
diff --git a/crates/nvisy-ner/src/backend/ner_backend.rs b/crates/nvisy-ner/src/backend/ner_backend.rs
deleted file mode 100644
index 1a5289210..000000000
--- a/crates/nvisy-ner/src/backend/ner_backend.rs
+++ /dev/null
@@ -1,107 +0,0 @@
-//! [`NerBackend`]: the unified per-call NER backend trait.
-//!
-//! Replaces the previous split between `GlinerBackend` (zero-shot,
-//! takes per-call kinds) and `NlpEngine`-produced NER spans
-//! (fixed-label, no per-call kinds). The `labels` field on
-//! [`NerRequest`] is `Option<&[&str]>`: `Some(...)` for zero-shot
-//! backends that take a label allowlist per call, `None` for
-//! fixed-label backends whose set of labels is baked into the
-//! model.
-//!
-//! Engines are called from inside [`NerRecognizer::recognize`] — no
-//! shared NLP pass, no orchestrator plumbing. Each recognizer holds
-//! its own backend and is self-contained.
-//!
-//! [`NerRecognizer::recognize`]: crate::NerRecognizer
-
-use nvisy_core::Result;
-use nvisy_core::entity::ModelProvenance;
-use nvisy_core::primitive::LanguageTag;
-use uuid::Uuid;
-
-use super::ner_span::RawNerSpan;
-
-/// One per-call NER request handed to a [`NerBackend`].
-#[derive(Debug, Clone)]
-pub struct NerRequest<'a> {
-    /// Source text to scan. Byte offsets in returned spans refer
-    /// back into this string.
-    pub text: &'a str,
-    /// Label names to detect when the backend supports per-call
-    /// label selection. `None` means the backend uses its built-in
-    /// fixed label set; `Some(slice)` means restrict detection to
-    /// the listed names. Empty slice short-circuits the call to no
-    /// work in the caller.
-    pub labels: Option<&'a [&'a str]>,
-    /// Caller-asserted language. Backends that support per-call
-    /// language hinting use this; backends that don't ignore it.
-    pub language: Option<&'a LanguageTag>,
-    /// Correlation UUID for tracing.
-    pub correlation_id: Option<Uuid>,
-}
-
-/// One per-call NER response from a [`NerBackend`].
-///
-/// Wraps the raw spans the backend produced. Pre-normalization:
-/// labels are still the backend's raw strings; the recognizer
-/// applies label-map + ignore-set + low-score demotion before
-/// emitting entities.
-#[derive(Debug, Clone, Default)]
-pub struct NerResponse {
-    /// Spans predicted for the request's text, in backend order.
-    pub spans: Vec<RawNerSpan>,
-}
-
-impl NerResponse {
-    /// Construct a response from raw spans.
-    #[must_use]
-    pub fn new(spans: Vec<RawNerSpan>) -> Self {
-        Self { spans }
-    }
-}
-
-/// Per-call NER backend.
-///
-/// Implemented by everything that turns `(text, kinds)` into raw
-/// NER spans — externalised inference services (BentoML), local
-/// model wrappers (future ORT/Candle backends), and the in-process
-/// no-op test stub.
-///
-/// Object-safe: recognizers hold `Arc<dyn NerBackend>` and dispatch
-/// per call.
-#[async_trait::async_trait]
-pub trait NerBackend: Send + Sync + 'static {
-    /// Backend identity (model / service name + provenance kind).
-    ///
-    /// Distinct from the recognizer's configured name: the
-    /// recognizer-level name (e.g. `"company-ner"`) labels the
-    /// configured slot, while [`provenance`] identifies the actual
-    /// model the backend wraps (e.g. `"noop-ner"`,
-    /// `"bento-ner"`).
-    ///
-    /// [`provenance`]: Self::provenance
-    fn provenance(&self) -> ModelProvenance;
-
-    /// Recognise spans for `request`. Returns raw
-    /// (pre-normalization) spans; the recognizer applies
-    /// label-map + ignore-set + low-score demotion on the way out.
-    ///
-    /// # Errors
-    ///
-    /// Returns the underlying transport / parse / inference error.
-    async fn recognize(&self, request: NerRequest<'_>) -> Result<NerResponse>;
-
-    /// Batched recognise. Defaults to a sequential fan-out;
-    /// backends with native batching should override.
-    ///
-    /// # Errors
-    ///
-    /// Returns the first error encountered.
-    async fn recognize_batch(&self, requests: &[NerRequest<'_>]) -> Result<Vec<NerResponse>> {
-        let mut out = Vec::with_capacity(requests.len());
-        for req in requests {
-            out.push(self.recognize(req.clone()).await?);
-        }
-        Ok(out)
-    }
-}
diff --git a/crates/nvisy-ner/src/backend/ner_span.rs b/crates/nvisy-ner/src/backend/ner_span.rs
deleted file mode 100644
index 63c21f133..000000000
--- a/crates/nvisy-ner/src/backend/ner_span.rs
+++ /dev/null
@@ -1,45 +0,0 @@
-//! [`RawNerSpan`]: pre-normalization NER prediction.
-//!
-//! Emitted by a [`NerBackend`] backend. The label is the *raw* string
-//! the model produced — `PER`, `LOC`, `ORG`, `B-PERSON`, etc. — left
-//! untranslated so [`NerRecognizer`] can apply its [`LabelMap`] and
-//! `labels_to_ignore` policy uniformly across engines.
-//!
-//! Score is the model's confidence in the raw `[0.0, 1.0]` range;
-//! the recognizer re-clamps to `Confidence` and may demote via the
-//! configured `low_score_multiplier`.
-//!
-//! [`NerBackend`]: super::NerBackend
-//! [`NerRecognizer`]: crate::NerRecognizer
-//! [`LabelMap`]: crate::LabelMap
-
-use std::ops::Range;
-
-/// One raw entity span predicted by a NER model.
-///
-/// Pre-normalization: the label is the model's string, not a
-/// canonical [`EntityLabelRef`]. Coordinate space is byte offsets
-/// into the source text the backend was called with.
-///
-/// [`EntityLabelRef`]: nvisy_core::entity::EntityLabelRef
-#[derive(Debug, Clone, PartialEq)]
-pub struct RawNerSpan {
-    /// Model-emitted label, verbatim.
-    pub label: String,
-    /// Model-emitted confidence, in `[0.0, 1.0]`. Out-of-range
-    /// values are the producer's bug to fix; consumers clamp.
-    pub score: f64,
-    /// Byte range of the prediction in the source text.
-    pub offset: Range<usize>,
-}
-
-impl RawNerSpan {
-    /// Construct a span.
-    pub fn new(label: impl Into<String>, score: f64, offset: Range<usize>) -> Self {
-        Self {
-            label: label.into(),
-            score,
-            offset,
-        }
-    }
-}
diff --git a/crates/nvisy-ner/src/backend/noop_backend.rs b/crates/nvisy-ner/src/backend/noop_backend.rs
deleted file mode 100644
index 3a0e8ed55..000000000
--- a/crates/nvisy-ner/src/backend/noop_backend.rs
+++ /dev/null
@@ -1,26 +0,0 @@
-//! [`NoopBackend`]: no-op [`NerBackend`] for tests + the default
-//! `NerBackend::Noop` config variant in `nvisy-engine`.
-
-use nvisy_core::Result;
-use nvisy_core::entity::ModelProvenance;
-
-use super::ner_backend::{NerBackend, NerRequest, NerResponse};
-
-/// No-op NER backend: every call returns an empty response.
-///
-/// Useful as a test stub and as the default `[detection.ner]` backend
-/// when the operator wants the NER recognizer wired but isn't ready to
-/// configure a real backend.
-#[derive(Debug, Default, Clone, Copy)]
-pub struct NoopBackend;
-
-#[async_trait::async_trait]
-impl NerBackend for NoopBackend {
-    fn provenance(&self) -> ModelProvenance {
-        ModelProvenance::new("noop-ner")
-    }
-
-    async fn recognize(&self, _request: NerRequest<'_>) -> Result<NerResponse> {
-        Ok(NerResponse::default())
-    }
-}
diff --git a/crates/nvisy-ner/src/lib.rs b/crates/nvisy-ner/src/lib.rs
deleted file mode 100644
index 82f30746a..000000000
--- a/crates/nvisy-ner/src/lib.rs
+++ /dev/null
@@ -1,11 +0,0 @@
-#![forbid(unsafe_code)]
-#![cfg_attr(docsrs, feature(doc_cfg))]
-#![doc = include_str!("../README.md")]
-
-pub mod backend;
-pub mod nlp;
-mod recognition;
-
-pub use self::recognition::{
-    LabelMap, NerModel, NerModelBuilder, NerRecognizer, NerRecognizerBuilder,
-};
diff --git a/crates/nvisy-ner/src/nlp/capabilities.rs b/crates/nvisy-ner/src/nlp/capabilities.rs
deleted file mode 100644
index 49d68ef3d..000000000
--- a/crates/nvisy-ner/src/nlp/capabilities.rs
+++ /dev/null
@@ -1,72 +0,0 @@
-//! [`NlpCapabilities`]: what an `NlpEngine` advertises it can
-//! produce.
-//!
-//! Composition-time contract between an `NlpEngine` and the
-//! recognizers / enhancer that read its artifacts. Lets the engine
-//! orchestrator refuse impossible asks at construction time — e.g.
-//! wiring a lemma-aware enhancer to a tokenizer-only
-//! engine that doesn't produce lemmas.
-//!
-//! Booleans rather than an enum because capabilities are
-//! independent — an engine may produce language only (Lingua),
-//! tokens + NER but no lemmas (a tokenizer + transformer model
-//! without a lemmatizer), or the full set (a hosted full-NLP
-//! service that includes lemmatization).
-
-/// Per-engine capability advertisement.
-///
-/// Fields are independent — each is `true` when the engine
-/// guarantees the corresponding artifact will be inserted into the
-/// shared `TypeMap` with meaningful data, `false` when the engine
-/// leaves it absent.
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub struct NlpCapabilities {
-    /// Engine emits tokens with byte offsets.
-    pub produces_tokens: bool,
-    /// Engine emits lemmas alongside tokens. Always implies
-    /// [`produces_tokens`].
-    ///
-    /// [`produces_tokens`]: Self::produces_tokens
-    pub produces_lemmas: bool,
-    /// Engine emits NER spans.
-    pub produces_ner: bool,
-    /// Engine emits a resolved stopword set.
-    pub produces_stopwords: bool,
-    /// Engine has native batch processing — calling `process_batch`
-    /// is more efficient than looping `process`.
-    pub batch_native: bool,
-}
-
-impl NlpCapabilities {
-    /// All capabilities off. Useful as a starting point for builder
-    /// patterns; for a truly capabilities-free engine, prefer
-    /// [`language_only`].
-    ///
-    /// [`language_only`]: Self::language_only
-    pub const NONE: Self = Self {
-        produces_tokens: false,
-        produces_lemmas: false,
-        produces_ner: false,
-        produces_stopwords: false,
-        batch_native: false,
-    };
-
-    /// Capabilities for an engine that only resolves language —
-    /// no tokens, no NER. Maps to `LinguaNlpEngine` in `nvisy-ner`.
-    pub const fn language_only() -> Self {
-        Self::NONE
-    }
-
-    /// Capabilities for an engine that produces tokens + lemmas +
-    /// NER + stopwords + native batching. Maps to a full
-    /// `BentoNlpEngine` (when the inference service supports it).
-    pub const fn full() -> Self {
-        Self {
-            produces_tokens: true,
-            produces_lemmas: true,
-            produces_ner: true,
-            produces_stopwords: true,
-            batch_native: true,
-        }
-    }
-}
diff --git a/crates/nvisy-ner/src/nlp/engine.rs b/crates/nvisy-ner/src/nlp/engine.rs
deleted file mode 100644
index 2766e9e5d..000000000
--- a/crates/nvisy-ner/src/nlp/engine.rs
+++ /dev/null
@@ -1,86 +0,0 @@
-//! [`NlpEngine`]: the producer-side trait that builds the
-//! shared-NLP-pass [`TypeMap`] for one or more texts.
-//!
-//! Engines stamp typed enrichment entries (`LanguageDetections`
-//! today; token artifacts when the upstream service supports
-//! them) into the returned `TypeMap`. An orchestrator that wants
-//! shared NLP runs `process` once per scan, wraps the result in
-//! [`Artifacts`], and attaches it to each [`RecognizerInput`] via
-//! [`RecognizerInput::with_artifacts`].
-//!
-//! [`Artifacts`]: nvisy_core::extraction::Artifacts
-//! [`RecognizerInput`]: nvisy_core::recognition::RecognizerInput
-//! [`RecognizerInput::with_artifacts`]: nvisy_core::recognition::RecognizerInput::with_artifacts
-//!
-//! Pluggable so different deployment shapes (pure language
-//! detection, hosted full-NLP service, future in-process model)
-//! can be wired interchangeably. The orchestrator calls `process`
-//! (or `process_batch`) once per scan; recognizers and the
-//! keyword-boost enhancer borrow the resulting map by reference.
-
-use nvisy_core::Result;
-use nvisy_core::primitive::LanguageTag;
-use type_map::concurrent::TypeMap;
-
-use super::capabilities::NlpCapabilities;
-
-/// Builds the shared-NLP-pass [`TypeMap`] for the orchestrator.
-///
-/// Engines advertise their capabilities via
-/// [`capabilities`] so the orchestrator can refuse impossible
-/// compositions at construction time (e.g. wiring a lemma-aware
-/// enhancer to an engine that doesn't produce lemmas). Engines also
-/// advertise the languages they support so a future per-language
-/// registry can route correctly.
-///
-/// `Send + Sync + 'static` — engines live behind `Arc<dyn _>` in
-/// the orchestrator and are shared across recognition tasks.
-///
-/// [`capabilities`]: Self::capabilities
-#[async_trait::async_trait]
-pub trait NlpEngine: Send + Sync + 'static {
-    /// Languages this engine can produce artifacts for. Empty when
-    /// the engine is language-agnostic (e.g. a tokenizer that
-    /// works on bytes alone).
-    fn supported_languages(&self) -> &[LanguageTag];
-
-    /// What the engine can produce. Advisory — consumers can still
-    /// call `process` even when capabilities are off, they just
-    /// get an empty map back.
-    fn capabilities(&self) -> NlpCapabilities;
-
-    /// Process one text. Returns a [`TypeMap`] populated with one
-    /// typed entry per enrichment the engine produced.
-    ///
-    /// `hint` is the caller-asserted language; engines that can
-    /// skip detection when given a hint should do so.
-    ///
-    /// # Errors
-    ///
-    /// Returns a runtime error when the underlying detection or
-    /// inference call fails. Empty input is not an error — engines
-    /// should return an empty (or sparsely populated) map.
-    async fn process(&self, text: &str, hint: Option<&LanguageTag>) -> Result<TypeMap>;
-
-    /// Process a batch of texts. The default fans out via
-    /// [`process`] concurrently; engines with
-    /// native batching (`capabilities().batch_native == true`)
-    /// should override.
-    ///
-    /// [`process`]: Self::process
-    ///
-    /// # Errors
-    ///
-    /// Returns the first error encountered.
-    async fn process_batch(
-        &self,
-        texts: &[&str],
-        hint: Option<&LanguageTag>,
-    ) -> Result<Vec<TypeMap>> {
-        let mut out = Vec::with_capacity(texts.len());
-        for text in texts {
-            out.push(self.process(text, hint).await?);
-        }
-        Ok(out)
-    }
-}
diff --git a/crates/nvisy-ner/src/nlp/lingua_detector.rs b/crates/nvisy-ner/src/nlp/lingua_detector.rs
deleted file mode 100644
index 30da92754..000000000
--- a/crates/nvisy-ner/src/nlp/lingua_detector.rs
+++ /dev/null
@@ -1,175 +0,0 @@
-//! [`LinguaDetector`]: thin wrapper around the
-//! [`lingua`] crate's
-//! `LanguageDetector`.
-//!
-//! Owns the lingua detector and exposes one method,
-//! [`detect`], that returns a
-//! [`Vec<LanguageDetection>`] in our ontology shape. Used by
-//! [`LinguaNlpEngine`].
-//!
-//! Construction takes either a candidate-language set or "all
-//! languages compiled into the lingua feature set" — the latter
-//! is the unrestricted fallback.
-//!
-//! [`lingua`]: https://crates.io/crates/lingua
-//! [`detect`]: LinguaDetector::detect
-//! [`LinguaNlpEngine`]: super::LinguaNlpEngine
-
-use std::collections::HashSet;
-use std::fmt;
-use std::str::FromStr;
-use std::sync::{Mutex, OnceLock};
-
-use lingua::{IsoCode639_1, Language, LanguageDetector as LinguaInner, LanguageDetectorBuilder};
-use nvisy_core::Result;
-use nvisy_core::primitive::{
-    Confidence, LanguageDetection, LanguageProvenance, LanguageSpan, LanguageTag,
-};
-
-/// Lingua-backed language detector.
-///
-/// Detects per-region languages — for mixed-language input,
-/// returns one [`LanguageDetection`] per detected region with a
-/// populated [`LanguageSpan`]. Monolingual input returns a single
-/// detection covering the whole text.
-pub struct LinguaDetector {
-    inner: LinguaInner,
-}
-
-impl LinguaDetector {
-    /// Construct a detector restricted to `tags`. Unrecognised
-    /// tags (no matching ISO 639-1 primary subtag in lingua) are
-    /// silently skipped. Returns `None` when no tag matched —
-    /// `LinguaNlpEngine` falls back to
-    /// [`for_all_languages`] in that
-    /// case.
-    ///
-    /// [`for_all_languages`]: Self::for_all_languages
-    pub(crate) fn for_languages(tags: &[LanguageTag]) -> Option<Self> {
-        let langs = tags_to_languages(tags);
-        if langs.is_empty() {
-            return None;
-        }
-        Some(Self {
-            inner: LanguageDetectorBuilder::from_languages(&langs).build(),
-        })
-    }
-
-    /// Construct a detector considering every language compiled
-    /// into the `lingua` crate's feature set.
-    pub(crate) fn for_all_languages() -> Self {
-        Self {
-            inner: LanguageDetectorBuilder::from_all_languages().build(),
-        }
-    }
-
-    /// Run detection on `text`.
-    ///
-    /// Empty for ambiguous inputs (lingua refused to commit), one
-    /// or more entries otherwise. Each entry has a populated
-    /// [`LanguageSpan`] so mixed-language input can be attributed
-    /// region-by-region.
-    pub fn detect(&self, text: &str) -> Result<Vec<LanguageDetection>> {
-        let detections = self
-            .inner
-            .detect_multiple_languages_of(text)
-            .into_iter()
-            .filter_map(|result| {
-                let language = lingua_to_tag(result.language())?;
-                let raw_confidence = self
-                    .inner
-                    .compute_language_confidence(text, result.language());
-                let confidence = Confidence::new(raw_confidence.clamp(0.0, 1.0));
-                Some(LanguageDetection {
-                    language,
-                    confidence,
-                    provenance: LanguageProvenance::Detected,
-                    span: Some(LanguageSpan {
-                        start: result.start_index(),
-                        end: result.end_index(),
-                    }),
-                })
-            })
-            .collect();
-        Ok(detections)
-    }
-}
-
-fn lingua_to_tag(lang: Language) -> Option<LanguageTag> {
-    let iso = lang.iso_code_639_1().to_string();
-    match iso.parse() {
-        Ok(tag) => Some(tag),
-        Err(e) => {
-            warn_once_unmappable(&iso, &e.to_string());
-            None
-        }
-    }
-}
-
-/// Cache of ISO codes we've already logged an "unmappable" warning
-/// for, so a hot detection loop doesn't spam the log with the same
-/// failure once per call. Lingua's code set is finite and fixed —
-/// real failures here are deterministic.
-fn warn_once_unmappable(iso: &str, error: &str) {
-    static SEEN: OnceLock<Mutex<HashSet<String>>> = OnceLock::new();
-    let seen = SEEN.get_or_init(|| Mutex::new(HashSet::new()));
-    let mut guard = match seen.lock() {
-        Ok(g) => g,
-        Err(poisoned) => poisoned.into_inner(),
-    };
-    if guard.insert(iso.to_owned()) {
-        tracing::warn!(
-            target: "nvisy_ner::nlp::lingua",
-            iso_code = %iso,
-            error = %error,
-            "lingua ISO 639-1 code did not parse as a BCP-47 LanguageTag (logged once per process)",
-        );
-    }
-}
-
-fn tags_to_languages(tags: &[LanguageTag]) -> Vec<Language> {
-    tags.iter()
-        .filter_map(|t| IsoCode639_1::from_str(t.primary_language()).ok())
-        .map(|iso| Language::from_iso_code_639_1(&iso))
-        .collect()
-}
-
-impl fmt::Debug for LinguaDetector {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("LinguaDetector").finish()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    fn english_only() -> LinguaDetector {
-        let tags = ["en".parse().unwrap()];
-        LinguaDetector::for_languages(&tags).expect("english is enabled")
-    }
-
-    #[test]
-    fn detects_english_sentence() {
-        let det = english_only();
-        let detections = det
-            .detect("The quick brown fox jumps over the lazy dog.")
-            .unwrap();
-        assert!(!detections.is_empty());
-        let first = &detections[0];
-        assert_eq!(first.language.primary_language(), "en");
-        assert!(first.span.is_some());
-    }
-
-    #[test]
-    fn empty_input_returns_empty_vec() {
-        let det = english_only();
-        assert!(det.detect("").unwrap().is_empty());
-    }
-
-    #[test]
-    fn rejects_construction_with_no_recognised_languages() {
-        let tags = ["xx".parse().unwrap()];
-        assert!(LinguaDetector::for_languages(&tags).is_none());
-    }
-}
diff --git a/crates/nvisy-ner/src/nlp/lingua_engine.rs b/crates/nvisy-ner/src/nlp/lingua_engine.rs
deleted file mode 100644
index de6ef0244..000000000
--- a/crates/nvisy-ner/src/nlp/lingua_engine.rs
+++ /dev/null
@@ -1,144 +0,0 @@
-//! [`LinguaNlpEngine`]: language-only [`NlpEngine`].
-//!
-//! Stamps a [`LanguageDetections`] entry onto the artifact
-//! [`TypeMap`]; never produces tokens, NER, or stopwords. Used by
-//! pattern-only pipelines that still want a resolved language
-//! carried on the artifact (for engine-level routing or for
-//! downstream language-aware policies).
-//!
-//! When `process` receives a `hint`, detection is skipped and the
-//! hint becomes a single-entry detection with provenance
-//! [`Asserted`].
-//!
-//! [`Asserted`]: nvisy_core::primitive::LanguageProvenance::Asserted
-
-use nvisy_core::Result;
-use nvisy_core::primitive::{
-    LanguageDetection, LanguageDetections, LanguageProvenance, LanguageTag,
-};
-use type_map::concurrent::TypeMap;
-
-use super::capabilities::NlpCapabilities;
-use super::engine::NlpEngine;
-use super::lingua_detector::LinguaDetector;
-
-/// Lingua-backed language-only NLP engine.
-///
-/// Stateless: every `process` call builds a fresh
-/// [`LinguaDetector`] for the requested language scope (the
-/// candidate set passed at construction, or "all languages"). The
-/// scope is locked in at construction time — there is no per-call
-/// scope override; pattern pipelines that need different scopes
-/// per call should hold multiple engines.
-#[derive(Debug, Clone)]
-pub struct LinguaNlpEngine {
-    candidates: Vec<LanguageTag>,
-}
-
-impl LinguaNlpEngine {
-    /// Construct an engine that considers every language compiled
-    /// into the `lingua` crate's feature set.
-    #[must_use]
-    pub fn unrestricted() -> Self {
-        Self {
-            candidates: Vec::new(),
-        }
-    }
-
-    /// Construct an engine restricted to `candidates`. Tags lingua
-    /// doesn't recognise are silently skipped at detector-build
-    /// time. An empty input is equivalent to
-    /// [`unrestricted`].
-    ///
-    /// [`unrestricted`]: Self::unrestricted
-    #[must_use]
-    pub fn with_candidates(candidates: impl IntoIterator<Item = LanguageTag>) -> Self {
-        Self {
-            candidates: candidates.into_iter().collect(),
-        }
-    }
-
-    fn detector(&self) -> LinguaDetector {
-        if self.candidates.is_empty() {
-            LinguaDetector::for_all_languages()
-        } else {
-            LinguaDetector::for_languages(&self.candidates)
-                .unwrap_or_else(LinguaDetector::for_all_languages)
-        }
-    }
-}
-
-impl Default for LinguaNlpEngine {
-    fn default() -> Self {
-        Self::unrestricted()
-    }
-}
-
-#[async_trait::async_trait]
-impl NlpEngine for LinguaNlpEngine {
-    fn supported_languages(&self) -> &[LanguageTag] {
-        &self.candidates
-    }
-
-    fn capabilities(&self) -> NlpCapabilities {
-        NlpCapabilities::language_only()
-    }
-
-    async fn process(&self, text: &str, hint: Option<&LanguageTag>) -> Result<TypeMap> {
-        let detections = if let Some(language) = hint {
-            vec![LanguageDetection {
-                language: language.clone(),
-                confidence: None,
-                provenance: LanguageProvenance::Asserted,
-                span: None,
-            }]
-        } else {
-            self.detector().detect(text)?
-        };
-        let mut artifacts = TypeMap::new();
-        artifacts.insert(LanguageDetections::new(detections));
-        Ok(artifacts)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[tokio::test]
-    async fn detects_english_without_hint() {
-        let engine = LinguaNlpEngine::unrestricted();
-        let artifacts = engine
-            .process("The quick brown fox jumps over the lazy dog.", None)
-            .await
-            .unwrap();
-        let langs = artifacts
-            .get::<LanguageDetections>()
-            .expect("language detection entry stamped");
-        let lang = langs.dominant().expect("language detected");
-        assert_eq!(lang.language.primary_language(), "en");
-        assert!(matches!(lang.provenance, LanguageProvenance::Detected));
-    }
-
-    #[tokio::test]
-    async fn asserted_hint_bypasses_detection() {
-        let engine = LinguaNlpEngine::unrestricted();
-        let asserted: LanguageTag = "de".parse().unwrap();
-        let artifacts = engine
-            .process("The quick brown fox", Some(&asserted))
-            .await
-            .unwrap();
-        let langs = artifacts.get::<LanguageDetections>().unwrap();
-        let lang = langs.dominant().unwrap();
-        assert_eq!(lang.language.primary_language(), "de");
-        assert!(matches!(lang.provenance, LanguageProvenance::Asserted));
-    }
-
-    #[tokio::test]
-    async fn capabilities_are_language_only() {
-        let engine = LinguaNlpEngine::unrestricted();
-        let caps = engine.capabilities();
-        assert!(!caps.produces_tokens);
-        assert!(!caps.produces_ner);
-    }
-}
diff --git a/crates/nvisy-ner/src/nlp/mod.rs b/crates/nvisy-ner/src/nlp/mod.rs
deleted file mode 100644
index ab265acc8..000000000
--- a/crates/nvisy-ner/src/nlp/mod.rs
+++ /dev/null
@@ -1,37 +0,0 @@
-//! Producer side of the shared-NLP-pass primitive.
-//!
-//! Consumer-side types live in their natural crates so any text
-//! consumer can read them without depending on this one:
-//! [`LanguageDetections`] sits with the language primitives in
-//! `nvisy-core`; the optional token artifact lives in
-//! `nvisy-context` next to its only consumer (the keyword-boost
-//! `Enhancer`). This module declares the [`NlpEngine`] trait and
-//! the engines that produce those artifacts into the shared
-//! `TypeMap` stamped on `RecognizerInput.artifacts`.
-//!
-//! One engine ships today:
-//! - [`LinguaNlpEngine`] — language-only NLP, backed by the
-//!   [`lingua`] crate. Used by pattern-only pipelines that still
-//!   want a resolved language carried on the artifact.
-//!
-//! NER lives in a separate trait-driven path ([`NerBackend`] +
-//! [`NerRecognizer`]) — zero-shot and fixed-label NER backends plug
-//! in there, not here.
-//!
-//! The trait is async because realistic implementations are
-//! HTTP-bound or otherwise yield.
-//!
-//! [`LanguageDetections`]: nvisy_core::primitive::LanguageDetections
-//! [`lingua`]: https://crates.io/crates/lingua
-//! [`NerBackend`]: crate::backend::NerBackend
-//! [`NerRecognizer`]: crate::NerRecognizer
-
-mod capabilities;
-mod engine;
-mod lingua_detector;
-mod lingua_engine;
-
-pub use self::capabilities::NlpCapabilities;
-pub use self::engine::NlpEngine;
-pub use self::lingua_detector::LinguaDetector;
-pub use self::lingua_engine::LinguaNlpEngine;
diff --git a/crates/nvisy-ner/src/recognition/aggregation.rs b/crates/nvisy-ner/src/recognition/aggregation.rs
deleted file mode 100644
index 7e5f59939..000000000
--- a/crates/nvisy-ner/src/recognition/aggregation.rs
+++ /dev/null
@@ -1,59 +0,0 @@
-//! [`AggregationStrategy`] and [`AlignmentMode`]: policies for
-//! collapsing per-token NER predictions into entity spans.
-//!
-//! The producer engine may apply them server-side (the Bento
-//! `inference-gliner` already returns aggregated spans), in which
-//! case the consumer-side knobs are advisory; or the producer may
-//! emit unaggregated token-classification output, in which case
-//! [`NerRecognizer`] applies them itself.
-//!
-//! [`NerRecognizer`]: super::NerRecognizer
-
-use serde::{Deserialize, Serialize};
-
-/// How adjacent token-level predictions of the same entity type are
-/// merged into one span. Names follow HuggingFace
-/// `transformers.pipeline(aggregation_strategy=...)` semantics.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
-#[serde(rename_all = "lowercase")]
-pub enum AggregationStrategy {
-    /// No aggregation: every token-level prediction becomes its own
-    /// span. Cheapest, noisiest.
-    None,
-    /// Merge adjacent tokens that share a label. The span's score
-    /// is the first token's. Useful when later tokens of a span
-    /// continue the same entity (BIO-style `B-PER` then `I-PER`).
-    First,
-    /// Merge adjacent same-label tokens; the span's score is the
-    /// maximum across constituent tokens. The default — works best
-    /// when models output per-token confidences without a strong
-    /// boundary signal.
-    #[default]
-    Max,
-    /// Merge adjacent same-label tokens; the span's score is the
-    /// arithmetic mean across constituent tokens. Smoother than
-    /// [`Max`], more lenient at span boundaries.
-    ///
-    /// [`Max`]: Self::Max
-    Average,
-    /// Merge based on `B-`/`I-`/`O` BIO tags rather than label
-    /// equality. The span's score is the first token's. Match
-    /// HuggingFace `"simple"` strategy.
-    Simple,
-}
-
-/// How a token-classification span snaps to character boundaries
-/// when the model emits sub-word predictions. Mirrors spaCy
-/// `Doc.char_span(alignment_mode=...)`.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
-#[serde(rename_all = "lowercase")]
-pub enum AlignmentMode {
-    /// Reject spans that don't land on a token boundary.
-    Strict,
-    /// Shrink the span to the next inner token boundary.
-    Contract,
-    /// Expand the span to the next outer token boundary. Default —
-    /// favors recall over precision.
-    #[default]
-    Expand,
-}
diff --git a/crates/nvisy-ner/src/recognition/config.rs b/crates/nvisy-ner/src/recognition/config.rs
deleted file mode 100644
index e8216d005..000000000
--- a/crates/nvisy-ner/src/recognition/config.rs
+++ /dev/null
@@ -1,102 +0,0 @@
-//! [`NerModel`]: client-side NER tuning knobs.
-//!
-//! Applied inside [`NerRecognizer`] before entities are emitted,
-//! so backends stay dumb and label normalization is uniform
-//! across them.
-//!
-//! Construct via [`NerModel::default`] for the canonical defaults
-//! (canonical label map, no ignored labels, score = 0.85, no
-//! demotion) — these match what the current `inference-gliner`
-//! service produces — or via [`NerModel::builder`] for an infallible
-//! chainable builder that overrides only the fields you care about.
-//!
-//! [`NerRecognizer`]: super::NerRecognizer
-
-use std::collections::HashSet;
-
-use derive_builder::Builder;
-use nvisy_core::primitive::Confidence;
-use nvisy_core::recognition::LabelMap;
-
-use super::aggregation::{AggregationStrategy, AlignmentMode};
-
-/// Per-recognizer NER policy.
-#[derive(Debug, Clone, Builder)]
-#[builder(
-    name = "NerModelBuilder",
-    pattern = "owned",
-    setter(into, prefix = "with"),
-    build_fn(skip)
-)]
-pub struct NerModel {
-    /// Translation from raw model labels to canonical entity
-    /// label names. Defaults to [`LabelMap::canonical`].
-    pub label_map: LabelMap,
-    /// Raw labels the adapter drops without translation. Useful
-    /// for filtering out labels the model emits but we don't care
-    /// about (e.g. `O` from BIO tagging, `MISC` from generic NER
-    /// schemas).
-    pub labels_to_ignore: HashSet<String>,
-    /// Fallback confidence when a backend emits a score outside
-    /// `[0.0, 1.0]` (treated as a bug; clamped + this used as the
-    /// safe default).
-    pub default_score: Confidence,
-    /// Entity label names whose emitted confidence is multiplied
-    /// by `low_score_multiplier` before being surfaced. Use for
-    /// noisy-but-high-recall labels.
-    pub low_score_labels: HashSet<String>,
-    /// Multiplier applied to `low_score_labels`. Must be in
-    /// `[0.0, 1.0]`.
-    pub low_score_multiplier: f64,
-    /// Aggregation policy for backends that emit token-level
-    /// predictions. Advisory for bac kends that aggregate
-    /// server-side.
-    pub aggregation: AggregationStrategy,
-    /// Alignment policy for sub-word predictions. Same advisory
-    /// status as `aggregation`.
-    pub alignment: AlignmentMode,
-}
-
-impl Default for NerModel {
-    fn default() -> Self {
-        Self {
-            label_map: LabelMap::canonical(),
-            labels_to_ignore: HashSet::new(),
-            default_score: Confidence::new(0.85).expect("0.85 in range"),
-            low_score_labels: HashSet::new(),
-            low_score_multiplier: 0.4,
-            aggregation: AggregationStrategy::Max,
-            alignment: AlignmentMode::Expand,
-        }
-    }
-}
-
-impl NerModel {
-    /// Start a chainable, infallible builder seeded from the
-    /// canonical defaults — every field has a sensible default, so
-    /// callers only override what they care about.
-    #[must_use]
-    pub fn builder() -> NerModelBuilder {
-        NerModelBuilder::default()
-    }
-}
-
-impl NerModelBuilder {
-    /// Finish the builder, filling every unset field with its
-    /// default. Infallible — no required fields.
-    #[must_use]
-    pub fn build(self) -> NerModel {
-        let defaults = NerModel::default();
-        NerModel {
-            label_map: self.label_map.unwrap_or(defaults.label_map),
-            labels_to_ignore: self.labels_to_ignore.unwrap_or(defaults.labels_to_ignore),
-            default_score: self.default_score.unwrap_or(defaults.default_score),
-            low_score_labels: self.low_score_labels.unwrap_or(defaults.low_score_labels),
-            low_score_multiplier: self
-                .low_score_multiplier
-                .unwrap_or(defaults.low_score_multiplier),
-            aggregation: self.aggregation.unwrap_or(defaults.aggregation),
-            alignment: self.alignment.unwrap_or(defaults.alignment),
-        }
-    }
-}
diff --git a/crates/nvisy-ner/src/recognition/mod.rs b/crates/nvisy-ner/src/recognition/mod.rs
deleted file mode 100644
index 295288f07..000000000
--- a/crates/nvisy-ner/src/recognition/mod.rs
+++ /dev/null
@@ -1,21 +0,0 @@
-//! Recognizer layer: the [`NerRecognizer`] that drives any
-//! [`NerBackend`] backend, the [`NerModel`] normalization knobs it
-//! applies to raw spans, and the [`LabelMap`] translation table
-//! shared across backends.
-//!
-//! Implements [`EntityRecognizer<Text>`] so it composes with the
-//! rest of the platform through the same trait every other text
-//! recognizer uses.
-//!
-//! [`NerBackend`]: crate::backend::NerBackend
-//! [`EntityRecognizer<Text>`]: nvisy_core::recognition::EntityRecognizer
-//! [`Text`]: nvisy_core::modality::Text
-
-mod aggregation;
-mod config;
-mod recognizer;
-
-pub use nvisy_core::recognition::LabelMap;
-
-pub use self::config::{NerModel, NerModelBuilder};
-pub use self::recognizer::{NerRecognizer, NerRecognizerBuilder};
diff --git a/crates/nvisy-ner/src/recognition/recognizer.rs b/crates/nvisy-ner/src/recognition/recognizer.rs
deleted file mode 100644
index 1c4647e66..000000000
--- a/crates/nvisy-ner/src/recognition/recognizer.rs
+++ /dev/null
@@ -1,204 +0,0 @@
-//! [`NerRecognizer`]: unified NER recognizer that drives any
-//! [`NerBackend`] backend.
-//!
-//! Holds an `Arc<dyn NerBackend>` plus a [`NerModel`] (label-map +
-//! ignore set + low-score demotion knobs) plus the recognizer's
-//! advertised [`supported_kinds`]. On each `recognize` call it asks
-//! the backend for spans (passing `Some(&supported_kinds)` when
-//! non-empty for zero-shot backends, `None` when empty for
-//! fixed-label backends), then normalizes through the model and
-//! emits entities.
-//!
-//! Implements [`EntityRecognizer<Text>`] so it composes with the
-//! rest of the platform through the same trait every other text
-//! recognizer uses.
-//!
-//! [`supported_kinds`]: NerRecognizer::supported_kinds
-
-use std::sync::Arc;
-
-use derive_builder::Builder;
-use nvisy_core::entity::{Entity, EntityLabelRef, ModelProvenance, TrailProvenance, TrailStep};
-use nvisy_core::modality::{Text, TextLocation};
-use nvisy_core::primitive::Confidence;
-use nvisy_core::recognition::{EntityRecognizer, RecognizerInput, RecognizerOutput};
-use nvisy_core::{Error, Result};
-
-use super::config::NerModel;
-use crate::backend::{NerBackend, NerRequest, RawNerSpan};
-
-/// Trait-driven NER recognizer.
-#[derive(Clone, Builder)]
-#[builder(
-    name = "NerRecognizerBuilder",
-    pattern = "owned",
-    setter(into, prefix = "with"),
-    build_fn(error = "Error", name = "try_build", private)
-)]
-pub struct NerRecognizer {
-    /// Recognizer name. Surfaced in trail provenance on every
-    /// emitted entity.
-    name: String,
-    /// Backend that turns `(text, kinds)` into raw spans. Required.
-    /// Set via [`with_engine`], which accepts any concrete
-    /// [`NerBackend`] impl by value and wraps it in `Arc` internally.
-    ///
-    /// [`with_engine`]: NerRecognizerBuilder::with_engine
-    #[builder(setter(custom))]
-    engine: Arc<dyn NerBackend>,
-    /// Labels the recognizer advertises. When non-empty, the
-    /// recognizer asks the backend for only this subset on every
-    /// call (zero-shot path). When empty, the backend is asked for
-    /// whatever it natively produces (fixed-label path).
-    #[builder(default)]
-    supported_labels: Vec<EntityLabelRef>,
-    /// Normalization knobs applied to the backend's raw output
-    /// before entities are emitted.
-    #[builder(default)]
-    model: NerModel,
-}
-
-impl NerRecognizer {
-    /// Start the chainable builder. `name` and `engine` are
-    /// required — calling [`build`] without them returns a
-    /// validation error.
-    ///
-    /// [`build`]: NerRecognizerBuilder::build
-    #[must_use]
-    pub fn builder() -> NerRecognizerBuilder {
-        NerRecognizerBuilder::default()
-    }
-
-    /// Recognizer name.
-    #[must_use]
-    pub fn name(&self) -> &str {
-        &self.name
-    }
-
-    /// Labels this recognizer advertises.
-    #[must_use]
-    pub fn supported_labels(&self) -> &[EntityLabelRef] {
-        &self.supported_labels
-    }
-
-    /// Borrow the normalization config.
-    #[must_use]
-    pub fn model(&self) -> &NerModel {
-        &self.model
-    }
-
-    fn build_entity(&self, span: &RawNerSpan, label: EntityLabelRef) -> Entity<Text> {
-        let raw_confidence =
-            Confidence::try_clamped(span.score).unwrap_or(self.model.default_score);
-        let confidence = if self.model.low_score_labels.contains(label.as_str()) {
-            let demoted = raw_confidence.get() * self.model.low_score_multiplier;
-            Confidence::try_clamped(demoted).unwrap_or(self.model.default_score)
-        } else {
-            raw_confidence
-        };
-        let provenance = TrailProvenance::Model(ModelProvenance::new(self.name.clone()));
-        let reason = format!("recognizer `{}` identified {label}", self.name);
-        let step = TrailStep::recognition("ner", confidence, provenance, reason);
-        Entity::builder()
-            .with_label(label)
-            .with_trail(vec![step])
-            .with_confidence(confidence)
-            .with_location(TextLocation::new(span.offset.start, span.offset.end))
-            .build()
-            .expect("required fields provided")
-    }
-}
-
-impl NerRecognizerBuilder {
-    /// Set the [`NerBackend`] backend that powers this recognizer.
-    /// Accepts any concrete impl by value and wraps it in `Arc`.
-    /// Required — `build` errors when this hasn't been called.
-    #[must_use]
-    pub fn with_engine<E: NerBackend>(mut self, engine: E) -> Self {
-        self.engine = Some(Arc::new(engine));
-        self
-    }
-
-    /// Finish the builder. Errors when `name` or `engine` is unset.
-    pub fn build(self) -> Result<NerRecognizer> {
-        self.try_build()
-    }
-}
-
-#[async_trait::async_trait]
-impl EntityRecognizer<Text> for NerRecognizer {
-    async fn recognize(&self, input: &RecognizerInput<Text>) -> Result<RecognizerOutput<Text>> {
-        let supported_borrowed: Vec<&str> = self
-            .supported_labels
-            .iter()
-            .map(EntityLabelRef::as_str)
-            .collect();
-        let labels = if supported_borrowed.is_empty() {
-            None
-        } else {
-            Some(supported_borrowed.as_slice())
-        };
-        let request = NerRequest {
-            text: input.data.text.as_str(),
-            labels,
-            language: input.language.as_ref(),
-            correlation_id: input.correlation_id,
-        };
-        let response = self.engine.recognize(request).await?;
-
-        let entities: Vec<Entity<Text>> = response
-            .spans
-            .iter()
-            .filter(|s| !self.model.labels_to_ignore.contains(s.label.as_str()))
-            .filter_map(|s| {
-                self.model
-                    .label_map
-                    .lookup(&s.label)
-                    .filter(|name| {
-                        self.supported_labels.is_empty()
-                            || self.supported_labels.iter().any(|sl| sl == *name)
-                    })
-                    .cloned()
-                    .map(|name| self.build_entity(s, name))
-            })
-            .collect();
-        Ok(RecognizerOutput::new(entities))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::builtins;
-    use nvisy_core::modality::TextData;
-
-    use super::*;
-    use crate::backend::NoopBackend;
-
-    #[tokio::test]
-    async fn noop_engine_yields_no_entities() {
-        let rec = NerRecognizer::builder()
-            .with_name("test")
-            .with_engine(NoopBackend)
-            .with_supported_labels(vec![
-                EntityLabelRef::from(builtins::PERSON_NAME.name.clone()),
-                EntityLabelRef::from(builtins::EMAIL_ADDRESS.name.clone()),
-            ])
-            .build()
-            .expect("builder succeeds");
-        let input = RecognizerInput::new(TextData::new("Alice Smith"));
-        let out = rec.recognize(&input).await.unwrap();
-        assert!(out.entities.is_empty());
-    }
-
-    #[tokio::test]
-    async fn empty_supported_labels_passes_none_to_engine() {
-        let rec = NerRecognizer::builder()
-            .with_name("test")
-            .with_engine(NoopBackend)
-            .build()
-            .expect("builder succeeds");
-        let input = RecognizerInput::new(TextData::new("Alice Smith"));
-        let out = rec.recognize(&input).await.unwrap();
-        assert!(out.entities.is_empty());
-    }
-}
diff --git a/crates/nvisy-ocr/Cargo.toml b/crates/nvisy-ocr/Cargo.toml
deleted file mode 100644
index dc8999fd6..000000000
--- a/crates/nvisy-ocr/Cargo.toml
+++ /dev/null
@@ -1,56 +0,0 @@
-# https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[package]
-name = "nvisy-ocr"
-description = "OCR provider integrations & backend abstraction for the Nvisy runtime"
-keywords = ["nvisy", "ocr", "vision"]
-categories = ["text-processing", "computer-vision"]
-readme = "README.md"
-
-version = { workspace = true }
-rust-version = { workspace = true }
-edition = { workspace = true }
-license = { workspace = true }
-publish = { workspace = true }
-
-authors = { workspace = true }
-repository = { workspace = true }
-homepage = { workspace = true }
-documentation = { workspace = true }
-
-[features]
-default = []
-
-## Enables [`BentoOcrBackend`]: externalised OCR over an HTTP
-## inference service (the `inference-ocr` Bento in
-## https://github.com/nvisycom/inference). Scaffolding only until
-## the wire contract is finalised upstream — see
-## https://github.com/nvisycom/runtime/issues/128.
-bento = ["dep:bentoml"]
-
-[package.metadata.docs.rs]
-all-features = true
-rustdoc-args = ["--cfg", "docsrs"]
-
-[dependencies]
-# Internal crates
-nvisy-core = { workspace = true, features = [] }
-
-# Primitive datatypes (UUIDv7 for per-call request IDs from the
-# Bento backend; behind the `bento` feature gate at use sites).
-uuid = { workspace = true, features = ["v4", "v7"] }
-bytes = { workspace = true, features = [] }
-
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-futures = { workspace = true, features = [] }
-tokio = { workspace = true, features = ["rt"] }
-
-# Observability
-tracing = { workspace = true, features = [] }
-
-# AI / LLM frameworks (feature-gated)
-bentoml = { workspace = true, optional = true }
-
-[dev-dependencies]
-tokio = { workspace = true, features = ["macros", "rt"] }
diff --git a/crates/nvisy-ocr/README.md b/crates/nvisy-ocr/README.md
deleted file mode 100644
index 5b72195cb..000000000
--- a/crates/nvisy-ocr/README.md
+++ /dev/null
@@ -1,41 +0,0 @@
-# nvisy-ocr
-
-[![Build](https://img.shields.io/github/actions/workflow/status/nvisycom/runtime/build.yml?branch=main&label=build%20%26%20test&style=flat-square)](https://github.com/nvisycom/runtime/actions/workflows/build.yml)
-
-OCR `Backend` trait and dispatch `Extractor` for the Nvisy runtime.
-
-## Overview
-
-`Extractor` wraps any `Backend` implementation as an
-`Extractor<Image>` consumed by the document orchestrator; each
-backend produces unified `OcrOutput` (text + line/word geometry).
-Shipped backends: `NoopBackend` (default, returns no results — used
-in tests and OCR-less deployments) and the feature-gated
-`BentoBackend`, an HTTP client for the externalised `inference-ocr`
-Bento in [`nvisycom/inference`] (scaffolding only; tracked under
-[#128]).
-
-VLM-mediated entity verification (the LLM-side counterpart that
-checks image-side entity proposals against the source image) lives
-in `nvisy-llm`, not here.
-
-[`nvisycom/inference`]: https://github.com/nvisycom/inference
-[#128]: https://github.com/nvisycom/runtime/issues/128
-
-## Documentation
-
-See [`docs/`](../../docs/) for architecture, security, and API documentation.
-
-## Changelog
-
-See [CHANGELOG.md](../../CHANGELOG.md) for release notes and version history.
-
-## License
-
-Apache 2.0 License, see [LICENSE.txt](../../LICENSE.txt)
-
-## Support
-
-- **Documentation**: [docs.nvisy.com](https://docs.nvisy.com)
-- **Issues**: [GitHub Issues](https://github.com/nvisycom/runtime/issues)
-- **Email**: [support@nvisy.com](mailto:support@nvisy.com)
diff --git a/crates/nvisy-ocr/src/backend/bento_backend.rs b/crates/nvisy-ocr/src/backend/bento_backend.rs
deleted file mode 100644
index a475328b1..000000000
--- a/crates/nvisy-ocr/src/backend/bento_backend.rs
+++ /dev/null
@@ -1,91 +0,0 @@
-//! [`BentoBackend`] — externalised OCR over an `inference-ocr`
-//! Bento in [`nvisycom/inference`].
-//!
-//! **Scaffolding only.** The wire contract has not been finalised
-//! upstream (tracked under [#128]). The struct, parameters, and
-//! [`OcrBackend`] impl exist so config wiring and feature gates can
-//! land now and the externalised path becomes a one-file change
-//! when the contract is ready. Until then [`OcrBackend::run`] returns
-//! a clear runtime error rather than pretending to call the
-//! service.
-//!
-//! Once the wire types are pinned, this file mirrors
-//! [`bento_types`] the same way `nvisy-ner`'s `BentoBackend`
-//! mirrors `nvisy_core.ner.v1`.
-//!
-//! [`OcrBackend`]: super::ocr_backend::OcrBackend
-//! [`bento_types`]: super::bento_types
-//! [`nvisycom/inference`]: https://github.com/nvisycom/inference
-//! [#128]: https://github.com/nvisycom/runtime/issues/128
-
-use bentoml::prelude::*;
-use nvisy_core::Error;
-use nvisy_core::entity::ModelProvenance;
-
-use super::ocr_backend::{OcrBackend, OcrRequest, OcrResponse};
-
-/// Parameters for [`BentoBackend`].
-#[derive(Debug, Clone)]
-pub struct BentoParams {
-    /// Base URL of the `inference-ocr` Bento (e.g. `http://localhost:3001`).
-    pub base_url: String,
-}
-
-impl BentoParams {
-    /// Construct with the given service URL.
-    pub fn new(base_url: impl Into<String>) -> Self {
-        Self {
-            base_url: base_url.into(),
-        }
-    }
-}
-
-/// [`OcrBackend`] that will call an externalised OCR Bento over HTTP.
-///
-/// **Not yet functional.** Both [`OcrBackend::extract`] and
-/// [`OcrBackend::extract_batch`] return a clear runtime error until
-/// the inference repo finalises the OCR wire contract — see [#128].
-///
-/// Constructing one still validates the configured `base_url`
-/// (via the underlying [`Client`] builder) so config errors
-/// surface at startup rather than at the first request, even
-/// though the request path itself is stubbed out.
-#[derive(Debug)]
-pub struct BentoBackend;
-
-impl BentoBackend {
-    /// Build a backend against the given parameters.
-    ///
-    /// # Errors
-    ///
-    /// Returns an error if the underlying HTTP client cannot be
-    /// constructed (invalid `base_url`).
-    pub fn new(params: BentoParams) -> Result<Self, Error> {
-        // Validate the URL eagerly so misconfiguration surfaces at
-        // startup. The constructed client is discarded today; once
-        // the wire contract lands it moves to a field and the
-        // request path picks it up.
-        Client::builder()
-            .with_base_url(&params.base_url)
-            .build()
-            .map_err(|e| Error::runtime(format!("bentoml client init: {e}"), "ocr-bento", false))?;
-        Ok(Self)
-    }
-}
-
-#[async_trait::async_trait]
-impl OcrBackend for BentoBackend {
-    fn provenance(&self) -> ModelProvenance {
-        ModelProvenance::new("bento-ocr")
-    }
-
-    async fn extract(&self, _request: OcrRequest<'_>) -> Result<OcrResponse, Error> {
-        Err(Error::runtime(
-            "BentoBackend is scaffolded; the inference-ocr wire \
-             contract has not been finalised yet — see \
-             https://github.com/nvisycom/runtime/issues/128",
-            "ocr-bento",
-            false,
-        ))
-    }
-}
diff --git a/crates/nvisy-ocr/src/backend/bento_types.rs b/crates/nvisy-ocr/src/backend/bento_types.rs
deleted file mode 100644
index 7c01b92ab..000000000
--- a/crates/nvisy-ocr/src/backend/bento_types.rs
+++ /dev/null
@@ -1,20 +0,0 @@
-//! Wire types for the externalised `inference-ocr` Bento.
-//!
-//! **Scaffolding only — the wire contract has not been finalised
-//! in [`nvisycom/inference`].** The types below are placeholders so
-//! `BentoBackend` can compile and round-trip through serde, but
-//! the real shape will be defined alongside the service and pulled
-//! in here once that work lands.
-//!
-//! When the contract finalises, this module will mirror
-//! `nvisy_core.ocr.v1` the same way `nvisy-ner`'s `backend/bento_types`
-//! mirrors `nvisy_core.ner.v1`: each struct documented in place,
-//! field names camelCase on the wire, schema version baked into
-//! the module name.
-//!
-//! [`nvisycom/inference`]: https://github.com/nvisycom/inference
-
-// TODO(#128): mirror `nvisy_core.ocr.v1` here once the inference
-// repo finalises the OCR wire contract. Until then `BentoBackend`
-// returns a clear runtime error from its trait methods rather than
-// pretending to call the service.
diff --git a/crates/nvisy-ocr/src/backend/mod.rs b/crates/nvisy-ocr/src/backend/mod.rs
deleted file mode 100644
index 213b764c9..000000000
--- a/crates/nvisy-ocr/src/backend/mod.rs
+++ /dev/null
@@ -1,39 +0,0 @@
-//! Backend layer: the [`OcrBackend`] trait and its shipped impls.
-//!
-//! Built-in [`NoopBackend`] (returns no blocks; test stub) and
-//! feature-gated [`BentoBackend`] (HTTP call into the externalised
-//! `inference-ocr` service, scaffolded against the future wire
-//! contract).
-//!
-//! Selecting which backend to use at deployment time is the
-//! consumer's concern — `nvisy-ocr` itself doesn't enumerate "all
-//! backends" in a closed enum.
-
-mod noop_backend;
-mod ocr_backend;
-
-#[cfg(feature = "bento")]
-mod bento_backend;
-#[cfg(feature = "bento")]
-mod bento_types;
-
-#[cfg(feature = "bento")]
-#[cfg_attr(docsrs, doc(cfg(feature = "bento")))]
-pub use self::bento_backend::{BentoBackend, BentoParams};
-pub use self::noop_backend::NoopBackend;
-pub use self::ocr_backend::{OcrBackend, OcrRequest, OcrResponse};
-
-#[cfg(test)]
-mod tests {
-    use bytes::Bytes;
-
-    use super::*;
-
-    #[tokio::test]
-    async fn noop_returns_empty() {
-        let backend = NoopBackend::new();
-        let image = Bytes::from(vec![0u8; 8]);
-        let response = backend.extract(OcrRequest::new(&image)).await.unwrap();
-        assert_eq!(response.blocks.len(), 0);
-    }
-}
diff --git a/crates/nvisy-ocr/src/backend/noop_backend.rs b/crates/nvisy-ocr/src/backend/noop_backend.rs
deleted file mode 100644
index e93b9cdbe..000000000
--- a/crates/nvisy-ocr/src/backend/noop_backend.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-//! [`NoopBackend`] — returns no OCR results. Default selection
-//! for deployments where the runtime should accept image content
-//! but isn't expected to recognise text in it.
-//!
-//! Useful in tests, as a placeholder while wiring up a real
-//! externalised backend, and for redaction pipelines that operate
-//! purely on metadata or on entities sourced from elsewhere.
-
-use nvisy_core::Error;
-use nvisy_core::entity::ModelProvenance;
-
-use super::ocr_backend::{OcrBackend, OcrRequest, OcrResponse};
-
-/// An [`OcrBackend`] that produces no OCR results.
-///
-/// Every call returns an empty response.
-#[derive(Debug, Default, Clone, Copy)]
-pub struct NoopBackend;
-
-impl NoopBackend {
-    /// Construct an empty backend.
-    pub fn new() -> Self {
-        Self
-    }
-}
-
-#[async_trait::async_trait]
-impl OcrBackend for NoopBackend {
-    fn provenance(&self) -> ModelProvenance {
-        ModelProvenance::new("noop-ocr")
-    }
-
-    async fn extract(&self, _request: OcrRequest<'_>) -> Result<OcrResponse, Error> {
-        Ok(OcrResponse::default())
-    }
-}
diff --git a/crates/nvisy-ocr/src/backend/ocr_backend.rs b/crates/nvisy-ocr/src/backend/ocr_backend.rs
deleted file mode 100644
index ceb8cf90a..000000000
--- a/crates/nvisy-ocr/src/backend/ocr_backend.rs
+++ /dev/null
@@ -1,155 +0,0 @@
-//! [`OcrBackend`]: the OCR backend contract.
-
-use bytes::Bytes;
-use nvisy_core::Error;
-use nvisy_core::entity::ModelProvenance;
-use nvisy_core::primitive::LanguageTag;
-use uuid::Uuid;
-
-use crate::types::RawOcrBlock;
-
-/// One per-call OCR request handed to an [`OcrBackend`].
-///
-/// Bundles the image bytes with advisory hints (language,
-/// correlation id). Borrowed (`OcrRequest<'a>`) so call sites that
-/// already own the underlying values hand them through without
-/// cloning.
-#[derive(Debug, Clone)]
-pub struct OcrRequest<'a> {
-    /// Raw image bytes. Returned span / region coordinates refer
-    /// back into the pixel space of this image.
-    pub image: &'a Bytes,
-    /// Caller-asserted language. Multilingual OCR engines may
-    /// ignore it; engines with a per-language model variant use it
-    /// to pick the right one.
-    pub language: Option<&'a LanguageTag>,
-    /// Per-call correlation id propagated to remote backends (as
-    /// the `x-request-id` header on the Bento backend). When
-    /// `None`, transports that need a request id generate a
-    /// UUIDv7 themselves so every request is traceable.
-    pub correlation_id: Option<Uuid>,
-}
-
-impl<'a> OcrRequest<'a> {
-    /// Construct a request with no advisory hints set.
-    pub fn new(image: &'a Bytes) -> Self {
-        Self {
-            image,
-            language: None,
-            correlation_id: None,
-        }
-    }
-
-    /// Builder-style setter for the language hint.
-    #[must_use]
-    pub fn with_language(mut self, language: &'a LanguageTag) -> Self {
-        self.language = Some(language);
-        self
-    }
-
-    /// Builder-style setter for the correlation id.
-    #[must_use]
-    pub fn with_correlation_id(mut self, correlation_id: Uuid) -> Self {
-        self.correlation_id = Some(correlation_id);
-        self
-    }
-}
-
-/// One per-call OCR response from an [`OcrBackend`].
-///
-/// Wraps the raw blocks the backend produced. Pre-normalization:
-/// the extractor wraps each block into the per-document block
-/// shape and stamps provenance onto the surrounding metadata.
-#[derive(Debug, Clone, Default)]
-pub struct OcrResponse {
-    /// Blocks predicted for the request's image, in backend order.
-    pub blocks: Vec<RawOcrBlock>,
-}
-
-impl OcrResponse {
-    /// Construct a response from raw blocks.
-    #[must_use]
-    pub fn new(blocks: Vec<RawOcrBlock>) -> Self {
-        Self { blocks }
-    }
-}
-
-/// Per-call OCR backend.
-///
-/// Implementations send an image to an OCR service and return an
-/// [`OcrResponse`] — one [`RawOcrBlock`] per page or text region,
-/// with per-word spans populated and bounding boxes preserved on
-/// each location.
-///
-/// Backends are source-agnostic: they take bytes + hints and
-/// return shape. Wrapping the blocks into a per-document container
-/// happens at the extractor boundary.
-///
-/// Confidence values **must** be normalised to `0.0..=1.0` before
-/// being placed on spans. Backends whose upstream API uses a
-/// different scale are responsible for converting.
-///
-/// Implementors **must** provide [`extract`]. The default
-/// [`extract_batch`] impl dispatches the inputs concurrently via
-/// `futures::join_all` and concatenates the per-image responses.
-/// Backends with a native batch API (such as a single network
-/// round-trip) override it to merge server-side.
-///
-/// Batch entries **are assumed to come from the same source** —
-/// the typical caller is a multi-page document split into
-/// per-page images, so the per-request hints apply uniformly and
-/// the per-image blocks can be merged without further bookkeeping.
-/// Mixed-source inputs should be issued as separate batches.
-///
-/// Per-image page numbering is returned as-is — if the caller
-/// needs them rebased onto a containing document, the caller knows
-/// the per-image page offsets and is responsible for that rebase.
-///
-/// Object-safe: extractors hold `Arc<dyn OcrBackend>` and dispatch
-/// per call.
-///
-/// [`extract`]: Self::extract
-/// [`extract_batch`]: Self::extract_batch
-#[async_trait::async_trait]
-pub trait OcrBackend: Send + Sync + 'static {
-    /// Backend identity (model / service name + provenance kind).
-    ///
-    /// The document-side extraction phase reads this after
-    /// recognition runs and stamps it into
-    /// [`ImageExtraction::Ocr`] on the document's metadata, so the
-    /// audit records *which* OCR pass produced the document.
-    ///
-    /// [`ImageExtraction::Ocr`]: nvisy_core::modality::ImageExtraction::Ocr
-    fn provenance(&self) -> ModelProvenance;
-
-    /// Extract OCR blocks for `request`.
-    ///
-    /// # Errors
-    ///
-    /// Returns the underlying transport / parse / inference error.
-    async fn extract(&self, request: OcrRequest<'_>) -> Result<OcrResponse, Error>;
-
-    /// Batched extract. Defaults to a concurrent fan-out via
-    /// `futures::join_all` and concatenates the per-image
-    /// responses into a single [`OcrResponse`]; backends with a
-    /// native batch API override it.
-    ///
-    /// `requests` is assumed to be slices of the same source (see
-    /// the trait-level docs).
-    ///
-    /// # Errors
-    ///
-    /// Returns the first error encountered.
-    async fn extract_batch(&self, requests: &[OcrRequest<'_>]) -> Result<OcrResponse, Error> {
-        let pending: Vec<_> = requests
-            .iter()
-            .map(|req| self.extract(req.clone()))
-            .collect();
-        let results: Vec<Result<OcrResponse, Error>> = futures::future::join_all(pending).await;
-        let mut merged: Vec<RawOcrBlock> = Vec::new();
-        for r in results {
-            merged.extend(r?.blocks);
-        }
-        Ok(OcrResponse::new(merged))
-    }
-}
diff --git a/crates/nvisy-ocr/src/extraction/extractor.rs b/crates/nvisy-ocr/src/extraction/extractor.rs
deleted file mode 100644
index 07a05bdd1..000000000
--- a/crates/nvisy-ocr/src/extraction/extractor.rs
+++ /dev/null
@@ -1,99 +0,0 @@
-//! [`OcrExtractor`]: type-erased OCR extractor wrapping any
-//! [`OcrBackend`].
-
-use std::fmt;
-use std::sync::Arc;
-
-use nvisy_core::entity::ModelProvenance;
-use nvisy_core::extraction::{Extractor as CoreExtractor, ExtractorOutput, Span};
-use nvisy_core::modality::{Image, ImageExtraction};
-use nvisy_core::{Error, Result};
-
-use crate::backend::{OcrBackend, OcrRequest, OcrResponse};
-
-const TARGET: &str = "nvisy_ocr::extraction";
-
-/// Type-erased OCR extractor wrapping any [`OcrBackend`]
-/// implementation.
-///
-/// Owns an `Arc<dyn OcrBackend>` and forwards OCR requests to it,
-/// providing a concrete, object-safe entry point without generics
-/// at every call site. The extractor is `Clone` — cloning shares
-/// the backend.
-#[derive(Clone)]
-pub struct OcrExtractor {
-    backend: Arc<dyn OcrBackend>,
-}
-
-impl fmt::Debug for OcrExtractor {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("OcrExtractor").finish_non_exhaustive()
-    }
-}
-
-impl OcrExtractor {
-    /// Create a new extractor from any [`OcrBackend`] implementation.
-    pub fn new(backend: impl OcrBackend) -> Self {
-        Self {
-            backend: Arc::new(backend),
-        }
-    }
-
-    /// Provenance of the wrapped backend, forwarded from
-    /// [`OcrBackend::provenance`].
-    pub fn provenance(&self) -> ModelProvenance {
-        self.backend.provenance()
-    }
-
-    /// Extract OCR blocks from a single image.
-    #[tracing::instrument(skip_all, fields(image_bytes = request.image.len()))]
-    pub async fn extract(&self, request: OcrRequest<'_>) -> Result<OcrResponse, Error> {
-        let response = self.backend.extract(request).await?;
-        tracing::debug!(
-            target: TARGET,
-            spans = response.blocks.iter().map(|b| b.spans.len()).sum::<usize>(),
-            "ocr complete",
-        );
-        Ok(response)
-    }
-
-    /// Extract OCR blocks from multiple images, concatenating the
-    /// per-image blocks. See [`OcrBackend::extract_batch`] for the
-    /// same-source assumption.
-    #[tracing::instrument(skip_all, fields(count = requests.len()))]
-    pub async fn extract_batch(&self, requests: &[OcrRequest<'_>]) -> Result<OcrResponse, Error> {
-        let response = self.backend.extract_batch(requests).await?;
-        tracing::debug!(
-            target: TARGET,
-            spans = response.blocks.iter().map(|b| b.spans.len()).sum::<usize>(),
-            "batch ocr complete",
-        );
-        Ok(response)
-    }
-}
-
-/// Bridge `nvisy_ocr::OcrExtractor` into the core-side
-/// [`Extractor<Image>`] surface. The extractor's output is the
-/// backend-shaped [`OcrResponse`]; consumers translate that into
-/// per-document `Block<Image>` values.
-///
-/// [`Extractor<Image>`]: nvisy_core::extraction::Extractor
-#[async_trait::async_trait]
-impl CoreExtractor<Image> for OcrExtractor {
-    type Output = OcrResponse;
-
-    async fn extract(&self, span: &Span<Image>) -> Result<ExtractorOutput<Image, Self::Output>> {
-        let mut request = OcrRequest::new(&span.data.bytes);
-        if let Some(ref lang) = span.language {
-            request = request.with_language(lang);
-        }
-        if let Some(corr_id) = span.correlation_id {
-            request = request.with_correlation_id(corr_id);
-        }
-        let response = self.backend.extract(request).await?;
-        Ok(ExtractorOutput::new(
-            response,
-            ImageExtraction::Ocr(self.provenance()),
-        ))
-    }
-}
diff --git a/crates/nvisy-ocr/src/extraction/mod.rs b/crates/nvisy-ocr/src/extraction/mod.rs
deleted file mode 100644
index c88631dbc..000000000
--- a/crates/nvisy-ocr/src/extraction/mod.rs
+++ /dev/null
@@ -1,14 +0,0 @@
-//! Extraction layer: the [`OcrExtractor`] that drives any
-//! [`OcrBackend`] backend through a type-erased
-//! [`Arc<dyn OcrBackend>`] slot.
-//!
-//! Implements [`Extractor<Image>`] (from `nvisy-core`) so it
-//! composes with the rest of the platform through the same trait
-//! every other image extractor uses.
-//!
-//! [`OcrBackend`]: crate::backend::OcrBackend
-//! [`Extractor<Image>`]: nvisy_core::extraction::Extractor
-
-mod extractor;
-
-pub use self::extractor::OcrExtractor;
diff --git a/crates/nvisy-ocr/src/lib.rs b/crates/nvisy-ocr/src/lib.rs
deleted file mode 100644
index af62104a4..000000000
--- a/crates/nvisy-ocr/src/lib.rs
+++ /dev/null
@@ -1,9 +0,0 @@
-#![forbid(unsafe_code)]
-#![cfg_attr(docsrs, feature(doc_cfg))]
-#![doc = include_str!("../README.md")]
-
-pub mod backend;
-mod extraction;
-pub mod types;
-
-pub use self::extraction::OcrExtractor;
diff --git a/crates/nvisy-ocr/src/types/mod.rs b/crates/nvisy-ocr/src/types/mod.rs
deleted file mode 100644
index dabb401d7..000000000
--- a/crates/nvisy-ocr/src/types/mod.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-//! Backend-shaped OCR output: [`RawOcrBlock`] and its parts.
-//!
-//! Pre-normalization shapes emitted by an [`OcrBackend`].
-//! Consumers wrap each block into the per-document block shape
-//! they need; this crate stays free of any orchestrator-level
-//! dependency.
-//!
-//! [`OcrBackend`]: crate::backend::OcrBackend
-
-mod raw_block;
-
-pub use self::raw_block::{OcrBlockKind, OcrSpan, RawOcrBlock};
diff --git a/crates/nvisy-ocr/src/types/raw_block.rs b/crates/nvisy-ocr/src/types/raw_block.rs
deleted file mode 100644
index 7e9831dc2..000000000
--- a/crates/nvisy-ocr/src/types/raw_block.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-//! [`RawOcrBlock`]: pre-normalization OCR backend output.
-//!
-//! Emitted by an [`OcrBackend`]: one block per recognized text
-//! region, with the per-modality kind, the per-word [`OcrSpan`]s,
-//! and a block-level confidence. Consumers wrap each block into the
-//! per-document block shape they need.
-//!
-//! [`OcrBackend`]: crate::backend::OcrBackend
-
-use nvisy_core::modality::ImageLocation;
-use nvisy_core::primitive::Confidence;
-
-/// One raw OCR block predicted by a backend.
-///
-/// Pre-normalization: spans + confidence come straight from the
-/// model; consumers translate the variant into their own per-block
-/// shape and decide how to use the confidence.
-#[derive(Debug, Clone)]
-pub struct RawOcrBlock {
-    /// The image-modality block payload — recognized text +
-    /// bounding region.
-    pub kind: OcrBlockKind,
-    /// Per-word source spans in the recognized text.
-    pub spans: Vec<OcrSpan>,
-    /// Block-level confidence.
-    pub confidence: Confidence,
-}
-
-/// Subset of `ImageBlock` variants OCR backends emit.
-///
-/// OCR doesn't produce `Figure`/`Separator`/`Background`/`Logo` —
-/// those are layout-analysis outputs. Only text-bearing variants
-/// land here.
-#[derive(Debug, Clone)]
-#[non_exhaustive]
-pub enum OcrBlockKind {
-    /// A region of recognized text.
-    Text {
-        /// Bounding box of the recognized region in image coordinates.
-        region: ImageLocation,
-        /// Recognized text content.
-        text: String,
-    },
-    /// A heading.
-    Heading {
-        /// Bounding box of the recognized region in image coordinates.
-        region: ImageLocation,
-        /// Recognized text content.
-        text: String,
-    },
-    /// A tabular region recognized in the image.
-    Table {
-        /// Bounding box of the recognized region in image coordinates.
-        region: ImageLocation,
-        /// Recognized text content (cells flattened in reading order).
-        text: String,
-    },
-}
-
-/// Per-word span emitted by an OCR backend.
-#[derive(Debug, Clone)]
-pub struct OcrSpan {
-    /// Byte offset where the word starts in the block text.
-    pub text_start: usize,
-    /// Byte offset where the word ends in the block text.
-    pub text_end: usize,
-    /// Modality-typed source location (per-word bounding region).
-    pub source: ImageLocation,
-    /// Per-word confidence.
-    pub confidence: Confidence,
-}
diff --git a/crates/nvisy-pattern/Cargo.toml b/crates/nvisy-pattern/Cargo.toml
deleted file mode 100644
index 9a505f144..000000000
--- a/crates/nvisy-pattern/Cargo.toml
+++ /dev/null
@@ -1,54 +0,0 @@
-# https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[package]
-name = "nvisy-pattern"
-description = "Pattern and dictionary recognizers for PII/PHI detection"
-keywords = ["nvisy", "pattern", "dictionary", "pii"]
-categories = ["text-processing"]
-readme = "README.md"
-
-version = { workspace = true }
-rust-version = { workspace = true }
-edition = { workspace = true }
-license = { workspace = true }
-publish = { workspace = true }
-
-authors = { workspace = true }
-repository = { workspace = true }
-homepage = { workspace = true }
-documentation = { workspace = true }
-
-[package.metadata.docs.rs]
-all-features = true
-rustdoc-args = ["--cfg", "docsrs"]
-
-[dependencies]
-# Internal crates
-nvisy-context = { workspace = true, features = [] }
-nvisy-core = { workspace = true, features = [] }
-
-# Serialization
-serde = { workspace = true, features = [] }
-toml = { workspace = true, features = ["parse"] }
-
-# Derive macros and error handling
-derive_builder = { workspace = true, features = [] }
-derive_more = { workspace = true, features = ["from"] }
-
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-
-# Text processing
-regex = { workspace = true, features = [] }
-aho-corasick = { workspace = true, features = [] }
-
-# Tabular document parsing
-csv = { workspace = true, features = [] }
-
-# Checksum / encoding
-bs58 = { workspace = true, features = ["check"] }
-iban_validate = { workspace = true }
-phonenumber = { workspace = true }
-
-[dev-dependencies]
-tokio = { workspace = true, features = ["macros", "rt"] }
diff --git a/crates/nvisy-pattern/README.md b/crates/nvisy-pattern/README.md
deleted file mode 100644
index fa8acd751..000000000
--- a/crates/nvisy-pattern/README.md
+++ /dev/null
@@ -1,58 +0,0 @@
-# nvisy-pattern
-
-[![Build](https://img.shields.io/github/actions/workflow/status/nvisycom/runtime/build.yml?branch=main&label=build%20%26%20test&style=flat-square)](https://github.com/nvisycom/runtime/actions/workflows/build.yml)
-
-Regex and dictionary recognizers for PII / PHI detection in the
-Nvisy runtime.
-
-## Overview
-
-`PatternRecognizer` compiles a set of `Regex` rules (each holding
-one or more regex `Variant`s grouped as a multi-strategy detector
-for one entity type) and `Dictionary` term lists into pooled
-scanners — one shared `regex::RegexSet` for the regex side and
-one shared `aho_corasick::AhoCorasick` automaton for the literal
-side. A single walk over the input runs both scanners and emits
-`Entity<Text>` values in modality-local byte coordinates.
-
-Rules may declare per-label context keywords. Calling
-`build_context_enhanced()` wraps the recognizer in a
-`nvisy_context::ContextEnhanced` layer that lifts confidence on
-matches whose neighbourhood contains a declared keyword;
-`build()` returns the bare recognizer.
-
-The built-in pattern + dictionary set lives as TOML under
-`assets/` and is embedded at compile time. The recognizer's
-builder accepts both built-ins and user-supplied rules:
-
-```rust
-use nvisy_pattern::PatternRecognizer;
-
-let recognizer = PatternRecognizer::builder()
-    .with_builtin_patterns()
-    .with_builtin_dictionaries()
-    .build()
-    .expect("built-in recognizer builds");
-```
-
-Regex variants can opt into a post-match validator by name
-(`"luhn"`, `"ssn"`, `"iban"`, `"phone"`, `"date"`); custom
-validators can be registered via `ValidatorRegistry::with`.
-
-## Documentation
-
-See [`docs/`](../../docs/) for architecture, security, and API documentation.
-
-## Changelog
-
-See [CHANGELOG.md](../../CHANGELOG.md) for release notes and version history.
-
-## License
-
-Apache 2.0 License, see [LICENSE.txt](../../LICENSE.txt)
-
-## Support
-
-- **Documentation**: [docs.nvisy.com](https://docs.nvisy.com)
-- **Issues**: [GitHub Issues](https://github.com/nvisycom/runtime/issues)
-- **Email**: [support@nvisy.com](mailto:support@nvisy.com)
diff --git a/crates/nvisy-pattern/assets/PRESIDIO.md b/crates/nvisy-pattern/assets/PRESIDIO.md
deleted file mode 100644
index 1b810fdb5..000000000
--- a/crates/nvisy-pattern/assets/PRESIDIO.md
+++ /dev/null
@@ -1,41 +0,0 @@
-# Presidio attribution
-
-Several shipped pattern TOMLs under `patterns/` carry regular
-expressions ported or adapted from [Microsoft Presidio][presidio]
-(`microsoft/presidio`, MIT-licensed) — specifically the
-`presidio-analyzer/presidio_analyzer/predefined_recognizers/`
-classes. Validators (Luhn, IBAN mod-97, ABA, DEA, NPI, NHS,
-NINO, etc.) were re-implemented in Rust from the same upstream
-algorithms.
-
-The Presidio MIT license text is reproduced below to satisfy its
-"include this permission notice in all copies or substantial
-portions" clause.
-
-[presidio]: https://github.com/microsoft/presidio
-
----
-
-```
-MIT License
-
-Copyright (c) Microsoft Corporation.
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
-```
diff --git a/crates/nvisy-pattern/assets/README.md b/crates/nvisy-pattern/assets/README.md
deleted file mode 100644
index a41af49be..000000000
--- a/crates/nvisy-pattern/assets/README.md
+++ /dev/null
@@ -1,94 +0,0 @@
-# Shipped asset tree
-
-The `nvisy-pattern` crate compiles every TOML and term-source
-file under this directory into the binary via `include_str!`,
-so adding a pattern or dictionary is as simple as:
-
-1. Drop the asset into the right subtree.
-2. Wire a `shipped_pattern!` / `shipped_dictionary!` accessor
-   in `src/shipped/{patterns,dictionaries}/<scope>.rs`.
-3. Append the accessor to the sub-module's `all()`.
-
-The recognizer's per-call language and country fields filter
-patterns + dictionaries at runtime — see
-[`RecognizerInput::applies_to_language`] and
-[`RecognizerInput::applies_to_country`].
-
-## Layout
-
-```
-assets/
-  patterns/
-    world/                 jurisdiction-agnostic regex patterns
-      contact/             email, phone, url
-      credentials/         aws, github, stripe, generic api, private key
-      finance/             credit card, iban, swift, btc, eth
-      network/             ipv4, ipv6, mac
-      personal/            date of birth, datetime
-    us/                    US-jurisdiction patterns
-      identity/            ssn, itin, drivers_license, passport, postal_code
-      finance/             bank_routing, bank_account
-      health/              npi, mbi, medical_license (DEA)
-    uk/                    UK-jurisdiction patterns
-      identity/            nhs, nino, driving_licence, passport
-      contact/             postcode
-      vehicle/             registration
-
-  dictionaries/
-    world/                 universal: brand names + codes
-      finance/             cryptocurrencies (BTC, ETH, Bitcoin, …)
-    en/                    English-language terms
-      finance/             currencies (USD, US Dollar, EUR, …)
-      personal/            languages, nationalities, religions
-```
-
-Each pattern is a TOML file (`<name>.toml`). Each dictionary
-pairs a TOML metadata sidecar with a term source:
-`.csv` for multi-column lists (term + alias columns with
-per-column scores), `.txt` for one-per-line lists.
-
-## Scoring conventions
-
-Scores are baseline confidence — the context enhancer (in
-`nvisy-context`) lifts them when configured keywords appear
-nearby. The toolkit's default confidence threshold is `0.35`;
-anything below needs context boost or an out-of-band hint
-(CSV column header, JSON object key, HTML parent text) to
-clear it.
-
-| Tier | Score | Use |
-|------|-------|-----|
-| Strong | 0.95–0.98 | Branded credential headers (`AKIA…`, `-----BEGIN PRIVATE KEY-----`, `gh[pousr]_…`) |
-| Solid  | 0.4–0.5   | Format with a checksum or restrictive structure (IBAN, NHS, NPI, MBI, IPv4, MAC, IPv6) |
-| Loose  | 0.3       | Brand-aware with weak structural specificity (credit_card, dictionaries) |
-| Weak   | 0.1       | Generic shape that *requires* context to clear threshold (passport, postal_code, DoB) |
-| Trace  | 0.05      | Last-resort generic regex (bank_account `\b\d{8,17}\b`) |
-
-The targets mirror Microsoft Presidio's deliberately-conservative
-baselines — most of Presidio's predefined recognizers sit in
-0.1–0.5 because the context enhancer is expected to lift hits
-to 0.6+ when surrounding tokens match.
-
-## Validators
-
-A pattern variant can declare `validator = "<name>"` to drop
-matches that fail a post-match structural check. Built-in
-names (resolved via `ValidatorRegistry::builtin`):
-
-- Universal: `luhn`, `iban`, `phone`, `date`, `crypto.btc`
-- US: `us.ssn`, `us.aba_routing`, `us.npi`, `us.dea_number`,
-  `us.postal_code`
-- UK: `uk.nhs`, `uk.nino`, `uk.driving_licence`,
-  `uk.vehicle_registration`
-
-Each lives in `src/validators/` under the matching submodule.
-
-## Attribution
-
-Many patterns + validators are ports of upstream Microsoft
-Presidio recognizers. See [`PRESIDIO.md`](PRESIDIO.md) for the
-MIT-license attribution and the upstream class references that
-each adapted TOML's leading comment links to.
-
-[`RecognizerInput::applies_to_language`]: ../../nvisy-core/src/recognition/input.rs
-[`RecognizerInput::applies_to_country`]: ../../nvisy-core/src/recognition/input.rs
diff --git a/crates/nvisy-pattern/assets/dictionaries/en/finance/currencies.csv b/crates/nvisy-pattern/assets/dictionaries/en/finance/currencies.csv
deleted file mode 100644
index b74543d83..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/en/finance/currencies.csv
+++ /dev/null
@@ -1,40 +0,0 @@
-US Dollar,USD
-Euro,EUR
-British Pound,GBP
-Japanese Yen,JPY
-Swiss Franc,CHF
-Canadian Dollar,CAD
-Australian Dollar,AUD
-New Zealand Dollar,NZD
-Chinese Yuan,CNY,Renminbi,RMB
-Indian Rupee,INR
-Brazilian Real,BRL
-South Korean Won,KRW
-Mexican Peso,MXN
-Singapore Dollar,SGD
-Hong Kong Dollar,HKD
-Norwegian Krone,NOK
-Swedish Krona,SEK
-Danish Krone,DKK
-South African Rand,ZAR
-Russian Ruble,RUB
-Turkish Lira,TRY
-Israeli Shekel,ILS
-Saudi Riyal,SAR
-UAE Dirham,AED
-Thai Baht,THB
-Indonesian Rupiah,IDR
-Malaysian Ringgit,MYR
-Philippine Peso,PHP
-Polish Zloty,PLN
-Czech Koruna,CZK
-Hungarian Forint,HUF
-Romanian Leu,RON
-Colombian Peso,COP
-Chilean Peso,CLP
-Peruvian Sol,PEN
-Argentine Peso,ARS
-Egyptian Pound,EGP
-Nigerian Naira,NGN
-Kenyan Shilling,KES
-Ghanaian Cedi,GHS
diff --git a/crates/nvisy-pattern/assets/dictionaries/en/finance/currencies.toml b/crates/nvisy-pattern/assets/dictionaries/en/finance/currencies.toml
deleted file mode 100644
index ee86a7d20..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/en/finance/currencies.toml
+++ /dev/null
@@ -1,4 +0,0 @@
-name = "currencies"
-label = "currency"
-languages = ["en"]
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/dictionaries/en/personal/languages.csv b/crates/nvisy-pattern/assets/dictionaries/en/personal/languages.csv
deleted file mode 100644
index c50684dac..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/en/personal/languages.csv
+++ /dev/null
@@ -1,104 +0,0 @@
-Afrikaans,af
-Albanian,sq
-Amharic,am
-Arabic,ar
-Armenian,hy
-Azerbaijani,az
-Basque,eu
-Belarusian,be
-Bengali,bn
-Bosnian,bs
-Bulgarian,bg
-Burmese,my
-Cantonese,yue
-Catalan,ca
-Cebuano,ceb
-Croatian,hr
-Czech,cs
-Danish,da
-Dutch,nl
-English,en
-Esperanto,eo
-Estonian,et
-Filipino,fil
-Finnish,fi
-French,fr
-Galician,gl
-Georgian,ka
-German,de
-Greek,el
-Gujarati,gu
-Haitian Creole,ht
-Hausa,ha
-Hebrew,he
-Hindi,hi
-Hmong,hmn
-Hungarian,hu
-Icelandic,is
-Igbo,ig
-Indonesian,id
-Irish,ga
-Italian,it
-Japanese,ja
-Javanese,jv
-Kannada,kn
-Kazakh,kk
-Khmer,km
-Korean,ko
-Kurdish,ku
-Kyrgyz,ky
-Lao,lo
-Latin,la
-Latvian,lv
-Lithuanian,lt
-Luxembourgish,lb
-Macedonian,mk
-Malagasy,mg
-Malay,ms
-Malayalam,ml
-Maltese,mt
-Mandarin,zh
-Maori,mi
-Marathi,mr
-Mongolian,mn
-Nepali,ne
-Norwegian,no
-Odia,or
-Pashto,ps
-Persian,fa,Farsi
-Polish,pl
-Portuguese,pt
-Punjabi,pa
-Quechua,qu
-Romanian,ro
-Russian,ru
-Samoan,sm
-Serbian,sr
-Shona,sn
-Sindhi,sd
-Sinhala,si
-Slovak,sk
-Slovenian,sl
-Somali,so
-Spanish,es
-Sundanese,su
-Swahili,sw
-Swedish,sv
-Tagalog,tl
-Tajik,tg
-Tamil,ta
-Tatar,tt
-Telugu,te
-Thai,th
-Tibetan,bo
-Turkish,tr
-Turkmen,tk
-Ukrainian,uk
-Urdu,ur
-Uzbek,uz
-Vietnamese,vi
-Welsh,cy
-Xhosa,xh
-Yiddish,yi
-Yoruba,yo
-Zulu,zu
diff --git a/crates/nvisy-pattern/assets/dictionaries/en/personal/languages.toml b/crates/nvisy-pattern/assets/dictionaries/en/personal/languages.toml
deleted file mode 100644
index 3f530fb50..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/en/personal/languages.toml
+++ /dev/null
@@ -1,8 +0,0 @@
-name = "languages"
-label = "language"
-languages = ["en"]
-
-# column 0 = long-form names (`English`, `Spanish`, ...)
-# column 1 = ISO 639-1 codes (`en`, `es`, ...)
-# column 2 = alternate long-form names (`Farsi` for Persian)
-score = [0.4, 0.2, 0.4]
diff --git a/crates/nvisy-pattern/assets/dictionaries/en/personal/nationalities.toml b/crates/nvisy-pattern/assets/dictionaries/en/personal/nationalities.toml
deleted file mode 100644
index 76f2e5535..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/en/personal/nationalities.toml
+++ /dev/null
@@ -1,4 +0,0 @@
-name = "nationalities"
-label = "nationality"
-languages = ["en"]
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/dictionaries/en/personal/nationalities.txt b/crates/nvisy-pattern/assets/dictionaries/en/personal/nationalities.txt
deleted file mode 100644
index e4172e999..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/en/personal/nationalities.txt
+++ /dev/null
@@ -1,152 +0,0 @@
-Afghan
-Albanian
-Algerian
-American
-Andorran
-Angolan
-Argentine
-Armenian
-Australian
-Austrian
-Azerbaijani
-Bahamian
-Bahraini
-Bangladeshi
-Barbadian
-Belarusian
-Belgian
-Belizean
-Beninese
-Bhutanese
-Bolivian
-Bosnian
-Brazilian
-British
-Bruneian
-Bulgarian
-Burkinabe
-Burmese
-Burundian
-Cambodian
-Cameroonian
-Canadian
-Chilean
-Chinese
-Colombian
-Congolese
-Costa Rican
-Croatian
-Cuban
-Cypriot
-Czech
-Danish
-Dominican
-Dutch
-Ecuadorian
-Egyptian
-Emirati
-Eritrean
-Estonian
-Ethiopian
-Fijian
-Filipino
-Finnish
-French
-Gabonese
-Georgian
-German
-Ghanaian
-Greek
-Guatemalan
-Guinean
-Guyanese
-Haitian
-Honduran
-Hungarian
-Icelandic
-Indian
-Indonesian
-Iranian
-Iraqi
-Irish
-Israeli
-Italian
-Ivorian
-Jamaican
-Japanese
-Jordanian
-Kazakh
-Kenyan
-Korean
-Kuwaiti
-Kyrgyz
-Laotian
-Latvian
-Lebanese
-Liberian
-Libyan
-Lithuanian
-Luxembourgish
-Macedonian
-Malagasy
-Malawian
-Malaysian
-Maldivian
-Malian
-Maltese
-Mexican
-Moldovan
-Mongolian
-Montenegrin
-Moroccan
-Mozambican
-Namibian
-Nepalese
-New Zealander
-Nicaraguan
-Nigerian
-Norwegian
-Omani
-Pakistani
-Palestinian
-Panamanian
-Paraguayan
-Peruvian
-Polish
-Portuguese
-Qatari
-Romanian
-Russian
-Rwandan
-Saudi
-Senegalese
-Serbian
-Singaporean
-Slovak
-Slovenian
-Somali
-South African
-Spanish
-Sri Lankan
-Sudanese
-Swedish
-Swiss
-Syrian
-Taiwanese
-Tajik
-Tanzanian
-Thai
-Togolese
-Trinidadian
-Tunisian
-Turkish
-Turkmen
-Ugandan
-Ukrainian
-Uruguayan
-Uzbek
-Venezuelan
-Vietnamese
-Yemeni
-Zambian
-Zimbabwean
diff --git a/crates/nvisy-pattern/assets/dictionaries/en/personal/religions.toml b/crates/nvisy-pattern/assets/dictionaries/en/personal/religions.toml
deleted file mode 100644
index 8205f97ea..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/en/personal/religions.toml
+++ /dev/null
@@ -1,4 +0,0 @@
-name = "religions"
-label = "religion"
-languages = ["en"]
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/dictionaries/en/personal/religions.txt b/crates/nvisy-pattern/assets/dictionaries/en/personal/religions.txt
deleted file mode 100644
index 81b7d5e30..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/en/personal/religions.txt
+++ /dev/null
@@ -1,34 +0,0 @@
-Buddhist
-Catholic
-Christian
-Hindu
-Jewish
-Mormon
-Muslim
-Orthodox
-Protestant
-Sikh
-Sunni
-Shia
-Evangelical
-Anglican
-Baptist
-Methodist
-Presbyterian
-Lutheran
-Pentecostal
-Adventist
-Jehovah's Witness
-Quaker
-Amish
-Mennonite
-Baha'i
-Jain
-Shinto
-Taoist
-Zoroastrian
-Rastafari
-Wiccan
-Pagan
-Atheist
-Agnostic
diff --git a/crates/nvisy-pattern/assets/dictionaries/world/finance/cryptocurrencies.csv b/crates/nvisy-pattern/assets/dictionaries/world/finance/cryptocurrencies.csv
deleted file mode 100644
index a6cffc3b9..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/world/finance/cryptocurrencies.csv
+++ /dev/null
@@ -1,19 +0,0 @@
-Bitcoin,BTC
-Ethereum,ETH
-Tether,USDT
-USD Coin,USDC
-Binance Coin,BNB
-Ripple,XRP
-Cardano,ADA
-Solana,SOL
-Dogecoin,DOGE
-Polkadot,DOT
-Avalanche,AVAX
-Chainlink,LINK
-Litecoin,LTC
-Polygon,MATIC
-Stellar,XLM
-Monero,XMR
-Uniswap,UNI
-Cosmos,ATOM
-Toncoin,TON
diff --git a/crates/nvisy-pattern/assets/dictionaries/world/finance/cryptocurrencies.toml b/crates/nvisy-pattern/assets/dictionaries/world/finance/cryptocurrencies.toml
deleted file mode 100644
index 2bbbc05d4..000000000
--- a/crates/nvisy-pattern/assets/dictionaries/world/finance/cryptocurrencies.toml
+++ /dev/null
@@ -1,3 +0,0 @@
-name = "cryptocurrencies"
-label = "currency"
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/patterns/au/finance/abn.toml b/crates/nvisy-pattern/assets/patterns/au/finance/abn.toml
deleted file mode 100644
index a316fbc3a..000000000
--- a/crates/nvisy-pattern/assets/patterns/au/finance/abn.toml
+++ /dev/null
@@ -1,25 +0,0 @@
-# Australian Business Number (ABN): 11-digit ID issued by the
-# Australian Business Register. The leading digit is rewritten
-# (-1 with 0→9) before a weighted sum that must be divisible by
-# 89. The conventional rendering is `NN NNN NNN NNN`.
-
-name = "au-abn"
-label = "company_id"
-countries = ["AU"]
-languages = ["en"]
-context = [
-  "australian business number",
-  "abn",
-  "abr",
-  "australian business register",
-]
-
-[[variants]]
-regex = '\b\d{2}\s\d{3}\s\d{3}\s\d{3}\b'
-score = 0.4
-validator = "au.abn"
-
-[[variants]]
-regex = '\b\d{11}\b'
-score = 0.2
-validator = "au.abn"
diff --git a/crates/nvisy-pattern/assets/patterns/au/finance/acn.toml b/crates/nvisy-pattern/assets/patterns/au/finance/acn.toml
deleted file mode 100644
index 714b6fcfd..000000000
--- a/crates/nvisy-pattern/assets/patterns/au/finance/acn.toml
+++ /dev/null
@@ -1,24 +0,0 @@
-# Australian Company Number (ACN): 9-digit ID issued by ASIC.
-# Weighted-sum mod-10 checksum, complement-to-10. Conventional
-# rendering is `NNN NNN NNN`.
-
-name = "au-acn"
-label = "company_id"
-countries = ["AU"]
-languages = ["en"]
-context = [
-  "australian company number",
-  "acn",
-  "asic",
-  "australian securities and investments commission",
-]
-
-[[variants]]
-regex = '\b\d{3}\s\d{3}\s\d{3}\b'
-score = 0.4
-validator = "au.acn"
-
-[[variants]]
-regex = '\b\d{9}\b'
-score = 0.1
-validator = "au.acn"
diff --git a/crates/nvisy-pattern/assets/patterns/au/health/medicare.toml b/crates/nvisy-pattern/assets/patterns/au/health/medicare.toml
deleted file mode 100644
index 768e1c953..000000000
--- a/crates/nvisy-pattern/assets/patterns/au/health/medicare.toml
+++ /dev/null
@@ -1,25 +0,0 @@
-# Australian Medicare card number: 10-digit ID where the first
-# digit is in 2-6 and the 9th digit is a mod-10 weighted check.
-# Rendering is `NNNN NNNNN N` (card number + individual reference)
-# or `NNNN NNNNN N N` (with issue number).
-
-name = "au-medicare"
-label = "insurance_id"
-countries = ["AU"]
-languages = ["en"]
-context = [
-  "medicare",
-  "medicare card",
-  "medicare number",
-  "services australia",
-]
-
-[[variants]]
-regex = '\b[2-6]\d{3}\s\d{5}\s\d\b'
-score = 0.4
-validator = "au.medicare"
-
-[[variants]]
-regex = '\b[2-6]\d{9}\b'
-score = 0.2
-validator = "au.medicare"
diff --git a/crates/nvisy-pattern/assets/patterns/au/identity/tfn.toml b/crates/nvisy-pattern/assets/patterns/au/identity/tfn.toml
deleted file mode 100644
index f048d3273..000000000
--- a/crates/nvisy-pattern/assets/patterns/au/identity/tfn.toml
+++ /dev/null
@@ -1,24 +0,0 @@
-# Australian Tax File Number (TFN): 9-digit ID issued by the
-# Australian Taxation Office. Weighted mod-11 checksum.
-# Rendering is `NNN NNN NNN`.
-
-name = "au-tfn"
-label = "tax_id"
-countries = ["AU"]
-languages = ["en"]
-context = [
-  "tax file number",
-  "tfn",
-  "australian taxation office",
-  "ato",
-]
-
-[[variants]]
-regex = '\b\d{3}\s\d{3}\s\d{3}\b'
-score = 0.4
-validator = "au.tfn"
-
-[[variants]]
-regex = '\b\d{9}\b'
-score = 0.1
-validator = "au.tfn"
diff --git a/crates/nvisy-pattern/assets/patterns/ca/contact/postal_code.toml b/crates/nvisy-pattern/assets/patterns/ca/contact/postal_code.toml
deleted file mode 100644
index bc388a612..000000000
--- a/crates/nvisy-pattern/assets/patterns/ca/contact/postal_code.toml
+++ /dev/null
@@ -1,20 +0,0 @@
-# Canadian postal code: `A1A 1A1` format where the letter
-# alphabet excludes D, F, I, O, Q, U (and W, Z don't appear as
-# the first letter — Canada Post Address Standard).
-
-name = "ca-postal-code"
-label = "postal_code"
-countries = ["CA"]
-languages = ["en", "fr"]
-context = [
-  "postal code",
-  "code postal",
-  "canada post",
-  "postes canada",
-  "mailing address",
-  "adresse postale",
-]
-
-[[variants]]
-regex = '\b[ABCEGHJ-NPRSTVXY]\d[A-CEGHJ-NPRSTV-Z][ -]?\d[A-CEGHJ-NPRSTV-Z]\d\b'
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/ca/identity/sin.toml b/crates/nvisy-pattern/assets/patterns/ca/identity/sin.toml
deleted file mode 100644
index d8707d977..000000000
--- a/crates/nvisy-pattern/assets/patterns/ca/identity/sin.toml
+++ /dev/null
@@ -1,32 +0,0 @@
-# Canadian Social Insurance Number (SIN): 9 digits issued by
-# Employment and Social Development Canada. First digit must be
-# in `[1-7, 9]` (0 and 8 reserved). Luhn checksum.
-
-name = "ca-sin"
-label = "government_id"
-countries = ["CA"]
-languages = ["en", "fr"]
-context = [
-  "sin",
-  "social insurance number",
-  "social insurance",
-  "nas",
-  "numéro d'assurance sociale",
-  "numéro nas",
-  "assurance sociale",
-]
-
-[[variants]]
-regex = '\b[1-79]\d{2} \d{3} \d{3}\b'
-score = 0.5
-validator = "ca.sin"
-
-[[variants]]
-regex = '\b[1-79]\d{2}-\d{3}-\d{3}\b'
-score = 0.5
-validator = "ca.sin"
-
-[[variants]]
-regex = '\b[1-79]\d{8}\b'
-score = 0.2
-validator = "ca.sin"
diff --git a/crates/nvisy-pattern/assets/patterns/de/contact/plz.toml b/crates/nvisy-pattern/assets/patterns/de/contact/plz.toml
deleted file mode 100644
index fc6b6d022..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/contact/plz.toml
+++ /dev/null
@@ -1,21 +0,0 @@
-# German Postleitzahl (PLZ): 5 digits, with `01000` and `99999`
-# reserved as test sentinels and never assigned.
-
-name = "de-plz"
-label = "postal_code"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "postleitzahl",
-  "plz",
-  "postanschrift",
-  "anschrift",
-  "lieferadresse",
-  "rechnungsadresse",
-  "postfach",
-]
-
-[[variants]]
-regex = '\b(0[1-9]\d{3}|[1-9]\d{4})\b'
-score = 0.3
-validator = "de.plz"
diff --git a/crates/nvisy-pattern/assets/patterns/de/finance/handelsregister.toml b/crates/nvisy-pattern/assets/patterns/de/finance/handelsregister.toml
deleted file mode 100644
index 46d9b7bd7..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/finance/handelsregister.toml
+++ /dev/null
@@ -1,23 +0,0 @@
-# German Handelsregisternummer: court-registered company ID with
-# section prefix `HRA` (partnerships) or `HRB` (capital companies)
-# followed by 1-6 digits.
-
-name = "de-handelsregister"
-label = "company_id"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "handelsregister",
-  "handelsregisternummer",
-  "registernummer",
-  "registergericht",
-  "amtsgericht",
-  "hrb",
-  "hra",
-  "handelsregister-nr",
-  "hr-nummer",
-]
-
-[[variants]]
-regex = '\bHR[AB]\s*\d{1,6}\b'
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/de/health/bsnr.toml b/crates/nvisy-pattern/assets/patterns/de/health/bsnr.toml
deleted file mode 100644
index fb98d7571..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/health/bsnr.toml
+++ /dev/null
@@ -1,29 +0,0 @@
-# German Betriebsstättennummer (BSNR): 9-digit practice/site-of-care
-# number assigned by the regional Kassenärztliche Vereinigung (KV)
-# under § 75 Abs. 7 SGB V. Appears on Rezepte and billing records.
-
-name = "de-bsnr"
-label = "medical_id"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "betriebsstättennummer",
-  "betriebsstätten-nummer",
-  "bsnr",
-  "betriebsstätte",
-  "praxisnummer",
-  "arztpraxis",
-  "praxis",
-  "kassenärztliche vereinigung",
-  "kv-nummer",
-  "kv nummer",
-  "praxisadresse",
-  "praxisstandort",
-  "behandlungsort",
-  "vertragsarztpraxis",
-]
-
-[[variants]]
-regex = '\b\d{9}\b'
-score = 0.2
-validator = "de.bsnr"
diff --git a/crates/nvisy-pattern/assets/patterns/de/health/health_insurance.toml b/crates/nvisy-pattern/assets/patterns/de/health/health_insurance.toml
deleted file mode 100644
index 8c094cc4a..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/health/health_insurance.toml
+++ /dev/null
@@ -1,33 +0,0 @@
-# German Krankenversicherungsnummer (KVNR): 10 characters —
-# leading letter + 9 digits. Printed on the eGK (elektronische
-# Gesundheitskarte). Check digit per § 290 SGB V Anlage 1.
-
-name = "de-health-insurance"
-label = "insurance_id"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "krankenversicherungsnummer",
-  "krankenversichertennummer",
-  "versichertennummer",
-  "kvnr",
-  "krankenkasse",
-  "krankenversicherung",
-  "gesundheitskarte",
-  "egk",
-  "elektronische gesundheitskarte",
-  "gkv",
-  "gesetzliche krankenversicherung",
-  "krankenversicherungsausweis",
-  "versichertenausweis",
-  "versichertenkarte",
-  "aok",
-  "tkk",
-  "barmer",
-  "dak",
-]
-
-[[variants]]
-regex = '\b[A-Z]\d{9}\b'
-score = 0.3
-validator = "de.health_insurance"
diff --git a/crates/nvisy-pattern/assets/patterns/de/health/lanr.toml b/crates/nvisy-pattern/assets/patterns/de/health/lanr.toml
deleted file mode 100644
index b41d523ba..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/health/lanr.toml
+++ /dev/null
@@ -1,31 +0,0 @@
-# German Lebenslange Arztnummer (LANR): 9-digit lifetime physician
-# number assigned by the Kassenärztliche Vereinigung. Pos 7 is a
-# check digit per the KBV Arztnummern-Richtlinie; pos 8-9 carry
-# the Fachgruppe (specialty) code.
-
-name = "de-lanr"
-label = "medical_id"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "arztnummer",
-  "lanr",
-  "lebenslange arztnummer",
-  "arzt-nr",
-  "arzt nr",
-  "arzt-nummer",
-  "vertragsarzt",
-  "kassenarzt",
-  "niedergelassener arzt",
-  "kbv",
-  "rezept",
-  "verschreibung",
-  "behandelnder arzt",
-  "hausarzt",
-  "facharzt",
-]
-
-[[variants]]
-regex = '\b\d{9}\b'
-score = 0.3
-validator = "de.lanr"
diff --git a/crates/nvisy-pattern/assets/patterns/de/identity/fuehrerschein.toml b/crates/nvisy-pattern/assets/patterns/de/identity/fuehrerschein.toml
deleted file mode 100644
index c18693bdf..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/identity/fuehrerschein.toml
+++ /dev/null
@@ -1,34 +0,0 @@
-# German Führerscheinnummer (driving licence number): 11
-# alphanumeric characters in the EU-harmonized post-2013 format
-# (FeV Anlage 8). Format is opaque — position 3 onward mixes
-# letters and digits per issuing authority. The trailing check
-# character's algorithm is not published by KBA, so no validator
-# is wired.
-
-name = "de-fuehrerschein"
-label = "drivers_license"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "führerscheinnummer",
-  "führerschein",
-  "fahrerlaubnis",
-  "fahrerlaubnisnummer",
-  "fahrerlaubnisklasse",
-  "führerscheininhaber",
-  "fev",
-  "kba",
-  "kraftfahrt-bundesamt",
-  "driving licence",
-  "driving license",
-  "driver's license",
-  "licence number",
-  "license number",
-  "dokument nr",
-  "dokument-nr",
-  "feld 5",
-]
-
-[[variants]]
-regex = '\b[A-Z]\d{2}[A-Z0-9]{6}\d[A-Z0-9]\b'
-score = 0.35
diff --git a/crates/nvisy-pattern/assets/patterns/de/identity/id_card.toml b/crates/nvisy-pattern/assets/patterns/de/identity/id_card.toml
deleted file mode 100644
index 85fb9a203..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/identity/id_card.toml
+++ /dev/null
@@ -1,36 +0,0 @@
-# German Personalausweis (national ID card). Two formats coexist:
-#
-# - nPA (since November 2010): 9 alphanumeric characters from the
-#   ICAO restricted charset (excludes A, B, D, E, I, O, Q, S, U)
-#   plus a trailing ICAO check digit.
-# - Legacy (pre-2010): letter `T` + 8 digits. No ICAO checksum.
-
-name = "de-id-card"
-label = "government_id"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "personalausweis",
-  "ausweis",
-  "personalausweisnummer",
-  "ausweisnummer",
-  "ausweisdokument",
-  "dokumentennummer",
-  "seriennummer",
-  "npa",
-  "neuer personalausweis",
-  "personalausweisgesetz",
-  "bundespersonalausweis",
-  "identity card",
-  "national id",
-]
-
-[[variants]]
-regex = '\b[CFGHJKLMNPRTVWXYZ][CFGHJKLMNPRTVWXYZ0-9]{7}[0-9]\b'
-score = 0.4
-validator = "de.id_card"
-
-[[variants]]
-regex = '\bT\d{8}\b'
-score = 0.5
-validator = "de.id_card"
diff --git a/crates/nvisy-pattern/assets/patterns/de/identity/passport.toml b/crates/nvisy-pattern/assets/patterns/de/identity/passport.toml
deleted file mode 100644
index 499b95a47..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/identity/passport.toml
+++ /dev/null
@@ -1,23 +0,0 @@
-# German Reisepass: 9-character serial from the ICAO restricted
-# charset (excludes A, B, D, E, I, O, Q, S, U) plus a trailing
-# ICAO Doc 9303 check digit. Issued by the Bundesdruckerei.
-
-name = "de-passport"
-label = "passport_number"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "passnummer",
-  "reisepassnummer",
-  "passport number",
-  "pass-nr",
-  "dokumentennummer",
-  "reisepass",
-  "reisedokument",
-  "bundesdruckerei",
-]
-
-[[variants]]
-regex = '\b[CFGHJKLMNPRTVWXYZ][CFGHJKLMNPRTVWXYZ0-9]{7}[0-9]\b'
-score = 0.4
-validator = "de.passport"
diff --git a/crates/nvisy-pattern/assets/patterns/de/identity/social_security.toml b/crates/nvisy-pattern/assets/patterns/de/identity/social_security.toml
deleted file mode 100644
index 2b1e55100..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/identity/social_security.toml
+++ /dev/null
@@ -1,25 +0,0 @@
-# German Rentenversicherungsnummer (RVNR / SVNR): 12 characters
-# encoding region + birth date + surname initial + serial + check
-# digit per VKVV § 4 (Deutsche Rentenversicherung).
-
-name = "de-social-security"
-label = "national_insurance_number"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "rentenversicherungsnummer",
-  "sozialversicherungsnummer",
-  "versicherungsnummer",
-  "rvnr",
-  "svnr",
-  "sv-nummer",
-  "rentenversicherung",
-  "deutsche rentenversicherung",
-  "drv",
-  "sozialversicherung",
-]
-
-[[variants]]
-regex = '\b\d{8}[A-Z]\d{3}\b'
-score = 0.5
-validator = "de.social_security"
diff --git a/crates/nvisy-pattern/assets/patterns/de/identity/tax_id.toml b/crates/nvisy-pattern/assets/patterns/de/identity/tax_id.toml
deleted file mode 100644
index 004c2922f..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/identity/tax_id.toml
+++ /dev/null
@@ -1,24 +0,0 @@
-# German Steueridentifikationsnummer (Steuer-IdNr): 11-digit
-# lifetime tax identifier per Bundeszentralamt für Steuern.
-# Check digit at pos 11 via ISO 7064 Mod 11, 10. No leading zero.
-
-name = "de-tax-id"
-label = "tax_id"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "steueridentifikationsnummer",
-  "steuerliche identifikationsnummer",
-  "persönliche identifikationsnummer",
-  "idnr",
-  "steuer-idnr",
-  "steuer idnr",
-  "steueridentnummer",
-  "bzst",
-  "bundeszentralamt für steuern",
-]
-
-[[variants]]
-regex = '\b[1-9]\d{10}\b'
-score = 0.4
-validator = "de.tax_id"
diff --git a/crates/nvisy-pattern/assets/patterns/de/identity/tax_number.toml b/crates/nvisy-pattern/assets/patterns/de/identity/tax_number.toml
deleted file mode 100644
index 66e42d533..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/identity/tax_number.toml
+++ /dev/null
@@ -1,34 +0,0 @@
-# German Steuernummer (regional tax number): 10 or 11 digits
-# allocated per Finanzamt. Format varies by Bundesland and the
-# Vereinheitlichtes Bundesschema, so three regex variants cover the
-# common renderings. No checksum is publicly documented.
-
-name = "de-tax-number"
-label = "tax_id"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "steuernummer",
-  "st-nr",
-  "stnr",
-  "st.-nr",
-  "steuer-nr",
-  "finanzamt",
-  "bundeseinheitliche steuernummer",
-  "vereinheitlichtes bundesschema",
-]
-
-# Vereinheitlichtes Bundesschema: 13 digits.
-[[variants]]
-regex = '\b\d{13}\b'
-score = 0.2
-
-# Common rendering: 3/4/5 grouped form.
-[[variants]]
-regex = '\b\d{3}/\d{4}/\d{4}\b'
-score = 0.4
-
-# Common rendering: 2/3/5 grouped form (some Länder use this).
-[[variants]]
-regex = '\b\d{2,3}/\d{3}/\d{4,5}\b'
-score = 0.3
diff --git a/crates/nvisy-pattern/assets/patterns/de/identity/vat_id.toml b/crates/nvisy-pattern/assets/patterns/de/identity/vat_id.toml
deleted file mode 100644
index 5c1549b4f..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/identity/vat_id.toml
+++ /dev/null
@@ -1,32 +0,0 @@
-# German Umsatzsteuer-Identifikationsnummer (USt-IdNr): `DE` prefix
-# + 9 digits. Used in B2B invoices and VAT registration. Validator
-# applies the community-documented ISO 7064 Mod 11, 10 heuristic.
-# Whitespace and dashes are stripped before validation.
-
-name = "de-vat-id"
-label = "tax_id"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "umsatzsteuer-identifikationsnummer",
-  "umsatzsteueridentifikationsnummer",
-  "ust-idnr",
-  "ustidnr",
-  "ust idnr",
-  "ust-id",
-  "umsatzsteuer",
-  "mehrwertsteuer",
-  "mwst-id",
-  "vat number",
-  "vat id",
-]
-
-[[variants]]
-regex = '\bDE\d{9}\b'
-score = 0.5
-validator = "de.vat_id"
-
-[[variants]]
-regex = '\bDE[\s.\-]?\d{3}[\s.\-]?\d{3}[\s.\-]?\d{3}\b'
-score = 0.4
-validator = "de.vat_id"
diff --git a/crates/nvisy-pattern/assets/patterns/de/vehicle/kfz.toml b/crates/nvisy-pattern/assets/patterns/de/vehicle/kfz.toml
deleted file mode 100644
index 61a7fe83d..000000000
--- a/crates/nvisy-pattern/assets/patterns/de/vehicle/kfz.toml
+++ /dev/null
@@ -1,44 +0,0 @@
-# German Kfz-Kennzeichen (license plate): 1-3 letter
-# Unterscheidungszeichen (city/district code), optional
-# Erkennungszeichen (1-2 letters), and 1-4 digit serial.
-# Multiple variants cover hyphen, space, and Umlaut renderings.
-
-name = "de-kfz"
-label = "license_plate"
-countries = ["DE"]
-languages = ["de"]
-context = [
-  "kennzeichen",
-  "kfz-kennzeichen",
-  "kfz kennzeichen",
-  "amtliches kennzeichen",
-  "fahrzeugschein",
-  "fahrzeugbrief",
-  "zulassung",
-  "nummernschild",
-]
-
-# Canonical hyphen-separated rendering: BB-AB 1234.
-[[variants]]
-regex = '\b[A-ZÄÖÜ]{1,3}[-\s][A-Z]{1,2}\s?\d{1,4}H?E?\b'
-score = 0.3
-
-# ASCII-only rendering (commonly typed without Umlaut).
-[[variants]]
-regex = '\b[A-Z]{1,3}[-\s][A-Z]{1,2}\s?\d{1,4}H?E?\b'
-score = 0.2
-
-# Glued form without spaces: BBAB1234.
-[[variants]]
-regex = '\b[A-ZÄÖÜ]{1,3}[A-Z]{1,2}\d{1,4}H?E?\b'
-score = 0.2
-
-# Sondercode for diplomats and federal authorities (leading zero).
-[[variants]]
-regex = '\b0\s?\d{1,4}\b'
-score = 0.1
-
-# Saisonkennzeichen with month range suffix.
-[[variants]]
-regex = '\b[A-ZÄÖÜ]{1,3}[-\s][A-Z]{1,2}\s?\d{1,4}\s\d{2}\b'
-score = 0.3
diff --git a/crates/nvisy-pattern/assets/patterns/es/contact/codigo_postal.toml b/crates/nvisy-pattern/assets/patterns/es/contact/codigo_postal.toml
deleted file mode 100644
index 6f29396c9..000000000
--- a/crates/nvisy-pattern/assets/patterns/es/contact/codigo_postal.toml
+++ /dev/null
@@ -1,19 +0,0 @@
-# Spanish código postal: 5 digits where the first two are the
-# province code (01-52, including Ceuta=51 and Melilla=52). Codes
-# 53-99 are unassigned.
-
-name = "es-codigo-postal"
-label = "postal_code"
-countries = ["ES"]
-languages = ["es"]
-context = [
-  "código postal",
-  "codigo postal",
-  "cp",
-  "c.p.",
-  "dirección postal",
-]
-
-[[variants]]
-regex = '\b(0[1-9]|[1-4]\d|5[0-2])\d{3}\b'
-score = 0.3
diff --git a/crates/nvisy-pattern/assets/patterns/es/finance/cif.toml b/crates/nvisy-pattern/assets/patterns/es/finance/cif.toml
deleted file mode 100644
index 6344a137c..000000000
--- a/crates/nvisy-pattern/assets/patterns/es/finance/cif.toml
+++ /dev/null
@@ -1,21 +0,0 @@
-# Spanish CIF (Código de Identificación Fiscal): entity-class
-# letter + 7 digits + control char (digit or letter depending on
-# entity class). Now formally subsumed by NIF for new entities,
-# but legacy IDs remain in wide circulation.
-
-name = "es-cif"
-label = "company_id"
-countries = ["ES"]
-languages = ["es"]
-context = [
-  "código de identificación fiscal",
-  "cif",
-  "identificación fiscal",
-  "sociedad",
-  "empresa",
-]
-
-[[variants]]
-regex = '\b[ABCDEFGHJNPQRSUVW]-?\d{7}-?[0-9A-J]\b'
-score = 0.5
-validator = "es.cif"
diff --git a/crates/nvisy-pattern/assets/patterns/es/identity/nie.toml b/crates/nvisy-pattern/assets/patterns/es/identity/nie.toml
deleted file mode 100644
index d4b8df286..000000000
--- a/crates/nvisy-pattern/assets/patterns/es/identity/nie.toml
+++ /dev/null
@@ -1,21 +0,0 @@
-# Spanish NIE (Número de Identidad de Extranjero): X, Y, or Z
-# prefix + 7 digits + Mod 23 control letter (X→0, Y→1, Z→2 before
-# the modulo).
-
-name = "es-nie"
-label = "government_id"
-countries = ["ES"]
-languages = ["es"]
-context = [
-  "número de identidad de extranjero",
-  "número de identificación de extranjero",
-  "nie",
-  "tarjeta de residencia",
-  "residencia",
-  "extranjero",
-]
-
-[[variants]]
-regex = '\b[XYZ]-?\d{7}-?[A-Z]\b'
-score = 0.5
-validator = "es.nie"
diff --git a/crates/nvisy-pattern/assets/patterns/es/identity/nif.toml b/crates/nvisy-pattern/assets/patterns/es/identity/nif.toml
deleted file mode 100644
index 629dd7381..000000000
--- a/crates/nvisy-pattern/assets/patterns/es/identity/nif.toml
+++ /dev/null
@@ -1,22 +0,0 @@
-# Spanish NIF / DNI: 8 digits + Mod 23 control letter. The legacy
-# DNI (citizens) and NIF (taxpayers) share the algorithm; the
-# leading `0` is sometimes omitted, so 7 digits + letter is also
-# accepted by the validator.
-
-name = "es-nif"
-label = "government_id"
-countries = ["ES"]
-languages = ["es"]
-context = [
-  "documento nacional de identidad",
-  "número de identificación fiscal",
-  "dni",
-  "nif",
-  "identificación",
-  "identidad",
-]
-
-[[variants]]
-regex = '\b\d{7,8}-?[A-Z]\b'
-score = 0.5
-validator = "es.nif"
diff --git a/crates/nvisy-pattern/assets/patterns/es/identity/passport.toml b/crates/nvisy-pattern/assets/patterns/es/identity/passport.toml
deleted file mode 100644
index c02c4686c..000000000
--- a/crates/nvisy-pattern/assets/patterns/es/identity/passport.toml
+++ /dev/null
@@ -1,18 +0,0 @@
-# Spanish passport: 3 letters + 6 digits (e.g. AAA123456). No
-# published checksum, so the score stays low and relies on the
-# context-keyword boost.
-
-name = "es-passport"
-label = "passport_number"
-countries = ["ES"]
-languages = ["es"]
-context = [
-  "pasaporte",
-  "número de pasaporte",
-  "passport",
-  "passport number",
-]
-
-[[variants]]
-regex = '\b[A-Z]{3}\d{6}\b'
-score = 0.1
diff --git a/crates/nvisy-pattern/assets/patterns/fi/identity/hetu.toml b/crates/nvisy-pattern/assets/patterns/fi/identity/hetu.toml
deleted file mode 100644
index 885c1c530..000000000
--- a/crates/nvisy-pattern/assets/patterns/fi/identity/hetu.toml
+++ /dev/null
@@ -1,24 +0,0 @@
-# Finnish Henkilötunnus (HETU): 11 chars encoding birth date,
-# century separator, serial, and control character. Issued by
-# Digi- ja väestötietovirasto (DVV). The control character is
-# validated by `fi.hetu` (mod-31 lookup into
-# `0123456789ABCDEFHJKLMNPRSTUVWXY`).
-
-name = "fi-hetu"
-label = "government_id"
-countries = ["FI"]
-languages = ["fi", "sv"]
-context = [
-  "henkilötunnus",
-  "hetu",
-  "personbeteckning",
-  "personbeteckningen",
-  "personal identity code",
-  "social security number",
-  "dvv",
-]
-
-[[variants]]
-regex = '\b\d{6}[+\-A-FYXWVU]\d{3}[0-9A-FHJKLMNPRSTUVWXY]\b'
-score = 0.5
-validator = "fi.hetu"
diff --git a/crates/nvisy-pattern/assets/patterns/in/finance/gstin.toml b/crates/nvisy-pattern/assets/patterns/in/finance/gstin.toml
deleted file mode 100644
index c0dfc3f5a..000000000
--- a/crates/nvisy-pattern/assets/patterns/in/finance/gstin.toml
+++ /dev/null
@@ -1,20 +0,0 @@
-# Indian Goods and Services Tax Identification Number (GSTIN):
-# 15 chars: state code (01-37) + 10-char PAN + registration
-# sequence + `Z` literal + base-36 weighted check digit.
-
-name = "in-gstin"
-label = "tax_id"
-countries = ["IN"]
-languages = ["en", "hi"]
-context = [
-  "gstin",
-  "gst",
-  "goods and services tax",
-  "gst number",
-  "gst registration",
-]
-
-[[variants]]
-regex = '\b(0[1-9]|[12]\d|3[0-7])[A-Z]{5}\d{4}[A-Z][A-Z0-9]Z[A-Z0-9]\b'
-score = 0.6
-validator = "in.gstin"
diff --git a/crates/nvisy-pattern/assets/patterns/in/identity/aadhaar.toml b/crates/nvisy-pattern/assets/patterns/in/identity/aadhaar.toml
deleted file mode 100644
index ff80b08f9..000000000
--- a/crates/nvisy-pattern/assets/patterns/in/identity/aadhaar.toml
+++ /dev/null
@@ -1,29 +0,0 @@
-# Indian Aadhaar number: 12-digit ID issued by UIDAI. Leading
-# digit ≥ 2 (UIDAI reserves 0xx and 1xx); the number must not be
-# a palindrome; Verhoeff checksum.
-
-name = "in-aadhaar"
-label = "government_id"
-countries = ["IN"]
-languages = ["en", "hi"]
-context = [
-  "aadhaar",
-  "uidai",
-  "unique identification authority",
-  "आधार",
-]
-
-[[variants]]
-regex = '\b[2-9]\d{3}\s\d{4}\s\d{4}\b'
-score = 0.4
-validator = "in.aadhaar"
-
-[[variants]]
-regex = '\b[2-9]\d{3}-\d{4}-\d{4}\b'
-score = 0.4
-validator = "in.aadhaar"
-
-[[variants]]
-regex = '\b[2-9]\d{11}\b'
-score = 0.2
-validator = "in.aadhaar"
diff --git a/crates/nvisy-pattern/assets/patterns/in/identity/pan.toml b/crates/nvisy-pattern/assets/patterns/in/identity/pan.toml
deleted file mode 100644
index 21bffa697..000000000
--- a/crates/nvisy-pattern/assets/patterns/in/identity/pan.toml
+++ /dev/null
@@ -1,19 +0,0 @@
-# Indian Permanent Account Number (PAN): 10-char alphanumeric
-# `AAAAA9999A` issued by the Income Tax Department. The 4th
-# character encodes the entity type per IT Department spec.
-
-name = "in-pan"
-label = "tax_id"
-countries = ["IN"]
-languages = ["en", "hi"]
-context = [
-  "pan",
-  "permanent account number",
-  "income tax",
-  "pan card",
-]
-
-[[variants]]
-regex = '\b[A-Z]{3}[ABCFGHJLPTabcfghjlpt][A-Z]\d{4}[A-Z]\b'
-score = 0.5
-validator = "in.pan"
diff --git a/crates/nvisy-pattern/assets/patterns/in/identity/passport.toml b/crates/nvisy-pattern/assets/patterns/in/identity/passport.toml
deleted file mode 100644
index cbc51b8b4..000000000
--- a/crates/nvisy-pattern/assets/patterns/in/identity/passport.toml
+++ /dev/null
@@ -1,19 +0,0 @@
-# Indian passport: 8-char alphanumeric in the form
-# `LDDDDDDDD` where the first char is a letter, position 2 is a
-# non-zero digit, and the last is a non-zero check digit. No
-# published checksum.
-
-name = "in-passport"
-label = "passport_number"
-countries = ["IN"]
-languages = ["en", "hi"]
-context = [
-  "passport",
-  "indian passport",
-  "passport number",
-  "ministry of external affairs",
-]
-
-[[variants]]
-regex = '\b[A-Z][1-9]\d\s?\d{4}[1-9]\b'
-score = 0.2
diff --git a/crates/nvisy-pattern/assets/patterns/in/identity/voter.toml b/crates/nvisy-pattern/assets/patterns/in/identity/voter.toml
deleted file mode 100644
index 94630faef..000000000
--- a/crates/nvisy-pattern/assets/patterns/in/identity/voter.toml
+++ /dev/null
@@ -1,25 +0,0 @@
-# Indian Voter ID (EPIC — Elector's Photo Identity Card):
-# 10-char alphanumeric issued by the Election Commission of
-# India. Two formats: 3 letters + 7 digits (older), or 1 letter
-# + restricted-set 2nd char + letter + 7 digits.
-
-name = "in-voter"
-label = "government_id"
-countries = ["IN"]
-languages = ["en", "hi"]
-context = [
-  "voter",
-  "voter id",
-  "epic",
-  "elector photo identity card",
-  "election commission",
-  "मतदाता",
-]
-
-[[variants]]
-regex = '\b[A-Z][ABCDGHJKMNPRSY][A-Z]\d{7}\b'
-score = 0.4
-
-[[variants]]
-regex = '\b[A-Z]{3}\d{7}\b'
-score = 0.3
diff --git a/crates/nvisy-pattern/assets/patterns/in/vehicle/registration.toml b/crates/nvisy-pattern/assets/patterns/in/vehicle/registration.toml
deleted file mode 100644
index 28f0b0775..000000000
--- a/crates/nvisy-pattern/assets/patterns/in/vehicle/registration.toml
+++ /dev/null
@@ -1,31 +0,0 @@
-# Indian vehicle registration: state code (2 letters) + RTO
-# district (2 digits) + series (1-3 letters) + serial (4 digits).
-# Per-state district codes are not validated here — relying on
-# the canonical structural pattern + RTO/registration context.
-
-name = "in-vehicle-registration"
-label = "license_plate"
-countries = ["IN"]
-languages = ["en", "hi"]
-context = [
-  "rto",
-  "vehicle",
-  "registration plate",
-  "vehicle number",
-  "regional transport office",
-]
-
-# Canonical: SS-DD-LL-NNNN.
-[[variants]]
-regex = '\b[A-Z]{2}\d{2}[A-Z]{1,2}\d{4}\b'
-score = 0.5
-
-# Older single-letter series: SS-D-LLL-NNNN.
-[[variants]]
-regex = '\b[A-Z]{2}\d[A-Z]{1,3}\d{4}\b'
-score = 0.4
-
-# Diplomatic / armed forces prefixes.
-[[variants]]
-regex = '\b\d{1,3}(CD|CC|UN)[1-9]\d{1,3}\b'
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/patterns/it/finance/vat_code.toml b/crates/nvisy-pattern/assets/patterns/it/finance/vat_code.toml
deleted file mode 100644
index 526e7fe0e..000000000
--- a/crates/nvisy-pattern/assets/patterns/it/finance/vat_code.toml
+++ /dev/null
@@ -1,20 +0,0 @@
-# Italian Partita IVA (P.IVA): 11 digits with a Luhn-like
-# checksum. Validator rejects the all-zero sentinel.
-
-name = "it-vat-code"
-label = "tax_id"
-countries = ["IT"]
-languages = ["it"]
-context = [
-  "partita iva",
-  "piva",
-  "p.iva",
-  "p. iva",
-  "numero iva",
-  "agenzia delle entrate",
-]
-
-[[variants]]
-regex = '\b\d{11}\b'
-score = 0.3
-validator = "it.vat_code"
diff --git a/crates/nvisy-pattern/assets/patterns/it/identity/driving_licence.toml b/crates/nvisy-pattern/assets/patterns/it/identity/driving_licence.toml
deleted file mode 100644
index 06e806bd9..000000000
--- a/crates/nvisy-pattern/assets/patterns/it/identity/driving_licence.toml
+++ /dev/null
@@ -1,22 +0,0 @@
-# Italian patente di guida: 2 letters + 7 digits + 1 letter
-# (classic format) or `U1` + 7 alnum + 1 letter (Motorizzazione
-# Civile format issued post-2013).
-
-name = "it-driving-licence"
-label = "drivers_license"
-countries = ["IT"]
-languages = ["it"]
-context = [
-  "patente di guida",
-  "patente",
-  "motorizzazione civile",
-  "numero patente",
-]
-
-[[variants]]
-regex = '(?i)\b[A-Z]{2}\d{7}[A-Z]\b'
-score = 0.2
-
-[[variants]]
-regex = '(?i)\bU1[A-HJ-NP-Z0-9]{7}[A-Z]\b'
-score = 0.3
diff --git a/crates/nvisy-pattern/assets/patterns/it/identity/fiscal_code.toml b/crates/nvisy-pattern/assets/patterns/it/identity/fiscal_code.toml
deleted file mode 100644
index 426d0881a..000000000
--- a/crates/nvisy-pattern/assets/patterns/it/identity/fiscal_code.toml
+++ /dev/null
@@ -1,23 +0,0 @@
-# Italian Codice Fiscale (CF): 16 alphanumeric chars encoding
-# surname/name initials, birth date, municipality, and control
-# letter. The regex covers omocodia (digit→letter rewrites used
-# when two people would collide) and structural sanity (month
-# letter, day range, sex offset). Final validation by checksum
-# in `it.fiscal_code`.
-
-name = "it-fiscal-code"
-label = "tax_id"
-countries = ["IT"]
-languages = ["it"]
-context = [
-  "codice fiscale",
-  "cf",
-  "c.f.",
-  "tessera sanitaria",
-  "agenzia delle entrate",
-]
-
-[[variants]]
-regex = '''(?i)\b(?:[A-Z][AEIOU][AEIOUX]|[AEIOU]X{2}|[B-DF-HJ-NP-TV-Z]{2}[A-Z]){2}(?:[\dLMNP-V]{2}(?:[A-EHLMPR-T](?:[04LQ][1-9MNP-V]|[15MR][\dLMNP-V]|[26NS][0-8LMNP-U])|[DHPS][37PT][0L]|[ACELMRT][37PT][01LM]|[AC-EHLMPR-T][26NS][9V])|(?:[02468LNQSU][048LQU]|[13579MPRTV][26NS])B[26NS][9V])(?:[A-MZ][1-9MNP-V][\dLMNP-V]{2}|[A-M][0L](?:[1-9MNP-V][\dLMNP-V]|[0L][1-9MNP-V]))[A-Z]\b'''
-score = 0.4
-validator = "it.fiscal_code"
diff --git a/crates/nvisy-pattern/assets/patterns/it/identity/identity_card.toml b/crates/nvisy-pattern/assets/patterns/it/identity/identity_card.toml
deleted file mode 100644
index ec22523ce..000000000
--- a/crates/nvisy-pattern/assets/patterns/it/identity/identity_card.toml
+++ /dev/null
@@ -1,33 +0,0 @@
-# Italian Carta d'Identità: paper-based (`AA0000000`),
-# Carta d'Identità Elettronica (CIE) 2.0 (`0000000AA`), and CIE
-# 3.0 (`AA00000AA`). No published checksum, so scores stay weak
-# and rely on the context-keyword boost.
-
-name = "it-identity-card"
-label = "government_id"
-countries = ["IT"]
-languages = ["it"]
-context = [
-  "carta d'identità",
-  "carta di identità",
-  "carta identità",
-  "cie",
-  "documento di identità",
-  "documento di riconoscimento",
-  "espatrio",
-]
-
-# Paper-based (issued through 2018, still valid until expiry).
-[[variants]]
-regex = '(?i)\b[A-Z]{2}\s?\d{7}\b'
-score = 0.05
-
-# CIE 2.0.
-[[variants]]
-regex = '(?i)\b\d{7}[A-Z]{2}\b'
-score = 0.05
-
-# CIE 3.0 (current standard).
-[[variants]]
-regex = '(?i)\b[A-Z]{2}\d{5}[A-Z]{2}\b'
-score = 0.05
diff --git a/crates/nvisy-pattern/assets/patterns/it/identity/passport.toml b/crates/nvisy-pattern/assets/patterns/it/identity/passport.toml
deleted file mode 100644
index 75356c3f2..000000000
--- a/crates/nvisy-pattern/assets/patterns/it/identity/passport.toml
+++ /dev/null
@@ -1,19 +0,0 @@
-# Italian passport: 2 letters + 7 digits (Polizia di Stato
-# format). No published checksum, so the score stays weak.
-
-name = "it-passport"
-label = "passport_number"
-countries = ["IT"]
-languages = ["it"]
-context = [
-  "passaporto",
-  "passaporto italiano",
-  "passaporto elettronico",
-  "documento di viaggio",
-  "polizia di stato",
-  "dogana",
-]
-
-[[variants]]
-regex = '(?i)\b[A-Z]{2}\d{7}\b'
-score = 0.05
diff --git a/crates/nvisy-pattern/assets/patterns/kr/finance/brn.toml b/crates/nvisy-pattern/assets/patterns/kr/finance/brn.toml
deleted file mode 100644
index 2e5e4906e..000000000
--- a/crates/nvisy-pattern/assets/patterns/kr/finance/brn.toml
+++ /dev/null
@@ -1,31 +0,0 @@
-# Korean Business Registration Number (BRN): 10 digits formatted
-# as `AAA-BB-CCCCC` (district office + business type + serial +
-# check digit). Magic-keys Mod-10 checksum per National Tax
-# Service spec.
-
-name = "kr-brn"
-label = "company_id"
-countries = ["KR"]
-languages = ["ko"]
-context = [
-  "brn",
-  "사업자등록번호",
-  "사업자번호",
-  "사업자",
-  "business registration number",
-  "korean brn",
-  "business number",
-  "tax registration number",
-]
-
-# Hyphen-separated rendering.
-[[variants]]
-regex = '\b\d{3}-\d{2}-\d{5}\b'
-score = 0.5
-validator = "kr.brn"
-
-# Glued rendering.
-[[variants]]
-regex = '\b\d{10}\b'
-score = 0.2
-validator = "kr.brn"
diff --git a/crates/nvisy-pattern/assets/patterns/kr/identity/driver_license.toml b/crates/nvisy-pattern/assets/patterns/kr/identity/driver_license.toml
deleted file mode 100644
index 87e3c333f..000000000
--- a/crates/nvisy-pattern/assets/patterns/kr/identity/driver_license.toml
+++ /dev/null
@@ -1,28 +0,0 @@
-# Korean driver's license number: 12 digits formatted as
-# `AA-BB-CCCCCC-DD` (regional office + year + serial + check).
-# Region code is validated against the published Doro-gyotongan
-# list by `kr.driver_license`; the check digit algorithm itself
-# is not publicly disclosed.
-
-name = "kr-driver-license"
-label = "drivers_license"
-countries = ["KR"]
-languages = ["ko"]
-context = [
-  "운전면허",
-  "운전면허번호",
-  "면허번호",
-  "korean driver license",
-  "korean driver's license",
-  "driver license number",
-]
-
-[[variants]]
-regex = '\b\d{2}-\d{2}-\d{6}-\d{2}\b'
-score = 0.4
-validator = "kr.driver_license"
-
-[[variants]]
-regex = '\b\d{12}\b'
-score = 0.2
-validator = "kr.driver_license"
diff --git a/crates/nvisy-pattern/assets/patterns/kr/identity/frn.toml b/crates/nvisy-pattern/assets/patterns/kr/identity/frn.toml
deleted file mode 100644
index 347cb7bc9..000000000
--- a/crates/nvisy-pattern/assets/patterns/kr/identity/frn.toml
+++ /dev/null
@@ -1,27 +0,0 @@
-# Korean Foreigner Registration Number (FRN): same shape as RRN
-# but the gender/century digit is in `[5-8]` and the checksum
-# formula adds 13 instead of 11.
-
-name = "kr-frn"
-label = "government_id"
-countries = ["KR"]
-languages = ["ko"]
-context = [
-  "frn",
-  "외국인등록번호",
-  "외국인번호",
-  "foreigner registration number",
-  "korean frn",
-]
-
-# Hyphen-separated rendering.
-[[variants]]
-regex = '\b\d{2}(0[1-9]|1[0-2])(0[1-9]|[12]\d|3[01])-[5-8]\d{6}\b'
-score = 0.5
-validator = "kr.frn"
-
-# Glued rendering.
-[[variants]]
-regex = '\b\d{2}(0[1-9]|1[0-2])(0[1-9]|[12]\d|3[01])[5-8]\d{6}\b'
-score = 0.4
-validator = "kr.frn"
diff --git a/crates/nvisy-pattern/assets/patterns/kr/identity/passport.toml b/crates/nvisy-pattern/assets/patterns/kr/identity/passport.toml
deleted file mode 100644
index 06b27aaac..000000000
--- a/crates/nvisy-pattern/assets/patterns/kr/identity/passport.toml
+++ /dev/null
@@ -1,27 +0,0 @@
-# Korean passport number. Current format: a leading letter from
-# `M/S/R/O/D` followed by 3 digits, one letter, and 4 digits.
-# Previous format: leading letter + 8 digits. No published
-# checksum.
-
-name = "kr-passport"
-label = "passport_number"
-countries = ["KR"]
-languages = ["ko"]
-context = [
-  "passport",
-  "passport number",
-  "korean passport",
-  "여권",
-  "대한민국 여권",
-  "여권번호",
-]
-
-# Current format `LDDDLDDDD`.
-[[variants]]
-regex = '\b[MSRODmsrod]\d{3}[A-Za-z]\d{4}\b'
-score = 0.2
-
-# Legacy 9-character format `LDDDDDDDD`.
-[[variants]]
-regex = '\b[MSRODmsrod]\d{8}\b'
-score = 0.1
diff --git a/crates/nvisy-pattern/assets/patterns/kr/identity/rrn.toml b/crates/nvisy-pattern/assets/patterns/kr/identity/rrn.toml
deleted file mode 100644
index 2efc9288a..000000000
--- a/crates/nvisy-pattern/assets/patterns/kr/identity/rrn.toml
+++ /dev/null
@@ -1,29 +0,0 @@
-# Korean Resident Registration Number (RRN): 13 digits encoding
-# birth date, gender/century, region, serial, and weighted Mod-11
-# check digit. Format `YYMMDD-GHIJKLX`. Pre-October 2020 numbers
-# carry the checksum; later issues use a random tail.
-
-name = "kr-rrn"
-label = "government_id"
-countries = ["KR"]
-languages = ["ko"]
-context = [
-  "rrn",
-  "주민등록번호",
-  "주민번호",
-  "korean rrn",
-  "korean resident registration number",
-  "resident registration number",
-]
-
-# Hyphen-separated rendering.
-[[variants]]
-regex = '\b\d{2}(0[1-9]|1[0-2])(0[1-9]|[12]\d|3[01])-[1-4]\d{6}\b'
-score = 0.5
-validator = "kr.rrn"
-
-# Glued rendering.
-[[variants]]
-regex = '\b\d{2}(0[1-9]|1[0-2])(0[1-9]|[12]\d|3[01])[1-4]\d{6}\b'
-score = 0.4
-validator = "kr.rrn"
diff --git a/crates/nvisy-pattern/assets/patterns/ng/identity/nin.toml b/crates/nvisy-pattern/assets/patterns/ng/identity/nin.toml
deleted file mode 100644
index 01cbf3018..000000000
--- a/crates/nvisy-pattern/assets/patterns/ng/identity/nin.toml
+++ /dev/null
@@ -1,22 +0,0 @@
-# Nigerian National Identification Number (NIN): 11-digit ID
-# issued by the National Identity Management Commission (NIMC).
-# The trailing digit is a Verhoeff checksum.
-
-name = "ng-nin"
-label = "government_id"
-countries = ["NG"]
-languages = ["en"]
-context = [
-  "nin",
-  "national identification number",
-  "national identity number",
-  "nimc",
-  "national identity",
-  "nigeria id",
-  "nigerian identification",
-]
-
-[[variants]]
-regex = '\b\d{11}\b'
-score = 0.3
-validator = "ng.nin"
diff --git a/crates/nvisy-pattern/assets/patterns/ng/vehicle/registration.toml b/crates/nvisy-pattern/assets/patterns/ng/vehicle/registration.toml
deleted file mode 100644
index b105b6e1e..000000000
--- a/crates/nvisy-pattern/assets/patterns/ng/vehicle/registration.toml
+++ /dev/null
@@ -1,20 +0,0 @@
-# Nigerian vehicle registration plate (2011-present format):
-# 3-letter LGA (Local Government Area) code + 3 digits + 2-letter
-# year/batch code. Optional hyphen or space separator.
-
-name = "ng-vehicle-registration"
-label = "license_plate"
-countries = ["NG"]
-languages = ["en"]
-context = [
-  "plate number",
-  "vehicle registration",
-  "license plate",
-  "number plate",
-  "frsc",
-  "federal road safety corps",
-]
-
-[[variants]]
-regex = '\b[A-Z]{3}[- ]?\d{3}[A-Z]{2}\b'
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/pl/contact/kod_pocztowy.toml b/crates/nvisy-pattern/assets/patterns/pl/contact/kod_pocztowy.toml
deleted file mode 100644
index e9a0e6471..000000000
--- a/crates/nvisy-pattern/assets/patterns/pl/contact/kod_pocztowy.toml
+++ /dev/null
@@ -1,18 +0,0 @@
-# Polish kod pocztowy: NN-NNN format introduced by Poczta Polska
-# in 1973. The first two digits encode the postal region (00-99,
-# all values are assigned).
-
-name = "pl-kod-pocztowy"
-label = "postal_code"
-countries = ["PL"]
-languages = ["pl"]
-context = [
-  "kod pocztowy",
-  "kod",
-  "poczta polska",
-  "adres pocztowy",
-]
-
-[[variants]]
-regex = '\b\d{2}-\d{3}\b'
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/patterns/pl/finance/nip.toml b/crates/nvisy-pattern/assets/patterns/pl/finance/nip.toml
deleted file mode 100644
index 8bef2a459..000000000
--- a/crates/nvisy-pattern/assets/patterns/pl/finance/nip.toml
+++ /dev/null
@@ -1,31 +0,0 @@
-# Polish NIP (Numer Identyfikacji Podatkowej): 10-digit tax ID
-# with mod-11 weighted checksum per Ustawa o zasadach ewidencji i
-# identyfikacji podatników. The regex accepts the canonical
-# `XXX-XXX-XX-XX` and `XXX-XX-XX-XXX` separator renderings.
-
-name = "pl-nip"
-label = "tax_id"
-countries = ["PL"]
-languages = ["pl"]
-context = [
-  "nip",
-  "numer identyfikacji podatkowej",
-  "identyfikacja podatkowa",
-  "urząd skarbowy",
-  "ministerstwo finansów",
-]
-
-[[variants]]
-regex = '\b\d{10}\b'
-score = 0.3
-validator = "pl.nip"
-
-[[variants]]
-regex = '\b\d{3}-\d{3}-\d{2}-\d{2}\b'
-score = 0.5
-validator = "pl.nip"
-
-[[variants]]
-regex = '\b\d{3}-\d{2}-\d{2}-\d{3}\b'
-score = 0.4
-validator = "pl.nip"
diff --git a/crates/nvisy-pattern/assets/patterns/pl/finance/regon.toml b/crates/nvisy-pattern/assets/patterns/pl/finance/regon.toml
deleted file mode 100644
index b4cfa4ca2..000000000
--- a/crates/nvisy-pattern/assets/patterns/pl/finance/regon.toml
+++ /dev/null
@@ -1,26 +0,0 @@
-# Polish REGON (Rejestr Gospodarki Narodowej): 9-digit
-# entity-level or 14-digit unit-level company registry number.
-# Both forms use a weighted mod-11 checksum, validated by
-# `pl.regon`.
-
-name = "pl-regon"
-label = "company_id"
-countries = ["PL"]
-languages = ["pl"]
-context = [
-  "regon",
-  "numer regon",
-  "rejestr gospodarki narodowej",
-  "gus",
-  "główny urząd statystyczny",
-]
-
-[[variants]]
-regex = '\b\d{9}\b'
-score = 0.3
-validator = "pl.regon"
-
-[[variants]]
-regex = '\b\d{14}\b'
-score = 0.4
-validator = "pl.regon"
diff --git a/crates/nvisy-pattern/assets/patterns/pl/identity/pesel.toml b/crates/nvisy-pattern/assets/patterns/pl/identity/pesel.toml
deleted file mode 100644
index 8daf03c5d..000000000
--- a/crates/nvisy-pattern/assets/patterns/pl/identity/pesel.toml
+++ /dev/null
@@ -1,21 +0,0 @@
-# Polish PESEL (Powszechny Elektroniczny System Ewidencji
-# Ludności): 11 digits encoding birth date + sex + serial +
-# weighted-mod-10 check digit. The regex enforces structural
-# sanity (day 01-31, month range using +20/+40/+60/+80
-# Ergänzung).
-
-name = "pl-pesel"
-label = "government_id"
-countries = ["PL"]
-languages = ["pl"]
-context = [
-  "pesel",
-  "numer pesel",
-  "numer ewidencyjny",
-  "ewidencja ludności",
-]
-
-[[variants]]
-regex = '\b\d{2}([02468][1-9]|[13579][012])(0[1-9]|1[0-9]|2[0-9]|3[01])\d{5}\b'
-score = 0.4
-validator = "pl.pesel"
diff --git a/crates/nvisy-pattern/assets/patterns/se/contact/postnummer.toml b/crates/nvisy-pattern/assets/patterns/se/contact/postnummer.toml
deleted file mode 100644
index b85ed9c4c..000000000
--- a/crates/nvisy-pattern/assets/patterns/se/contact/postnummer.toml
+++ /dev/null
@@ -1,23 +0,0 @@
-# Swedish postnummer: 5-digit postal code in `XXX XX` rendering
-# (Postnord standard). Codes start in `10-99x` range; `00xxx` is
-# reserved.
-
-name = "se-postnummer"
-label = "postal_code"
-countries = ["SE"]
-languages = ["sv"]
-context = [
-  "postnummer",
-  "post",
-  "postnord",
-  "postadress",
-  "postal code",
-]
-
-[[variants]]
-regex = '\b[1-9]\d{2}\s\d{2}\b'
-score = 0.5
-
-[[variants]]
-regex = '\b[1-9]\d{4}\b'
-score = 0.2
diff --git a/crates/nvisy-pattern/assets/patterns/se/finance/organisationsnummer.toml b/crates/nvisy-pattern/assets/patterns/se/finance/organisationsnummer.toml
deleted file mode 100644
index 298c5d12f..000000000
--- a/crates/nvisy-pattern/assets/patterns/se/finance/organisationsnummer.toml
+++ /dev/null
@@ -1,21 +0,0 @@
-# Swedish Organisationsnummer: 10 digits issued by Bolagsverket
-# (Swedish Companies Registration Office). Third digit must be
-# ≥ 2 to distinguish from personnummer; Luhn checksum.
-
-name = "se-organisationsnummer"
-label = "company_id"
-countries = ["SE"]
-languages = ["sv"]
-context = [
-  "organisationsnummer",
-  "orgnr",
-  "org nr",
-  "org.nr",
-  "företagsnummer",
-  "bolagsverket",
-]
-
-[[variants]]
-regex = '\b\d{6}[-]?\d{4}\b'
-score = 0.6
-validator = "se.organisationsnummer"
diff --git a/crates/nvisy-pattern/assets/patterns/se/identity/personnummer.toml b/crates/nvisy-pattern/assets/patterns/se/identity/personnummer.toml
deleted file mode 100644
index 033581eea..000000000
--- a/crates/nvisy-pattern/assets/patterns/se/identity/personnummer.toml
+++ /dev/null
@@ -1,29 +0,0 @@
-# Swedish Personnummer: `YYMMDD-XXXX` (10-digit) or
-# `YYYYMMDD-XXXX` (12-digit) personal identity number.
-# Samordningsnummer adds 60 to the day field. Luhn checksum
-# applies to the last 10 digits.
-
-name = "se-personnummer"
-label = "government_id"
-countries = ["SE"]
-languages = ["sv"]
-context = [
-  "personnummer",
-  "svenskt personnummer",
-  "personlig identitet",
-  "samordningsnummer",
-  "ssn",
-  "personal identity number",
-]
-
-# 12-digit form with optional dash/plus separator.
-[[variants]]
-regex = '\b\d{8}[+\-]?\d{4}\b'
-score = 0.5
-validator = "se.personnummer"
-
-# 10-digit form with optional dash/plus separator.
-[[variants]]
-regex = '\b\d{6}[+\-]?\d{4}\b'
-score = 0.5
-validator = "se.personnummer"
diff --git a/crates/nvisy-pattern/assets/patterns/sg/contact/postal_code.toml b/crates/nvisy-pattern/assets/patterns/sg/contact/postal_code.toml
deleted file mode 100644
index f3fedd0ea..000000000
--- a/crates/nvisy-pattern/assets/patterns/sg/contact/postal_code.toml
+++ /dev/null
@@ -1,18 +0,0 @@
-# Singapore postal code: 6 digits where the first two encode the
-# delivery sector (01-82 in current Singapore Post assignments).
-# No published checksum.
-
-name = "sg-postal-code"
-label = "postal_code"
-countries = ["SG"]
-languages = ["en"]
-context = [
-  "postal code",
-  "singapore post",
-  "singpost",
-  "mailing address",
-]
-
-[[variants]]
-regex = '\b(0[1-9]|[1-7]\d|8[0-2])\d{4}\b'
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/patterns/sg/finance/uen.toml b/crates/nvisy-pattern/assets/patterns/sg/finance/uen.toml
deleted file mode 100644
index efb229ab7..000000000
--- a/crates/nvisy-pattern/assets/patterns/sg/finance/uen.toml
+++ /dev/null
@@ -1,37 +0,0 @@
-# Singapore Unique Entity Number (UEN) issued by ACRA. Three
-# formats:
-#
-# - Format A: 9 chars (8 digits + check letter).
-# - Format B: 10 chars (YYYY year + 4 digits + check letter).
-# - Format C: 10 chars (T/S/R + 2 digits + 2-letter entity type
-#   + 4 digits + check letter).
-
-name = "sg-uen"
-label = "company_id"
-countries = ["SG"]
-languages = ["en"]
-context = [
-  "uen",
-  "unique entity number",
-  "acra",
-  "business registration",
-  "company registration",
-]
-
-# Format A.
-[[variants]]
-regex = '\b\d{8}[A-Z]\b'
-score = 0.3
-validator = "sg.uen"
-
-# Format B (must start with a 4-digit year).
-[[variants]]
-regex = '\b\d{9}[A-Z]\b'
-score = 0.3
-validator = "sg.uen"
-
-# Format C.
-[[variants]]
-regex = '\b[TSR]\d{2}[A-Z]{2}\d{4}[A-Z]\b'
-score = 0.5
-validator = "sg.uen"
diff --git a/crates/nvisy-pattern/assets/patterns/sg/identity/nric.toml b/crates/nvisy-pattern/assets/patterns/sg/identity/nric.toml
deleted file mode 100644
index 6d681a1c9..000000000
--- a/crates/nvisy-pattern/assets/patterns/sg/identity/nric.toml
+++ /dev/null
@@ -1,31 +0,0 @@
-# Singapore NRIC / FIN: 9 characters — prefix letter (S/T for
-# Singapore citizens or PRs, F/G/M for foreigners) + 7 digits +
-# check letter. Validator applies the published weighted Mod-11
-# algorithm with prefix-specific letter tables.
-
-name = "sg-nric"
-label = "government_id"
-countries = ["SG"]
-languages = ["en"]
-context = [
-  "nric",
-  "nric#",
-  "fin",
-  "fin#",
-  "national registration",
-  "national registration identity",
-  "ica",
-]
-
-# Tight form anchored to the published prefix set.
-[[variants]]
-regex = '\b[STFGM]\d{7}[A-Z]\b'
-score = 0.5
-validator = "sg.nric"
-
-# Loose form (any leading letter); relies on the validator's
-# prefix check.
-[[variants]]
-regex = '\b[A-Z]\d{7}[A-Z]\b'
-score = 0.2
-validator = "sg.nric"
diff --git a/crates/nvisy-pattern/assets/patterns/th/contact/postal_code.toml b/crates/nvisy-pattern/assets/patterns/th/contact/postal_code.toml
deleted file mode 100644
index 5043f5d3c..000000000
--- a/crates/nvisy-pattern/assets/patterns/th/contact/postal_code.toml
+++ /dev/null
@@ -1,19 +0,0 @@
-# Thai postal code (รหัสไปรษณีย์): 5 digits where the first two
-# encode the province (10-96 in current Thailand Post
-# assignments). No published checksum.
-
-name = "th-postal-code"
-label = "postal_code"
-countries = ["TH"]
-languages = ["th"]
-context = [
-  "รหัสไปรษณีย์",
-  "ไปรษณีย์",
-  "ที่อยู่",
-  "postal code",
-  "thailand post",
-]
-
-[[variants]]
-regex = '\b(1\d|[2-8]\d|9[0-6])\d{3}\b'
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/patterns/th/identity/national_id.toml b/crates/nvisy-pattern/assets/patterns/th/identity/national_id.toml
deleted file mode 100644
index 64836d22e..000000000
--- a/crates/nvisy-pattern/assets/patterns/th/identity/national_id.toml
+++ /dev/null
@@ -1,30 +0,0 @@
-# Thai National ID (เลขประจำตัวประชาชน): 13-digit identifier
-# issued by the Department of Provincial Administration. The
-# first digit encodes the citizenship category (1-8) and the
-# 13th is a weighted Mod-11 check digit.
-
-name = "th-national-id"
-label = "government_id"
-countries = ["TH"]
-languages = ["th"]
-context = [
-  "เลขประจำตัวประชาชน",
-  "เลขบัตรประชาชน",
-  "บัตรประจำตัวประชาชน",
-  "national id",
-  "thai national id",
-  "thai id",
-  "id card",
-]
-
-# Grouped rendering: 1-2345-67890-12-1.
-[[variants]]
-regex = '\b[1-8]-\d{4}-\d{5}-\d{2}-\d\b'
-score = 0.5
-validator = "th.national_id"
-
-# Glued rendering.
-[[variants]]
-regex = '\b[1-8]\d{12}\b'
-score = 0.3
-validator = "th.national_id"
diff --git a/crates/nvisy-pattern/assets/patterns/tr/contact/posta_kodu.toml b/crates/nvisy-pattern/assets/patterns/tr/contact/posta_kodu.toml
deleted file mode 100644
index 5a972255b..000000000
--- a/crates/nvisy-pattern/assets/patterns/tr/contact/posta_kodu.toml
+++ /dev/null
@@ -1,18 +0,0 @@
-# Turkish posta kodu: 5-digit postal code. The first two digits
-# encode the province (01-81), matching the license-plate
-# province code.
-
-name = "tr-posta-kodu"
-label = "postal_code"
-countries = ["TR"]
-languages = ["tr"]
-context = [
-  "posta kodu",
-  "ptt",
-  "adres",
-  "postal code",
-]
-
-[[variants]]
-regex = '\b(0[1-9]|[1-7]\d|8[01])\d{3}\b'
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/patterns/tr/identity/tckn.toml b/crates/nvisy-pattern/assets/patterns/tr/identity/tckn.toml
deleted file mode 100644
index 24de81293..000000000
--- a/crates/nvisy-pattern/assets/patterns/tr/identity/tckn.toml
+++ /dev/null
@@ -1,24 +0,0 @@
-# Turkish T.C. Kimlik No (TCKN): 11-digit national identifier
-# issued by Nüfus ve Vatandaşlık İşleri (NVI) with a two-step
-# weighted checksum. Required on KVKK compliance forms.
-
-name = "tr-tckn"
-label = "government_id"
-countries = ["TR"]
-languages = ["tr"]
-context = [
-  "tc kimlik",
-  "kimlik no",
-  "kimlik numarası",
-  "tckn",
-  "tc no",
-  "nüfus cüzdanı",
-  "national id",
-  "türk kimlik",
-  "turkish id",
-]
-
-[[variants]]
-regex = '\b[1-9]\d{10}\b'
-score = 0.4
-validator = "tr.tckn"
diff --git a/crates/nvisy-pattern/assets/patterns/tr/vehicle/license_plate.toml b/crates/nvisy-pattern/assets/patterns/tr/vehicle/license_plate.toml
deleted file mode 100644
index 2d2af803c..000000000
--- a/crates/nvisy-pattern/assets/patterns/tr/vehicle/license_plate.toml
+++ /dev/null
@@ -1,28 +0,0 @@
-# Turkish license plate (plaka): province code 01-81 + 1-3
-# letters (Turkish alphabet excludes Q, W, X) + 2-4 digits.
-# Civilian format per Karayolları Trafik Kanunu Madde 23.
-
-name = "tr-license-plate"
-label = "license_plate"
-countries = ["TR"]
-languages = ["tr"]
-context = [
-  "plaka",
-  "araç plakası",
-  "plaka numarası",
-  "kayıt plakası",
-  "taşıt plakası",
-  "tr plaka",
-  "license plate",
-  "number plate",
-]
-
-# Space-separated rendering: `34 ABC 1234`.
-[[variants]]
-regex = '\b(0[1-9]|[1-7]\d|8[01])\s?[A-PR-VY-Z]{1,3}\s?\d{2,4}\b'
-score = 0.4
-
-# Hyphen-separated rendering: `34-ABC-1234`.
-[[variants]]
-regex = '\b(0[1-9]|[1-7]\d|8[01])-[A-PR-VY-Z]{1,3}-\d{2,4}\b'
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/uk/contact/postcode.toml b/crates/nvisy-pattern/assets/patterns/uk/contact/postcode.toml
deleted file mode 100644
index 8c0db110e..000000000
--- a/crates/nvisy-pattern/assets/patterns/uk/contact/postcode.toml
+++ /dev/null
@@ -1,23 +0,0 @@
-# UK Postcode.
-#
-# Reference: https://en.wikipedia.org/wiki/Postcodes_in_the_United_Kingdom
-
-name = "uk-postcode"
-label = "postal_code"
-countries = ["GB"]
-languages = ["en"]
-context = [
-  "postcode",
-  "post code",
-  "postal code",
-  "zip",
-  "address",
-  "delivery",
-  "mailing",
-  "shipping",
-  "correspondence",
-]
-
-[[variants]]
-regex = '\b(GIR\s?0AA|[A-PR-UWYZ][0-9][ABCDEFGHJKPSTUW]?\s?[0-9][ABD-HJLNP-UW-Z]{2}|[A-PR-UWYZ][0-9]{2}\s?[0-9][ABD-HJLNP-UW-Z]{2}|[A-PR-UWYZ][A-HK-Y][0-9][ABEHMNPRVWXY]?\s?[0-9][ABD-HJLNP-UW-Z]{2}|[A-PR-UWYZ][A-HK-Y][0-9]{2}\s?[0-9][ABD-HJLNP-UW-Z]{2})\b'
-score = 0.1
diff --git a/crates/nvisy-pattern/assets/patterns/uk/identity/driving_licence.toml b/crates/nvisy-pattern/assets/patterns/uk/identity/driving_licence.toml
deleted file mode 100644
index 85f2e3882..000000000
--- a/crates/nvisy-pattern/assets/patterns/uk/identity/driving_licence.toml
+++ /dev/null
@@ -1,25 +0,0 @@
-# UK Driving Licence Number (DVLA).
-#
-# Format: 16-char alphanumeric — 5-char surname (letters padded
-# with 9s) + decade digit + month (01-12 or 51-62 for female) +
-# day + year digit + 2 initial chars (or 9) + check + 2 letters.
-
-name = "uk-driving-licence"
-label = "drivers_license"
-countries = ["GB"]
-languages = ["en"]
-context = [
-  "driving licence",
-  "driving license",
-  "driver's licence",
-  "driver's license",
-  "dvla",
-  "dl number",
-  "licence number",
-  "license number",
-]
-
-[[variants]]
-regex = '\b[A-Z9]{5}[0-9](?:0[1-9]|1[0-2]|5[1-9]|6[0-2])(?:0[1-9]|[12][0-9]|3[01])[0-9][A-Z9]{2}[A-Z0-9][A-Z]{2}\b'
-score = 0.5
-validator = "uk.driving_licence"
diff --git a/crates/nvisy-pattern/assets/patterns/uk/identity/nhs.toml b/crates/nvisy-pattern/assets/patterns/uk/identity/nhs.toml
deleted file mode 100644
index 8c39ee2ee..000000000
--- a/crates/nvisy-pattern/assets/patterns/uk/identity/nhs.toml
+++ /dev/null
@@ -1,17 +0,0 @@
-# UK National Health Service number.
-
-name = "uk-nhs"
-label = "medical_id"
-countries = ["GB"]
-languages = ["en"]
-context = [
-  "nhs",
-  "national health service",
-  "health services authority",
-  "health authority",
-]
-
-[[variants]]
-regex = '\b([0-9]{3})[- ]?([0-9]{3})[- ]?([0-9]{4})\b'
-score = 0.5
-validator = "uk.nhs"
diff --git a/crates/nvisy-pattern/assets/patterns/uk/identity/nino.toml b/crates/nvisy-pattern/assets/patterns/uk/identity/nino.toml
deleted file mode 100644
index 8de7d04e3..000000000
--- a/crates/nvisy-pattern/assets/patterns/uk/identity/nino.toml
+++ /dev/null
@@ -1,16 +0,0 @@
-# UK National Insurance Number.
-#
-# Rust's `regex` crate doesn't support look-around, so the
-# reserved-prefix exclusions live in the `uk.nino` validator
-# rather than the regex.
-
-name = "uk-nino"
-label = "national_insurance_number"
-countries = ["GB"]
-languages = ["en"]
-context = ["national insurance", "ni number", "nino"]
-
-[[variants]]
-regex = '\b([a-ceghj-pr-tw-zA-CEGHJ-PR-TW-Z][a-ceghj-npr-tw-zA-CEGHJ-NPR-TW-Z]) ?([0-9]{2}) ?([0-9]{2}) ?([0-9]{2}) ?([a-dA-D])\b'
-score = 0.5
-validator = "uk.nino"
diff --git a/crates/nvisy-pattern/assets/patterns/uk/identity/passport.toml b/crates/nvisy-pattern/assets/patterns/uk/identity/passport.toml
deleted file mode 100644
index 03b8b9a48..000000000
--- a/crates/nvisy-pattern/assets/patterns/uk/identity/passport.toml
+++ /dev/null
@@ -1,26 +0,0 @@
-# UK Passport Number (post-2015 format).
-#
-# 2 letters + 7 digits. Same shape as several other country
-# passports — without context-keyword boost the FP rate is high.
-# Low score by default so callers rely on context to lift
-# legitimate matches.
-
-name = "uk-passport"
-label = "passport_number"
-countries = ["GB"]
-languages = ["en"]
-context = [
-  "passport",
-  "passport number",
-  "travel document",
-  "uk passport",
-  "british passport",
-  "her majesty",
-  "his majesty",
-  "hm passport",
-  "hmpo",
-]
-
-[[variants]]
-regex = '\b[A-Z]{2}\d{7}\b'
-score = 0.1
diff --git a/crates/nvisy-pattern/assets/patterns/uk/vehicle/registration.toml b/crates/nvisy-pattern/assets/patterns/uk/vehicle/registration.toml
deleted file mode 100644
index d23047eeb..000000000
--- a/crates/nvisy-pattern/assets/patterns/uk/vehicle/registration.toml
+++ /dev/null
@@ -1,40 +0,0 @@
-# UK Vehicle Registration Number.
-#
-# Covers three eras still seen on the road:
-# - Current (2001+): 2 area letters + 2-digit age + 3 random
-#   letters, e.g. "AB51 ABC"
-# - Prefix (1983-2001): year letter + 1-3 digits + 3 letters
-# - Suffix (1963-1983): 3 letters + 1-3 digits + year letter
-
-name = "uk-vehicle-registration"
-label = "license_plate"
-countries = ["GB"]
-languages = ["en"]
-context = [
-  "vehicle",
-  "registration",
-  "number plate",
-  "licence plate",
-  "license plate",
-  "reg",
-  "vrn",
-  "dvla",
-  "v5c",
-  "logbook",
-  "mot",
-  "car",
-  "insured vehicle",
-]
-
-[[variants]]
-regex = '\b[A-HJ-PR-Y][A-HJ-PR-Y](?:0[1-9]|[1-7][0-9])[- ]?[A-HJ-PR-Z]{3}\b'
-score = 0.3
-validator = "uk.vehicle_registration"
-
-[[variants]]
-regex = '\b[A-HJ-NPR-TV-Y]\d{1,3}[- ]?[A-HJ-PR-Y][A-HJ-PR-Z]{2}\b'
-score = 0.2
-
-[[variants]]
-regex = '\b[A-HJ-PR-Z]{3}[- ]?\d{1,3}[- ]?[A-HJ-NPR-TV-Y]\b'
-score = 0.15
diff --git a/crates/nvisy-pattern/assets/patterns/us/finance/bank_account.toml b/crates/nvisy-pattern/assets/patterns/us/finance/bank_account.toml
deleted file mode 100644
index 03dd30efa..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/finance/bank_account.toml
+++ /dev/null
@@ -1,26 +0,0 @@
-# US Bank Account Number.
-#
-# 8-17 digit run with no checksum — usable only with strong
-# surrounding context. The very-weak score relies on the
-# context-keyword boost to lift legitimate matches above
-# downstream confidence thresholds.
-
-name = "us-bank-account"
-label = "bank_account"
-countries = ["US"]
-languages = ["en"]
-context = [
-  "check",
-  "checking account",
-  "account",
-  "account#",
-  "acct",
-  "bank",
-  "save",
-  "savings",
-  "debit",
-]
-
-[[variants]]
-regex = '\b\d{8,17}\b'
-score = 0.05
diff --git a/crates/nvisy-pattern/assets/patterns/us/finance/bank_routing.toml b/crates/nvisy-pattern/assets/patterns/us/finance/bank_routing.toml
deleted file mode 100644
index 687512aee..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/finance/bank_routing.toml
+++ /dev/null
@@ -1,32 +0,0 @@
-# US ABA Routing Number.
-#
-# Two variants:
-# - 9-digit unbroken form, prefix-validated to the Fed routing
-#   ranges (01-12, 21-32) and checksum-validated by `us.aba_routing`.
-# - Dashed form `\b[0123678]\d{3}-\d{4}-\d\b` with the same
-#   checksum.
-
-name = "us-bank-routing"
-label = "bank_routing"
-countries = ["US"]
-languages = ["en"]
-context = [
-  "routing",
-  "aba",
-  "rtn",
-  "transit",
-  "bank",
-  "deposit",
-  "wire",
-  "ach",
-]
-
-[[variants]]
-regex = '\b(?:0[1-9]|[12]\d|3[0-2])\d{7}\b'
-score = 0.5
-validator = "us.aba_routing"
-
-[[variants]]
-regex = '\b[0123678]\d{3}-\d{4}-\d\b'
-score = 0.4
-validator = "us.aba_routing"
diff --git a/crates/nvisy-pattern/assets/patterns/us/health/mbi.toml b/crates/nvisy-pattern/assets/patterns/us/health/mbi.toml
deleted file mode 100644
index 1359e09d4..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/health/mbi.toml
+++ /dev/null
@@ -1,27 +0,0 @@
-# US Medicare Beneficiary Identifier (MBI).
-#
-# 11-character CMS identifier with position-specific
-# character-class rules and letters restricted to
-# `ACDEFGHJKMNPQRTUVWXY` (excludes S, L, O, I, B, Z).
-
-name = "us-mbi"
-label = "medical_id"
-countries = ["US"]
-languages = ["en"]
-context = [
-  "mbi",
-  "medicare beneficiary",
-  "medicare",
-  "beneficiary",
-  "cms",
-]
-
-# 11-character form (no dashes), positions: N A AN N A AN N A A N N.
-[[variants]]
-regex = '\b[0-9][ACDEFGHJKMNPQRTUVWXY][0-9ACDEFGHJKMNPQRTUVWXY][0-9][ACDEFGHJKMNPQRTUVWXY][0-9ACDEFGHJKMNPQRTUVWXY][0-9][ACDEFGHJKMNPQRTUVWXY][ACDEFGHJKMNPQRTUVWXY][0-9][0-9]\b'
-score = 0.3
-
-# Dashed display form `XXXX-XXX-XXXX`.
-[[variants]]
-regex = '\b[0-9][ACDEFGHJKMNPQRTUVWXY][0-9ACDEFGHJKMNPQRTUVWXY][0-9]-[ACDEFGHJKMNPQRTUVWXY][0-9ACDEFGHJKMNPQRTUVWXY][0-9]-[ACDEFGHJKMNPQRTUVWXY][ACDEFGHJKMNPQRTUVWXY][0-9][0-9]\b'
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/us/health/medical_license.toml b/crates/nvisy-pattern/assets/patterns/us/health/medical_license.toml
deleted file mode 100644
index 3859126e8..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/health/medical_license.toml
+++ /dev/null
@@ -1,36 +0,0 @@
-# US DEA (Drug Enforcement Administration) registration number,
-# used as a medical license identifier.
-#
-# 9-character format: two letters (DEA registration type plus
-# surname initial) followed by 7 digits, the last of which is a
-# checksum. Without context the pattern has a high false-positive
-# rate (matches arbitrary letter-pair plus digit-run strings), so
-# the regex score is low and the boost layer is expected to lift
-# legitimate matches.
-
-name = "us-medical-license"
-label = "medical_id"
-countries = ["US"]
-languages = ["en"]
-context = [
-  "medical",
-  "license",
-  "licence",
-  "certificate",
-  "dea",
-  "controlled substance",
-  "prescription",
-  "prescriber",
-]
-
-# DEA registration type letters: A, B, C, D, E, F, G, H, J, K, L,
-# M, P, R, S, T, U, X (plus mid-2000s practitioner-9 series).
-[[variants]]
-regex = '\b[abcdefghjklmprstuxABCDEFGHJKLMPRSTUX][a-zA-Z]\d{7}\b'
-score = 0.4
-validator = "us.dea_number"
-
-[[variants]]
-regex = '\b[abcdefghjklmprstuxABCDEFGHJKLMPRSTUX]9\d{7}\b'
-score = 0.4
-validator = "us.dea_number"
diff --git a/crates/nvisy-pattern/assets/patterns/us/health/npi.toml b/crates/nvisy-pattern/assets/patterns/us/health/npi.toml
deleted file mode 100644
index 0c6572099..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/health/npi.toml
+++ /dev/null
@@ -1,28 +0,0 @@
-# US National Provider Identifier (NPI).
-#
-# 10-digit number assigned by CMS; validated via Luhn checksum on
-# `80840` + the 10 digits.
-
-name = "us-npi"
-label = "medical_id"
-countries = ["US"]
-languages = ["en"]
-context = [
-  "npi",
-  "national provider",
-  "provider",
-  "npi number",
-  "provider id",
-  "provider identifier",
-  "taxonomy",
-]
-
-[[variants]]
-regex = '\b[12]\d{9}\b'
-score = 0.1
-validator = "us.npi"
-
-[[variants]]
-regex = '\b[12]\d{3}[ -]\d{3}[ -]\d{3}\b'
-score = 0.4
-validator = "us.npi"
diff --git a/crates/nvisy-pattern/assets/patterns/us/identity/drivers_license.toml b/crates/nvisy-pattern/assets/patterns/us/identity/drivers_license.toml
deleted file mode 100644
index a72501ee4..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/identity/drivers_license.toml
+++ /dev/null
@@ -1,26 +0,0 @@
-# US Driver's License (state-shape union).
-#
-# Omits a bare-digit weak variant (`\b[0-9]{6,14}\b`) since it
-# matches phone numbers, dates, etc. and is unusable without
-# strong context.
-
-name = "us-drivers-license"
-label = "drivers_license"
-countries = ["US"]
-languages = ["en"]
-context = [
-  "driver",
-  "drivers",
-  "driver's",
-  "license",
-  "licence",
-  "dl",
-  "dl#",
-  "dlnum",
-]
-
-# Alphanumeric shapes — most US state license formats start with
-# letters, often with a hyphenated structure (e.g. MI: D123-4567-8901).
-[[variants]]
-regex = '\b(?:[A-Z]\d{3}-\d{4}-\d{4}|[A-Z]\d{3,6}|[A-Z]\d{5,9}|[A-Z]\d{6,8}|[A-Z]\d{4,8}|[A-Z]\d{9,11}|[A-Z]{1,2}\d{5,6}|H\d{8}|V\d{6}|X\d{8}|[A-Z]{2}\d{2,5}|[A-Z]{2}\d{3,7}|\d{2}[A-Z]{3}\d{5,6}|[A-Z]\d{13,14}|[A-Z]\d{18}|[A-Z]\d{6}R|[A-Z]\d{9}|[A-Z]\d{1,12}|\d{9}[A-Z]|[A-Z]{2}\d{6}[A-Z]|\d{8}[A-Z]{2}|\d{3}[A-Z]{2}\d{4}|[A-Z]\d[A-Z]\d[A-Z]|\d{7,8}[A-Z])\b'
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/patterns/us/identity/itin.toml b/crates/nvisy-pattern/assets/patterns/us/identity/itin.toml
deleted file mode 100644
index 726861f58..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/identity/itin.toml
+++ /dev/null
@@ -1,28 +0,0 @@
-# US Individual Taxpayer Identification Number (ITIN).
-#
-# Format: 9NN-MM-NNNN where the middle group restricts to the
-# IRS-published ranges. No checksum.
-
-name = "us-itin"
-label = "tax_id"
-countries = ["US"]
-languages = ["en"]
-context = [
-  "individual",
-  "taxpayer",
-  "itin",
-  "tax",
-  "payer",
-  "taxid",
-  "tin",
-]
-
-# Medium: dashes / spaces around the IRS-published middle range.
-[[variants]]
-regex = '\b9\d{2}[- ](?:5\d|6[0-5]|7\d|8[0-8]|9(?:[0-2]|[4-9]))[- ]\d{4}\b'
-score = 0.5
-
-# Weak: unbroken 9-digit form.
-[[variants]]
-regex = '\b9\d{2}(?:5\d|6[0-5]|7\d|8[0-8]|9(?:[0-2]|[4-9]))\d{4}\b'
-score = 0.3
diff --git a/crates/nvisy-pattern/assets/patterns/us/identity/passport.toml b/crates/nvisy-pattern/assets/patterns/us/identity/passport.toml
deleted file mode 100644
index 7c6508479..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/identity/passport.toml
+++ /dev/null
@@ -1,14 +0,0 @@
-name = "us-passport"
-label = "passport_number"
-countries = ["US"]
-context = [
-  "passport",
-  "passport#",
-  "travel document",
-  "us passport",
-  "united states passport",
-]
-
-[[variants]]
-regex = "\\b[A-Z]\\d{8}\\b"
-score = 0.1
diff --git a/crates/nvisy-pattern/assets/patterns/us/identity/postal_code.toml b/crates/nvisy-pattern/assets/patterns/us/identity/postal_code.toml
deleted file mode 100644
index f6fd66f96..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/identity/postal_code.toml
+++ /dev/null
@@ -1,17 +0,0 @@
-name = "us-postal-code"
-label = "postal_code"
-countries = ["US"]
-context = [
-  "zip",
-  "zip code",
-  "zipcode",
-  "postal",
-  "postal code",
-  "address",
-  "mailing",
-]
-
-[[variants]]
-regex = "\\b\\d{5}(?:-\\d{4})?\\b"
-score = 0.1
-validator = "us.postal_code"
diff --git a/crates/nvisy-pattern/assets/patterns/us/identity/ssn.toml b/crates/nvisy-pattern/assets/patterns/us/identity/ssn.toml
deleted file mode 100644
index 18b110346..000000000
--- a/crates/nvisy-pattern/assets/patterns/us/identity/ssn.toml
+++ /dev/null
@@ -1,14 +0,0 @@
-# US Social Security Number.
-#
-# Omits very-weak variants (bare 9 digits, off-position splits).
-# Context-keyword boost is expected to lift legitimate matches.
-
-name = "ssn"
-label = "government_id"
-countries = ["US"]
-context = ["social security", "ssn", "tax id", "taxpayer identification"]
-
-[[variants]]
-regex = '\b(\d{3})[- .](\d{2})[- .](\d{4})\b'
-score = 0.5
-validator = "us.ssn"
diff --git a/crates/nvisy-pattern/assets/patterns/world/contact/email.toml b/crates/nvisy-pattern/assets/patterns/world/contact/email.toml
deleted file mode 100644
index 68114a4c1..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/contact/email.toml
+++ /dev/null
@@ -1,12 +0,0 @@
-# Email address (RFC 5322-loose).
-#
-# Accepts the RFC 5322 atom characters in the local part
-# (`!#$%&'*+\-/=?^_`{|}~`) so plus-addressing, dotted locals, and
-# similar are caught.
-
-name = "email"
-label = "email_address"
-
-[[variants]]
-regex = '''\b(?:[!#$%&'*+\-/=?^_`{|}~\w]|[!#$%&'*+\-/=?^_`{|}~\w][!#$%&'*+\-/=?^_`{|}~.\w]{0,}[!#$%&'*+\-/=?^_`{|}~\w])@\w+(?:[-.]\w+)*\.\w+(?:[-.]\w+)*\b'''
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/world/contact/phone.toml b/crates/nvisy-pattern/assets/patterns/world/contact/phone.toml
deleted file mode 100644
index 83e0bc963..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/contact/phone.toml
+++ /dev/null
@@ -1,13 +0,0 @@
-name = "phone"
-label = "phone_number"
-
-[context]
-en = ["phone", "call", "mobile", "tel", "fax", "contact"]
-es = ["teléfono", "telefono", "llamar", "móvil", "movil", "celular", "tel", "fax", "contacto"]
-de = ["telefon", "anruf", "mobil", "handy", "tel", "fax", "kontakt"]
-fr = ["téléphone", "telephone", "appel", "mobile", "portable", "tel", "fax", "contact"]
-
-[[variants]]
-regex = "(?:\\+\\d{1,3}[\\s.\\-]?)?\\(?\\d{2,4}\\)?[\\s.\\-]?\\d{3,4}[\\s.\\-]?\\d{4}\\b"
-score = 0.4
-validator = "phone"
diff --git a/crates/nvisy-pattern/assets/patterns/world/contact/url.toml b/crates/nvisy-pattern/assets/patterns/world/contact/url.toml
deleted file mode 100644
index 7096570a0..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/contact/url.toml
+++ /dev/null
@@ -1,6 +0,0 @@
-name = "url"
-label = "url"
-
-[[variants]]
-regex = "\\bhttps?://[^\\s/$.?#][^\\s]*\\b"
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/world/credentials/aws_key.toml b/crates/nvisy-pattern/assets/patterns/world/credentials/aws_key.toml
deleted file mode 100644
index b4320c3cb..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/credentials/aws_key.toml
+++ /dev/null
@@ -1,26 +0,0 @@
-name = "aws-key"
-label = "api_key"
-context = [
-  "aws",
-  "amazon",
-  "access key",
-  "secret access key",
-  "aws_access_key_id",
-  "aws_secret_access_key",
-]
-
-# AWS access key ID — 20 chars, fixed 4-letter principal prefix
-# plus 16 base32-ish chars. Prefixes per AWS docs: AKIA (IAM
-# user), ASIA (STS temporary), AIDA (IAM user identifier), AROA
-# (IAM role), ANPA / AGPA (managed policy / group), AIPA (EC2
-# instance profile).
-[[variants]]
-regex = "\\b(?:AKIA|ASIA|AIDA|AROA|ANPA|AGPA|AIPA)[0-9A-Z]{16}\\b"
-score = 0.95
-
-# AWS secret access key — 40 chars of base64url. Without context
-# this collides with any 40-char base64 string (hashes, tokens),
-# so the score is modest; the boost layer lifts colocated hits.
-[[variants]]
-regex = "\\b[A-Za-z0-9/+=]{40}\\b"
-score = 0.3
diff --git a/crates/nvisy-pattern/assets/patterns/world/credentials/generic_api_key.toml b/crates/nvisy-pattern/assets/patterns/world/credentials/generic_api_key.toml
deleted file mode 100644
index c67d49568..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/credentials/generic_api_key.toml
+++ /dev/null
@@ -1,10 +0,0 @@
-name = "generic-api-key"
-label = "api_key"
-
-# Separator after the keyword is either `:` / `=` (assignment
-# style: `api_key="…"`) or one-or-more spaces (header style:
-# `Authorization: Bearer <token>`). Accepting whitespace-only is
-# what catches the dominant Authorization-header leak form.
-[[variants]]
-regex = "(?i)(?:api[_\\-]?key|api[_\\-]?secret|access[_\\-]?token|secret[_\\-]?key|bearer)(?:\\s*[:=]\\s*|\\s+)[\"']?([a-zA-Z0-9_\\-\\.]{20,})[\"']?"
-score = 0.4
diff --git a/crates/nvisy-pattern/assets/patterns/world/credentials/github_token.toml b/crates/nvisy-pattern/assets/patterns/world/credentials/github_token.toml
deleted file mode 100644
index a1a76b41d..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/credentials/github_token.toml
+++ /dev/null
@@ -1,14 +0,0 @@
-name = "github-token"
-label = "auth_token"
-
-# Classic prefix tokens — gh{p,o,u,s,r}_ for PAT, OAuth,
-# user-to-server, server-to-server, refresh.
-[[variants]]
-regex = "\\bgh[pousr]_[a-zA-Z0-9]{36}\\b"
-score = 0.95
-
-# Fine-grained personal access token (introduced 2022) — current
-# recommended PAT form, 82-char body over [A-Z0-9_].
-[[variants]]
-regex = "\\bgithub_pat_[A-Z0-9_]{82}\\b"
-score = 0.95
diff --git a/crates/nvisy-pattern/assets/patterns/world/credentials/private_key.toml b/crates/nvisy-pattern/assets/patterns/world/credentials/private_key.toml
deleted file mode 100644
index 7f0aaf597..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/credentials/private_key.toml
+++ /dev/null
@@ -1,17 +0,0 @@
-name = "private-key"
-label = "private_key"
-
-[[variants]]
-regex = "-----BEGIN (?:RSA |EC |DSA |OPENSSH |ENCRYPTED |PGP )?PRIVATE KEY(?: BLOCK)?-----[\\s\\S]*?-----END (?:RSA |EC |DSA |OPENSSH |ENCRYPTED |PGP )?PRIVATE KEY(?: BLOCK)?-----"
-score = 0.98
-
-# RFC 4716 SSH2 private key (Tectia, older SSH implementations).
-[[variants]]
-regex = "---- BEGIN SSH2 ENCRYPTED PRIVATE KEY ----[\\s\\S]*?---- END SSH2 ENCRYPTED PRIVATE KEY ----"
-score = 0.98
-
-# PuTTY .ppk (PuTTY-User-Key-File-2 or -3). Body extends to the
-# `Private-MAC:` trailer which closes the keystore.
-[[variants]]
-regex = "PuTTY-User-Key-File-[23]:[\\s\\S]*?Private-MAC: [0-9a-f]+"
-score = 0.98
diff --git a/crates/nvisy-pattern/assets/patterns/world/credentials/stripe_key.toml b/crates/nvisy-pattern/assets/patterns/world/credentials/stripe_key.toml
deleted file mode 100644
index 127517f68..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/credentials/stripe_key.toml
+++ /dev/null
@@ -1,6 +0,0 @@
-name = "stripe-key"
-label = "api_key"
-
-[[variants]]
-regex = "\\bsk_(live|test)_[a-zA-Z0-9]{24,}\\b"
-score = 0.95
diff --git a/crates/nvisy-pattern/assets/patterns/world/finance/bitcoin_address.toml b/crates/nvisy-pattern/assets/patterns/world/finance/bitcoin_address.toml
deleted file mode 100644
index c4c3118c9..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/finance/bitcoin_address.toml
+++ /dev/null
@@ -1,17 +0,0 @@
-name = "bitcoin-address"
-label = "crypto_address"
-
-# Legacy Base58 addresses (P2PKH `1…`, P2SH `3…`) carry a
-# four-byte double-SHA256 checksum that the `crypto.btc`
-# validator verifies.
-[[variants]]
-regex = "\\b[13][a-km-zA-HJ-NP-Z1-9]{25,34}\\b"
-score = 0.5
-validator = "crypto.btc"
-
-# Bech32 / Bech32m segwit + Taproot. Length window covers
-# v0 P2WPKH/P2WSH (42/62 chars total, body 39/59) and
-# v1 P2TR Taproot (62 chars total, body 59).
-[[variants]]
-regex = "\\bbc1[a-z0-9]{25,59}\\b"
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/world/finance/credit_card.toml b/crates/nvisy-pattern/assets/patterns/world/finance/credit_card.toml
deleted file mode 100644
index a06bd0fec..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/finance/credit_card.toml
+++ /dev/null
@@ -1,22 +0,0 @@
-# Credit / debit card number.
-#
-# Brand-prefix variants matching Visa, Mastercard, Discover, etc.
-# Luhn-validated. The Luhn validator drops false positives that
-# a length-based fallback regex would otherwise pick up.
-
-name = "credit-card"
-label = "payment_card"
-
-[context]
-en = ["card", "credit", "debit", "payment", "visa", "mastercard", "amex"]
-es = ["tarjeta", "crédito", "credito", "débito", "debito", "pago", "visa", "mastercard", "amex"]
-de = ["karte", "kredit", "kreditkarte", "debit", "zahlung", "visa", "mastercard", "amex"]
-fr = ["carte", "crédit", "credit", "débit", "debit", "paiement", "visa", "mastercard", "amex"]
-
-# Brand BIN ranges: Visa (4), Mastercard (51-55 and 2-series
-# 2221-2720 introduced 2017), Discover (6), Amex/Diners (3),
-# and a loose 1xxx catch-all that Luhn filters down.
-[[variants]]
-regex = '\b(?:(?:4\d{3})|(?:5[0-5]\d{2})|(?:2[2-7]\d{2})|(?:6\d{3})|(?:1\d{3})|(?:3\d{3}))[- ]?(?:\d{3,4})[- ]?(?:\d{3,4})[- ]?(?:\d{3,5})\b'
-score = 0.3
-validator = "luhn"
diff --git a/crates/nvisy-pattern/assets/patterns/world/finance/ethereum_address.toml b/crates/nvisy-pattern/assets/patterns/world/finance/ethereum_address.toml
deleted file mode 100644
index 939ed8188..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/finance/ethereum_address.toml
+++ /dev/null
@@ -1,6 +0,0 @@
-name = "ethereum-address"
-label = "crypto_address"
-
-[[variants]]
-regex = "\\b0x[0-9a-fA-F]{40}\\b"
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/world/finance/iban.toml b/crates/nvisy-pattern/assets/patterns/world/finance/iban.toml
deleted file mode 100644
index 79d2bf30c..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/finance/iban.toml
+++ /dev/null
@@ -1,8 +0,0 @@
-name = "iban"
-label = "iban"
-context = ["iban", "bank", "account", "transfer", "swift"]
-
-[[variants]]
-regex = "\\b[A-Z]{2}\\d{2}[\\s\\-]?[A-Z0-9]{4}[\\s\\-]?(?:[A-Z0-9]{4}[\\s\\-]?){2,7}[A-Z0-9]{1,4}\\b"
-score = 0.5
-validator = "iban"
diff --git a/crates/nvisy-pattern/assets/patterns/world/finance/swift_code.toml b/crates/nvisy-pattern/assets/patterns/world/finance/swift_code.toml
deleted file mode 100644
index e1f1a60c9..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/finance/swift_code.toml
+++ /dev/null
@@ -1,6 +0,0 @@
-name = "swift-code"
-label = "swift_code"
-
-[[variants]]
-regex = "\\b[A-Z]{4}[A-Z]{2}[A-Z0-9]{2}(?:[A-Z0-9]{3})?\\b"
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/world/network/ipv4.toml b/crates/nvisy-pattern/assets/patterns/world/network/ipv4.toml
deleted file mode 100644
index 926d88904..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/network/ipv4.toml
+++ /dev/null
@@ -1,8 +0,0 @@
-# IPv4 address with optional CIDR suffix.
-
-name = "ipv4"
-label = "ip_address"
-
-[[variants]]
-regex = '\b(?:25[0-5]|2[0-4]\d|[01]?\d\d?)\.(?:25[0-5]|2[0-4]\d|[01]?\d\d?)\.(?:25[0-5]|2[0-4]\d|[01]?\d\d?)\.(?:25[0-5]|2[0-4]\d|[01]?\d\d?)(?:/(?:[0-2]?\d|3[0-2]))?\b'
-score = 0.6
diff --git a/crates/nvisy-pattern/assets/patterns/world/network/ipv6.toml b/crates/nvisy-pattern/assets/patterns/world/network/ipv6.toml
deleted file mode 100644
index 5ff023176..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/network/ipv6.toml
+++ /dev/null
@@ -1,25 +0,0 @@
-# IPv6 address (comprehensive form set).
-#
-# Rust's `regex` crate doesn't support look-around, so the
-# variants use `\b` boundaries to limit match overlap inside
-# larger IPv6-shaped strings. Over-matching at sub-token
-# boundaries is rare in practice and gets collapsed by the entity
-# dedup pipeline downstream.
-
-name = "ipv6"
-label = "ip_address"
-
-# IPv4-mapped IPv6 (::ffff:0:0:1.2.3.4 et al).
-[[variants]]
-regex = '\b::(?:ffff(?::0{1,4})?:)?(?:(?:25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)\.){3}(?:25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)(?:/(?:12[0-8]|1[01]\d|[1-9]?\d))?\b'
-score = 0.6
-
-# IPv4-embedded IPv6 (X:X:…:1.2.3.4).
-[[variants]]
-regex = '\b(?:(?:[0-9A-Fa-f]{1,4}:){1,5}:(?:[0-9A-Fa-f]{1,4}:){0,4}|(?:[0-9A-Fa-f]{1,4}:){6})(?:(?:25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)\.){3}(?:25[0-5]|2[0-4][0-9]|[01]?[0-9][0-9]?)(?:/(?:12[0-8]|1[01]\d|[1-9]?\d))?\b'
-score = 0.6
-
-# Standard IPv6 with all compressed forms.
-[[variants]]
-regex = '\b(?:(?:[0-9A-Fa-f]{1,4}:){7}[0-9A-Fa-f]{1,4}|(?:[0-9A-Fa-f]{1,4}:){1,7}:|:(?::[0-9A-Fa-f]{1,4}){1,7}|(?:[0-9A-Fa-f]{1,4}:){1,6}:[0-9A-Fa-f]{1,4}|(?:[0-9A-Fa-f]{1,4}:){1,5}(?::[0-9A-Fa-f]{1,4}){1,2}|(?:[0-9A-Fa-f]{1,4}:){1,4}(?::[0-9A-Fa-f]{1,4}){1,3}|(?:[0-9A-Fa-f]{1,4}:){1,3}(?::[0-9A-Fa-f]{1,4}){1,4}|(?:[0-9A-Fa-f]{1,4}:){1,2}(?::[0-9A-Fa-f]{1,4}){1,5}|[0-9A-Fa-f]{1,4}:(?::[0-9A-Fa-f]{1,4}){1,6}|:(?::[0-9A-Fa-f]{1,4}){1,6})(?:%[0-9a-zA-Z]+)?(?:/(?:12[0-8]|1[01]\d|[1-9]?\d))?\b'
-score = 0.6
diff --git a/crates/nvisy-pattern/assets/patterns/world/network/mac_address.toml b/crates/nvisy-pattern/assets/patterns/world/network/mac_address.toml
deleted file mode 100644
index c74250507..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/network/mac_address.toml
+++ /dev/null
@@ -1,16 +0,0 @@
-# MAC (Ethernet) address.
-#
-# Two variants:
-# - Colon/hyphen form: `00:1A:2B:3C:4D:5E` or `00-1A-2B-3C-4D-5E`.
-# - Cisco dot-form: `0000.0000.0000` (three 4-hex-digit groups).
-
-name = "mac-address"
-label = "mac_address"
-
-[[variants]]
-regex = "\\b(?:[0-9A-Fa-f]{2}[:\\-]){5}[0-9A-Fa-f]{2}\\b"
-score = 0.5
-
-[[variants]]
-regex = "\\b[0-9A-Fa-f]{4}\\.[0-9A-Fa-f]{4}\\.[0-9A-Fa-f]{4}\\b"
-score = 0.5
diff --git a/crates/nvisy-pattern/assets/patterns/world/personal/date_of_birth.toml b/crates/nvisy-pattern/assets/patterns/world/personal/date_of_birth.toml
deleted file mode 100644
index 0045927ba..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/personal/date_of_birth.toml
+++ /dev/null
@@ -1,13 +0,0 @@
-name = "date_of_birth"
-label = "date_of_birth"
-
-[context]
-en = ["birth", "born", "dob", "birthday", "date of birth"]
-es = ["nacimiento", "nacido", "nacida", "fecha de nacimiento", "cumpleaños", "cumpleanos"]
-de = ["geburt", "geboren", "geburtsdatum", "geburtstag"]
-fr = ["naissance", "né", "nee", "née", "date de naissance", "anniversaire"]
-
-[[variants]]
-regex = "\\b(?:(?:0[1-9]|1[0-2]|[1-9])[/\\-](?:0[1-9]|[12]\\d|3[01]|[1-9])[/\\-](?:19|20)\\d{2}|(?:19|20)\\d{2}[/\\-](?:0[1-9]|1[0-2])[/\\-](?:0[1-9]|[12]\\d|3[01]))\\b"
-score = 0.1
-validator = "date"
diff --git a/crates/nvisy-pattern/assets/patterns/world/personal/datetime.toml b/crates/nvisy-pattern/assets/patterns/world/personal/datetime.toml
deleted file mode 100644
index 76d0e1df2..000000000
--- a/crates/nvisy-pattern/assets/patterns/world/personal/datetime.toml
+++ /dev/null
@@ -1,12 +0,0 @@
-name = "datetime"
-label = "date_time"
-
-[context]
-en = ["timestamp", "created", "modified", "logged", "at", "time"]
-es = ["marca de tiempo", "creado", "creada", "modificado", "modificada", "registrado", "a las", "hora", "fecha"]
-de = ["zeitstempel", "erstellt", "geändert", "geandert", "protokolliert", "um", "uhrzeit", "zeit"]
-fr = ["horodatage", "créé", "cree", "créée", "creee", "modifié", "modifie", "à", "heure", "date"]
-
-[[variants]]
-regex = "\\b(?:19|20)\\d{2}[/\\-](?:0[1-9]|1[0-2])[/\\-](?:0[1-9]|[12]\\d|3[01])[T ](?:[01]\\d|2[0-3]):[0-5]\\d(?::[0-5]\\d)?(?:Z|[+\\-]\\d{2}:?\\d{2})?\\b"
-score = 0.1
diff --git a/crates/nvisy-pattern/src/lib.rs b/crates/nvisy-pattern/src/lib.rs
deleted file mode 100644
index ca00f5f91..000000000
--- a/crates/nvisy-pattern/src/lib.rs
+++ /dev/null
@@ -1,13 +0,0 @@
-#![forbid(unsafe_code)]
-#![cfg_attr(docsrs, feature(doc_cfg))]
-#![doc = include_str!("../README.md")]
-
-mod recognition;
-mod shipped;
-pub mod validators;
-
-pub use self::recognition::{
-    Context, Dictionary, DictionaryBuilder, PatternRecognizer, PatternRecognizerBuilder, Regex,
-    RegexBuilder, Scoring, Term, Variant,
-};
-pub use self::shipped::{dictionaries, patterns};
diff --git a/crates/nvisy-pattern/src/recognition/compiled.rs b/crates/nvisy-pattern/src/recognition/compiled.rs
deleted file mode 100644
index c9fe8d6e2..000000000
--- a/crates/nvisy-pattern/src/recognition/compiled.rs
+++ /dev/null
@@ -1,172 +0,0 @@
-//! Compiled, recognizer-ready forms of [`Regex`] rules and
-//! [`Dictionary`]s.
-//!
-//! [`PatternRecognizerBuilder::build`] compiles each regex variant
-//! into a [`::regex::Regex`] and folds every dictionary's terms
-//! into a shared [`AhoCorasick`] automaton, then stores the
-//! per-rule emission metadata next to those scanners. This module
-//! holds the per-rule metadata structs ([`CompiledPattern`],
-//! [`CompiledDictionary`]) and their `build_entity` constructors —
-//! the bits that turn a regex / Aho-Corasick hit into an
-//! `Entity<Text>`.
-//!
-//! [`Regex`]: super::Regex
-//! [`Dictionary`]: super::Dictionary
-//! [`AhoCorasick`]: aho_corasick::AhoCorasick
-//! [`PatternRecognizerBuilder::build`]: super::PatternRecognizerBuilder::build
-
-use std::sync::Arc;
-
-use nvisy_core::entity::{Entity, EntityLabelRef, PatternProvenance, TrailProvenance, TrailStep};
-use nvisy_core::modality::{Text, TextLocation};
-use nvisy_core::primitive::{Confidence, CountryCode, LanguageTag};
-use regex::Regex;
-
-use crate::validators::Validator;
-
-/// One compiled regex slot: a single `(pattern, variant)` pair,
-/// keyed in the shared `RegexSet` by its position in
-/// `PatternRecognizer.patterns`. Pattern-level metadata (name,
-/// label, languages) is repeated across the pattern's variants so
-/// the dispatch loop has everything it needs without a second
-/// indirection.
-///
-/// `context` is intentionally not stored on compiled state — the
-/// recognizer's wrapping `ContextEnhanced` layer harvests keywords from
-/// the source patterns at build time.
-pub(super) struct CompiledPattern {
-    /// Pattern name (e.g. `"ssn"`). Surfaced in trail provenance.
-    pub pattern_name: String,
-    pub label: EntityLabelRef,
-    pub regex: Regex,
-    pub score: Confidence,
-    pub validator: Option<Arc<dyn Validator>>,
-    /// Languages the parent pattern applies to.
-    /// Empty means "any language".
-    pub languages: Vec<LanguageTag>,
-    /// Countries the parent pattern applies to.
-    /// Empty means "any country".
-    pub countries: Vec<CountryCode>,
-}
-
-impl CompiledPattern {
-    /// Emit an `Entity<Text>` for a regex match at `[start, end)`
-    /// in modality-local byte coordinates. The recognizer phase
-    /// lifts the location to absolute document coordinates after
-    /// dispatch.
-    pub(super) fn build_entity(&self, start: usize, end: usize) -> Entity<Text> {
-        let provenance = TrailProvenance::Pattern(PatternProvenance::Regex {
-            name: self.pattern_name.clone(),
-            regex: Some(self.regex.as_str().to_owned()),
-            validator: self.validator.as_ref().map(|_| self.pattern_name.clone()),
-            contextual: false,
-        });
-        let step = TrailStep::recognition(
-            "pattern",
-            self.score,
-            provenance,
-            format!("pattern `{}` matched", self.pattern_name),
-        );
-        Entity::builder()
-            .with_label(self.label.clone())
-            .with_trail(vec![step])
-            .with_confidence(self.score)
-            .with_location(TextLocation::new(start, end))
-            .build()
-            .expect("required fields provided")
-    }
-}
-
-/// Source of truth for one runtime dictionary: its term range
-/// inside the shared Aho-Corasick automaton, plus per-dictionary
-/// emission metadata.
-pub(super) struct CompiledDictionary {
-    pub name: String,
-    pub label: EntityLabelRef,
-    /// First term-id (inclusive) for this dictionary inside the
-    /// shared automaton.
-    pub term_start: usize,
-    /// One past the last term-id for this dictionary inside the
-    /// shared automaton.
-    pub term_end: usize,
-    /// Per-term confidence, indexed by `term_id - term_start`.
-    /// Resolved at compile time from the dictionary's `scoring`
-    /// policy and any per-term overrides.
-    pub term_scores: Vec<Confidence>,
-    /// Languages this dictionary applies to. Empty means "any
-    /// language".
-    pub languages: Vec<LanguageTag>,
-    /// Countries this dictionary applies to. Empty means "any
-    /// country".
-    pub countries: Vec<CountryCode>,
-    /// Reject matches whose immediate neighbours are word
-    /// characters (alphanumeric or `_`). Mirrors regex `\b`.
-    pub word_boundary: bool,
-}
-
-impl CompiledDictionary {
-    /// Emit an `Entity<Text>` for an Aho-Corasick hit at
-    /// `[start, end)` in modality-local byte coordinates. `score`
-    /// is the per-term confidence resolved at recognizer-build
-    /// time (the dictionary's `scoring` policy or per-term
-    /// override).
-    pub(super) fn build_entity(&self, score: Confidence, start: usize, end: usize) -> Entity<Text> {
-        let provenance = TrailProvenance::Pattern(PatternProvenance::Dictionary {
-            name: self.name.clone(),
-            contextual: false,
-        });
-        let step = TrailStep::recognition(
-            "pattern",
-            score,
-            provenance,
-            format!("dictionary `{}` matched", self.name),
-        );
-        Entity::builder()
-            .with_label(self.label.clone())
-            .with_trail(vec![step])
-            .with_confidence(score)
-            .with_location(TextLocation::new(start, end))
-            .build()
-            .expect("required fields provided")
-    }
-}
-
-/// Mirror of regex `\b` for the byte range `text[start..end]`:
-/// the immediate neighbour characters (or start/end of input)
-/// must not be word characters. A word character here is Unicode
-/// alphanumeric or `_`, matching the conventional regex
-/// definition.
-///
-/// Operates on `char` boundaries, not raw bytes, so multibyte
-/// codepoints don't trigger false rejections (`é` is one char,
-/// not two).
-pub(super) fn has_word_boundaries(text: &str, start: usize, end: usize) -> bool {
-    let left_is_word = text[..start].chars().next_back().is_some_and(is_word_char);
-    let right_is_word = text[end..].chars().next().is_some_and(is_word_char);
-    !left_is_word && !right_is_word
-}
-
-fn is_word_char(c: char) -> bool {
-    c.is_alphanumeric() || c == '_'
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn has_word_boundaries_handles_edges_and_unicode() {
-        // Match touches both edges of the input → boundaries OK.
-        assert!(has_word_boundaries("hello", 0, 5));
-        // Match preceded by a word char → not a boundary.
-        assert!(!has_word_boundaries("example", 5, 7));
-        // Match followed by a word char → not a boundary.
-        assert!(!has_word_boundaries("amount", 0, 2));
-        // Space surround → boundaries OK.
-        assert!(has_word_boundaries(" am ", 1, 3));
-        // Unicode word char on the left → not a boundary.
-        assert!(!has_word_boundaries("café_am", 5, 7));
-        // Punctuation around → boundaries OK.
-        assert!(has_word_boundaries("(am)", 1, 3));
-    }
-}
diff --git a/crates/nvisy-pattern/src/recognition/context.rs b/crates/nvisy-pattern/src/recognition/context.rs
deleted file mode 100644
index 90f375040..000000000
--- a/crates/nvisy-pattern/src/recognition/context.rs
+++ /dev/null
@@ -1,163 +0,0 @@
-//! [`Context`]: per-rule keyword set used by the post-recognition
-//! [`ContextEnhanced`] layer.
-//!
-//! Two shapes:
-//!
-//! - [`Global`] — one flat keyword list applied regardless of the
-//!   per-call language hint.
-//! - [`PerLanguage`] — keyword lists keyed by [`LanguageTag`]; the
-//!   enhancer picks the entry matching `RecognizerInput.language`.
-//!   When no language hint is set, the union of every per-language
-//!   keyword fires (matches the crate's "missing language = any"
-//!   theme used by [`Regex::languages`] / [`Dictionary::languages`]).
-//!
-//! [`Global`]: Context::Global
-//! [`PerLanguage`]: Context::PerLanguage
-//! [`ContextEnhanced`]: nvisy_context::ContextEnhanced
-//! [`Regex::languages`]: super::Regex::languages
-//! [`Dictionary::languages`]: super::Dictionary::languages
-
-use std::collections::HashMap;
-use std::collections::hash_map::Iter;
-
-use derive_more::From;
-use nvisy_core::primitive::LanguageTag;
-use serde::Deserialize;
-
-/// Per-rule context keyword set.
-///
-/// Either a single flat list ([`Global`]) or a map keyed by
-/// language ([`PerLanguage`]).
-///
-/// [`Global`]: Self::Global
-/// [`PerLanguage`]: Self::PerLanguage
-#[derive(Debug, Clone, PartialEq, Eq, Deserialize, From)]
-#[serde(untagged)]
-pub enum Context {
-    /// One flat keyword list applied regardless of the per-call
-    /// language hint.
-    Global(Vec<String>),
-    /// Per-language keyword lists. The enhancer picks the entry
-    /// matching `RecognizerInput.language`, or unions every list
-    /// when no hint is set.
-    PerLanguage(HashMap<LanguageTag, Vec<String>>),
-}
-
-impl Context {
-    /// Return `true` when no keywords are declared in any scope.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        match self {
-            Self::Global(kws) => kws.is_empty(),
-            Self::PerLanguage(map) => map.values().all(Vec::is_empty),
-        }
-    }
-
-    /// Iterate over `(language, keywords)` pairs.
-    ///
-    /// [`Global`] yields one entry with `language = None`;
-    /// [`PerLanguage`] yields one entry per language.
-    ///
-    /// [`Global`]: Self::Global
-    /// [`PerLanguage`]: Self::PerLanguage
-    pub fn iter(&self) -> ContextIter<'_> {
-        match self {
-            Self::Global(kws) => ContextIter::Global(Some(kws.as_slice())),
-            Self::PerLanguage(map) => ContextIter::PerLanguage(map.iter()),
-        }
-    }
-}
-
-impl Default for Context {
-    fn default() -> Self {
-        Self::Global(Vec::new())
-    }
-}
-
-/// Iterator returned by [`Context::iter`].
-pub enum ContextIter<'a> {
-    Global(Option<&'a [String]>),
-    PerLanguage(Iter<'a, LanguageTag, Vec<String>>),
-}
-
-impl<'a> Iterator for ContextIter<'a> {
-    type Item = (Option<&'a LanguageTag>, &'a [String]);
-
-    fn next(&mut self) -> Option<Self::Item> {
-        match self {
-            Self::Global(slot) => slot.take().map(|kws| (None, kws)),
-            Self::PerLanguage(it) => it.next().map(|(lang, kws)| (Some(lang), kws.as_slice())),
-        }
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[derive(Deserialize)]
-    struct Wrap {
-        context: Context,
-    }
-
-    #[test]
-    fn parses_flat_array_as_global() {
-        let toml = r#"context = ["a", "b"]"#;
-        let w: Wrap = toml::from_str(toml).unwrap();
-        assert_eq!(w.context, Context::Global(vec!["a".into(), "b".into()]));
-    }
-
-    #[test]
-    fn parses_table_as_per_language() {
-        let toml = r#"
-            [context]
-            en = ["card"]
-            es = ["tarjeta"]
-        "#;
-        let w: Wrap = toml::from_str(toml).unwrap();
-        let map = match w.context {
-            Context::PerLanguage(m) => m,
-            _ => panic!("expected PerLanguage"),
-        };
-        assert_eq!(map.len(), 2);
-        assert_eq!(
-            map.get(&LanguageTag::new("en").unwrap()).unwrap(),
-            &vec!["card".to_owned()]
-        );
-        assert_eq!(
-            map.get(&LanguageTag::new("es").unwrap()).unwrap(),
-            &vec!["tarjeta".to_owned()]
-        );
-    }
-
-    #[test]
-    fn iter_global_yields_one_none_entry() {
-        let ctx = Context::Global(vec!["a".into(), "b".into()]);
-        let collected: Vec<_> = ctx
-            .iter()
-            .map(|(lang, kws)| (lang.cloned(), kws.to_vec()))
-            .collect();
-        assert_eq!(collected.len(), 1);
-        assert!(collected[0].0.is_none());
-        assert_eq!(collected[0].1, vec!["a".to_owned(), "b".to_owned()]);
-    }
-
-    #[test]
-    fn iter_per_language_yields_one_entry_per_language() {
-        let mut map = HashMap::new();
-        map.insert(LanguageTag::new("en").unwrap(), vec!["card".into()]);
-        map.insert(LanguageTag::new("es").unwrap(), vec!["tarjeta".into()]);
-        let ctx = Context::PerLanguage(map);
-        let collected: Vec<_> = ctx
-            .iter()
-            .map(|(lang, kws)| (lang.unwrap().to_string(), kws.to_vec()))
-            .collect();
-        assert_eq!(collected.len(), 2);
-    }
-
-    #[test]
-    fn default_is_empty_global() {
-        let ctx = Context::default();
-        assert!(ctx.is_empty());
-    }
-}
diff --git a/crates/nvisy-pattern/src/recognition/dictionary.rs b/crates/nvisy-pattern/src/recognition/dictionary.rs
deleted file mode 100644
index e5acdce54..000000000
--- a/crates/nvisy-pattern/src/recognition/dictionary.rs
+++ /dev/null
@@ -1,246 +0,0 @@
-//! [`Dictionary`]: literal-term detection rule.
-
-use derive_builder::Builder;
-use nvisy_core::Error;
-use nvisy_core::entity::EntityLabelRef;
-use nvisy_core::primitive::{Confidence, CountryCode, LanguageTag};
-use serde::Deserialize;
-
-use super::context::Context;
-use super::term::Term;
-
-/// Confidence policy for a [`Dictionary`]'s matches.
-///
-/// Either every term gets the same score ([`Uniform`]), or scores
-/// vary by CSV source column ([`PerColumn`]). The untagged serde
-/// representation accepts a bare number for the uniform case and
-/// an array for the per-column case:
-///
-/// ```toml
-/// score = 0.9              # Uniform
-/// score = [0.85, 0.30]     # PerColumn
-/// ```
-///
-/// [`Uniform`]: Scoring::Uniform
-/// [`PerColumn`]: Scoring::PerColumn
-#[derive(Debug, Clone, PartialEq, Deserialize)]
-#[serde(untagged)]
-pub enum Scoring {
-    /// One confidence stamped on every match — the common case.
-    Uniform(Confidence),
-    /// Per-column confidence vector. Entry `i` is the score for
-    /// terms loaded from CSV column `i`. A term from a column past
-    /// the end of this vector causes a recognizer-build error, so
-    /// callers must declare one score per source column.
-    PerColumn(Vec<Confidence>),
-}
-
-impl Scoring {
-    /// Return `Ok(())` when the policy can resolve a score for at
-    /// least one input.
-    ///
-    /// [`PerColumn`] with an empty vector can never resolve and is
-    /// rejected here; the recognizer surfaces the error at build
-    /// time.
-    ///
-    /// # Errors
-    ///
-    /// Returns a human-readable reason when the policy is invalid.
-    ///
-    /// [`PerColumn`]: Self::PerColumn
-    pub fn validate(&self) -> Result<(), &'static str> {
-        match self {
-            Self::Uniform(_) => Ok(()),
-            Self::PerColumn(scores) if scores.is_empty() => {
-                Err("PerColumn scoring with no scores can never resolve")
-            }
-            Self::PerColumn(_) => Ok(()),
-        }
-    }
-
-    /// Resolve a score for the given source `column`.
-    ///
-    /// [`Uniform`] ignores `column` and always returns its score;
-    /// [`PerColumn`] returns the entry at `column`, or `None` when
-    /// `column` is `None` or out of range. Callers decide the
-    /// fall-back policy (per-term override, hard error, …).
-    ///
-    /// [`Uniform`]: Self::Uniform
-    /// [`PerColumn`]: Self::PerColumn
-    #[must_use]
-    pub fn get(&self, column: Option<u16>) -> Option<Confidence> {
-        match self {
-            Self::Uniform(s) => Some(*s),
-            Self::PerColumn(scores) => column.and_then(|c| scores.get(c as usize).copied()),
-        }
-    }
-}
-
-impl Default for Scoring {
-    fn default() -> Self {
-        Self::Uniform(Confidence::MAX)
-    }
-}
-
-/// Literal-term detection rule.
-///
-/// Scans for a fixed list of literals using a shared Aho-Corasick
-/// automaton. Unlike [`Regex`], a dictionary has no regex engine,
-/// no validator, and a [`Scoring`] policy shared across its terms.
-///
-/// # Examples
-///
-/// ```
-/// use nvisy_core::entity::builtins;
-/// use nvisy_pattern::{Dictionary, Term};
-///
-/// let dictionary = Dictionary::builder()
-///     .with_name("nationalities")
-///     .with_label(builtins::NATIONALITY.label_ref())
-///     .with_terms(vec![
-///         Term::new("German"),
-///         Term::new("French"),
-///         Term::new("Italian"),
-///     ])
-///     .build()
-///     .expect("nationalities dictionary builds");
-/// ```
-///
-/// [`Regex`]: crate::Regex
-#[derive(Debug, Clone, PartialEq, Builder, Deserialize)]
-#[builder(
-    name = "DictionaryBuilder",
-    pattern = "owned",
-    setter(into, strip_option, prefix = "with"),
-    build_fn(error = "Error")
-)]
-pub struct Dictionary {
-    /// Human-readable identifier surfaced in trail provenance
-    /// (e.g. `"nationalities"`).
-    pub name: String,
-    /// Entity label every match emits.
-    pub label: EntityLabelRef,
-    /// Literal terms to scan for. Compiled into the shared
-    /// Aho-Corasick automaton at recognizer-build time.
-    pub terms: Vec<Term>,
-    /// Confidence policy resolved against each term at
-    /// recognizer-build time. Defaults to [`Scoring::Uniform`]
-    /// with [`Confidence::MAX`].
-    #[builder(default)]
-    #[serde(default, rename = "score")]
-    pub scoring: Scoring,
-    /// Context keywords that lift confidence when one of them
-    /// appears near a match. Either a flat list applied
-    /// regardless of language, or a per-language map.
-    #[builder(default)]
-    #[serde(default)]
-    pub context: Context,
-    /// BCP-47 language tags the dictionary applies to. Empty means
-    /// "any language"; otherwise the recognizer skips the
-    /// dictionary when the per-call language hint is not in the
-    /// list.
-    #[builder(default)]
-    #[serde(default)]
-    pub languages: Vec<LanguageTag>,
-    /// ISO 3166-1 alpha-2 country codes the dictionary applies
-    /// to. Empty means "any country" — the dictionary fires
-    /// regardless of the per-call jurisdiction hint.
-    #[builder(default)]
-    #[serde(default)]
-    pub countries: Vec<CountryCode>,
-    /// Require word-boundary surroundings on every match.
-    ///
-    /// With the default of `true`, the term `"am"` matches the
-    /// word `"am"` but not the `"am"` inside `"example"`. Word
-    /// characters are Unicode alphanumerics and `_`. Set to
-    /// `false` to allow substring matches (e.g. scanning for
-    /// embedded credentials).
-    #[builder(default = "true")]
-    #[serde(default = "default_word_boundary")]
-    pub word_boundary: bool,
-}
-
-fn default_word_boundary() -> bool {
-    true
-}
-
-impl Dictionary {
-    /// Start a chainable builder.
-    ///
-    /// Required fields: `name`, `label`, `terms`.
-    #[must_use]
-    pub fn builder() -> DictionaryBuilder {
-        DictionaryBuilder::default()
-    }
-
-    /// Parse a self-contained dictionary from a TOML source.
-    ///
-    /// The TOML must include a `terms` field; for metadata-only
-    /// TOML paired with a separate term source, use
-    /// [`metadata_from_toml`] instead.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when the TOML is malformed or
-    /// missing required fields.
-    ///
-    /// [`metadata_from_toml`]: Self::metadata_from_toml
-    pub fn from_toml(raw: &str) -> Result<Self, Error> {
-        toml::from_str(raw)
-            .map_err(|e| Error::validation(format!("dictionary TOML: {e}"), "nvisy-pattern"))
-    }
-
-    /// Parse dictionary metadata from a sidecar TOML source.
-    ///
-    /// The returned [`DictionaryBuilder`] is seeded with every
-    /// field except `terms`; callers chain [`with_terms`] (e.g.
-    /// loaded from a paired CSV/TXT) before [`build`].
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when the TOML is malformed or
-    /// missing required metadata fields.
-    ///
-    /// [`with_terms`]: DictionaryBuilder::with_terms
-    /// [`build`]: DictionaryBuilder::build
-    pub fn metadata_from_toml(raw: &str) -> Result<DictionaryBuilder, Error> {
-        let metadata: DictionaryMetadata = toml::from_str(raw).map_err(|e| {
-            Error::validation(format!("dictionary metadata TOML: {e}"), "nvisy-pattern")
-        })?;
-        let mut builder = Dictionary::builder()
-            .with_name(metadata.name)
-            .with_label(metadata.label);
-        if let Some(scoring) = metadata.score {
-            builder = builder.with_scoring(scoring);
-        }
-        if let Some(context) = metadata.context {
-            builder = builder.with_context(context);
-        }
-        if let Some(wb) = metadata.word_boundary {
-            builder = builder.with_word_boundary(wb);
-        }
-        if let Some(languages) = metadata.languages {
-            builder = builder.with_languages(languages);
-        }
-        if let Some(countries) = metadata.countries {
-            builder = builder.with_countries(countries);
-        }
-        Ok(builder)
-    }
-}
-
-#[derive(Debug, Clone, Deserialize)]
-struct DictionaryMetadata {
-    name: String,
-    label: EntityLabelRef,
-    #[serde(default)]
-    score: Option<Scoring>,
-    #[serde(default)]
-    context: Option<Context>,
-    #[serde(default)]
-    word_boundary: Option<bool>,
-    #[serde(default)]
-    languages: Option<Vec<LanguageTag>>,
-    #[serde(default)]
-    countries: Option<Vec<CountryCode>>,
-}
diff --git a/crates/nvisy-pattern/src/recognition/mod.rs b/crates/nvisy-pattern/src/recognition/mod.rs
deleted file mode 100644
index 7dff66a5e..000000000
--- a/crates/nvisy-pattern/src/recognition/mod.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-//! Recognition primitives.
-//!
-//! Holds the rule shapes ([`Regex`] + its [`Variant`]s,
-//! [`Dictionary`]), their building blocks ([`Term`]), and the
-//! runtime [`PatternRecognizer`] that compiles them into pooled
-//! scanners. Per-rule and per-dictionary `context` keyword lists
-//! are harvested by [`PatternRecognizerBuilder::build_context_enhanced`]
-//! into a wrapping `ContextEnhanced` layer that lifts confidence
-//! on matches near a declared keyword.
-
-mod compiled;
-mod context;
-mod dictionary;
-mod recognizer;
-mod regex;
-mod term;
-
-pub use self::context::Context;
-pub use self::dictionary::{Dictionary, DictionaryBuilder, Scoring};
-pub use self::recognizer::{PatternRecognizer, PatternRecognizerBuilder};
-pub use self::regex::{Regex, RegexBuilder, Variant};
-pub use self::term::Term;
diff --git a/crates/nvisy-pattern/src/recognition/recognizer.rs b/crates/nvisy-pattern/src/recognition/recognizer.rs
deleted file mode 100644
index 5a01dd2ff..000000000
--- a/crates/nvisy-pattern/src/recognition/recognizer.rs
+++ /dev/null
@@ -1,858 +0,0 @@
-//! [`PatternRecognizer`] and its builder.
-
-use aho_corasick::{AhoCorasick, MatchKind};
-use nvisy_context::{BoostRule, ContextEnhanced, Enhancer, SubstringMatcher};
-use nvisy_core::entity::{Entity, EntityLabelCatalog, EntityLabelRef};
-use nvisy_core::modality::Text;
-use nvisy_core::primitive::LanguageTag;
-use nvisy_core::recognition::{EntityRecognizer, RecognizerInput, RecognizerOutput};
-use nvisy_core::{Error, Result};
-use regex::RegexSet;
-
-use super::compiled::{CompiledDictionary, CompiledPattern, has_word_boundaries};
-use super::dictionary::Dictionary;
-use super::regex::Regex;
-use crate::shipped;
-use crate::validators::{ValidationContext, ValidatorRegistry};
-
-/// Runtime text recognizer composed of a regex pool and an
-/// Aho-Corasick automaton.
-///
-/// Every registered [`Regex`] variant goes into one
-/// [`::regex::RegexSet`] for a single one-pass scan across every
-/// regex; every [`Dictionary`] term goes into one
-/// [`::aho_corasick::AhoCorasick`] automaton for a single one-pass
-/// scan across every literal. Both passes share one walk over the
-/// input and emit entities in modality-local byte coordinates.
-///
-/// Construct via [`PatternRecognizer::builder`]. [`build`]
-/// returns the bare recognizer; [`build_context_enhanced`] wraps
-/// it in a [`ContextEnhanced`] layer that lifts confidence on
-/// matches whose neighbourhood contains a per-label context
-/// keyword.
-///
-/// # Examples
-///
-/// ```
-/// use nvisy_pattern::PatternRecognizer;
-///
-/// let recognizer = PatternRecognizer::builder()
-///     .with_builtin_patterns()
-///     .with_builtin_dictionaries()
-///     .build()
-///     .expect("built-in recognizer builds");
-/// ```
-///
-/// [`Regex`]: super::Regex
-/// [`Dictionary`]: super::Dictionary
-/// [`build`]: PatternRecognizerBuilder::build
-/// [`build_context_enhanced`]: PatternRecognizerBuilder::build_context_enhanced
-pub struct PatternRecognizer {
-    patterns: Vec<CompiledPattern>,
-    regex_set: Option<RegexSet>,
-    dictionaries: Vec<CompiledDictionary>,
-    aho: Option<AhoCorasick>,
-}
-
-impl PatternRecognizer {
-    /// Start a chainable builder.
-    ///
-    /// A recognizer built with no patterns and no dictionaries is
-    /// valid — it emits zero entities on every call.
-    #[must_use]
-    pub fn builder() -> PatternRecognizerBuilder {
-        PatternRecognizerBuilder::default()
-    }
-
-    fn dictionary_owning_term(&self, term_id: usize) -> Option<&CompiledDictionary> {
-        self.dictionaries
-            .iter()
-            .find(|d| term_id >= d.term_start && term_id < d.term_end)
-    }
-}
-
-/// Accumulator of rules + validator registry for
-/// [`PatternRecognizer`].
-///
-/// Patterns and dictionaries are stored as authored — compilation
-/// into the pooled scanners happens in [`build`].
-///
-/// [`build`]: Self::build
-#[derive(Debug, Clone, Default)]
-pub struct PatternRecognizerBuilder {
-    patterns: Vec<Regex>,
-    dictionaries: Vec<Dictionary>,
-    validators: Option<ValidatorRegistry>,
-}
-
-impl PatternRecognizerBuilder {
-    /// Construct an empty builder.
-    #[must_use]
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Pre-seed with the shipped built-in patterns and
-    /// dictionaries.
-    ///
-    /// Shorthand for
-    /// `Self::new().with_builtin_patterns().with_builtin_dictionaries()`.
-    #[must_use]
-    pub fn builtin() -> Self {
-        Self::new()
-            .with_builtin_patterns()
-            .with_builtin_dictionaries()
-    }
-
-    /// Register one pattern; patterns accumulate in registration
-    /// order.
-    #[must_use]
-    pub fn with_pattern(mut self, pattern: Regex) -> Self {
-        self.patterns.push(pattern);
-        self
-    }
-
-    /// Register one dictionary; dictionaries accumulate in
-    /// registration order.
-    #[must_use]
-    pub fn with_dictionary(mut self, dictionary: Dictionary) -> Self {
-        self.dictionaries.push(dictionary);
-        self
-    }
-
-    /// Register every shipped built-in pattern.
-    #[must_use]
-    pub fn with_builtin_patterns(mut self) -> Self {
-        self.patterns.extend(shipped::patterns::all());
-        self
-    }
-
-    /// Register every shipped built-in dictionary.
-    #[must_use]
-    pub fn with_builtin_dictionaries(mut self) -> Self {
-        self.dictionaries.extend(shipped::dictionaries::all());
-        self
-    }
-
-    /// Override the validator registry used to resolve variant
-    /// validator names.
-    ///
-    /// Defaults to [`ValidatorRegistry::builtin`] when unset.
-    #[must_use]
-    pub fn with_validators(mut self, registry: ValidatorRegistry) -> Self {
-        self.validators = Some(registry);
-        self
-    }
-
-    /// Drop every pattern and dictionary whose label is not
-    /// declared in `catalog`.
-    ///
-    /// The engine uses this to build a per-request recognizer from
-    /// a workspace-wide template — rules that would emit labels no
-    /// policy declared never run.
-    #[must_use]
-    pub fn filter_by_catalog(mut self, catalog: &EntityLabelCatalog) -> Self {
-        self.patterns
-            .retain(|p| catalog.lookup(p.label.as_str()).is_some());
-        self.dictionaries
-            .retain(|d| catalog.lookup(d.label.as_str()).is_some());
-        self
-    }
-
-    /// Return `true` when no patterns and no dictionaries are
-    /// registered.
-    ///
-    /// The engine uses this to skip the per-request recognizer
-    /// entirely after a catalog filter dropped every rule.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.patterns.is_empty() && self.dictionaries.is_empty()
-    }
-
-    /// Borrow the accumulated patterns.
-    #[must_use]
-    pub fn patterns(&self) -> &[Regex] {
-        &self.patterns
-    }
-
-    /// Borrow the accumulated dictionaries.
-    #[must_use]
-    pub fn dictionaries(&self) -> &[Dictionary] {
-        &self.dictionaries
-    }
-
-    /// Compile every rule into the pooled scanners and return the
-    /// bare recognizer.
-    ///
-    /// Per-rule `context` keywords are ignored on the emission
-    /// path; the recognizer emits raw confidence as authored by
-    /// each rule. Wrap the result with [`build_context_enhanced`]
-    /// (or compose with [`ContextEnhanced`] manually) to lift
-    /// confidence on matches near a declared keyword.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when a pattern variant's regex
-    /// fails to compile, when a variant references an unknown
-    /// validator name, when a dictionary's `scoring` is invalid
-    /// or under-declared for some term's source column, or when
-    /// the shared automata cannot be constructed.
-    ///
-    /// [`build_context_enhanced`]: Self::build_context_enhanced
-    pub fn build(self) -> Result<PatternRecognizer> {
-        let validators = self
-            .validators
-            .clone()
-            .unwrap_or_else(ValidatorRegistry::builtin);
-        let (compiled_patterns, regex_set) = self.compile_patterns(&validators)?;
-        let (compiled_dicts, aho) = self.compile_dictionaries()?;
-
-        Ok(PatternRecognizer {
-            patterns: compiled_patterns,
-            regex_set,
-            dictionaries: compiled_dicts,
-            aho,
-        })
-    }
-
-    /// Compile every rule and wrap the recognizer in a
-    /// [`ContextEnhanced`] layer.
-    ///
-    /// Context keywords from every pattern and dictionary are
-    /// harvested into per-label [`BoostRule`]s that lift confidence
-    /// on matches whose neighbourhood contains a declared keyword.
-    ///
-    /// # Errors
-    ///
-    /// See [`build`].
-    ///
-    /// [`build`]: Self::build
-    pub fn build_context_enhanced(self) -> Result<ContextEnhanced<PatternRecognizer>> {
-        let enhancer = self.build_enhancer();
-        let recognizer = self.build()?;
-        Ok(ContextEnhanced::new(recognizer, enhancer))
-    }
-
-    /// Compile every `(pattern, variant)` pair into a
-    /// [`CompiledPattern`] keyed by its slot in the shared
-    /// [`RegexSet`].
-    fn compile_patterns(
-        &self,
-        validators: &ValidatorRegistry,
-    ) -> Result<(Vec<CompiledPattern>, Option<RegexSet>)> {
-        let variant_total: usize = self.patterns.iter().map(|p| p.variants.len()).sum();
-        let mut compiled = Vec::with_capacity(variant_total);
-        let mut regex_sources = Vec::with_capacity(variant_total);
-
-        for pattern in &self.patterns {
-            for variant in &pattern.variants {
-                let regex = ::regex::Regex::new(&variant.regex).map_err(|e| {
-                    Error::validation(
-                        format!("pattern `{}`: invalid regex: {e}", pattern.name),
-                        "nvisy-pattern",
-                    )
-                })?;
-                let validator = match variant.validator.as_deref() {
-                    None => None,
-                    Some(name) => Some(validators.resolve(name).ok_or_else(|| {
-                        Error::validation(
-                            format!("pattern `{}`: unknown validator `{}`", pattern.name, name),
-                            "nvisy-pattern",
-                        )
-                    })?),
-                };
-                regex_sources.push(variant.regex.clone());
-                compiled.push(CompiledPattern {
-                    pattern_name: pattern.name.clone(),
-                    label: pattern.label.clone(),
-                    regex,
-                    score: variant.score,
-                    validator,
-                    languages: pattern.languages.clone(),
-                    countries: pattern.countries.clone(),
-                });
-            }
-        }
-
-        let regex_set = if regex_sources.is_empty() {
-            None
-        } else {
-            Some(RegexSet::new(&regex_sources).map_err(|e| {
-                Error::validation(format!("compiling regex set: {e}"), "nvisy-pattern")
-            })?)
-        };
-        Ok((compiled, regex_set))
-    }
-
-    /// Compile every dictionary into a [`CompiledDictionary`]
-    /// with its term-id range inside the shared Aho-Corasick
-    /// automaton, plus per-term confidences resolved from the
-    /// dictionary's `scoring` policy (with per-term overrides
-    /// taking precedence).
-    fn compile_dictionaries(&self) -> Result<(Vec<CompiledDictionary>, Option<AhoCorasick>)> {
-        let mut compiled = Vec::with_capacity(self.dictionaries.len());
-        let mut all_terms: Vec<String> = Vec::new();
-
-        for dict in &self.dictionaries {
-            if let Err(reason) = dict.scoring.validate() {
-                return Err(Error::validation(
-                    format!("dictionary `{}`: {reason}", dict.name),
-                    "nvisy-pattern",
-                ));
-            }
-            let term_start = all_terms.len();
-            let mut term_scores = Vec::with_capacity(dict.terms.len());
-            for entry in &dict.terms {
-                all_terms.push(entry.term.clone());
-                // Per-term `score` wins when set; otherwise ask
-                // the dictionary's `Scoring` to resolve against
-                // the term's source column. `None` means the
-                // column didn't map to a declared score —
-                // surfaced as a hard build error so silent
-                // misconfiguration can't happen.
-                let score = entry
-                    .score
-                    .or_else(|| dict.scoring.get(entry.column))
-                    .ok_or_else(|| {
-                        let column_desc = entry
-                            .column
-                            .map_or_else(|| "no column".to_owned(), |c| format!("column {c}"));
-                        Error::validation(
-                            format!(
-                                "dictionary `{}`: term `{}` ({column_desc}) has no score in \
-                                 dictionary scoring",
-                                dict.name, entry.term,
-                            ),
-                            "nvisy-pattern",
-                        )
-                    })?;
-                term_scores.push(score);
-            }
-            let term_end = all_terms.len();
-            compiled.push(CompiledDictionary {
-                name: dict.name.clone(),
-                label: dict.label.clone(),
-                term_start,
-                term_end,
-                term_scores,
-                languages: dict.languages.clone(),
-                countries: dict.countries.clone(),
-                word_boundary: dict.word_boundary,
-            });
-        }
-
-        let aho = if all_terms.is_empty() {
-            None
-        } else {
-            Some(
-                AhoCorasick::builder()
-                    .ascii_case_insensitive(false)
-                    // Longest-match-at-position: when both `en`
-                    // and `English` start at the same offset,
-                    // return `English`. Without this, the short
-                    // ISO code would win and word-boundary
-                    // post-filtering would then reject it,
-                    // dropping the legitimate long-form match.
-                    .match_kind(MatchKind::LeftmostLongest)
-                    .build(&all_terms)
-                    .map_err(|e| {
-                        Error::validation(
-                            format!("compiling dictionary automaton: {e}"),
-                            "nvisy-pattern",
-                        )
-                    })?,
-            )
-        };
-        Ok((compiled, aho))
-    }
-
-    /// Build the wrapping [`Enhancer`] from per-pattern and
-    /// per-dictionary context keywords.
-    ///
-    /// Per-rule [`Context`] produces one [`BoostRule`] per
-    /// language scope (global rules carry
-    /// `language = None`; per-language rules carry the language
-    /// tag). The enhancer keys these by label and filters them
-    /// against the per-call language hint at apply time.
-    ///
-    /// [`Context`]: super::Context
-    fn build_enhancer(&self) -> Enhancer {
-        let boost_rules: Vec<BoostRule> = self
-            .context_keywords()
-            .map(|(label, language, keywords)| {
-                let rule = BoostRule::for_label(label.clone(), keywords.iter().cloned());
-                match language {
-                    Some(lang) => rule.with_language(lang.clone()),
-                    None => rule,
-                }
-            })
-            .collect();
-        Enhancer::new(boost_rules, Box::new(SubstringMatcher))
-    }
-
-    /// Yield `(label, language, keywords)` for every pattern and
-    /// dictionary that declares a non-empty context. Global
-    /// keywords carry `language = None`; per-language keywords
-    /// carry `Some(tag)`.
-    fn context_keywords(
-        &self,
-    ) -> impl Iterator<Item = (&EntityLabelRef, Option<&LanguageTag>, &[String])> {
-        let pattern_keywords = self
-            .patterns
-            .iter()
-            .filter(|p| !p.context.is_empty())
-            .flat_map(|p| {
-                p.context
-                    .iter()
-                    .map(move |(lang, kws)| (&p.label, lang, kws))
-            });
-        let dict_keywords = self
-            .dictionaries
-            .iter()
-            .filter(|d| !d.context.is_empty())
-            .flat_map(|d| {
-                d.context
-                    .iter()
-                    .map(move |(lang, kws)| (&d.label, lang, kws))
-            });
-        pattern_keywords.chain(dict_keywords)
-    }
-}
-
-#[async_trait::async_trait]
-impl EntityRecognizer<Text> for PatternRecognizer {
-    async fn recognize(&self, input: &RecognizerInput<Text>) -> Result<RecognizerOutput<Text>> {
-        let text = input.data.text.as_str();
-        let mut entities: Vec<Entity<Text>> = Vec::new();
-
-        if let Some(set) = self.regex_set.as_ref() {
-            for pattern_id in set.matches(text).into_iter() {
-                let pat = &self.patterns[pattern_id];
-                if !input.applies_to_language(&pat.languages) {
-                    continue;
-                }
-                if !input.applies_to_country(&pat.countries) {
-                    continue;
-                }
-                let ctx = ValidationContext {
-                    country: input.country,
-                    language: input.language.clone(),
-                };
-                for m in pat.regex.find_iter(text) {
-                    if let Some(validator) = pat.validator.as_ref()
-                        && !validator.validate(m.as_str(), &ctx)
-                    {
-                        continue;
-                    }
-                    entities.push(pat.build_entity(m.start(), m.end()));
-                }
-            }
-        }
-
-        if let Some(aho) = self.aho.as_ref() {
-            for mat in aho.find_iter(text) {
-                let term_id = mat.pattern().as_usize();
-                let Some(dict) = self.dictionary_owning_term(term_id) else {
-                    continue;
-                };
-                if !input.applies_to_language(&dict.languages) {
-                    continue;
-                }
-                if !input.applies_to_country(&dict.countries) {
-                    continue;
-                }
-                if dict.word_boundary && !has_word_boundaries(text, mat.start(), mat.end()) {
-                    continue;
-                }
-                let score = dict.term_scores[term_id - dict.term_start];
-                entities.push(dict.build_entity(score, mat.start(), mat.end()));
-            }
-        }
-
-        Ok(RecognizerOutput::new(entities))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::collections::HashMap;
-
-    use nvisy_core::entity::{Entity, EntityLabelRef, builtins};
-    use nvisy_core::modality::{Text, TextData};
-    use nvisy_core::primitive::{Confidence, CountryCode};
-    use nvisy_core::recognition::RecognizerInput;
-
-    use super::*;
-    use crate::Dictionary;
-    use crate::recognition::term::Term;
-
-    fn dict(name: &str, terms: &[&str], word_boundary: bool) -> Dictionary {
-        Dictionary::builder()
-            .with_name(name.to_owned())
-            .with_label(EntityLabelRef::from(builtins::LANGUAGE.name.clone()))
-            .with_terms(terms.iter().copied().map(Term::new).collect::<Vec<_>>())
-            .with_word_boundary(word_boundary)
-            .build()
-            .expect("dictionary builds")
-    }
-
-    async fn run(recognizer: &impl EntityRecognizer<Text>, text: &str) -> Vec<Entity<Text>> {
-        let input = RecognizerInput::new(TextData::new(text.to_owned()));
-        recognizer
-            .recognize(&input)
-            .await
-            .expect("recognize succeeds")
-            .entities
-    }
-
-    #[tokio::test]
-    async fn word_boundary_rejects_substring_matches() {
-        let recognizer = PatternRecognizer::builder()
-            .with_dictionary(dict("langs", &["am", "or"], true))
-            .build()
-            .expect("recognizer builds");
-
-        let entities = run(&recognizer, "the example or a candidate").await;
-        let matched: Vec<&str> = entities
-            .iter()
-            .map(|e| &"the example or a candidate"[e.location.start..e.location.end])
-            .collect();
-
-        // "am" inside "example" and "or" inside "candidate" are
-        // substring matches and must be rejected. The standalone
-        // "or" between two spaces must be kept.
-        assert_eq!(matched, vec!["or"]);
-    }
-
-    #[tokio::test]
-    async fn word_boundary_disabled_keeps_substring_matches() {
-        let recognizer = PatternRecognizer::builder()
-            .with_dictionary(dict("langs", &["am"], false))
-            .build()
-            .expect("recognizer builds");
-
-        let entities = run(&recognizer, "example").await;
-        assert_eq!(entities.len(), 1, "substring match must be kept");
-    }
-
-    #[test]
-    fn regex_parses_flat_context_as_global() {
-        let toml = r#"
-            name = "x"
-            label = "government_id"
-            context = ["ssn", "social security"]
-            [[variants]]
-            regex = "\\d+"
-        "#;
-        let regex = crate::Regex::from_toml(toml).expect("flat-context TOML parses");
-        assert!(matches!(regex.context, crate::Context::Global(_)));
-    }
-
-    #[test]
-    fn regex_parses_table_context_as_per_language() {
-        let toml = r#"
-            name = "x"
-            label = "payment_card"
-            [context]
-            en = ["card", "credit"]
-            es = ["tarjeta", "crédito"]
-            [[variants]]
-            regex = "\\d+"
-        "#;
-        let regex = crate::Regex::from_toml(toml).expect("table-context TOML parses");
-        let map = match regex.context {
-            crate::Context::PerLanguage(m) => m,
-            _ => panic!("expected PerLanguage"),
-        };
-        assert_eq!(map.len(), 2);
-    }
-
-    #[test]
-    fn regex_omits_countries_by_default() {
-        let toml = r#"
-            name = "x"
-            label = "government_id"
-            [[variants]]
-            regex = "\\d+"
-        "#;
-        let regex = crate::Regex::from_toml(toml).expect("TOML parses");
-        assert!(
-            regex.countries.is_empty(),
-            "default countries must be empty"
-        );
-    }
-
-    #[test]
-    fn regex_parses_countries_field() {
-        let toml = r#"
-            name = "ssn"
-            label = "government_id"
-            countries = ["US"]
-            [[variants]]
-            regex = "\\d+"
-        "#;
-        let regex = crate::Regex::from_toml(toml).expect("TOML parses");
-        assert_eq!(regex.countries.len(), 1);
-        assert_eq!(regex.countries[0].as_str(), "US");
-    }
-
-    #[test]
-    fn regex_parses_multiple_countries() {
-        let toml = r#"
-            name = "eu-vat"
-            label = "tax_id"
-            countries = ["de", "FR", "iT"]
-            [[variants]]
-            regex = "\\d+"
-        "#;
-        let regex = crate::Regex::from_toml(toml).expect("TOML parses");
-        assert_eq!(regex.countries.len(), 3);
-        // Construction normalises to uppercase.
-        let codes: Vec<&str> = regex.countries.iter().map(CountryCode::as_str).collect();
-        assert_eq!(codes, vec!["DE", "FR", "IT"]);
-    }
-
-    #[test]
-    fn regex_rejects_invalid_country() {
-        let toml = r#"
-            name = "x"
-            label = "government_id"
-            countries = ["XZ"]
-            [[variants]]
-            regex = "\\d+"
-        "#;
-        assert!(
-            crate::Regex::from_toml(toml).is_err(),
-            "unassigned country code must error",
-        );
-    }
-
-    #[test]
-    fn regex_builder_accepts_countries() {
-        let variant = crate::Variant::new(r"\d{3}-\d{2}-\d{4}").unwrap();
-        let regex = crate::Regex::builder()
-            .with_name("ssn")
-            .with_label(builtins::GOVERNMENT_ID.label_ref())
-            .with_variants(vec![variant])
-            .with_countries(vec![CountryCode::new("US").unwrap()])
-            .build()
-            .expect("regex builds");
-        assert_eq!(regex.countries.len(), 1);
-        assert_eq!(regex.countries[0].as_str(), "US");
-    }
-
-    async fn run_with_language(
-        recognizer: &impl EntityRecognizer<Text>,
-        text: &str,
-        language: Option<&str>,
-    ) -> Vec<Entity<Text>> {
-        let mut input = RecognizerInput::new(TextData::new(text.to_owned()));
-        if let Some(lang) = language {
-            input = input.with_language(LanguageTag::new(lang).expect("language tag parses"));
-        }
-        recognizer
-            .recognize(&input)
-            .await
-            .expect("recognize succeeds")
-            .entities
-    }
-
-    async fn run_with_country(
-        recognizer: &impl EntityRecognizer<Text>,
-        text: &str,
-        country: Option<&str>,
-    ) -> Vec<Entity<Text>> {
-        let mut input = RecognizerInput::new(TextData::new(text.to_owned()));
-        if let Some(c) = country {
-            input = input.with_country(CountryCode::new(c).expect("country code parses"));
-        }
-        recognizer
-            .recognize(&input)
-            .await
-            .expect("recognize succeeds")
-            .entities
-    }
-
-    fn us_ssn_regex() -> crate::Regex {
-        let variant = crate::Variant::new(r"\b\d{3}-\d{2}-\d{4}\b")
-            .expect("variant builds")
-            .with_score(Confidence::clamped(0.5));
-        crate::Regex::builder()
-            .with_name("ssn")
-            .with_label(builtins::GOVERNMENT_ID.label_ref())
-            .with_variants(vec![variant])
-            .with_countries(vec![CountryCode::new("US").unwrap()])
-            .build()
-            .expect("regex builds")
-    }
-
-    #[tokio::test]
-    async fn country_scoped_rule_fires_under_matching_hint() {
-        let recognizer = PatternRecognizer::builder()
-            .with_pattern(us_ssn_regex())
-            .build()
-            .expect("recognizer builds");
-        let entities = run_with_country(&recognizer, "SSN: 123-45-6789", Some("US")).await;
-        assert_eq!(entities.len(), 1, "US-scoped rule must fire under US hint");
-    }
-
-    #[tokio::test]
-    async fn country_scoped_rule_skipped_under_non_matching_hint() {
-        let recognizer = PatternRecognizer::builder()
-            .with_pattern(us_ssn_regex())
-            .build()
-            .expect("recognizer builds");
-        let entities = run_with_country(&recognizer, "SSN: 123-45-6789", Some("GB")).await;
-        assert!(
-            entities.is_empty(),
-            "US-scoped rule must not fire under GB hint",
-        );
-    }
-
-    #[tokio::test]
-    async fn country_scoped_rule_fires_without_hint() {
-        // Permissive fallback: missing hint shouldn't drop the
-        // detection. Matches the existing `applies_to_language`
-        // semantic.
-        let recognizer = PatternRecognizer::builder()
-            .with_pattern(us_ssn_regex())
-            .build()
-            .expect("recognizer builds");
-        let entities = run_with_country(&recognizer, "SSN: 123-45-6789", None).await;
-        assert_eq!(
-            entities.len(),
-            1,
-            "missing country hint must permit US-scoped rule to run",
-        );
-    }
-
-    fn per_language_credit_card_regex() -> crate::Regex {
-        let variant = crate::Variant::new(r"\b\d{16}\b")
-            .expect("variant builds")
-            .with_score(Confidence::clamped(0.5));
-        let mut context = HashMap::new();
-        context.insert(
-            LanguageTag::new("en").unwrap(),
-            vec!["credit".to_owned(), "card".to_owned()],
-        );
-        context.insert(
-            LanguageTag::new("es").unwrap(),
-            vec!["tarjeta".to_owned(), "crédito".to_owned()],
-        );
-        crate::Regex::builder()
-            .with_name("credit_card")
-            .with_label(builtins::PAYMENT_CARD.label_ref())
-            .with_context(crate::Context::PerLanguage(context))
-            .with_variants(vec![variant])
-            .build()
-            .expect("regex builds")
-    }
-
-    #[tokio::test]
-    async fn per_language_boost_fires_for_matching_language() {
-        let recognizer = PatternRecognizer::builder()
-            .with_pattern(per_language_credit_card_regex())
-            .build_context_enhanced()
-            .expect("recognizer builds");
-
-        let text = "Pay with your credit card 4111111111111111 today";
-        let entities = run_with_language(&recognizer, text, Some("en")).await;
-        let card = entities
-            .iter()
-            .find(|e| &text[e.location.start..e.location.end] == "4111111111111111")
-            .expect("card match present");
-        assert!(
-            card.confidence.get() > 0.5,
-            "English keyword `credit` should boost under en hint",
-        );
-    }
-
-    #[tokio::test]
-    async fn per_language_boost_fires_for_regional_variant() {
-        // Pattern is scoped `en`; hint is `en-US`. Primary subtag
-        // matches, so the boost must fire.
-        let recognizer = PatternRecognizer::builder()
-            .with_pattern(per_language_credit_card_regex())
-            .build_context_enhanced()
-            .expect("recognizer builds");
-
-        let text = "Pay with your credit card 4111111111111111 today";
-        let entities = run_with_language(&recognizer, text, Some("en-US")).await;
-        let card = entities
-            .iter()
-            .find(|e| &text[e.location.start..e.location.end] == "4111111111111111")
-            .expect("card match present");
-        assert!(
-            card.confidence.get() > 0.5,
-            "`en-US` hint should fire the `en`-scoped boost",
-        );
-    }
-
-    #[tokio::test]
-    async fn rule_language_filter_accepts_regional_variant() {
-        // Pattern is scoped `languages = ["en"]`; the per-call
-        // hint is `en-US`. The rule must still run.
-        let variant = crate::Variant::new(r"\b\d{3}-\d{2}-\d{4}\b")
-            .expect("variant builds")
-            .with_score(Confidence::clamped(0.5));
-        let regex = crate::Regex::builder()
-            .with_name("ssn")
-            .with_label(builtins::GOVERNMENT_ID.label_ref())
-            .with_variants(vec![variant])
-            .with_languages(vec![LanguageTag::new("en").unwrap()])
-            .build()
-            .expect("regex builds");
-
-        let recognizer = PatternRecognizer::builder()
-            .with_pattern(regex)
-            .build()
-            .expect("recognizer builds");
-
-        let entities = run_with_language(&recognizer, "SSN: 123-45-6789", Some("en-US")).await;
-        assert_eq!(
-            entities.len(),
-            1,
-            "`en`-scoped rule must run for `en-US` input",
-        );
-    }
-
-    #[tokio::test]
-    async fn per_language_boost_skipped_for_non_matching_language() {
-        let recognizer = PatternRecognizer::builder()
-            .with_pattern(per_language_credit_card_regex())
-            .build_context_enhanced()
-            .expect("recognizer builds");
-
-        // English keywords near the match, but caller asserted Spanish.
-        let text = "Pay with your credit card 4111111111111111 today";
-        let entities = run_with_language(&recognizer, text, Some("es")).await;
-        let card = entities
-            .iter()
-            .find(|e| &text[e.location.start..e.location.end] == "4111111111111111")
-            .expect("card match present");
-        assert!(
-            (card.confidence.get() - 0.5).abs() < f64::EPSILON,
-            "English keywords must not boost under es hint",
-        );
-    }
-
-    #[tokio::test]
-    async fn no_language_hint_unions_per_language_keywords() {
-        let recognizer = PatternRecognizer::builder()
-            .with_pattern(per_language_credit_card_regex())
-            .build_context_enhanced()
-            .expect("recognizer builds");
-
-        // English keyword near the match, no language hint set.
-        let text = "Pay with your credit card 4111111111111111 today";
-        let entities = run_with_language(&recognizer, text, None).await;
-        let card = entities
-            .iter()
-            .find(|e| &text[e.location.start..e.location.end] == "4111111111111111")
-            .expect("card match present");
-        assert!(
-            card.confidence.get() > 0.5,
-            "missing language hint should permit any per-language keyword to boost",
-        );
-    }
-}
diff --git a/crates/nvisy-pattern/src/recognition/regex.rs b/crates/nvisy-pattern/src/recognition/regex.rs
deleted file mode 100644
index 0690ce34d..000000000
--- a/crates/nvisy-pattern/src/recognition/regex.rs
+++ /dev/null
@@ -1,181 +0,0 @@
-//! [`Regex`] rule and its [`Variant`]s.
-
-use derive_builder::Builder;
-use nvisy_core::Error;
-use nvisy_core::entity::EntityLabelRef;
-use nvisy_core::primitive::{Confidence, CountryCode, LanguageTag};
-use serde::Deserialize;
-
-use super::context::Context;
-
-/// One regex strategy inside a [`Regex`] rule.
-///
-/// A variant pairs a regex source with the confidence stamped on
-/// every match it produces and, optionally, a validator name
-/// resolved against the [`ValidatorRegistry`] at recognizer-build
-/// time so structurally-suspect matches can be dropped.
-///
-/// [`ValidatorRegistry`]: crate::validators::ValidatorRegistry
-#[derive(Debug, Clone, PartialEq, Deserialize)]
-pub struct Variant {
-    /// Regex source. Compiled to a [`::regex::Regex`] by
-    /// [`PatternRecognizer::build`].
-    ///
-    /// [`PatternRecognizer::build`]: super::PatternRecognizer
-    pub regex: String,
-    /// Confidence stamped on every match, before any
-    /// post-recognition keyword boost.
-    #[serde(default = "default_score")]
-    pub score: Confidence,
-    /// Validator name resolved against the [`ValidatorRegistry`].
-    /// Matches that fail validation are dropped.
-    ///
-    /// [`ValidatorRegistry`]: crate::validators::ValidatorRegistry
-    #[serde(default)]
-    pub validator: Option<String>,
-}
-
-impl Variant {
-    /// Construct a variant from a regex source.
-    ///
-    /// `score` defaults to [`Confidence::MAX`] and `validator` to
-    /// `None`; override with [`with_score`] / [`with_validator`].
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when `regex` is not a valid
-    /// regular expression.
-    ///
-    /// [`with_score`]: Self::with_score
-    /// [`with_validator`]: Self::with_validator
-    pub fn new(regex: impl Into<String>) -> Result<Self, Error> {
-        let regex = regex.into();
-        if let Err(e) = ::regex::Regex::new(&regex) {
-            return Err(Error::validation(
-                format!("invalid regex: {e}"),
-                "nvisy-pattern",
-            ));
-        }
-        Ok(Self {
-            regex,
-            score: Confidence::MAX,
-            validator: None,
-        })
-    }
-
-    /// Set the per-match confidence score.
-    #[must_use]
-    pub fn with_score(mut self, score: Confidence) -> Self {
-        self.score = score;
-        self
-    }
-
-    /// Set the validator name to run on every match.
-    ///
-    /// The name is resolved against the [`ValidatorRegistry`] when
-    /// the parent [`PatternRecognizer`] is built; unknown names
-    /// surface as a build-time error.
-    ///
-    /// [`ValidatorRegistry`]: crate::validators::ValidatorRegistry
-    /// [`PatternRecognizer`]: super::PatternRecognizer
-    #[must_use]
-    pub fn with_validator(mut self, name: impl Into<String>) -> Self {
-        self.validator = Some(name.into());
-        self
-    }
-}
-
-fn default_score() -> Confidence {
-    Confidence::MAX
-}
-
-/// Regex detection rule: one label, optional keyword boosts, and
-/// one or more [`Variant`]s.
-///
-/// A rule groups several regex strategies under a single entity
-/// type plus a shared context-keyword list. Every variant emits
-/// the same [`label`]; context keywords are harvested by
-/// [`PatternRecognizer`] into a wrapping boost layer and are
-/// never read by the rule itself.
-///
-/// # Examples
-///
-/// ```
-/// use nvisy_core::entity::builtins;
-/// use nvisy_core::primitive::Confidence;
-/// use nvisy_pattern::{Regex, Variant};
-///
-/// let variant = Variant::new(r"\b\d{3}-\d{2}-\d{4}\b")
-///     .expect("ssn variant builds")
-///     .with_score(Confidence::clamped(0.9))
-///     .with_validator("ssn");
-///
-/// let ssn = Regex::builder()
-///     .with_name("ssn")
-///     .with_label(builtins::GOVERNMENT_ID.label_ref())
-///     .with_context(vec!["ssn".to_owned(), "social security".to_owned()])
-///     .with_variants(vec![variant])
-///     .build()
-///     .expect("ssn rule builds");
-/// ```
-///
-/// [`label`]: Regex::label
-/// [`PatternRecognizer`]: super::PatternRecognizer
-#[derive(Debug, Clone, PartialEq, Builder, Deserialize)]
-#[builder(
-    name = "RegexBuilder",
-    pattern = "owned",
-    setter(into, strip_option, prefix = "with"),
-    build_fn(error = "Error")
-)]
-pub struct Regex {
-    /// Human-readable identifier surfaced in trail provenance (e.g.
-    /// `"ssn"`, `"credit_card"`).
-    pub name: String,
-    /// Entity label every variant emits.
-    pub label: EntityLabelRef,
-    /// Context keywords that lift confidence when one of them
-    /// appears near a match. Either a flat list applied
-    /// regardless of language, or a per-language map.
-    #[builder(default)]
-    #[serde(default)]
-    pub context: Context,
-    /// Regex variants. At least one is required to produce matches;
-    /// the recognizer skips rules with an empty variant list.
-    pub variants: Vec<Variant>,
-    /// BCP-47 language tags the rule applies to. Empty means "any
-    /// language"; otherwise the recognizer skips the rule when the
-    /// per-call language hint is not in the list.
-    #[builder(default)]
-    #[serde(default)]
-    pub languages: Vec<LanguageTag>,
-    /// ISO 3166-1 alpha-2 country codes the rule applies to.
-    /// Empty means "any country" — the rule fires regardless of
-    /// the per-call jurisdiction hint. Use this to scope a
-    /// pattern to specific national formats (e.g. `["US"]` for
-    /// the SSN regex).
-    #[builder(default)]
-    #[serde(default)]
-    pub countries: Vec<CountryCode>,
-}
-
-impl Regex {
-    /// Start a chainable builder.
-    ///
-    /// Required fields: `name`, `label`, `variants`.
-    #[must_use]
-    pub fn builder() -> RegexBuilder {
-        RegexBuilder::default()
-    }
-
-    /// Parse a rule from a TOML source.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when the TOML is malformed or
-    /// missing required fields.
-    pub fn from_toml(raw: &str) -> Result<Self, Error> {
-        toml::from_str(raw)
-            .map_err(|e| Error::validation(format!("regex rule TOML: {e}"), "nvisy-pattern"))
-    }
-}
diff --git a/crates/nvisy-pattern/src/recognition/term.rs b/crates/nvisy-pattern/src/recognition/term.rs
deleted file mode 100644
index 613dbe0ec..000000000
--- a/crates/nvisy-pattern/src/recognition/term.rs
+++ /dev/null
@@ -1,104 +0,0 @@
-//! [`Term`]: one literal entry inside a [`Dictionary`].
-//!
-//! [`Dictionary`]: crate::Dictionary
-
-use nvisy_core::Error;
-use nvisy_core::primitive::Confidence;
-use serde::Deserialize;
-
-/// One literal scanned for by a [`Dictionary`].
-///
-/// The `column` field is `Some(i)` for CSV-loaded terms and `None`
-/// for plain-text or programmatic sources. The `score` field
-/// overrides the dictionary's [`Scoring`] for this single term
-/// when set — useful for one-off exceptions in an otherwise
-/// uniform list.
-///
-/// [`Dictionary`]: crate::Dictionary
-/// [`Scoring`]: crate::Scoring
-#[derive(Debug, Clone, PartialEq, Deserialize)]
-pub struct Term {
-    /// The literal scanned for.
-    pub term: String,
-    /// CSV source-column index when loaded via [`Term::from_csv`];
-    /// `None` otherwise.
-    #[serde(default)]
-    pub column: Option<u16>,
-    /// Per-term score override. When `Some`, the recognizer
-    /// stamps this score on every match; when `None`, falls back
-    /// to the dictionary's [`Scoring`] resolved against [`column`].
-    ///
-    /// [`Scoring`]: crate::Scoring
-    /// [`column`]: Self::column
-    #[serde(default)]
-    pub score: Option<Confidence>,
-}
-
-impl Term {
-    /// Construct a term with no column and no score override.
-    #[must_use]
-    pub fn new(term: impl Into<String>) -> Self {
-        Self {
-            term: term.into(),
-            column: None,
-            score: None,
-        }
-    }
-
-    /// Attach a CSV source-column index.
-    #[must_use]
-    pub fn with_column(mut self, column: u16) -> Self {
-        self.column = Some(column);
-        self
-    }
-
-    /// Set a per-term score override.
-    #[must_use]
-    pub fn with_score(mut self, score: Confidence) -> Self {
-        self.score = Some(score);
-        self
-    }
-
-    /// Parse a list of terms from plain text — one term per line.
-    ///
-    /// Each line is trimmed; empty lines and lines starting with
-    /// `#` are skipped. Plain-text terms carry no column.
-    pub fn from_text(raw: &str) -> Vec<Self> {
-        raw.lines()
-            .map(str::trim)
-            .filter(|line| !line.is_empty() && !line.starts_with('#'))
-            .map(Term::new)
-            .collect()
-    }
-
-    /// Parse a list of terms from CSV.
-    ///
-    /// Every non-empty cell becomes a term tagged with its 0-based
-    /// source-column index. The dictionary's [`Scoring::PerColumn`]
-    /// uses that index to resolve a per-column confidence.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when the CSV is malformed.
-    ///
-    /// [`Scoring::PerColumn`]: crate::Scoring::PerColumn
-    pub fn from_csv(raw: &str) -> Result<Vec<Self>, Error> {
-        let mut reader = csv::ReaderBuilder::new()
-            .has_headers(false)
-            .flexible(true)
-            .from_reader(raw.as_bytes());
-        let mut entries = Vec::new();
-        for row in reader.records() {
-            let row =
-                row.map_err(|e| Error::validation(format!("terms CSV: {e}"), "nvisy-pattern"))?;
-            for (col_idx, cell) in row.iter().enumerate() {
-                let trimmed = cell.trim();
-                if !trimmed.is_empty() {
-                    let column = u16::try_from(col_idx).unwrap_or(u16::MAX);
-                    entries.push(Term::new(trimmed).with_column(column));
-                }
-            }
-        }
-        Ok(entries)
-    }
-}
diff --git a/crates/nvisy-pattern/src/shipped/dictionaries/en.rs b/crates/nvisy-pattern/src/shipped/dictionaries/en.rs
deleted file mode 100644
index 407311e6b..000000000
--- a/crates/nvisy-pattern/src/shipped/dictionaries/en.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-//! English-language dictionaries — terms written in English
-//! and meant to fire on English documents.
-//!
-//! Demonyms, religion names, and language names translate
-//! across locales (`American` → `américain`, `Catholic` →
-//! `catolico`, `English` → `englisch`). The terms here are the
-//! English forms; a non-English document needs its own locale
-//! sub-module (`fr`, `es`, …) once those land. Runtime
-//! filtering by `RecognizerInput.language` keeps these
-//! dictionaries from firing when the caller asserts a
-//! non-English document.
-
-use crate::{__shipped_dictionary as shipped_dictionary, Dictionary};
-
-shipped_dictionary!(
-    /// Fiat currency names and ISO 4217 codes (USD, US Dollar,
-    /// EUR, Euro, …). Long-form names are English; ISO codes
-    /// happen to match in non-English text too but the dictionary
-    /// is scoped to `en` so the long-form bulk doesn't fire on
-    /// French / German documents.
-    fn currencies
-        from "../../../assets/dictionaries/en/finance/currencies.toml"
-        with csv "../../../assets/dictionaries/en/finance/currencies.csv"
-);
-shipped_dictionary!(
-    /// English names of human languages plus ISO 639-1 codes
-    /// (English, en, French, fr, …).
-    fn languages
-        from "../../../assets/dictionaries/en/personal/languages.toml"
-        with csv "../../../assets/dictionaries/en/personal/languages.csv"
-);
-shipped_dictionary!(
-    /// English demonyms and nationality terms (American, French,
-    /// Japanese, …).
-    fn nationalities
-        from "../../../assets/dictionaries/en/personal/nationalities.toml"
-        with text "../../../assets/dictionaries/en/personal/nationalities.txt"
-);
-shipped_dictionary!(
-    /// English names of religious affiliations (Christian, Muslim,
-    /// Buddhist, …).
-    fn religions
-        from "../../../assets/dictionaries/en/personal/religions.toml"
-        with text "../../../assets/dictionaries/en/personal/religions.txt"
-);
-
-/// Every English-scoped built-in dictionary.
-#[must_use]
-pub fn all() -> Vec<Dictionary> {
-    vec![currencies(), languages(), nationalities(), religions()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/dictionaries/mod.rs b/crates/nvisy-pattern/src/shipped/dictionaries/mod.rs
deleted file mode 100644
index b0c20b1b3..000000000
--- a/crates/nvisy-pattern/src/shipped/dictionaries/mod.rs
+++ /dev/null
@@ -1,80 +0,0 @@
-//! Built-in [`Dictionary`]s, embedded at compile time.
-//!
-//! Accessors are grouped by *scope*:
-//!
-//! - `world::*` — universal: brand names, ISO codes (crypto,
-//!   currencies) that work in any language.
-//! - `<lang>::*` — locale-specific: terms written in `<lang>`
-//!   that translate when the document language changes
-//!   (`en::nationalities`, future `fr::nationalites`, etc.).
-//!
-//! Each pairs a TOML metadata sidecar
-//! (`assets/dictionaries/<scope>/<domain>/*.toml`) with a term
-//! source (`*.csv` for multi-column lists, `*.txt` for
-//! one-per-line), merging them via
-//! [`Dictionary::metadata_from_toml`] + [`crate::Term::from_csv`] /
-//! [`crate::Term::from_text`].
-//!
-//! [`Dictionary`]: crate::Dictionary
-
-pub mod en;
-pub mod world;
-
-use crate::Dictionary;
-
-/// Helper used by every per-region sub-module to define a shipped
-/// dictionary accessor.
-///
-/// Paths are resolved with `include_str!` against the path of the
-/// file that *expands* the macro, so callers in sub-modules pass
-/// paths relative to themselves.
-#[doc(hidden)]
-#[macro_export]
-macro_rules! __shipped_dictionary {
-    ($(#[$meta:meta])* fn $name:ident from $meta_path:literal with csv $terms:literal) => {
-        $(#[$meta])*
-        #[must_use]
-        pub fn $name() -> $crate::Dictionary {
-            let terms = $crate::Term::from_csv(include_str!($terms))
-                .expect(concat!("shipped term source `", $terms, "` parses"));
-            $crate::Dictionary::metadata_from_toml(include_str!($meta_path))
-                .expect(concat!("shipped metadata `", $meta_path, "` is well-formed"))
-                .with_terms(terms)
-                .build()
-                .expect(concat!("shipped dictionary `", $meta_path, "` builds"))
-        }
-    };
-    ($(#[$meta:meta])* fn $name:ident from $meta_path:literal with text $terms:literal) => {
-        $(#[$meta])*
-        #[must_use]
-        pub fn $name() -> $crate::Dictionary {
-            let terms = $crate::Term::from_text(include_str!($terms));
-            $crate::Dictionary::metadata_from_toml(include_str!($meta_path))
-                .expect(concat!("shipped metadata `", $meta_path, "` is well-formed"))
-                .with_terms(terms)
-                .build()
-                .expect(concat!("shipped dictionary `", $meta_path, "` builds"))
-        }
-    };
-}
-
-/// Every built-in dictionary shipped by this crate, regardless of
-/// scope.
-#[must_use]
-pub fn all() -> Vec<Dictionary> {
-    let mut dicts = world::all();
-    dicts.extend(en::all());
-    dicts
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn every_shipped_dictionary_parses() {
-        for dict in all() {
-            assert!(!dict.terms.is_empty(), "{} has no terms", dict.name);
-        }
-    }
-}
diff --git a/crates/nvisy-pattern/src/shipped/dictionaries/world.rs b/crates/nvisy-pattern/src/shipped/dictionaries/world.rs
deleted file mode 100644
index 08ab2a279..000000000
--- a/crates/nvisy-pattern/src/shipped/dictionaries/world.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-//! Universal dictionaries — apply regardless of jurisdiction or
-//! language.
-//!
-//! These are brand-name lists whose terms transcend locale.
-//! Cryptocurrency names (`Bitcoin`, `Ethereum`) and tickers
-//! (`BTC`, `ETH`) are the same string in every language.
-
-use crate::{__shipped_dictionary as shipped_dictionary, Dictionary};
-
-shipped_dictionary!(
-    /// Cryptocurrency names and ticker symbols (BTC, Bitcoin, ETH,
-    /// Ethereum, …).
-    fn cryptocurrencies
-        from "../../../assets/dictionaries/world/finance/cryptocurrencies.toml"
-        with csv "../../../assets/dictionaries/world/finance/cryptocurrencies.csv"
-);
-
-/// Every world-scoped built-in dictionary.
-#[must_use]
-pub fn all() -> Vec<Dictionary> {
-    vec![cryptocurrencies()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/mod.rs b/crates/nvisy-pattern/src/shipped/mod.rs
deleted file mode 100644
index db9c7fc5e..000000000
--- a/crates/nvisy-pattern/src/shipped/mod.rs
+++ /dev/null
@@ -1,14 +0,0 @@
-//! Built-in [`Regex`] rules and [`Dictionary`]s shipped with this
-//! crate.
-//!
-//! Each accessor parses an asset embedded via [`include_str!`] and
-//! returns a fresh value. Dictionaries split metadata into a TOML
-//! sidecar paired with a CSV/TXT term source; regex rules are
-//! self-contained TOML. Call [`patterns::all`] / [`dictionaries::all`]
-//! to load the full set, or pick individual accessors.
-//!
-//! [`Regex`]: crate::Regex
-//! [`Dictionary`]: crate::Dictionary
-
-pub mod dictionaries;
-pub mod patterns;
diff --git a/crates/nvisy-pattern/src/shipped/patterns/au.rs b/crates/nvisy-pattern/src/shipped/patterns/au.rs
deleted file mode 100644
index 446096943..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/au.rs
+++ /dev/null
@@ -1,32 +0,0 @@
-//! Australia — patterns scoped to AU jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// ABN — 11-digit Australian Business Number with mod-89
-    /// weighted checksum.
-    fn abn from "../../../assets/patterns/au/finance/abn.toml"
-);
-shipped_pattern!(
-    /// ACN — 9-digit Australian Company Number with mod-10
-    /// weighted checksum.
-    fn acn from "../../../assets/patterns/au/finance/acn.toml"
-);
-shipped_pattern!(
-    /// Medicare — 10-digit Australian Medicare card number
-    /// (prefix 2-6, mod-10 weighted check).
-    fn medicare from "../../../assets/patterns/au/health/medicare.toml"
-);
-shipped_pattern!(
-    /// TFN — 9-digit Australian Tax File Number with mod-11
-    /// weighted checksum.
-    fn tfn from "../../../assets/patterns/au/identity/tfn.toml"
-);
-
-/// Every AU-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![abn(), acn(), medicare(), tfn()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/ca.rs b/crates/nvisy-pattern/src/shipped/patterns/ca.rs
deleted file mode 100644
index 41c284072..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/ca.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-//! Canada — patterns scoped to CA jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// SIN — 9-digit Social Insurance Number with Luhn checksum
-    /// (first digit in `[1-7, 9]`).
-    fn sin from "../../../assets/patterns/ca/identity/sin.toml"
-);
-shipped_pattern!(
-    /// Canadian postal code — `A1A 1A1` (Canada Post Address
-    /// Standard letter alphabet).
-    fn postal_code from "../../../assets/patterns/ca/contact/postal_code.toml"
-);
-
-/// Every CA-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![sin(), postal_code()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/de.rs b/crates/nvisy-pattern/src/shipped/patterns/de.rs
deleted file mode 100644
index b2532f4f6..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/de.rs
+++ /dev/null
@@ -1,87 +0,0 @@
-//! Germany — patterns scoped to DE jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// Betriebsstättennummer (BSNR) — 9-digit clinic/practice ID.
-    fn bsnr from "../../../assets/patterns/de/health/bsnr.toml"
-);
-shipped_pattern!(
-    /// Lebenslange Arztnummer (LANR) — 9-digit physician ID with
-    /// KBV checksum.
-    fn lanr from "../../../assets/patterns/de/health/lanr.toml"
-);
-shipped_pattern!(
-    /// Krankenversichertennummer (KVNR) — statutory health
-    /// insurance ID per §290 SGB V.
-    fn health_insurance from "../../../assets/patterns/de/health/health_insurance.toml"
-);
-shipped_pattern!(
-    /// Personalausweisnummer — nPA (post-2010 ICAO) and legacy
-    /// `T`+8-digit ID card formats.
-    fn id_card from "../../../assets/patterns/de/identity/id_card.toml"
-);
-shipped_pattern!(
-    /// Reisepassnummer — ICAO Doc 9303 passport serial.
-    fn passport from "../../../assets/patterns/de/identity/passport.toml"
-);
-shipped_pattern!(
-    /// Rentenversicherungsnummer (RVNR/SVNR) — Deutsche
-    /// Rentenversicherung pension/social-security ID per VKVV § 4.
-    fn social_security from "../../../assets/patterns/de/identity/social_security.toml"
-);
-shipped_pattern!(
-    /// Steueridentifikationsnummer (Steuer-IdNr) — 11-digit
-    /// lifetime tax identifier with ISO 7064 Mod 11, 10 checksum.
-    fn tax_id from "../../../assets/patterns/de/identity/tax_id.toml"
-);
-shipped_pattern!(
-    /// Steuernummer — regional Finanzamt tax number.
-    fn tax_number from "../../../assets/patterns/de/identity/tax_number.toml"
-);
-shipped_pattern!(
-    /// Umsatzsteuer-Identifikationsnummer (USt-IdNr) — VAT
-    /// identification number.
-    fn vat_id from "../../../assets/patterns/de/identity/vat_id.toml"
-);
-shipped_pattern!(
-    /// Führerscheinnummer — EU-harmonized post-2013 driving
-    /// licence number.
-    fn fuehrerschein from "../../../assets/patterns/de/identity/fuehrerschein.toml"
-);
-shipped_pattern!(
-    /// Kfz-Kennzeichen — license plate (Unterscheidungszeichen +
-    /// Erkennungszeichen).
-    fn kfz from "../../../assets/patterns/de/vehicle/kfz.toml"
-);
-shipped_pattern!(
-    /// Postleitzahl (PLZ) — 5-digit postal code.
-    fn plz from "../../../assets/patterns/de/contact/plz.toml"
-);
-shipped_pattern!(
-    /// Handelsregisternummer — court-registered company ID with
-    /// `HRA`/`HRB` section prefix.
-    fn handelsregister from "../../../assets/patterns/de/finance/handelsregister.toml"
-);
-
-/// Every DE-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![
-        bsnr(),
-        lanr(),
-        health_insurance(),
-        id_card(),
-        passport(),
-        social_security(),
-        tax_id(),
-        tax_number(),
-        vat_id(),
-        fuehrerschein(),
-        kfz(),
-        plz(),
-        handelsregister(),
-    ]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/es.rs b/crates/nvisy-pattern/src/shipped/patterns/es.rs
deleted file mode 100644
index 5495b4f56..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/es.rs
+++ /dev/null
@@ -1,34 +0,0 @@
-//! Spain — patterns scoped to ES jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// NIF / DNI — 8-digit national ID + Mod 23 letter.
-    fn nif from "../../../assets/patterns/es/identity/nif.toml"
-);
-shipped_pattern!(
-    /// NIE — foreign-resident ID with `X`/`Y`/`Z` prefix.
-    fn nie from "../../../assets/patterns/es/identity/nie.toml"
-);
-shipped_pattern!(
-    /// Spanish passport — 3 letters + 6 digits.
-    fn passport from "../../../assets/patterns/es/identity/passport.toml"
-);
-shipped_pattern!(
-    /// CIF — company tax ID with entity-class letter + 7 digits +
-    /// control char (digit or letter per class).
-    fn cif from "../../../assets/patterns/es/finance/cif.toml"
-);
-shipped_pattern!(
-    /// Código postal — 5-digit postal code (province 01-52 in
-    /// the leading pair).
-    fn codigo_postal from "../../../assets/patterns/es/contact/codigo_postal.toml"
-);
-
-/// Every ES-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![nif(), nie(), passport(), cif(), codigo_postal()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/fi.rs b/crates/nvisy-pattern/src/shipped/patterns/fi.rs
deleted file mode 100644
index 0cc0a62f1..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/fi.rs
+++ /dev/null
@@ -1,17 +0,0 @@
-//! Finland — patterns scoped to FI jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// HETU — Finnish personal identity code with mod-31 control
-    /// character.
-    fn hetu from "../../../assets/patterns/fi/identity/hetu.toml"
-);
-
-/// Every FI-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![hetu()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/in.rs b/crates/nvisy-pattern/src/shipped/patterns/in.rs
deleted file mode 100644
index f3d44dbb4..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/in.rs
+++ /dev/null
@@ -1,49 +0,0 @@
-//! India — patterns scoped to IN jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// Aadhaar — 12-digit UIDAI ID with Verhoeff checksum and
-    /// no-palindrome rule.
-    fn aadhaar from "../../../assets/patterns/in/identity/aadhaar.toml"
-);
-shipped_pattern!(
-    /// PAN — 10-char Permanent Account Number with entity-type
-    /// letter at position 4.
-    fn pan from "../../../assets/patterns/in/identity/pan.toml"
-);
-shipped_pattern!(
-    /// GSTIN — 15-char Goods and Services Tax ID with base-36
-    /// weighted check digit.
-    fn gstin from "../../../assets/patterns/in/finance/gstin.toml"
-);
-shipped_pattern!(
-    /// Indian passport — 8-char alphanumeric (letter + non-zero
-    /// digit + 5 digits + non-zero digit).
-    fn passport from "../../../assets/patterns/in/identity/passport.toml"
-);
-shipped_pattern!(
-    /// EPIC voter ID — 10-char alphanumeric issued by the
-    /// Election Commission of India.
-    fn voter from "../../../assets/patterns/in/identity/voter.toml"
-);
-shipped_pattern!(
-    /// Indian vehicle registration — state + RTO district +
-    /// series + 4-digit serial.
-    fn vehicle_registration from "../../../assets/patterns/in/vehicle/registration.toml"
-);
-
-/// Every IN-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![
-        aadhaar(),
-        pan(),
-        gstin(),
-        passport(),
-        voter(),
-        vehicle_registration(),
-    ]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/it.rs b/crates/nvisy-pattern/src/shipped/patterns/it.rs
deleted file mode 100644
index 4ea108642..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/it.rs
+++ /dev/null
@@ -1,43 +0,0 @@
-//! Italy — patterns scoped to IT jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// Codice Fiscale — 16-character personal tax/health ID with
-    /// odd/even-mapping checksum, omocodia-aware.
-    fn fiscal_code from "../../../assets/patterns/it/identity/fiscal_code.toml"
-);
-shipped_pattern!(
-    /// Carta d'Identità — paper-based, CIE 2.0, and CIE 3.0
-    /// renderings.
-    fn identity_card from "../../../assets/patterns/it/identity/identity_card.toml"
-);
-shipped_pattern!(
-    /// Italian passport — 2 letters + 7 digits (Polizia di Stato
-    /// format).
-    fn passport from "../../../assets/patterns/it/identity/passport.toml"
-);
-shipped_pattern!(
-    /// Patente di guida — classic + Motorizzazione Civile
-    /// `U1`-prefixed format.
-    fn driving_licence from "../../../assets/patterns/it/identity/driving_licence.toml"
-);
-shipped_pattern!(
-    /// Partita IVA (P.IVA) — 11-digit VAT identifier with
-    /// Luhn-like checksum.
-    fn vat_code from "../../../assets/patterns/it/finance/vat_code.toml"
-);
-
-/// Every IT-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![
-        fiscal_code(),
-        identity_card(),
-        passport(),
-        driving_licence(),
-        vat_code(),
-    ]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/kr.rs b/crates/nvisy-pattern/src/shipped/patterns/kr.rs
deleted file mode 100644
index 2d0f95743..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/kr.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-//! South Korea — patterns scoped to KR jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// RRN — 13-digit Resident Registration Number with
-    /// weighted Mod-11 checksum (pre-Oct 2020) or random tail.
-    fn rrn from "../../../assets/patterns/kr/identity/rrn.toml"
-);
-shipped_pattern!(
-    /// FRN — Foreigner Registration Number (RRN shape with
-    /// gender/century digit 5-8 and (13 - sum) Mod-10 checksum).
-    fn frn from "../../../assets/patterns/kr/identity/frn.toml"
-);
-shipped_pattern!(
-    /// BRN — Business Registration Number with magic-keys
-    /// Mod-10 checksum.
-    fn brn from "../../../assets/patterns/kr/finance/brn.toml"
-);
-shipped_pattern!(
-    /// Korean passport — current (`LDDDLDDDD`) and legacy
-    /// (`LDDDDDDDD`) formats.
-    fn passport from "../../../assets/patterns/kr/identity/passport.toml"
-);
-shipped_pattern!(
-    /// Driver's license — 12-digit with region-code allowlist.
-    fn driver_license from "../../../assets/patterns/kr/identity/driver_license.toml"
-);
-
-/// Every KR-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![rrn(), frn(), brn(), passport(), driver_license()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/mod.rs b/crates/nvisy-pattern/src/shipped/patterns/mod.rs
deleted file mode 100644
index 86319d14c..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/mod.rs
+++ /dev/null
@@ -1,358 +0,0 @@
-//! Built-in [`Regex`] rules, embedded at compile time.
-//!
-//! Accessors are grouped by region — `world::*` for universal
-//! patterns, `<country>::*` (e.g. `us::*`, `uk::*`) for
-//! country-specific ones. Each returns a fresh [`Regex`] parsed
-//! from a TOML definition under
-//! `assets/patterns/<region>/<domain>/`. The parse happens on
-//! every call — rules are cheap to construct since
-//! [`PatternRecognizer::build`] does the heavy compilation.
-//!
-//! [`Regex`]: crate::Regex
-//! [`PatternRecognizer::build`]: crate::PatternRecognizer
-
-pub mod au;
-pub mod ca;
-pub mod de;
-pub mod es;
-pub mod fi;
-pub mod r#in;
-pub mod it;
-pub mod kr;
-pub mod ng;
-pub mod pl;
-pub mod se;
-pub mod sg;
-pub mod th;
-pub mod tr;
-pub mod uk;
-pub mod us;
-pub mod world;
-
-use crate::Regex;
-
-/// Helper used by every per-region sub-module to define a shipped
-/// pattern accessor.
-///
-/// The `$path` is resolved with `include_str!` against the path
-/// of the file that *expands* the macro, so callers in sub-modules
-/// (e.g. `world.rs`) pass paths relative to themselves.
-#[doc(hidden)]
-#[macro_export]
-macro_rules! __shipped_pattern {
-    ($(#[$meta:meta])* fn $name:ident from $path:literal) => {
-        $(#[$meta])*
-        #[must_use]
-        pub fn $name() -> $crate::Regex {
-            $crate::Regex::from_toml(include_str!($path))
-                .expect(concat!("shipped pattern `", $path, "` is well-formed"))
-        }
-    };
-}
-
-/// Every built-in pattern shipped by this crate, regardless of
-/// region.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    let mut out = world::all();
-    out.extend(us::all());
-    out.extend(uk::all());
-    out.extend(de::all());
-    out.extend(es::all());
-    out.extend(it::all());
-    out.extend(pl::all());
-    out.extend(au::all());
-    out.extend(ca::all());
-    out.extend(fi::all());
-    out.extend(se::all());
-    out.extend(r#in::all());
-    out.extend(kr::all());
-    out.extend(sg::all());
-    out.extend(tr::all());
-    out.extend(ng::all());
-    out.extend(th::all());
-    out
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn every_shipped_pattern_has_variants() {
-        for pattern in all() {
-            assert!(
-                !pattern.variants.is_empty(),
-                "pattern `{}` has no variants",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn world_patterns_have_no_country_scope() {
-        for pattern in world::all() {
-            assert!(
-                pattern.countries.is_empty(),
-                "world-scoped pattern `{}` must not declare countries",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn us_patterns_are_country_scoped_to_us() {
-        for pattern in us::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["US"],
-                "US-scoped pattern `{}` must declare countries = [US]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn uk_patterns_are_country_scoped_to_gb() {
-        for pattern in uk::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["GB"],
-                "UK-scoped pattern `{}` must declare countries = [GB]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn de_patterns_are_country_scoped_to_de() {
-        for pattern in de::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["DE"],
-                "DE-scoped pattern `{}` must declare countries = [DE]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn es_patterns_are_country_scoped_to_es() {
-        for pattern in es::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["ES"],
-                "ES-scoped pattern `{}` must declare countries = [ES]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn it_patterns_are_country_scoped_to_it() {
-        for pattern in it::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["IT"],
-                "IT-scoped pattern `{}` must declare countries = [IT]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn pl_patterns_are_country_scoped_to_pl() {
-        for pattern in pl::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["PL"],
-                "PL-scoped pattern `{}` must declare countries = [PL]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn au_patterns_are_country_scoped_to_au() {
-        for pattern in au::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["AU"],
-                "AU-scoped pattern `{}` must declare countries = [AU]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn ca_patterns_are_country_scoped_to_ca() {
-        for pattern in ca::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["CA"],
-                "CA-scoped pattern `{}` must declare countries = [CA]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn fi_patterns_are_country_scoped_to_fi() {
-        for pattern in fi::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["FI"],
-                "FI-scoped pattern `{}` must declare countries = [FI]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn se_patterns_are_country_scoped_to_se() {
-        for pattern in se::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["SE"],
-                "SE-scoped pattern `{}` must declare countries = [SE]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn in_patterns_are_country_scoped_to_in() {
-        for pattern in r#in::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["IN"],
-                "IN-scoped pattern `{}` must declare countries = [IN]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn kr_patterns_are_country_scoped_to_kr() {
-        for pattern in kr::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["KR"],
-                "KR-scoped pattern `{}` must declare countries = [KR]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn sg_patterns_are_country_scoped_to_sg() {
-        for pattern in sg::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["SG"],
-                "SG-scoped pattern `{}` must declare countries = [SG]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn tr_patterns_are_country_scoped_to_tr() {
-        for pattern in tr::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["TR"],
-                "TR-scoped pattern `{}` must declare countries = [TR]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn ng_patterns_are_country_scoped_to_ng() {
-        for pattern in ng::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["NG"],
-                "NG-scoped pattern `{}` must declare countries = [NG]",
-                pattern.name,
-            );
-        }
-    }
-
-    #[test]
-    fn th_patterns_are_country_scoped_to_th() {
-        for pattern in th::all() {
-            assert_eq!(
-                pattern
-                    .countries
-                    .iter()
-                    .map(|c| c.as_str())
-                    .collect::<Vec<_>>(),
-                vec!["TH"],
-                "TH-scoped pattern `{}` must declare countries = [TH]",
-                pattern.name,
-            );
-        }
-    }
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/ng.rs b/crates/nvisy-pattern/src/shipped/patterns/ng.rs
deleted file mode 100644
index fd2fc8b54..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/ng.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-//! Nigeria — patterns scoped to NG jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// NIN — 11-digit National Identification Number with
-    /// Verhoeff checksum.
-    fn nin from "../../../assets/patterns/ng/identity/nin.toml"
-);
-shipped_pattern!(
-    /// Vehicle registration — 3-letter LGA + 3 digits + 2-letter
-    /// year/batch (current 2011+ format).
-    fn vehicle_registration from "../../../assets/patterns/ng/vehicle/registration.toml"
-);
-
-/// Every NG-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![nin(), vehicle_registration()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/pl.rs b/crates/nvisy-pattern/src/shipped/patterns/pl.rs
deleted file mode 100644
index 26e169890..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/pl.rs
+++ /dev/null
@@ -1,32 +0,0 @@
-//! Poland — patterns scoped to PL jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// PESEL — 11-digit personal ID with date + sex + serial +
-    /// weighted-mod-10 check digit.
-    fn pesel from "../../../assets/patterns/pl/identity/pesel.toml"
-);
-shipped_pattern!(
-    /// NIP — 10-digit taxpayer identification number, mod-11
-    /// weighted checksum.
-    fn nip from "../../../assets/patterns/pl/finance/nip.toml"
-);
-shipped_pattern!(
-    /// REGON — 9 or 14-digit company registry number, mod-11
-    /// weighted checksum.
-    fn regon from "../../../assets/patterns/pl/finance/regon.toml"
-);
-shipped_pattern!(
-    /// Kod pocztowy — `NN-NNN` postal code (Poczta Polska,
-    /// 1973-present).
-    fn kod_pocztowy from "../../../assets/patterns/pl/contact/kod_pocztowy.toml"
-);
-
-/// Every PL-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![pesel(), nip(), regon(), kod_pocztowy()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/se.rs b/crates/nvisy-pattern/src/shipped/patterns/se.rs
deleted file mode 100644
index aa5f275ac..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/se.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-//! Sweden — patterns scoped to SE jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// Personnummer — 10/12-digit personal identity number with
-    /// date validity (incl. samordningsnummer) and Luhn checksum.
-    fn personnummer from "../../../assets/patterns/se/identity/personnummer.toml"
-);
-shipped_pattern!(
-    /// Organisationsnummer — 10-digit Bolagsverket company ID
-    /// with third digit ≥ 2 and Luhn checksum.
-    fn organisationsnummer from "../../../assets/patterns/se/finance/organisationsnummer.toml"
-);
-shipped_pattern!(
-    /// Postnummer — 5-digit postal code in `XXX XX` rendering
-    /// (Postnord standard).
-    fn postnummer from "../../../assets/patterns/se/contact/postnummer.toml"
-);
-
-/// Every SE-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![personnummer(), organisationsnummer(), postnummer()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/sg.rs b/crates/nvisy-pattern/src/shipped/patterns/sg.rs
deleted file mode 100644
index 62349df79..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/sg.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-//! Singapore — patterns scoped to SG jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// NRIC / FIN — Singapore National Registration Identity
-    /// Card / Foreign Identification Number with weighted Mod-11
-    /// checksum.
-    fn nric from "../../../assets/patterns/sg/identity/nric.toml"
-);
-shipped_pattern!(
-    /// UEN — Unique Entity Number issued by ACRA (formats A, B,
-    /// and C, each with its own checksum).
-    fn uen from "../../../assets/patterns/sg/finance/uen.toml"
-);
-shipped_pattern!(
-    /// Singapore postal code — 6-digit Singapore Post code.
-    fn postal_code from "../../../assets/patterns/sg/contact/postal_code.toml"
-);
-
-/// Every SG-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![nric(), uen(), postal_code()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/th.rs b/crates/nvisy-pattern/src/shipped/patterns/th.rs
deleted file mode 100644
index 48fb43d4f..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/th.rs
+++ /dev/null
@@ -1,23 +0,0 @@
-//! Thailand — patterns scoped to TH jurisdictional formats.
-//!
-//! Algorithm sourced from the Department of Provincial
-//! Administration spec (not from the Presidio bundle).
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// National ID (เลขประจำตัวประชาชน) — 13-digit ID with
-    /// weighted Mod-11 check digit; first digit 1-8.
-    fn national_id from "../../../assets/patterns/th/identity/national_id.toml"
-);
-shipped_pattern!(
-    /// Thai postal code (รหัสไปรษณีย์) — 5-digit Thailand Post
-    /// code with province prefix 10-96.
-    fn postal_code from "../../../assets/patterns/th/contact/postal_code.toml"
-);
-
-/// Every TH-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![national_id(), postal_code()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/tr.rs b/crates/nvisy-pattern/src/shipped/patterns/tr.rs
deleted file mode 100644
index 47615a863..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/tr.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-//! Turkey — patterns scoped to TR jurisdictional formats.
-//!
-//! See `assets/PRESIDIO.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// TCKN — 11-digit T.C. Kimlik No with two-step weighted
-    /// checksum.
-    fn tckn from "../../../assets/patterns/tr/identity/tckn.toml"
-);
-shipped_pattern!(
-    /// Turkish license plate (plaka) — space- and
-    /// hyphen-separated renderings, province codes 01-81.
-    fn license_plate from "../../../assets/patterns/tr/vehicle/license_plate.toml"
-);
-shipped_pattern!(
-    /// Turkish posta kodu — 5-digit postal code (province
-    /// prefix 01-81).
-    fn posta_kodu from "../../../assets/patterns/tr/contact/posta_kodu.toml"
-);
-
-/// Every TR-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![tckn(), license_plate(), posta_kodu()]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/uk.rs b/crates/nvisy-pattern/src/shipped/patterns/uk.rs
deleted file mode 100644
index d5e0fff06..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/uk.rs
+++ /dev/null
@@ -1,45 +0,0 @@
-//! United Kingdom — patterns scoped to UK jurisdictional formats.
-//!
-//! See `assets/NOTICE.md` for third-party attribution.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// UK NHS numbers (10-digit, mod-11 validated).
-    fn nhs from "../../../assets/patterns/uk/identity/nhs.toml"
-);
-shipped_pattern!(
-    /// UK National Insurance numbers (NINO).
-    fn nino from "../../../assets/patterns/uk/identity/nino.toml"
-);
-shipped_pattern!(
-    /// UK driving licence numbers (DVLA 16-character).
-    fn driving_licence from "../../../assets/patterns/uk/identity/driving_licence.toml"
-);
-shipped_pattern!(
-    /// UK postcodes (BS7666 format, plus GIR 0AA).
-    fn postcode from "../../../assets/patterns/uk/contact/postcode.toml"
-);
-shipped_pattern!(
-    /// UK vehicle registration numbers (current, prefix, and
-    /// suffix eras).
-    fn vehicle_registration from "../../../assets/patterns/uk/vehicle/registration.toml"
-);
-shipped_pattern!(
-    /// UK passport numbers (post-2015 format). Weak score; relies
-    /// on the context-keyword boost.
-    fn passport from "../../../assets/patterns/uk/identity/passport.toml"
-);
-
-/// Every UK-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![
-        nhs(),
-        nino(),
-        driving_licence(),
-        postcode(),
-        vehicle_registration(),
-        passport(),
-    ]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/us.rs b/crates/nvisy-pattern/src/shipped/patterns/us.rs
deleted file mode 100644
index afde142fe..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/us.rs
+++ /dev/null
@@ -1,63 +0,0 @@
-//! United States — patterns scoped to US jurisdictional formats.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// US bank routing numbers (ABA RTN, mod-10 validated).
-    fn bank_routing from "../../../assets/patterns/us/finance/bank_routing.toml"
-);
-shipped_pattern!(
-    /// US Social Security numbers (AAA-GG-SSSS).
-    fn ssn from "../../../assets/patterns/us/identity/ssn.toml"
-);
-shipped_pattern!(
-    /// US driver's license numbers (state-shape union).
-    fn drivers_license from "../../../assets/patterns/us/identity/drivers_license.toml"
-);
-shipped_pattern!(
-    /// US passport numbers.
-    fn passport from "../../../assets/patterns/us/identity/passport.toml"
-);
-shipped_pattern!(
-    /// US ZIP and ZIP+4 postal codes.
-    fn postal_code from "../../../assets/patterns/us/identity/postal_code.toml"
-);
-shipped_pattern!(
-    /// US Individual Taxpayer Identification Number (ITIN).
-    fn itin from "../../../assets/patterns/us/identity/itin.toml"
-);
-shipped_pattern!(
-    /// US National Provider Identifier (NPI, Luhn-on-80840 validated).
-    fn npi from "../../../assets/patterns/us/health/npi.toml"
-);
-shipped_pattern!(
-    /// US Medicare Beneficiary Identifier (MBI).
-    fn mbi from "../../../assets/patterns/us/health/mbi.toml"
-);
-shipped_pattern!(
-    /// US generic bank account number (8-17 digits, no checksum).
-    /// Very weak score; relies on the context-keyword boost.
-    fn bank_account from "../../../assets/patterns/us/finance/bank_account.toml"
-);
-shipped_pattern!(
-    /// US DEA registration number (medical license,
-    /// checksum-validated).
-    fn medical_license from "../../../assets/patterns/us/health/medical_license.toml"
-);
-
-/// Every US-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![
-        bank_routing(),
-        ssn(),
-        drivers_license(),
-        passport(),
-        postal_code(),
-        itin(),
-        npi(),
-        mbi(),
-        bank_account(),
-        medical_license(),
-    ]
-}
diff --git a/crates/nvisy-pattern/src/shipped/patterns/world.rs b/crates/nvisy-pattern/src/shipped/patterns/world.rs
deleted file mode 100644
index c023f5401..000000000
--- a/crates/nvisy-pattern/src/shipped/patterns/world.rs
+++ /dev/null
@@ -1,105 +0,0 @@
-//! Universal patterns — apply regardless of jurisdiction.
-
-use crate::{__shipped_pattern as shipped_pattern, Regex};
-
-shipped_pattern!(
-    /// Email address (RFC-loose).
-    fn email from "../../../assets/patterns/world/contact/email.toml"
-);
-shipped_pattern!(
-    /// International phone numbers.
-    fn phone from "../../../assets/patterns/world/contact/phone.toml"
-);
-shipped_pattern!(
-    /// URLs (HTTP/HTTPS/FTP).
-    fn url from "../../../assets/patterns/world/contact/url.toml"
-);
-
-shipped_pattern!(
-    /// AWS access key IDs.
-    fn aws_key from "../../../assets/patterns/world/credentials/aws_key.toml"
-);
-shipped_pattern!(
-    /// Heuristic generic API key.
-    fn generic_api_key from "../../../assets/patterns/world/credentials/generic_api_key.toml"
-);
-shipped_pattern!(
-    /// GitHub personal access tokens.
-    fn github_token from "../../../assets/patterns/world/credentials/github_token.toml"
-);
-shipped_pattern!(
-    /// PEM-formatted private keys.
-    fn private_key from "../../../assets/patterns/world/credentials/private_key.toml"
-);
-shipped_pattern!(
-    /// Stripe live/test secret keys.
-    fn stripe_key from "../../../assets/patterns/world/credentials/stripe_key.toml"
-);
-
-shipped_pattern!(
-    /// Bitcoin (legacy + bech32) addresses.
-    fn bitcoin_address from "../../../assets/patterns/world/finance/bitcoin_address.toml"
-);
-shipped_pattern!(
-    /// Credit-card numbers, Luhn-validated.
-    fn credit_card from "../../../assets/patterns/world/finance/credit_card.toml"
-);
-shipped_pattern!(
-    /// Ethereum addresses.
-    fn ethereum_address from "../../../assets/patterns/world/finance/ethereum_address.toml"
-);
-shipped_pattern!(
-    /// International Bank Account Numbers.
-    fn iban from "../../../assets/patterns/world/finance/iban.toml"
-);
-shipped_pattern!(
-    /// SWIFT / BIC codes.
-    fn swift_code from "../../../assets/patterns/world/finance/swift_code.toml"
-);
-
-shipped_pattern!(
-    /// IPv4 addresses.
-    fn ipv4 from "../../../assets/patterns/world/network/ipv4.toml"
-);
-shipped_pattern!(
-    /// IPv6 addresses.
-    fn ipv6 from "../../../assets/patterns/world/network/ipv6.toml"
-);
-shipped_pattern!(
-    /// MAC (Ethernet) addresses.
-    fn mac_address from "../../../assets/patterns/world/network/mac_address.toml"
-);
-
-shipped_pattern!(
-    /// Date of birth in common written formats.
-    fn date_of_birth from "../../../assets/patterns/world/personal/date_of_birth.toml"
-);
-shipped_pattern!(
-    /// Date + time stamps in ISO-like formats.
-    fn datetime from "../../../assets/patterns/world/personal/datetime.toml"
-);
-
-/// Every world-scoped built-in pattern.
-#[must_use]
-pub fn all() -> Vec<Regex> {
-    vec![
-        email(),
-        phone(),
-        url(),
-        aws_key(),
-        generic_api_key(),
-        github_token(),
-        private_key(),
-        stripe_key(),
-        bitcoin_address(),
-        credit_card(),
-        ethereum_address(),
-        iban(),
-        swift_code(),
-        ipv4(),
-        ipv6(),
-        mac_address(),
-        date_of_birth(),
-        datetime(),
-    ]
-}
diff --git a/crates/nvisy-pattern/src/validators/au/abn.rs b/crates/nvisy-pattern/src/validators/au/abn.rs
deleted file mode 100644
index bbbfc0f17..000000000
--- a/crates/nvisy-pattern/src/validators/au/abn.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-//! Australian Business Number (ABN) validator.
-//!
-//! 11 digits issued by the Australian Business Register (ABR).
-//! Algorithm: subtract 1 from the leading digit (wrap 0→9),
-//! then weighted sum with `[10, 1, 3, 5, 7, 9, 11, 13, 15, 17, 19]`
-//! must be divisible by 89.
-
-const WEIGHTS: [u32; 11] = [10, 1, 3, 5, 7, 9, 11, 13, 15, 17, 19];
-
-/// Return `true` when `value` is a valid 11-digit ABN. Whitespace
-/// and dash separators in the canonical `NN NNN NNN NNN` rendering
-/// are stripped before validation.
-pub fn abn(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 11 || extras > 0 {
-        return false;
-    }
-    let mut adjusted = digits;
-    adjusted[0] = if adjusted[0] == 0 { 9 } else { adjusted[0] - 1 };
-    let sum: u32 = adjusted.iter().zip(WEIGHTS).map(|(d, w)| d * w).sum();
-    sum.is_multiple_of(89)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_abn() {
-        // 51 824 753 556 — Australian Taxation Office sample.
-        assert!(abn("51824753556"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(abn("51 824 753 556"));
-        assert!(abn("51-824-753-556"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        // 53 004 085 616 — Telstra Corporation, public ABN.
-        assert!(abn("53004085616"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!abn("11000000000"));
-        assert!(!abn("51824753557"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!abn("5182475355"));
-        assert!(!abn("518247535566"));
-        assert!(!abn(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!abn("5182475355A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/au/acn.rs b/crates/nvisy-pattern/src/validators/au/acn.rs
deleted file mode 100644
index 054fbec8b..000000000
--- a/crates/nvisy-pattern/src/validators/au/acn.rs
+++ /dev/null
@@ -1,63 +0,0 @@
-//! Australian Company Number (ACN) validator.
-//!
-//! 9 digits issued by ASIC. Algorithm: weighted sum of the first
-//! 8 digits with `[8, 7, 6, 5, 4, 3, 2, 1]`; check digit equals
-//! `(10 - sum mod 10) mod 10`.
-
-const WEIGHTS: [u32; 8] = [8, 7, 6, 5, 4, 3, 2, 1];
-
-/// Return `true` when `value` is a valid 9-digit ACN. Whitespace
-/// and dash separators in the canonical `NNN NNN NNN` rendering
-/// are stripped before validation.
-pub fn acn(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 9 || extras > 0 {
-        return false;
-    }
-    let sum: u32 = digits[..8].iter().zip(WEIGHTS).map(|(d, w)| d * w).sum();
-    let check = (10 - sum % 10) % 10;
-    check == digits[8]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_acn() {
-        // Body 12345678 → check 0.
-        assert!(acn("123456780"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(acn("123 456 780"));
-        assert!(acn("123-456-780"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        // Body 00400000 → check 6.
-        assert!(acn("004000006"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!acn("123456789"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!acn("12345678"));
-        assert!(!acn("1234567890"));
-        assert!(!acn(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/au/medicare.rs b/crates/nvisy-pattern/src/validators/au/medicare.rs
deleted file mode 100644
index d78e4e1cb..000000000
--- a/crates/nvisy-pattern/src/validators/au/medicare.rs
+++ /dev/null
@@ -1,76 +0,0 @@
-//! Australian Medicare number validator.
-//!
-//! Card numbers are 10 or 11 digits where the first 9 form the
-//! identifier (first digit in `[2..=6]`) followed by a check
-//! digit; the trailing digit(s) encode the individual reference
-//! and issue number (not part of the checksum).
-//!
-//! Algorithm: weighted sum of the first 8 digits with
-//! `[1, 3, 7, 9, 1, 3, 7, 9]` mod 10 equals the 9th digit.
-
-const WEIGHTS: [u32; 8] = [1, 3, 7, 9, 1, 3, 7, 9];
-
-/// Return `true` when the first 9 digits of `value` form a
-/// Medicare number whose checksum matches. Whitespace and dash
-/// separators are stripped before validation. Trailing digits
-/// (individual reference + issue) are accepted but ignored.
-pub fn medicare(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if !(9..=11).contains(&digits.len()) || extras > 0 {
-        return false;
-    }
-    if !(2..=6).contains(&digits[0]) {
-        return false;
-    }
-    let sum: u32 = digits[..8].iter().zip(WEIGHTS).map(|(d, w)| d * w).sum();
-    sum % 10 == digits[8]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_medicare() {
-        // Body 22281236 → check 6.
-        assert!(medicare("222812366"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(medicare("2228 12366"));
-        assert!(medicare("2228 1236 6"));
-    }
-
-    #[test]
-    fn accepts_with_individual_reference() {
-        // 10-digit form: 9-digit Medicare + individual reference.
-        assert!(medicare("2228123661"));
-    }
-
-    #[test]
-    fn rejects_wrong_prefix() {
-        // Body must start with 2-6; 1xxx is invalid.
-        assert!(!medicare("122812366"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!medicare("222812360"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!medicare("22281236"));
-        assert!(!medicare("222812366111"));
-        assert!(!medicare(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/au/mod.rs b/crates/nvisy-pattern/src/validators/au/mod.rs
deleted file mode 100644
index efe67b81c..000000000
--- a/crates/nvisy-pattern/src/validators/au/mod.rs
+++ /dev/null
@@ -1,17 +0,0 @@
-//! Australia-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"au.abn"`, `"au.acn"`, `"au.medicare"`,
-//! `"au.tfn"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod abn;
-mod acn;
-mod medicare;
-mod tfn;
-
-pub use self::abn::abn;
-pub use self::acn::acn;
-pub use self::medicare::medicare;
-pub use self::tfn::tfn;
diff --git a/crates/nvisy-pattern/src/validators/au/tfn.rs b/crates/nvisy-pattern/src/validators/au/tfn.rs
deleted file mode 100644
index ae16cf138..000000000
--- a/crates/nvisy-pattern/src/validators/au/tfn.rs
+++ /dev/null
@@ -1,61 +0,0 @@
-//! Australian Tax File Number (TFN) validator.
-//!
-//! 9 digits issued by the Australian Taxation Office. Algorithm:
-//! weighted sum with `[1, 4, 3, 7, 5, 8, 6, 9, 10]` over all 9
-//! digits must be divisible by 11.
-
-const WEIGHTS: [u32; 9] = [1, 4, 3, 7, 5, 8, 6, 9, 10];
-
-/// Return `true` when `value` is a valid 9-digit TFN. Whitespace
-/// and dash separators in the canonical `NNN NNN NNN` rendering
-/// are stripped before validation.
-pub fn tfn(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 9 || extras > 0 {
-        return false;
-    }
-    let sum: u32 = digits.iter().zip(WEIGHTS).map(|(d, w)| d * w).sum();
-    sum.is_multiple_of(11)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_tfn() {
-        // 123 456 782 — ATO documented test value.
-        assert!(tfn("123456782"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(tfn("123 456 782"));
-        assert!(tfn("123-456-782"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        assert!(tfn("100000001"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!tfn("123456789"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!tfn("12345678"));
-        assert!(!tfn("1234567823"));
-        assert!(!tfn(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/btc.rs b/crates/nvisy-pattern/src/validators/btc.rs
deleted file mode 100644
index 7abc6d679..000000000
--- a/crates/nvisy-pattern/src/validators/btc.rs
+++ /dev/null
@@ -1,53 +0,0 @@
-//! Bitcoin legacy-address (Base58Check) checksum validator.
-//!
-//! Validates P2PKH (`1…`) and P2SH (`3…`) addresses by decoding
-//! the Base58 payload and verifying its trailing four-byte
-//! double-SHA256 checksum. Bech32 / Bech32m addresses (`bc1…`)
-//! are not handled here — those use a different polynomial check.
-
-/// Return `true` if `value` is a structurally valid Base58Check
-/// Bitcoin address.
-///
-/// Accepts P2PKH (version byte `0x00`, `1…`) and P2SH
-/// (version byte `0x05`, `3…`) on mainnet. Rejects mismatched
-/// version bytes, broken Base58, and bad checksums.
-pub fn btc(value: &str) -> bool {
-    match bs58::decode(value.trim()).with_check(None).into_vec() {
-        Ok(bytes) if bytes.len() == 21 => matches!(bytes[0], 0x00 | 0x05),
-        _ => false,
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_known_p2pkh() {
-        // Satoshi's genesis-block coinbase address.
-        assert!(btc("1A1zP1eP5QGefi2DMPTfTL5SLmv7DivfNa"));
-    }
-
-    #[test]
-    fn accepts_known_p2sh() {
-        assert!(btc("3J98t1WpEZ73CNmQviecrnyiWrnqRhWNLy"));
-    }
-
-    #[test]
-    fn rejects_bad_checksum() {
-        // Final char flipped.
-        assert!(!btc("1A1zP1eP5QGefi2DMPTfTL5SLmv7DivfNb"));
-    }
-
-    #[test]
-    fn rejects_non_base58() {
-        assert!(!btc("1A1zP1eP5QGefi2DMPTfTL5SLmv7Divf0a"));
-        assert!(!btc(""));
-    }
-
-    #[test]
-    fn rejects_unknown_version() {
-        // Bitcoin testnet P2PKH (version byte `0x6f`).
-        assert!(!btc("mipcBbFg9gMiCh81Kj8tqqdgoZub1ZJRfn"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/ca/mod.rs b/crates/nvisy-pattern/src/validators/ca/mod.rs
deleted file mode 100644
index e905d0c39..000000000
--- a/crates/nvisy-pattern/src/validators/ca/mod.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-//! Canada-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"ca.sin"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod sin;
-
-pub use self::sin::sin;
diff --git a/crates/nvisy-pattern/src/validators/ca/sin.rs b/crates/nvisy-pattern/src/validators/ca/sin.rs
deleted file mode 100644
index aac70e62e..000000000
--- a/crates/nvisy-pattern/src/validators/ca/sin.rs
+++ /dev/null
@@ -1,77 +0,0 @@
-//! Canadian Social Insurance Number (SIN) validator.
-//!
-//! 9 digits with a Luhn check digit over the first 8. Numbers
-//! beginning with `0` or `8` are reserved by Employment and
-//! Social Development Canada (ESDC) and never assigned.
-
-/// Return `true` when `value` is a valid Canadian SIN. Whitespace
-/// and dash separators in the canonical `NNN NNN NNN` /
-/// `NNN-NNN-NNN` renderings are stripped before validation.
-pub fn sin(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 9 || extras > 0 {
-        return false;
-    }
-    if digits[0] == 0 || digits[0] == 8 {
-        return false;
-    }
-    let mut sum: u32 = 0;
-    for (i, d) in digits.iter().rev().enumerate() {
-        if i.is_multiple_of(2) {
-            sum += d;
-        } else {
-            let m = d * 2;
-            sum += if m > 9 { m - 9 } else { m };
-        }
-    }
-    sum.is_multiple_of(10)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_sin() {
-        // 123 456 782 — widely-quoted ESDC test value.
-        assert!(sin("123456782"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(sin("123 456 782"));
-        assert!(sin("123-456-782"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        assert!(sin("100000009"));
-    }
-
-    #[test]
-    fn rejects_reserved_prefix() {
-        // 0xxxxxxxx and 8xxxxxxxx are reserved.
-        assert!(!sin("012345670"));
-        assert!(!sin("812345674"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!sin("123456789"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!sin("12345678"));
-        assert!(!sin("1234567823"));
-        assert!(!sin(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/date.rs b/crates/nvisy-pattern/src/validators/date.rs
deleted file mode 100644
index 7d35d0d1f..000000000
--- a/crates/nvisy-pattern/src/validators/date.rs
+++ /dev/null
@@ -1,135 +0,0 @@
-//! Calendar-date structural validator.
-
-/// Return `true` if `value` is a real calendar date in one of the
-/// supported written formats.
-///
-/// Supported formats are `MM/DD/YYYY`, `DD/MM/YYYY`, `YYYY-MM-DD`,
-/// and `YYYY/MM/DD`, with `/` or `-` as separators. Leap years
-/// are honoured and the year must fall in `1900..=2100`.
-///
-/// # Ambiguity
-///
-/// When both interpretations are valid (e.g. `02/03/1999` could
-/// mean Feb 3 or 3 Mar), the validator prefers `MM/DD/YYYY` and
-/// only falls back to `DD/MM/YYYY` when the first part is not a
-/// valid month. Locale disambiguation is out of scope.
-pub fn date(value: &str) -> bool {
-    let parts: Vec<&str> = value.split(['/', '-']).collect();
-    if parts.len() != 3 {
-        return false;
-    }
-
-    let (year, month, day) = if parts[0].len() == 4 {
-        // YYYY-MM-DD or YYYY/MM/DD
-        match (parts[0].parse(), parts[1].parse(), parts[2].parse()) {
-            (Ok(y), Ok(m), Ok(d)) => (y, m, d),
-            _ => return false,
-        }
-    } else if parts[2].len() == 4 {
-        // Could be MM/DD/YYYY or DD/MM/YYYY.
-        let a: u32 = match parts[0].parse() {
-            Ok(v) => v,
-            Err(_) => return false,
-        };
-        let b: u32 = match parts[1].parse() {
-            Ok(v) => v,
-            Err(_) => return false,
-        };
-        let y: u32 = match parts[2].parse() {
-            Ok(v) => v,
-            Err(_) => return false,
-        };
-
-        // Try MM/DD/YYYY first, fall back to DD/MM/YYYY.
-        if (1..=12).contains(&a) && is_valid_date(y, a, b) {
-            (y, a, b)
-        } else if (1..=12).contains(&b) && is_valid_date(y, b, a) {
-            (y, b, a)
-        } else {
-            return false;
-        }
-    } else {
-        return false;
-    };
-
-    is_valid_date(year, month, day)
-}
-
-fn is_valid_date(year: u32, month: u32, day: u32) -> bool {
-    if !(1900..=2100).contains(&year) {
-        return false;
-    }
-    if !(1..=12).contains(&month) {
-        return false;
-    }
-
-    let max_day = match month {
-        1 | 3 | 5 | 7 | 8 | 10 | 12 => 31,
-        4 | 6 | 9 | 11 => 30,
-        2 => {
-            if is_leap_year(year) {
-                29
-            } else {
-                28
-            }
-        }
-        _ => return false,
-    };
-
-    (1..=max_day).contains(&day)
-}
-
-fn is_leap_year(year: u32) -> bool {
-    (year.is_multiple_of(4) && !year.is_multiple_of(100)) || year.is_multiple_of(400)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn mm_dd_yyyy() {
-        assert!(date("01/15/1990"));
-        assert!(date("12-31-2000"));
-    }
-
-    #[test]
-    fn yyyy_mm_dd() {
-        assert!(date("1990-01-15"));
-        assert!(date("2000/12/31"));
-    }
-
-    #[test]
-    fn leap_year() {
-        assert!(date("02/29/2000"));
-        assert!(date("2000-02-29"));
-        assert!(!date("02/29/2001"));
-    }
-
-    #[test]
-    fn invalid_day() {
-        assert!(!date("04/31/1990"));
-        assert!(!date("01/32/1990"));
-        assert!(!date("01/00/1990"));
-    }
-
-    #[test]
-    fn invalid_month() {
-        // 13/01/1990 is valid as DD/MM/YYYY (Jan 13)
-        assert!(date("13/01/1990"));
-        // YYYY-MM-DD format: month 13 is invalid
-        assert!(!date("1990-13-01"));
-    }
-
-    #[test]
-    fn invalid_year() {
-        assert!(!date("01/01/1899"));
-        assert!(!date("1899-01-01"));
-    }
-
-    #[test]
-    fn dd_mm_yyyy_ambiguous() {
-        // 15/01/1990: first part > 12 so must be DD/MM
-        assert!(date("15/01/1990"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/de/bsnr.rs b/crates/nvisy-pattern/src/validators/de/bsnr.rs
deleted file mode 100644
index 90ea239b8..000000000
--- a/crates/nvisy-pattern/src/validators/de/bsnr.rs
+++ /dev/null
@@ -1,51 +0,0 @@
-//! German Betriebsstättennummer (BSNR) structural validator.
-//!
-//! BSNR is a 9-digit practice/site-of-care number assigned by the
-//! regional Kassenärztliche Vereinigung (KV). There is no
-//! published Prüfziffer algorithm, so this validator only drops
-//! obvious garbage (wrong length, non-digit, all-zero); the
-//! `\b\d{9}\b` regex is too broad to promote a 2-digit prefix
-//! whitelist into a high-confidence signal, so the upstream
-//! `valid_kv_codes` table is left out — context keywords
-//! ("Betriebsstättennummer", "Praxis", …) drive final confidence
-//! via the enhancer.
-
-/// Return `true` when `value` is a structurally-plausible BSNR.
-///
-/// Rejects: wrong length, non-digit characters, all-zero string.
-pub fn bsnr(value: &str) -> bool {
-    let digits: String = value.chars().filter(|c| !c.is_ascii_whitespace()).collect();
-    if digits.len() != 9 || !digits.chars().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    !digits.chars().all(|c| c == '0')
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_plausible_shapes() {
-        assert!(bsnr("021234568"));
-        assert!(bsnr("381789045"));
-        assert!(bsnr("721234567"));
-    }
-
-    #[test]
-    fn rejects_all_zero() {
-        assert!(!bsnr("000000000"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!bsnr("12345678"));
-        assert!(!bsnr("1234567890"));
-        assert!(!bsnr(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!bsnr("12345678A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/de/health_insurance.rs b/crates/nvisy-pattern/src/validators/de/health_insurance.rs
deleted file mode 100644
index 2e6ca1022..000000000
--- a/crates/nvisy-pattern/src/validators/de/health_insurance.rs
+++ /dev/null
@@ -1,81 +0,0 @@
-//! German Krankenversicherungsnummer (KVNR) checksum validator.
-//!
-//! 10-character KVNR: leading letter + 8 digits + check digit.
-//! GKV §290 SGB V Anlage 1 expands the letter to its 2-digit
-//! 1-based ordinal (A→01, B→02, …, Z→26), concatenates with the
-//! 8 data digits, weights the resulting 10 digits with the
-//! alternating factors `[1,2,1,2,1,2,1,2,1,2]`, cross-sums any
-//! product ≥ 10, and asserts the total mod 10 equals the check
-//! digit.
-
-/// Return `true` when `value` is a valid 10-character KVNR
-/// (letter + 9 digits) per GKV §290 SGB V.
-pub fn health_insurance(value: &str) -> bool {
-    let trimmed = value.trim().to_ascii_uppercase();
-    let chars: Vec<char> = trimmed.chars().collect();
-    if chars.len() != 10 {
-        return false;
-    }
-    let Some(letter) = chars.first().copied() else {
-        return false;
-    };
-    if !letter.is_ascii_uppercase() {
-        return false;
-    }
-    if !chars[1..].iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-
-    // Letter → 2-digit ordinal (A=01, …, Z=26) then 8 data digits.
-    let letter_val = (letter as u32) - ('A' as u32) + 1;
-    let mut effective: Vec<u32> = vec![letter_val / 10, letter_val % 10];
-    for c in chars.iter().skip(1).take(8) {
-        effective.push(c.to_digit(10).unwrap());
-    }
-
-    let weights = [1, 2, 1, 2, 1, 2, 1, 2, 1, 2];
-    let total: u32 = effective
-        .iter()
-        .zip(weights)
-        .map(|(d, w)| {
-            let p = d * w;
-            if p >= 10 { (p / 10) + (p % 10) } else { p }
-        })
-        .sum();
-
-    let check = chars[9].to_digit(10).unwrap();
-    total % 10 == check
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_spec_example() {
-        // From § 290 SGB V Anlage 1, Stand 02.01.2023.
-        assert!(health_insurance("A000500015"));
-    }
-
-    #[test]
-    fn rejects_wrong_check_digit() {
-        assert!(!health_insurance("A000500016"));
-    }
-
-    #[test]
-    fn rejects_missing_letter_prefix() {
-        assert!(!health_insurance("0000500015"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!health_insurance("A00050001"));
-        assert!(!health_insurance("A0005000150"));
-        assert!(!health_insurance(""));
-    }
-
-    #[test]
-    fn rejects_non_digit_payload() {
-        assert!(!health_insurance("A00050001A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/de/icao.rs b/crates/nvisy-pattern/src/validators/de/icao.rs
deleted file mode 100644
index 9d2bf452e..000000000
--- a/crates/nvisy-pattern/src/validators/de/icao.rs
+++ /dev/null
@@ -1,37 +0,0 @@
-//! ICAO Doc 9303 check-digit helpers shared by German document
-//! validators ([`passport`], [`id_card`]).
-//!
-//! Both nPA (Personalausweis since 2010) and German
-//! Reisepass document numbers carry a 7-3-1 weighted check
-//! digit at position 9, computed over the first 8 alphanumeric
-//! characters with letters mapped `A`=10, `B`=11, …, `Z`=35.
-//!
-//! ICAO also restricts the serial-charset to letters excluding
-//! `A`, `B`, `D`, `E`, `I`, `O`, `Q`, `S`, `U` (visually
-//! ambiguous). Callers verify that themselves through the regex
-//! character class; this helper only computes the checksum.
-//!
-//! [`passport`]: super::passport
-//! [`id_card`]: super::id_card
-
-/// Compute the ICAO Doc 9303 check digit over an 8-character
-/// alphanumeric serial. Returns `None` when a character is not
-/// `0`–`9` or `A`–`Z`.
-pub(super) fn mrz_check_digit(serial: &str) -> Option<u32> {
-    if serial.len() != 8 {
-        return None;
-    }
-    let weights = [7, 3, 1];
-    let mut total: u32 = 0;
-    for (i, c) in serial.chars().enumerate() {
-        let value = if c.is_ascii_digit() {
-            c.to_digit(10).unwrap()
-        } else if c.is_ascii_uppercase() {
-            (c as u32) - ('A' as u32) + 10
-        } else {
-            return None;
-        };
-        total += value * weights[i % 3];
-    }
-    Some(total % 10)
-}
diff --git a/crates/nvisy-pattern/src/validators/de/id_card.rs b/crates/nvisy-pattern/src/validators/de/id_card.rs
deleted file mode 100644
index 63ca8ac51..000000000
--- a/crates/nvisy-pattern/src/validators/de/id_card.rs
+++ /dev/null
@@ -1,85 +0,0 @@
-//! German Personalausweis (national ID card) validator.
-//!
-//! Two formats coexist on issued cards in 2026:
-//!
-//! - **nPA** (neuer Personalausweis), issued since November 2010:
-//!   9 alphanumeric characters, ICAO Doc 9303 charset
-//!   (excludes A, B, D, E, I, O, Q, S, U) plus a trailing check
-//!   digit. The check digit uses 7-3-1 weights — same as
-//!   [`super::passport`].
-//! - **Legacy** `T`-prefix card, issued before 2010: letter `T`
-//!   followed by 8 digits. The trailing digit is part of the
-//!   serial, not a checksum. Accepted at pattern confidence
-//!   because there is no structural check to apply.
-
-use super::icao::mrz_check_digit;
-
-const FORBIDDEN_LETTERS: &str = "ABDEIOQSU";
-
-/// Return `true` when `value` is a structurally-plausible
-/// German Personalausweis number — either an nPA serial with a
-/// valid ICAO Doc 9303 check digit, or a legacy `T` + 8-digit
-/// number.
-pub fn id_card(value: &str) -> bool {
-    let trimmed = value.trim().to_ascii_uppercase();
-    if trimmed.len() != 9 {
-        return false;
-    }
-
-    // Legacy T-format: `T` followed by 8 digits. No checksum.
-    if trimmed.starts_with('T') && trimmed[1..].chars().all(|c| c.is_ascii_digit()) {
-        return true;
-    }
-
-    // nPA: ICAO 7-3-1 over first 8 chars, no forbidden letters.
-    let (serial, check) = trimmed.split_at(8);
-    let Some(check_digit) = check.chars().next().and_then(|c| c.to_digit(10)) else {
-        return false;
-    };
-    if serial.chars().any(|c| FORBIDDEN_LETTERS.contains(c)) {
-        return false;
-    }
-    mrz_check_digit(serial) == Some(check_digit)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_legacy_t_format() {
-        assert!(id_card("T22000124"));
-        // No checksum to verify, so any T+8-digit string passes.
-        assert!(id_card("T00000000"));
-    }
-
-    #[test]
-    fn rejects_legacy_with_letter_payload() {
-        assert!(!id_card("T2200012A"));
-    }
-
-    #[test]
-    fn accepts_npa_with_valid_check() {
-        // Serial `L01X00T44` — known nPA sample (legal text in
-        // PassG references).
-        assert!(id_card("L01X00T44"));
-    }
-
-    #[test]
-    fn rejects_npa_with_forbidden_letter() {
-        // `B` is in the forbidden ICAO charset.
-        assert!(!id_card("LB1X00T44"));
-    }
-
-    #[test]
-    fn rejects_npa_with_invalid_check() {
-        assert!(!id_card("L01X00T45"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!id_card("L01X00T4"));
-        assert!(!id_card("L01X00T440"));
-        assert!(!id_card(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/de/lanr.rs b/crates/nvisy-pattern/src/validators/de/lanr.rs
deleted file mode 100644
index 9faee4c43..000000000
--- a/crates/nvisy-pattern/src/validators/de/lanr.rs
+++ /dev/null
@@ -1,67 +0,0 @@
-//! German Lebenslange Arztnummer (LANR) checksum validator.
-//!
-//! 9-digit lifetime physician number. Check digit at position 7
-//! derives from positions 1–6 via the KBV Arztnummern-Richtlinie:
-//! sum of `[4, 9, 4, 9, 4, 9]`-weighted digits, then the complement
-//! to 10. Positions 8–9 carry the physician's Fachgruppe and are
-//! not part of the checksum.
-
-/// Return `true` when `value` is a 9-digit LANR with a valid
-/// position-7 check digit per KBV.
-pub fn lanr(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace())
-        .map(|c| c.to_digit(10))
-        .collect::<Option<Vec<_>>>()
-        .unwrap_or_default();
-    if digits.len() != 9 {
-        return false;
-    }
-    let total: u32 = digits
-        .iter()
-        .take(6)
-        .zip([4, 9, 4, 9, 4, 9])
-        .map(|(d, w)| d * w)
-        .sum();
-    let expected = (10 - total % 10) % 10;
-    digits[6] == expected
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn worked_example() {
-        // Per Presidio docstring: physician digits `123456` produce
-        // check digit 6, so 123456601 is a valid LANR.
-        assert!(lanr("123456601"));
-    }
-
-    #[test]
-    fn additional_test_vectors() {
-        // sum 234567 * [4,9,4,9,4,9] = 8+27+16+45+24+63 = 183; 183 mod 10 = 3; check = 7.
-        assert!(lanr("234567701"));
-        // sum 100000 * [4,9,4,9,4,9] = 4; 10-4 = 6.
-        assert!(lanr("100000601"));
-    }
-
-    #[test]
-    fn rejects_wrong_check_digit() {
-        assert!(!lanr("123456701"));
-        assert!(!lanr("123456501"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!lanr("12345670"));
-        assert!(!lanr("1234566010"));
-        assert!(!lanr(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!lanr("12345670A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/de/mod.rs b/crates/nvisy-pattern/src/validators/de/mod.rs
deleted file mode 100644
index 2d76df63a..000000000
--- a/crates/nvisy-pattern/src/validators/de/mod.rs
+++ /dev/null
@@ -1,29 +0,0 @@
-//! Germany-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"de.bsnr"`, `"de.lanr"`, `"de.passport"`,
-//! `"de.id_card"`, `"de.health_insurance"`, `"de.social_security"`,
-//! `"de.tax_id"`, `"de.vat_id"`, `"de.plz"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod bsnr;
-mod health_insurance;
-mod icao;
-mod id_card;
-mod lanr;
-mod passport;
-mod plz;
-mod social_security;
-mod tax_id;
-mod vat_id;
-
-pub use self::bsnr::bsnr;
-pub use self::health_insurance::health_insurance;
-pub use self::id_card::id_card;
-pub use self::lanr::lanr;
-pub use self::passport::passport;
-pub use self::plz::plz;
-pub use self::social_security::social_security;
-pub use self::tax_id::tax_id;
-pub use self::vat_id::vat_id;
diff --git a/crates/nvisy-pattern/src/validators/de/passport.rs b/crates/nvisy-pattern/src/validators/de/passport.rs
deleted file mode 100644
index 1b4289fd5..000000000
--- a/crates/nvisy-pattern/src/validators/de/passport.rs
+++ /dev/null
@@ -1,67 +0,0 @@
-//! German Reisepass (passport) ICAO Doc 9303 check-digit
-//! validator.
-//!
-//! 9-character serial; the trailing digit is the check digit
-//! computed via 7-3-1 weighting over the first 8 characters.
-//! Letters `A`, `B`, `D`, `E`, `I`, `O`, `Q`, `S`, `U` are
-//! visually ambiguous and never appear in ICAO travel-document
-//! serials — reject outright so a lucky checksum can't promote a
-//! non-passport string.
-
-use super::icao::mrz_check_digit;
-
-const FORBIDDEN_LETTERS: &str = "ABDEIOQSU";
-
-/// Return `true` when `value` is a 9-character German passport
-/// serial whose final digit matches the ICAO Doc 9303 check.
-pub fn passport(value: &str) -> bool {
-    let trimmed = value.trim().to_ascii_uppercase();
-    if trimmed.len() != 9 {
-        return false;
-    }
-    let (serial, check) = trimmed.split_at(8);
-    let Some(check_digit) = check.chars().next().and_then(|c| c.to_digit(10)) else {
-        return false;
-    };
-    if serial.chars().any(|c| FORBIDDEN_LETTERS.contains(c)) {
-        return false;
-    }
-    mrz_check_digit(serial) == Some(check_digit)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn rejects_forbidden_letters() {
-        // `B` is in the forbidden set.
-        assert!(!passport("B1234567X"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!passport("C1234567"));
-        assert!(!passport("C12345678X"));
-        assert!(!passport(""));
-    }
-
-    #[test]
-    fn rejects_non_digit_check() {
-        assert!(!passport("C1234567X"));
-    }
-
-    #[test]
-    fn accepts_with_valid_checksum() {
-        // Serial `C0J9H58P`: compute mrz_check_digit manually.
-        // C=12, 0=0, J=19, 9=9, H=17, 5=5, 8=8, P=25.
-        // weights [7,3,1,7,3,1,7,3].
-        // 84 + 0 + 19 + 63 + 51 + 5 + 56 + 75 = 353; 353 mod 10 = 3.
-        assert!(passport("C0J9H58P3"));
-    }
-
-    #[test]
-    fn rejects_invalid_checksum() {
-        assert!(!passport("C0J9H58P4"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/de/plz.rs b/crates/nvisy-pattern/src/validators/de/plz.rs
deleted file mode 100644
index 1c08b3c18..000000000
--- a/crates/nvisy-pattern/src/validators/de/plz.rs
+++ /dev/null
@@ -1,50 +0,0 @@
-//! German Postleitzahl (PLZ) validator.
-//!
-//! Rejects the two sentinel ranges that Deutsche Post reserves and
-//! never assigns: `01000` (Briefzentrum-Sortierung test) and
-//! `99999` (catch-all routing test).
-
-const SENTINELS: &[&str] = &["01000", "99999"];
-
-/// Return `true` when `value` is a 5-digit PLZ outside the
-/// reserved sentinel ranges.
-pub fn plz(value: &str) -> bool {
-    let trimmed = value.trim();
-    if trimmed.len() != 5 || !trimmed.chars().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    if trimmed.starts_with('0') && trimmed.chars().nth(1).is_none_or(|c| c == '0') {
-        return false;
-    }
-    !SENTINELS.contains(&trimmed)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_assigned_codes() {
-        assert!(plz("10117"));
-        assert!(plz("01067"));
-        assert!(plz("80331"));
-    }
-
-    #[test]
-    fn rejects_sentinels() {
-        assert!(!plz("01000"));
-        assert!(!plz("99999"));
-    }
-
-    #[test]
-    fn rejects_leading_zero_block() {
-        assert!(!plz("00123"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!plz("1234"));
-        assert!(!plz("123456"));
-        assert!(!plz(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/de/social_security.rs b/crates/nvisy-pattern/src/validators/de/social_security.rs
deleted file mode 100644
index 9494c1095..000000000
--- a/crates/nvisy-pattern/src/validators/de/social_security.rs
+++ /dev/null
@@ -1,100 +0,0 @@
-//! German Rentenversicherungsnummer (RVNR / SVNR) validator.
-//!
-//! 12 characters: 8 digits + letter + 3 digits. VKVV §4
-//! (Deutsche Rentenversicherung) defines the structure:
-//!
-//! - Pos 1–2: regional Bereichsnummer
-//! - Pos 3–4: birth day (`01`–`31` or `51`–`81` with the
-//!   `+50` Ergänzungsmerkmal that disambiguates duplicates)
-//! - Pos 5–6: birth month (`01`–`12`)
-//! - Pos 7–8: birth year (last two digits)
-//! - Pos 9:   first letter of birth surname (A–Z)
-//! - Pos 10–11: serial number
-//! - Pos 12:  check digit (mod 10 with cross-sum)
-//!
-//! Checksum: expand the letter at pos 9 to its 2-digit ordinal,
-//! interleave with the surrounding digits, apply weights
-//! `[2,1,2,5,7,1,2,1,2,1,2,1]`, cross-sum each product (sum of
-//! its two decimal digits), accept when total mod 10 == check.
-
-/// Return `true` when `value` is a valid 12-character RVNR per
-/// VKVV §4.
-pub fn social_security(value: &str) -> bool {
-    let trimmed = value.trim().to_ascii_uppercase();
-    let chars: Vec<char> = trimmed.chars().collect();
-    if chars.len() != 12 {
-        return false;
-    }
-    if !chars[..8].iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    if !chars[8].is_ascii_uppercase() {
-        return false;
-    }
-    if !chars[9..12].iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-
-    // Structural date sanity: day 01-31 or 51-81; month 01-12.
-    let day: u32 = chars[2..4].iter().collect::<String>().parse().unwrap_or(0);
-    let month: u32 = chars[4..6].iter().collect::<String>().parse().unwrap_or(0);
-    if !((1..=31).contains(&day) || (51..=81).contains(&day)) {
-        return false;
-    }
-    if !(1..=12).contains(&month) {
-        return false;
-    }
-
-    // Letter at pos 9 → 2-digit ordinal.
-    let letter_val = (chars[8] as u32) - ('A' as u32) + 1;
-    let mut effective: Vec<u32> = chars[..8].iter().map(|c| c.to_digit(10).unwrap()).collect();
-    effective.push(letter_val / 10);
-    effective.push(letter_val % 10);
-    for c in &chars[9..11] {
-        effective.push(c.to_digit(10).unwrap());
-    }
-
-    let weights = [2, 1, 2, 5, 7, 1, 2, 1, 2, 1, 2, 1];
-    let total: u32 = effective
-        .iter()
-        .zip(weights)
-        .map(|(d, w)| {
-            let p = d * w;
-            (p / 10) + (p % 10)
-        })
-        .sum();
-
-    let check = chars[11].to_digit(10).unwrap();
-    total % 10 == check
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!social_security("12345678A12"));
-        assert!(!social_security("12345678A1234"));
-        assert!(!social_security(""));
-    }
-
-    #[test]
-    fn rejects_invalid_day() {
-        // Day 32: impossible.
-        assert!(!social_security("12320178A123"));
-        // Day 82: in the +50 forbidden range.
-        assert!(!social_security("12820178A123"));
-    }
-
-    #[test]
-    fn rejects_invalid_month() {
-        // Month 13: impossible.
-        assert!(!social_security("12121378A123"));
-    }
-
-    #[test]
-    fn rejects_non_digit_in_serial() {
-        assert!(!social_security("12010178AAAA"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/de/tax_id.rs b/crates/nvisy-pattern/src/validators/de/tax_id.rs
deleted file mode 100644
index f97cfe992..000000000
--- a/crates/nvisy-pattern/src/validators/de/tax_id.rs
+++ /dev/null
@@ -1,87 +0,0 @@
-//! German Steueridentifikationsnummer (Steuer-IdNr) checksum
-//! validator.
-//!
-//! 11-digit lifetime tax identifier per Bundeszentralamt für
-//! Steuern (BZSt). Check digit at pos 11 derives from pos 1–10
-//! via ISO 7064 Mod 11, 10.
-//!
-//! Post-2016 BZSt rule: no digit may appear more than three
-//! times within pos 1–10. Also rules out the all-identical-digit
-//! degenerate case the pre-2016 rule forbade.
-
-/// Return `true` when `value` is a valid 11-digit German tax ID
-/// per BZSt.
-pub fn tax_id(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace())
-        .map(|c| c.to_digit(10))
-        .collect::<Option<Vec<_>>>()
-        .unwrap_or_default();
-    if digits.len() != 11 {
-        return false;
-    }
-    if digits[0] == 0 {
-        return false;
-    }
-
-    // No digit appears more than 3 times in positions 1-10.
-    let mut counts = [0u32; 10];
-    for d in &digits[..10] {
-        counts[*d as usize] += 1;
-    }
-    if counts.iter().any(|&c| c > 3) {
-        return false;
-    }
-
-    // ISO 7064 Mod 11, 10.
-    let mut product = 10u32;
-    for d in digits.iter().take(10) {
-        let mut total = (d + product) % 10;
-        if total == 0 {
-            total = 10;
-        }
-        product = (total * 2) % 11;
-    }
-    let mut check = 11 - product;
-    if check == 10 {
-        check = 0;
-    }
-    check == digits[10]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_known_test_vectors() {
-        // BZSt-published example IdNr (used in documentation).
-        assert!(tax_id("65929970489"));
-        // ELSTER demo IdNr from the official Steuer-IdNr leaflet.
-        assert!(tax_id("36574261809"));
-    }
-
-    #[test]
-    fn rejects_leading_zero() {
-        assert!(!tax_id("06592997048"));
-    }
-
-    #[test]
-    fn rejects_wrong_check_digit() {
-        assert!(!tax_id("65929970480"));
-    }
-
-    #[test]
-    fn rejects_digit_repeated_four_times() {
-        // Four `1`s in positions 1-10.
-        assert!(!tax_id("11112345601"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!tax_id("123456789"));
-        assert!(!tax_id("123456789012"));
-        assert!(!tax_id(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/de/vat_id.rs b/crates/nvisy-pattern/src/validators/de/vat_id.rs
deleted file mode 100644
index e2b0d353d..000000000
--- a/crates/nvisy-pattern/src/validators/de/vat_id.rs
+++ /dev/null
@@ -1,85 +0,0 @@
-//! German Umsatzsteuer-Identifikationsnummer (USt-IdNr) validator.
-//!
-//! 11 characters: `DE` prefix + 9 digits. The check digit at
-//! position 9 (last) derives from positions 1–8 via the
-//! community-documented ISO 7064 Mod 11, 10 heuristic. BZSt has
-//! not published the official algorithm; tools across the EU
-//! converge on this formulation, so we use it as the validator
-//! and accept the rare false-negative trade-off.
-//!
-//! Whitespace, dots, and dashes are stripped before validation,
-//! so `"DE 123 456 789"`, `"DE-123-456-789"`, and `"DE.123.456.789"`
-//! all normalize to `"DE123456789"`.
-
-/// Return `true` when `value` is a valid German USt-IdNr per the
-/// ISO 7064 Mod 11,10 heuristic.
-pub fn vat_id(value: &str) -> bool {
-    let normalized: String = value
-        .chars()
-        .filter(|c| !matches!(c, ' ' | '.' | '-' | '\t'))
-        .collect::<String>()
-        .to_ascii_uppercase();
-    if normalized.len() != 11 || !normalized.starts_with("DE") {
-        return false;
-    }
-    let digits_str = &normalized[2..];
-    if !digits_str.chars().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    let digits: Vec<u32> = digits_str
-        .chars()
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-
-    let mut product = 10u32;
-    for d in digits.iter().take(8) {
-        let mut total = (d + product) % 10;
-        if total == 0 {
-            total = 10;
-        }
-        product = (total * 2) % 11;
-    }
-    let mut check = 11 - product;
-    if check == 10 {
-        check = 0;
-    }
-    check == digits[8]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_known_vat_ids() {
-        // Public BMW Group USt-IdNr.
-        assert!(vat_id("DE129273398"));
-        // Public Siemens AG USt-IdNr.
-        assert!(vat_id("DE129273398"));
-    }
-
-    #[test]
-    fn strips_separators() {
-        assert!(vat_id("DE 129 273 398"));
-        assert!(vat_id("DE-129-273-398"));
-        assert!(vat_id("DE.129.273.398"));
-    }
-
-    #[test]
-    fn rejects_wrong_check_digit() {
-        assert!(!vat_id("DE129273390"));
-    }
-
-    #[test]
-    fn rejects_missing_de_prefix() {
-        assert!(!vat_id("FR129273398"));
-        assert!(!vat_id("129273398"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!vat_id("DE12345678"));
-        assert!(!vat_id("DE1234567890"));
-        assert!(!vat_id(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/es/cif.rs b/crates/nvisy-pattern/src/validators/es/cif.rs
deleted file mode 100644
index 5bdf14452..000000000
--- a/crates/nvisy-pattern/src/validators/es/cif.rs
+++ /dev/null
@@ -1,119 +0,0 @@
-//! Spanish CIF (Código de Identificación Fiscal) validator.
-//!
-//! 9 chars: entity-class letter + 7 digits + control char. The
-//! control char's representation (digit vs. letter) depends on
-//! the entity class — orgs starting with `P`, `Q`, `R`, `S`,
-//! `N`, `W`, or `K` always carry a letter; `A`, `B`, `E`, `H`
-//! always carry a digit; the rest accept either.
-//!
-//! Checksum:
-//! 1. Double each digit at odd 1-indexed positions (positions 1,
-//!    3, 5, 7 of the 7-digit body), sum the resulting decimal
-//!    digits.
-//! 2. Add raw digits at even 1-indexed positions (2, 4, 6).
-//! 3. Compute `c = (10 - total mod 10) mod 10`.
-//! 4. The control matches `c` (digit form) or
-//!    `"JABCDEFGHI"[c]` (letter form).
-
-const ENTITY_LETTERS: &str = "ABCDEFGHJNPQRSUVW";
-const LETTER_ONLY: &str = "PQRSNWK";
-const DIGIT_ONLY: &str = "ABEH";
-const LETTER_TABLE: &[u8; 10] = b"JABCDEFGHI";
-
-/// Return `true` when `value` is a structurally valid CIF.
-pub fn cif(value: &str) -> bool {
-    let normalized: String = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .map(|c| c.to_ascii_uppercase())
-        .collect();
-    let chars: Vec<char> = normalized.chars().collect();
-    if chars.len() != 9 {
-        return false;
-    }
-    let entity = chars[0];
-    if !ENTITY_LETTERS.contains(entity) {
-        return false;
-    }
-    if !chars[1..8].iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-
-    let mut total: u32 = 0;
-    for (idx, ch) in chars[1..8].iter().enumerate() {
-        let d = ch.to_digit(10).unwrap();
-        if idx % 2 == 0 {
-            let doubled = d * 2;
-            total += (doubled / 10) + (doubled % 10);
-        } else {
-            total += d;
-        }
-    }
-    let c = (10 - total % 10) % 10;
-    let control = chars[8];
-
-    let letter_form = LETTER_TABLE[c as usize] as char;
-    let digit_form = char::from(b'0' + c as u8);
-
-    if LETTER_ONLY.contains(entity) {
-        control == letter_form
-    } else if DIGIT_ONLY.contains(entity) {
-        control == digit_form
-    } else {
-        control == letter_form || control == digit_form
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_digit_form_a_class() {
-        // Body 1234567: doubled-odd-positions cross-sum + raw
-        // evens = 2+2+6+4+1+6+5 = 26 → c = 4. `A` is digit-only.
-        assert!(cif("A12345674"));
-    }
-
-    #[test]
-    fn accepts_letter_form_p_class() {
-        // Same body; `P` is letter-only → LETTER_TABLE[4] = 'D'.
-        assert!(cif("P1234567D"));
-    }
-
-    #[test]
-    fn accepts_either_form_for_mixed_class() {
-        // `C` accepts both digit and letter forms.
-        assert!(cif("C12345674"));
-        assert!(cif("C1234567D"));
-    }
-
-    #[test]
-    fn rejects_wrong_entity_letter() {
-        // `I` is not a valid CIF entity letter.
-        assert!(!cif("I12345674"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!cif("A12345670"));
-    }
-
-    #[test]
-    fn rejects_digit_form_for_letter_class() {
-        // `P` requires a letter control; digit form must fail.
-        assert!(!cif("P12345674"));
-    }
-
-    #[test]
-    fn rejects_letter_form_for_digit_class() {
-        // `A` requires a digit control; letter form must fail.
-        assert!(!cif("A1234567D"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!cif("A1234567"));
-        assert!(!cif(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/es/mod.rs b/crates/nvisy-pattern/src/validators/es/mod.rs
deleted file mode 100644
index bdb7acac9..000000000
--- a/crates/nvisy-pattern/src/validators/es/mod.rs
+++ /dev/null
@@ -1,14 +0,0 @@
-//! Spain-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"es.nif"`, `"es.nie"`, `"es.cif"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod cif;
-mod nie;
-mod nif;
-
-pub use self::cif::cif;
-pub use self::nie::nie;
-pub use self::nif::nif;
diff --git a/crates/nvisy-pattern/src/validators/es/nie.rs b/crates/nvisy-pattern/src/validators/es/nie.rs
deleted file mode 100644
index 45b32a786..000000000
--- a/crates/nvisy-pattern/src/validators/es/nie.rs
+++ /dev/null
@@ -1,83 +0,0 @@
-//! Spanish NIE (Número de Identidad de Extranjero) validator.
-//!
-//! Format `[XYZ]` + 7 digits + control letter. The first letter
-//! is replaced by its position in `XYZ` (`X→0`, `Y→1`, `Z→2`),
-//! then the same Mod 23 check used by [`super::nif`] applies.
-
-use super::nif::LETTERS;
-
-/// Return `true` when `value` is a valid NIE — `X`, `Y`, or `Z`
-/// prefix + 7 digits + Mod 23 control letter, with optional `-`.
-pub fn nie(value: &str) -> bool {
-    let normalized: String = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .map(|c| c.to_ascii_uppercase())
-        .collect();
-    let chars: Vec<char> = normalized.chars().collect();
-    if chars.len() != 9 {
-        return false;
-    }
-    let prefix_pos = match chars[0] {
-        'X' => 0,
-        'Y' => 1,
-        'Z' => 2,
-        _ => return false,
-    };
-    if !chars[1..8].iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    let letter = chars[8];
-    if !letter.is_ascii_uppercase() {
-        return false;
-    }
-    let body: String = std::iter::once(char::from(b'0' + prefix_pos))
-        .chain(chars[1..8].iter().copied())
-        .collect();
-    let number: u64 = body.parse().unwrap_or(0);
-    LETTERS[(number % 23) as usize] as char == letter
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_nie() {
-        // X1234567 → 01234567 mod 23 = 12 → LETTERS[12] = 'L'.
-        assert!(nie("X1234567L"));
-    }
-
-    #[test]
-    fn accepts_dash_separator() {
-        assert!(nie("X-1234567-L"));
-    }
-
-    #[test]
-    fn accepts_y_prefix() {
-        // Y1234567 → 11234567 mod 23 = 10 → LETTERS[10] = 'X'.
-        assert!(nie("Y1234567X"));
-    }
-
-    #[test]
-    fn accepts_z_prefix() {
-        // Z1234567 → 21234567 mod 23 = 1 → LETTERS[1] = 'R'.
-        assert!(nie("Z1234567R"));
-    }
-
-    #[test]
-    fn rejects_wrong_prefix() {
-        assert!(!nie("A1234567L"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!nie("X1234567A"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!nie("X123456L"));
-        assert!(!nie(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/es/nif.rs b/crates/nvisy-pattern/src/validators/es/nif.rs
deleted file mode 100644
index b48157929..000000000
--- a/crates/nvisy-pattern/src/validators/es/nif.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-//! Spanish NIF / DNI checksum validator.
-//!
-//! 8 digits + control letter computed as `LETTERS[n mod 23]` over
-//! the table `TRWAGMYFPDXBNJZSQVHLCKE` (Real Decreto 338/1990).
-//! Older DNIs may be issued with a leading `0` truncated, so 7
-//! digits + letter is also accepted; the modulo is taken over the
-//! numeric value, so leading zeros don't matter.
-
-pub(super) const LETTERS: &[u8; 23] = b"TRWAGMYFPDXBNJZSQVHLCKE";
-
-/// Return `true` when `value` is a valid NIF (DNI) — 7 or 8
-/// digits + Mod 23 control letter, with optional `-` separator.
-pub fn nif(value: &str) -> bool {
-    let normalized: String = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .collect();
-    let chars: Vec<char> = normalized.chars().collect();
-    if !matches!(chars.len(), 8 | 9) {
-        return false;
-    }
-    let (digits, letter_char) = chars.split_at(chars.len() - 1);
-    if !digits.iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    let letter = letter_char[0].to_ascii_uppercase();
-    if !letter.is_ascii_uppercase() {
-        return false;
-    }
-    let number: u64 = digits.iter().collect::<String>().parse().unwrap_or(0);
-    LETTERS[(number % 23) as usize] as char == letter
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_nif() {
-        // 12345678 mod 23 = 14 → LETTERS[14] = 'Z'.
-        assert!(nif("12345678Z"));
-    }
-
-    #[test]
-    fn accepts_dash_separator() {
-        assert!(nif("12345678-Z"));
-    }
-
-    #[test]
-    fn accepts_7_digit_nif() {
-        // 1234567 mod 23 = 19 → LETTERS[19] = 'L'.
-        assert!(nif("1234567L"));
-    }
-
-    #[test]
-    fn rejects_wrong_letter() {
-        assert!(!nif("12345678A"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!nif("123456Z"));
-        assert!(!nif("123456789Z"));
-        assert!(!nif(""));
-    }
-
-    #[test]
-    fn rejects_non_digit_body() {
-        assert!(!nif("1234567AZ"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/fi/hetu.rs b/crates/nvisy-pattern/src/validators/fi/hetu.rs
deleted file mode 100644
index f5c6b56e9..000000000
--- a/crates/nvisy-pattern/src/validators/fi/hetu.rs
+++ /dev/null
@@ -1,90 +0,0 @@
-//! Finnish Henkilötunnus (HETU) personal identity code
-//! validator.
-//!
-//! 11 characters: 6-digit date, century separator, 3-digit
-//! serial, and a control character. The century separator
-//! encodes the century of birth (1800s, 1900s, 2000s) via the
-//! sets documented by Digi- ja väestötietovirasto. The control
-//! character is the index of `int(date concatenated with serial)
-//! mod 31` into the alphabet `0123456789ABCDEFHJKLMNPRSTUVWXY`.
-
-const CONTROL_TABLE: &[u8; 31] = b"0123456789ABCDEFHJKLMNPRSTUVWXY";
-
-fn is_separator(c: char) -> bool {
-    matches!(
-        c,
-        '+' | '-' | 'A' | 'B' | 'C' | 'D' | 'E' | 'F' | 'Y' | 'X' | 'W' | 'V' | 'U'
-    )
-}
-
-/// Return `true` when `value` is a valid 11-character HETU.
-pub fn hetu(value: &str) -> bool {
-    let trimmed = value.trim().to_ascii_uppercase();
-    let chars: Vec<char> = trimmed.chars().collect();
-    if chars.len() != 11 {
-        return false;
-    }
-    if !chars[..6].iter().all(|c| c.is_ascii_digit())
-        || !is_separator(chars[6])
-        || !chars[7..10].iter().all(|c| c.is_ascii_digit())
-        || !chars[10].is_ascii_alphanumeric()
-    {
-        return false;
-    }
-
-    let day: u32 = chars[0..2].iter().collect::<String>().parse().unwrap();
-    let month: u32 = chars[2..4].iter().collect::<String>().parse().unwrap();
-    if !(1..=31).contains(&day) || !(1..=12).contains(&month) {
-        return false;
-    }
-
-    let date_serial: String = chars[..6].iter().chain(chars[7..10].iter()).collect();
-    let n: u64 = date_serial.parse().unwrap();
-    CONTROL_TABLE[(n % 31) as usize] as char == chars[10]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_hetu() {
-        // 010170-123F per CONTROL_TABLE[(010170123) mod 31].
-        assert!(hetu("010170-123F"));
-    }
-
-    #[test]
-    fn accepts_2000s_century() {
-        // Same digits with century char `A` (2000s) → recomputed.
-        // 010101A123 → date+serial 010101123 mod 31 = ?; compute
-        // and assert what falls out.
-        let n: u64 = 10_101_123;
-        let expected = CONTROL_TABLE[(n % 31) as usize] as char;
-        assert!(hetu(&format!("010101A123{expected}")));
-    }
-
-    #[test]
-    fn rejects_invalid_separator() {
-        assert!(!hetu("010170Z123F"));
-    }
-
-    #[test]
-    fn rejects_invalid_date() {
-        // Day 32.
-        assert!(!hetu("320170-123F"));
-        // Month 13.
-        assert!(!hetu("011370-123F"));
-    }
-
-    #[test]
-    fn rejects_wrong_control_character() {
-        assert!(!hetu("010170-123G"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!hetu("010170-123"));
-        assert!(!hetu("010170-123FX"));
-        assert!(!hetu(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/fi/mod.rs b/crates/nvisy-pattern/src/validators/fi/mod.rs
deleted file mode 100644
index 658c4eddb..000000000
--- a/crates/nvisy-pattern/src/validators/fi/mod.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-//! Finland-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"fi.hetu"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod hetu;
-
-pub use self::hetu::hetu;
diff --git a/crates/nvisy-pattern/src/validators/iban.rs b/crates/nvisy-pattern/src/validators/iban.rs
deleted file mode 100644
index 13241bdcc..000000000
--- a/crates/nvisy-pattern/src/validators/iban.rs
+++ /dev/null
@@ -1,67 +0,0 @@
-//! ISO 13616 IBAN validator backed by the [`iban_validate`]
-//! crate, which ships the SWIFT IBAN registry (per-country
-//! length and BBAN structure) on top of the mod-97 checksum.
-//!
-//! [`iban_validate`]: https://crates.io/crates/iban_validate
-
-use iban::Iban;
-
-/// Return `true` when `value` is a valid IBAN — both the mod-97
-/// checksum and the country-specific length/BBAN structure must
-/// match the SWIFT registry. Whitespace and dashes are stripped
-/// before validation.
-pub fn iban(value: &str) -> bool {
-    let cleaned: String = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .collect();
-    cleaned.parse::<Iban>().is_ok()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn valid_ibans() {
-        assert!(iban("GB29 NWBK 6016 1331 9268 19"));
-        assert!(iban("DE89370400440532013000"));
-        assert!(iban("FR76 3000 6000 0112 3456 7890 189"));
-    }
-
-    #[test]
-    fn invalid_check_digits() {
-        assert!(!iban("GB29 NWBK 6016 1331 9268 18"));
-        assert!(!iban("DE00370400440532013000"));
-    }
-
-    #[test]
-    fn rejects_wrong_country_length() {
-        // German IBAN must be 22 chars; trimming the last block
-        // leaves a mod-97-valid string that the registry rejects.
-        assert!(!iban("DE89370400440532013"));
-    }
-
-    #[test]
-    fn rejects_unknown_country_code() {
-        // `XX` is not in the SWIFT registry.
-        assert!(!iban("XX29NWBK60161331926819"));
-    }
-
-    #[test]
-    fn too_short() {
-        assert!(!iban("GB29"));
-        assert!(!iban(""));
-    }
-
-    #[test]
-    fn non_alphanumeric() {
-        assert!(!iban("GB29!NWBK60161331926819"));
-    }
-
-    #[test]
-    fn strips_whitespace_and_dashes() {
-        assert!(iban("GB29-NWBK-6016-1331-9268-19"));
-        assert!(iban("  GB29 NWBK 6016 1331 9268 19  "));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/in/aadhaar.rs b/crates/nvisy-pattern/src/validators/in/aadhaar.rs
deleted file mode 100644
index 2640416d3..000000000
--- a/crates/nvisy-pattern/src/validators/in/aadhaar.rs
+++ /dev/null
@@ -1,72 +0,0 @@
-//! Indian Aadhaar number validator.
-//!
-//! 12 digits issued by UIDAI. Structural rules: leading digit
-//! ≥ 2 (UIDAI reserves 0xx and 1xx); the number must not be a
-//! palindrome; Verhoeff checksum over all 12 digits.
-
-use super::super::verhoeff::verhoeff;
-
-/// Return `true` when `value` is a valid 12-digit Aadhaar.
-/// Whitespace and `-`/`:` separators are stripped before
-/// validation.
-pub fn aadhaar(value: &str) -> bool {
-    let digits: String = value.chars().filter(|c| c.is_ascii_digit()).collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-' && *c != ':')
-        .count();
-    if digits.len() != 12 || extras > 0 {
-        return false;
-    }
-    let first = digits.chars().next().unwrap().to_digit(10).unwrap();
-    if first < 2 {
-        return false;
-    }
-    let reversed: String = digits.chars().rev().collect();
-    if reversed == digits {
-        return false;
-    }
-    verhoeff(&digits)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_aadhaar() {
-        // 234123412346 — widely-quoted Verhoeff-valid test value.
-        assert!(aadhaar("234123412346"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(aadhaar("2341 2341 2346"));
-        assert!(aadhaar("2341-2341-2346"));
-        assert!(aadhaar("2341:2341:2346"));
-    }
-
-    #[test]
-    fn rejects_leading_digit_below_two() {
-        assert!(!aadhaar("134123412346"));
-        assert!(!aadhaar("034123412346"));
-    }
-
-    #[test]
-    fn rejects_palindrome() {
-        // Palindrome would be rejected even with valid checksum.
-        assert!(!aadhaar("212121212121"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!aadhaar("234123412340"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!aadhaar("23412341234"));
-        assert!(!aadhaar("2341234123466"));
-        assert!(!aadhaar(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/in/gstin.rs b/crates/nvisy-pattern/src/validators/in/gstin.rs
deleted file mode 100644
index e226c7ccd..000000000
--- a/crates/nvisy-pattern/src/validators/in/gstin.rs
+++ /dev/null
@@ -1,89 +0,0 @@
-//! Indian Goods and Services Tax Identification Number (GSTIN)
-//! validator.
-//!
-//! 15 chars: state code (01-37) + 10-char PAN + 13th char
-//! (registration sequence) + `Z` literal at position 14 + check
-//! digit at position 15. The check digit uses a base-36 weighted
-//! sum per GST Network spec.
-
-const BASE36: &[u8; 36] = b"0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ";
-
-fn base36_value(c: char) -> Option<u32> {
-    BASE36.iter().position(|&b| b == c as u8).map(|p| p as u32)
-}
-
-/// Return `true` when `value` is a valid 15-char GSTIN.
-pub fn gstin(value: &str) -> bool {
-    let normalized = value.trim().to_ascii_uppercase();
-    let chars: Vec<char> = normalized.chars().collect();
-    if chars.len() != 15 {
-        return false;
-    }
-    if !chars[..2].iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    let state: u32 = chars[..2].iter().collect::<String>().parse().unwrap();
-    if !(1..=37).contains(&state) {
-        return false;
-    }
-    if chars[13] != 'Z' {
-        return false;
-    }
-    if !chars.iter().all(|c| c.is_ascii_alphanumeric()) {
-        return false;
-    }
-
-    // GSTIN checksum: each position's base-36 value is multiplied
-    // by a factor alternating 1, 2, 1, 2, …; if the product is
-    // ≥ 36, sum its base-36 digits (i.e. quotient + remainder).
-    // The sum mod 36 gives a number `c`; check digit = (36 - c) mod 36.
-    let mut total = 0u32;
-    for (i, ch) in chars[..14].iter().enumerate() {
-        let v = base36_value(*ch).unwrap_or(0);
-        let factor = if i % 2 == 0 { 1 } else { 2 };
-        let p = v * factor;
-        total += (p / 36) + (p % 36);
-    }
-    let check_value = (36 - (total % 36)) % 36;
-    base36_value(chars[14]) == Some(check_value)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_gstin() {
-        // 27AAAPL1234C1ZE — Maharashtra state 27, PAN AAAPL1234C,
-        // reg 1, Z marker, base-36 weighted check digit `E`.
-        assert!(gstin("27AAAPL1234C1ZE"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        // 29ABCDE1234F1ZW — Karnataka state 29, check `W`.
-        assert!(gstin("29ABCDE1234F1ZW"));
-    }
-
-    #[test]
-    fn rejects_invalid_state_code() {
-        assert!(!gstin("00AAAPL1234C1ZE"));
-        assert!(!gstin("99AAAPL1234C1ZE"));
-    }
-
-    #[test]
-    fn rejects_missing_z_marker() {
-        assert!(!gstin("27AAAPL1234C1AE"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!gstin("27AAAPL1234C1Z0"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!gstin("27AAAPL1234C1Z"));
-        assert!(!gstin(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/in/mod.rs b/crates/nvisy-pattern/src/validators/in/mod.rs
deleted file mode 100644
index bc7044d94..000000000
--- a/crates/nvisy-pattern/src/validators/in/mod.rs
+++ /dev/null
@@ -1,14 +0,0 @@
-//! India-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"in.aadhaar"`, `"in.pan"`, `"in.gstin"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod aadhaar;
-mod gstin;
-mod pan;
-
-pub use self::aadhaar::aadhaar;
-pub use self::gstin::gstin;
-pub use self::pan::pan;
diff --git a/crates/nvisy-pattern/src/validators/in/pan.rs b/crates/nvisy-pattern/src/validators/in/pan.rs
deleted file mode 100644
index 2310a867a..000000000
--- a/crates/nvisy-pattern/src/validators/in/pan.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-//! Indian Permanent Account Number (PAN) validator.
-//!
-//! 10 chars in the structural format `AAAAA9999A`. The 4th
-//! character encodes the entity type per the Income Tax
-//! Department: `P` individual, `C` company, `H` HUF, `F` firm,
-//! `A` AOP, `T` trust, `B` BOI, `L` local authority, `J`
-//! artificial juridical, `G` government. No published checksum.
-
-const ENTITY_TYPES: &str = "PCHFATBLJG";
-
-/// Return `true` when `value` is a structurally valid PAN.
-pub fn pan(value: &str) -> bool {
-    let normalized = value.trim().to_ascii_uppercase();
-    let chars: Vec<char> = normalized.chars().collect();
-    if chars.len() != 10 {
-        return false;
-    }
-    if !chars[..3].iter().all(|c| c.is_ascii_uppercase()) {
-        return false;
-    }
-    if !ENTITY_TYPES.contains(chars[3]) {
-        return false;
-    }
-    if !chars[4].is_ascii_uppercase() {
-        return false;
-    }
-    if !chars[5..9].iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    chars[9].is_ascii_uppercase()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_pan() {
-        // ABCPK1234E — individual entity (`P` at position 4).
-        assert!(pan("ABCPK1234E"));
-    }
-
-    #[test]
-    fn accepts_lowercase_input() {
-        assert!(pan("abcpk1234e"));
-    }
-
-    #[test]
-    fn accepts_company_pan() {
-        // ABCCD1234E — company entity (`C` at position 4).
-        assert!(pan("ABCCD1234E"));
-    }
-
-    #[test]
-    fn rejects_invalid_entity_type() {
-        // `X` is not a valid entity type at position 4.
-        assert!(!pan("ABCXK1234E"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!pan("ABCPK1234"));
-        assert!(!pan("ABCPK1234EE"));
-        assert!(!pan(""));
-    }
-
-    #[test]
-    fn rejects_wrong_digit_section() {
-        assert!(!pan("ABCPK12A4E"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/it/fiscal_code.rs b/crates/nvisy-pattern/src/validators/it/fiscal_code.rs
deleted file mode 100644
index 880f251b6..000000000
--- a/crates/nvisy-pattern/src/validators/it/fiscal_code.rs
+++ /dev/null
@@ -1,109 +0,0 @@
-//! Italian Codice Fiscale (CF) checksum validator.
-//!
-//! 16 characters: 6 letters (surname/name initials) + 2 digits
-//! (year) + 1 letter (month) + 2 digits (day, +40 for female) +
-//! 1 letter + 3 digits (municipality code) + 1 control letter.
-//!
-//! Omocodia: when two people would receive the same CF, certain
-//! digit positions get rewritten as specific letters
-//! (`0→L, 1→M, 2→N, 3→P, 4→Q, 5→R, 6→S, 7→T, 8→U, 9→V`). The
-//! checksum operates over the original alphanumeric (the regex
-//! gate already allowed both forms).
-//!
-//! Checksum: split body (first 15 chars) into odd-indexed (1st,
-//! 3rd, …) and even-indexed characters. Odd characters get a
-//! lookup-table value, even characters get their plain
-//! letter/digit value. Total mod 26 indexes into `A-Z` for the
-//! control letter.
-
-const ODD_TABLE: [u32; 36] = [
-    // 0-9
-    1, 0, 5, 7, 9, 13, 15, 17, 19, 21, // A-Z
-    1, 0, 5, 7, 9, 13, 15, 17, 19, 21, 2, 4, 18, 20, 11, 3, 6, 8, 12, 14, 16, 10, 22, 25, 24, 23,
-];
-
-fn table_index(c: char) -> Option<usize> {
-    match c {
-        '0'..='9' => Some(c as usize - '0' as usize),
-        'A'..='Z' => Some(10 + (c as usize - 'A' as usize)),
-        _ => None,
-    }
-}
-
-fn even_value(c: char) -> u32 {
-    match c {
-        '0'..='9' => c as u32 - '0' as u32,
-        'A'..='Z' => c as u32 - 'A' as u32,
-        _ => 0,
-    }
-}
-
-/// Return `true` when `value` is a 16-character Codice Fiscale
-/// whose control letter matches the computed checksum.
-pub fn fiscal_code(value: &str) -> bool {
-    let normalized = value.trim().to_ascii_uppercase();
-    let chars: Vec<char> = normalized.chars().collect();
-    if chars.len() != 16 || !chars.iter().all(|c| c.is_ascii_alphanumeric()) {
-        return false;
-    }
-
-    let mut sum: u32 = 0;
-    for (idx, ch) in chars[..15].iter().enumerate() {
-        // CF positions are 1-indexed: even 0-index = odd
-        // CF-position, so it consults ODD_TABLE.
-        if idx % 2 == 0 {
-            let table_idx = match table_index(*ch) {
-                Some(v) => v,
-                None => return false,
-            };
-            sum += ODD_TABLE[table_idx];
-        } else {
-            sum += even_value(*ch);
-        }
-    }
-
-    let expected = char::from(b'A' + (sum % 26) as u8);
-    chars[15] == expected
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_cf() {
-        // Surname RSS, name MRA, born 1970-01-01 in Rome (H501),
-        // control letter S per Ministero delle Finanze D.M.
-        // 23/12/1976 Allegato 1.
-        assert!(fiscal_code("RSSMRA70A01H501S"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        // Surname MRT, name MTT, born 1925-04-09 in Florence
-        // (F205), control letter Z.
-        assert!(fiscal_code("MRTMTT25D09F205Z"));
-    }
-
-    #[test]
-    fn accepts_lowercase_input() {
-        assert!(fiscal_code("rssmra70a01h501s"));
-    }
-
-    #[test]
-    fn rejects_wrong_control_letter() {
-        assert!(!fiscal_code("RSSMRA70A01H501Y"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!fiscal_code("RSSMRA70A01H501"));
-        assert!(!fiscal_code("RSSMRA70A01H501XX"));
-        assert!(!fiscal_code(""));
-    }
-
-    #[test]
-    fn rejects_non_alphanumeric() {
-        assert!(!fiscal_code("RSSMRA70A01H501-"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/it/mod.rs b/crates/nvisy-pattern/src/validators/it/mod.rs
deleted file mode 100644
index 56d0a77ff..000000000
--- a/crates/nvisy-pattern/src/validators/it/mod.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-//! Italy-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"it.fiscal_code"`, `"it.vat_code"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod fiscal_code;
-mod vat_code;
-
-pub use self::fiscal_code::fiscal_code;
-pub use self::vat_code::vat_code;
diff --git a/crates/nvisy-pattern/src/validators/it/vat_code.rs b/crates/nvisy-pattern/src/validators/it/vat_code.rs
deleted file mode 100644
index f0f1a7979..000000000
--- a/crates/nvisy-pattern/src/validators/it/vat_code.rs
+++ /dev/null
@@ -1,76 +0,0 @@
-//! Italian Partita IVA (P.IVA) checksum validator.
-//!
-//! 11 digits. Algorithm: split into 5 pairs of (odd, even)
-//! positions (1-indexed); raw-sum the odd digits, then for each
-//! even digit double it and subtract 9 when the result is ≥10,
-//! sum those. Total mod 10 should equal the complement of the
-//! 11th (check) digit. The all-zero string passes the math but
-//! is reserved — reject explicitly.
-
-/// Return `true` when `value` is a valid 11-digit P.IVA.
-pub fn vat_code(value: &str) -> bool {
-    let digits: String = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-' && *c != '_')
-        .collect();
-    let chars: Vec<char> = digits.chars().collect();
-    if chars.len() != 11 || !chars.iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    if digits == "00000000000" {
-        return false;
-    }
-
-    let mut x: u32 = 0;
-    let mut y: u32 = 0;
-    for i in 0..5 {
-        x += chars[2 * i].to_digit(10).unwrap();
-        let doubled = chars[2 * i + 1].to_digit(10).unwrap() * 2;
-        y += if doubled > 9 { doubled - 9 } else { doubled };
-    }
-    let t = (x + y) % 10;
-    let c = (10 - t) % 10;
-    c == chars[10].to_digit(10).unwrap()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_vat() {
-        // Body 0015498056 → odd-sum 18, even-doubled-sum 13,
-        // total mod 10 = 1, complement = 9.
-        assert!(vat_code("00154980569"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(vat_code("00154-980-569"));
-        assert!(vat_code("00 154 980 569"));
-        assert!(vat_code("00_154_980_569"));
-    }
-
-    #[test]
-    fn rejects_all_zero_sentinel() {
-        // Passes the checksum math but is reserved.
-        assert!(!vat_code("00000000000"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!vat_code("00154980560"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!vat_code("0015498056"));
-        assert!(!vat_code("001549805688"));
-        assert!(!vat_code(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!vat_code("0015498056A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/kr/brn.rs b/crates/nvisy-pattern/src/validators/kr/brn.rs
deleted file mode 100644
index 4ae4da1c5..000000000
--- a/crates/nvisy-pattern/src/validators/kr/brn.rs
+++ /dev/null
@@ -1,68 +0,0 @@
-//! Korean Business Registration Number (BRN) validator.
-//!
-//! 10 digits formatted as `AAA-BB-CCCCC`. Checksum uses magic
-//! keys `[1, 3, 7, 1, 3, 7, 1, 3, 5]` over the first 9 digits,
-//! with the 9th digit getting an extra `(digit * 5) / 10` term
-//! added to the sum. Check digit = `(10 - sum mod 10) mod 10`.
-
-const MAGIC: [u32; 9] = [1, 3, 7, 1, 3, 7, 1, 3, 5];
-
-/// Return `true` when `value` is a valid 10-digit BRN. Hyphen
-/// separators are stripped before validation.
-pub fn brn(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 10 || extras > 0 {
-        return false;
-    }
-
-    let mut total = 0u32;
-    for i in 0..8 {
-        total += digits[i] * MAGIC[i];
-    }
-    let last = digits[8] * MAGIC[8];
-    total += (last / 10) + last;
-    let check = (10 - total % 10) % 10;
-    check == digits[9]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_brn() {
-        // 123456789 + computed check digit 1.
-        assert!(brn("1234567891"));
-    }
-
-    #[test]
-    fn accepts_with_dash_separator() {
-        assert!(brn("123-45-67891"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        // 219810378 + check digit 3.
-        assert!(brn("2198103783"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!brn("1234567890"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!brn("123456789"));
-        assert!(!brn("12345678901"));
-        assert!(!brn(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/kr/driver_license.rs b/crates/nvisy-pattern/src/validators/kr/driver_license.rs
deleted file mode 100644
index 83a0f189e..000000000
--- a/crates/nvisy-pattern/src/validators/kr/driver_license.rs
+++ /dev/null
@@ -1,56 +0,0 @@
-//! Korean driver's license number validator.
-//!
-//! 12 digits formatted as `AA-BB-CCCCCC-DD`. The first two
-//! digits are the regional issuing office code (one of the
-//! published Doro-gyotongan list). The 2-digit check digits at
-//! positions 11-12 use an undisclosed algorithm, so only
-//! structural + region-code validation is performed here.
-
-const REGION_CODES: &[&str] = &[
-    "11", "12", "13", "14", "15", "16", "17", "18", "19", "20", "21", "22", "23", "24", "25", "26",
-    "28",
-];
-
-/// Return `true` when `value` is a 12-digit Korean driver
-/// license number with a valid region code. Hyphen and space
-/// separators are stripped before validation.
-pub fn driver_license(value: &str) -> bool {
-    let digits: String = value.chars().filter(|c| c.is_ascii_digit()).collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 12 || extras > 0 {
-        return false;
-    }
-    REGION_CODES.contains(&&digits[..2])
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_license() {
-        // Region 11 (Seoul) + year 20 + serial 123456 + check 78.
-        assert!(driver_license("112012345678"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(driver_license("11-20-123456-78"));
-        assert!(driver_license("11 20 123456 78"));
-    }
-
-    #[test]
-    fn rejects_unknown_region_code() {
-        // Region 99 is not in the published list.
-        assert!(!driver_license("992012345678"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!driver_license("11201234567"));
-        assert!(!driver_license(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/kr/frn.rs b/crates/nvisy-pattern/src/validators/kr/frn.rs
deleted file mode 100644
index 749c4e486..000000000
--- a/crates/nvisy-pattern/src/validators/kr/frn.rs
+++ /dev/null
@@ -1,70 +0,0 @@
-//! Korean Foreigner Registration Number (FRN) validator.
-//!
-//! Same structure as the RRN but the gender/century digit (G)
-//! is in `[5-8]` and the checksum formula is
-//! `(13 - sum mod 11) mod 10`.
-
-use super::rrn::{WEIGHTS, structural_ok};
-
-/// Return `true` when `value` is a valid FRN.
-pub fn frn(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 13 || extras > 0 {
-        return false;
-    }
-    if !structural_ok(&digits) {
-        return false;
-    }
-    if !(5..=8).contains(&digits[6]) {
-        return false;
-    }
-    let region = digits[7] * 10 + digits[8];
-    if region > 95 {
-        return false;
-    }
-    let sum: u32 = digits[..12].iter().zip(WEIGHTS).map(|(d, w)| d * w).sum();
-    let check = (13 - (sum % 11)) % 10;
-    check == digits[12]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_frn() {
-        // 900101 + gender 5 (1900s male foreigner) + region 11 +
-        // 234 + computed check digit 4.
-        assert!(frn("9001015112344"));
-    }
-
-    #[test]
-    fn accepts_with_dash_separator() {
-        assert!(frn("900101-5112344"));
-    }
-
-    #[test]
-    fn rejects_non_foreigner_gender_digit() {
-        // Gender 1 → not FRN, expects 5-8.
-        assert!(!frn("9001011112344"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!frn("9001015112340"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!frn("900101511234"));
-        assert!(!frn(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/kr/mod.rs b/crates/nvisy-pattern/src/validators/kr/mod.rs
deleted file mode 100644
index 34c9cfcf9..000000000
--- a/crates/nvisy-pattern/src/validators/kr/mod.rs
+++ /dev/null
@@ -1,17 +0,0 @@
-//! South Korea-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"kr.rrn"`, `"kr.frn"`, `"kr.brn"`,
-//! `"kr.driver_license"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod brn;
-mod driver_license;
-mod frn;
-mod rrn;
-
-pub use self::brn::brn;
-pub use self::driver_license::driver_license;
-pub use self::frn::frn;
-pub use self::rrn::rrn;
diff --git a/crates/nvisy-pattern/src/validators/kr/rrn.rs b/crates/nvisy-pattern/src/validators/kr/rrn.rs
deleted file mode 100644
index 1c40c386c..000000000
--- a/crates/nvisy-pattern/src/validators/kr/rrn.rs
+++ /dev/null
@@ -1,86 +0,0 @@
-//! Korean Resident Registration Number (RRN) validator.
-//!
-//! 13 digits in the form `YYMMDD-GHIJKLX` per the Ministry of
-//! the Interior and Safety spec. Pre-October 2020 numbers carry
-//! a checksum X computed from the first 12 digits with the
-//! weights `[2,3,4,5,6,7,8,9,2,3,4,5]`; the checksum equals
-//! `(11 - sum mod 11) mod 10`. Post-October 2020 numbers carry
-//! a random tail and pass structural checks only.
-
-pub(super) const WEIGHTS: [u32; 12] = [2, 3, 4, 5, 6, 7, 8, 9, 2, 3, 4, 5];
-
-/// Return `true` when `value` is a valid RRN. Hyphen separators
-/// are stripped before validation.
-pub fn rrn(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 13 || extras > 0 {
-        return false;
-    }
-    if !structural_ok(&digits) {
-        return false;
-    }
-    let region = digits[7] * 10 + digits[8];
-    if region > 95 {
-        return false;
-    }
-    let sum: u32 = digits[..12].iter().zip(WEIGHTS).map(|(d, w)| d * w).sum();
-    let check = (11 - (sum % 11)) % 10;
-    check == digits[12]
-}
-
-pub(super) fn structural_ok(digits: &[u32]) -> bool {
-    let month = digits[2] * 10 + digits[3];
-    let day = digits[4] * 10 + digits[5];
-    (1..=12).contains(&month) && (1..=31).contains(&day)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_rrn() {
-        // 800101 (DOB 1980-01-01) + gender 1 + region 11 + 234 +
-        // computed check digit 3.
-        assert!(rrn("8001011112343"));
-    }
-
-    #[test]
-    fn accepts_with_dash_separator() {
-        assert!(rrn("800101-1112343"));
-    }
-
-    #[test]
-    fn rejects_invalid_month() {
-        assert!(!rrn("8013011112343"));
-    }
-
-    #[test]
-    fn rejects_invalid_day() {
-        assert!(!rrn("8001321112343"));
-    }
-
-    #[test]
-    fn rejects_invalid_region() {
-        assert!(!rrn("8001011962343"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!rrn("8001011112340"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!rrn("80010111123"));
-        assert!(!rrn(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/luhn.rs b/crates/nvisy-pattern/src/validators/luhn.rs
deleted file mode 100644
index 878728ac7..000000000
--- a/crates/nvisy-pattern/src/validators/luhn.rs
+++ /dev/null
@@ -1,92 +0,0 @@
-//! Luhn checksum validator for credit-card and similar identifier
-//! numbers.
-
-/// Return `true` if `num` passes the [Luhn algorithm] checksum.
-///
-/// Spaces and dashes are stripped before validation, so
-/// `"4539 1488 0343 6467"`, `"4539-1488-0343-6467"`, and
-/// `"4539148803436467"` are equivalent inputs.
-///
-/// Returns `false` when the input is empty or contains any
-/// character other than digits, spaces, and dashes.
-///
-/// [Luhn algorithm]: https://en.wikipedia.org/wiki/Luhn_algorithm
-pub fn luhn(num: &str) -> bool {
-    if num.is_empty() {
-        return false;
-    }
-
-    // Reject anything that isn't a digit, space, or dash.
-    if !num
-        .chars()
-        .all(|c| c.is_ascii_digit() || c == ' ' || c == '-')
-    {
-        return false;
-    }
-
-    let digits: Vec<u32> = num.chars().filter_map(|c| c.to_digit(10)).collect();
-
-    if digits.is_empty() {
-        return false;
-    }
-
-    let mut sum = 0u32;
-    let mut alternate = false;
-    for &n in digits.iter().rev() {
-        let mut d = n;
-        if alternate {
-            d *= 2;
-            if d > 9 {
-                d -= 9;
-            }
-        }
-        sum += d;
-        alternate = !alternate;
-    }
-    sum.is_multiple_of(10)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn valid_card_numbers() {
-        assert!(luhn("4539 1488 0343 6467"));
-        assert!(luhn("4539148803436467"));
-        assert!(luhn("4539-1488-0343-6467"));
-    }
-
-    #[test]
-    fn invalid_card_numbers() {
-        assert!(!luhn("4539 1488 0343 6466"));
-        assert!(!luhn("1234567890123456"));
-    }
-
-    #[test]
-    fn empty_input() {
-        assert!(!luhn(""));
-    }
-
-    #[test]
-    fn non_digit_input() {
-        assert!(!luhn("abcdef"));
-    }
-
-    #[test]
-    fn mixed_alpha_digit_rejected() {
-        assert!(!luhn("45abc39"));
-        assert!(!luhn("4539 14X8 0343 6467"));
-    }
-
-    #[test]
-    fn single_zero() {
-        assert!(luhn("0"));
-    }
-
-    #[test]
-    fn only_separators_rejected() {
-        assert!(!luhn("   "));
-        assert!(!luhn("---"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/mod.rs b/crates/nvisy-pattern/src/validators/mod.rs
deleted file mode 100644
index 9c00208d2..000000000
--- a/crates/nvisy-pattern/src/validators/mod.rs
+++ /dev/null
@@ -1,281 +0,0 @@
-//! Post-match validators for regex-detected entity values.
-//!
-//! A [`Variant`] inside a [`Regex`] rule may name a validator
-//! (e.g. `validator: Some("luhn")`); the recognizer resolves the
-//! name against a [`ValidatorRegistry`] at build time and drops
-//! matches that fail the resolved check. Use validators to weed
-//! out structurally-suspect false positives that a regex alone
-//! can't.
-//!
-//! [`ValidatorRegistry::builtin`] ships universal validators
-//! ([`luhn`], [`iban`], [`phone`], [`date`], [`btc`]) plus
-//! jurisdiction-scoped sets re-exported from [`us`] (`"us.ssn"`,
-//! `"us.aba_routing"`, `"us.npi"`, `"us.dea_number"`,
-//! `"us.postal_code"`), [`uk`]
-//! (`"uk.nhs"`, `"uk.nino"`, `"uk.driving_licence"`,
-//! `"uk.vehicle_registration"`), [`de`] (`"de.bsnr"`,
-//! `"de.lanr"`, `"de.passport"`, `"de.id_card"`,
-//! `"de.health_insurance"`, `"de.social_security"`,
-//! `"de.tax_id"`, `"de.vat_id"`, `"de.plz"`), [`es`]
-//! (`"es.nif"`, `"es.nie"`, `"es.cif"`), [`it`]
-//! (`"it.fiscal_code"`, `"it.vat_code"`), [`pl`]
-//! (`"pl.pesel"`, `"pl.nip"`, `"pl.regon"`), [`au`]
-//! (`"au.abn"`, `"au.acn"`, `"au.medicare"`, `"au.tfn"`), [`ca`]
-//! (`"ca.sin"`), [`fi`] (`"fi.hetu"`), [`se`]
-//! (`"se.personnummer"`, `"se.organisationsnummer"`), and the
-//! India module (`"in.aadhaar"`, `"in.pan"`, `"in.gstin"`).
-//! Each validator is also re-exported as a free function so
-//! consumers can compose a custom registry without taking the
-//! full set.
-//!
-//! [`Variant`]: crate::Variant
-//! [`Regex`]: crate::Regex
-
-mod btc;
-mod date;
-mod iban;
-mod luhn;
-mod phone;
-mod verhoeff;
-
-pub mod au;
-pub mod ca;
-pub mod de;
-pub mod es;
-pub mod fi;
-pub mod r#in;
-pub mod it;
-pub mod kr;
-pub mod ng;
-pub mod pl;
-pub mod se;
-pub mod sg;
-pub mod th;
-pub mod tr;
-pub mod uk;
-pub mod us;
-
-use std::borrow::Cow;
-use std::collections::HashMap;
-use std::fmt;
-use std::sync::Arc;
-
-use nvisy_core::primitive::{CountryCode, LanguageTag};
-
-pub use self::btc::btc;
-pub use self::date::date;
-pub use self::iban::iban;
-pub use self::luhn::luhn;
-pub use self::phone::phone;
-
-/// Per-call hints supplied to validators alongside the matched
-/// string.
-///
-/// Carries the caller's [`RecognizerInput`] jurisdiction and
-/// language so validators that need region-aware semantics
-/// (e.g. `phone`) can honour the caller's intent instead of
-/// guessing across a fixed fallback set. Validators that don't
-/// need either field can ignore it via `_ctx`.
-///
-/// [`RecognizerInput`]: nvisy_core::recognition::RecognizerInput
-#[derive(Debug, Clone, Default)]
-pub struct ValidationContext {
-    /// ISO 3166-1 alpha-2 jurisdiction associated with the input,
-    /// when the caller specified one.
-    pub country: Option<CountryCode>,
-    /// BCP-47 language tag associated with the input, when the
-    /// caller specified one.
-    pub language: Option<LanguageTag>,
-}
-
-/// Post-match validator returning whether a matched string is
-/// structurally valid.
-///
-/// Implemented by every `Fn(&str, &ValidationContext) -> bool +
-/// Send + Sync` via the blanket impl, so plain function pointers
-/// slot in without a wrapper type. Implement directly for types
-/// that need to carry state (e.g. a remote-lookup client).
-pub trait Validator: Send + Sync {
-    /// Return `true` to keep the match, `false` to drop it.
-    fn validate(&self, matched: &str, ctx: &ValidationContext) -> bool;
-}
-
-impl<F> Validator for F
-where
-    F: Fn(&str, &ValidationContext) -> bool + Send + Sync,
-{
-    fn validate(&self, matched: &str, ctx: &ValidationContext) -> bool {
-        self(matched, ctx)
-    }
-}
-
-/// Name → validator resolver consulted at recognizer-build time.
-///
-/// Keys are [`Cow<'static, str>`] so a `&'static str` literal stays
-/// borrowed while a runtime-built name flows through as an owned
-/// `String`.
-#[derive(Clone, Default)]
-pub struct ValidatorRegistry {
-    table: HashMap<Cow<'static, str>, Arc<dyn Validator>>,
-}
-
-impl ValidatorRegistry {
-    /// Construct an empty registry.
-    ///
-    /// Any [`Variant`] referencing a validator name will fail to
-    /// resolve at recognizer-build time.
-    ///
-    /// [`Variant`]: crate::Variant
-    #[must_use]
-    pub fn empty() -> Self {
-        Self::default()
-    }
-
-    /// Construct a registry pre-loaded with the shipped built-in
-    /// validators.
-    ///
-    /// Universal keys: `"luhn"`, `"iban"`, `"phone"`, `"date"`,
-    /// `"crypto.btc"`.
-    ///
-    /// US-scoped: `"us.ssn"`, `"us.aba_routing"`, `"us.npi"`,
-    /// `"us.dea_number"`, `"us.postal_code"`.
-    ///
-    /// UK-scoped: `"uk.nhs"`, `"uk.nino"`,
-    /// `"uk.driving_licence"`, `"uk.vehicle_registration"`.
-    ///
-    /// DE-scoped: `"de.bsnr"`, `"de.lanr"`, `"de.passport"`,
-    /// `"de.id_card"`, `"de.health_insurance"`,
-    /// `"de.social_security"`, `"de.tax_id"`, `"de.vat_id"`,
-    /// `"de.plz"`.
-    ///
-    /// ES-scoped: `"es.nif"`, `"es.nie"`, `"es.cif"`.
-    ///
-    /// IT-scoped: `"it.fiscal_code"`, `"it.vat_code"`.
-    ///
-    /// PL-scoped: `"pl.pesel"`, `"pl.nip"`, `"pl.regon"`.
-    ///
-    /// AU-scoped: `"au.abn"`, `"au.acn"`, `"au.medicare"`,
-    /// `"au.tfn"`.
-    ///
-    /// CA-scoped: `"ca.sin"`.
-    ///
-    /// FI-scoped: `"fi.hetu"`.
-    ///
-    /// SE-scoped: `"se.personnummer"`,
-    /// `"se.organisationsnummer"`.
-    ///
-    /// IN-scoped: `"in.aadhaar"`, `"in.pan"`, `"in.gstin"`.
-    ///
-    /// KR-scoped: `"kr.rrn"`, `"kr.frn"`, `"kr.brn"`,
-    /// `"kr.driver_license"`.
-    ///
-    /// SG-scoped: `"sg.nric"`, `"sg.uen"`.
-    ///
-    /// TR-scoped: `"tr.tckn"`.
-    ///
-    /// NG-scoped: `"ng.nin"`.
-    ///
-    /// TH-scoped: `"th.national_id"`.
-    #[must_use]
-    pub fn builtin() -> Self {
-        Self::empty()
-            .with_simple("luhn", luhn)
-            .with_simple("iban", iban)
-            .with("phone", phone)
-            .with_simple("date", date)
-            .with_simple("crypto.btc", btc)
-            .with_simple("us.ssn", us::ssn)
-            .with_simple("us.aba_routing", us::aba_routing)
-            .with_simple("us.npi", us::npi)
-            .with_simple("us.dea_number", us::dea_number)
-            .with_simple("us.postal_code", us::postal_code)
-            .with_simple("uk.nhs", uk::nhs)
-            .with_simple("uk.nino", uk::nino)
-            .with_simple("uk.driving_licence", uk::driving_licence)
-            .with_simple("uk.vehicle_registration", uk::vehicle_registration)
-            .with_simple("de.bsnr", de::bsnr)
-            .with_simple("de.lanr", de::lanr)
-            .with_simple("de.passport", de::passport)
-            .with_simple("de.id_card", de::id_card)
-            .with_simple("de.health_insurance", de::health_insurance)
-            .with_simple("de.social_security", de::social_security)
-            .with_simple("de.tax_id", de::tax_id)
-            .with_simple("de.vat_id", de::vat_id)
-            .with_simple("de.plz", de::plz)
-            .with_simple("es.nif", es::nif)
-            .with_simple("es.nie", es::nie)
-            .with_simple("es.cif", es::cif)
-            .with_simple("it.fiscal_code", it::fiscal_code)
-            .with_simple("it.vat_code", it::vat_code)
-            .with_simple("pl.pesel", pl::pesel)
-            .with_simple("pl.nip", pl::nip)
-            .with_simple("pl.regon", pl::regon)
-            .with_simple("au.abn", au::abn)
-            .with_simple("au.acn", au::acn)
-            .with_simple("au.medicare", au::medicare)
-            .with_simple("au.tfn", au::tfn)
-            .with_simple("ca.sin", ca::sin)
-            .with_simple("fi.hetu", fi::hetu)
-            .with_simple("se.personnummer", se::personnummer)
-            .with_simple("se.organisationsnummer", se::organisationsnummer)
-            .with_simple("in.aadhaar", r#in::aadhaar)
-            .with_simple("in.pan", r#in::pan)
-            .with_simple("in.gstin", r#in::gstin)
-            .with_simple("kr.rrn", kr::rrn)
-            .with_simple("kr.frn", kr::frn)
-            .with_simple("kr.brn", kr::brn)
-            .with_simple("kr.driver_license", kr::driver_license)
-            .with_simple("sg.nric", sg::nric)
-            .with_simple("sg.uen", sg::uen)
-            .with_simple("tr.tckn", tr::tckn)
-            .with_simple("ng.nin", ng::nin)
-            .with_simple("th.national_id", th::national_id)
-    }
-
-    /// Register a context-aware `validator` under `name`,
-    /// overwriting any previous entry with the same key.
-    ///
-    /// Override a built-in by registering under the same name
-    /// (e.g. `"phone"`).
-    #[must_use]
-    pub fn with<N, V>(mut self, name: N, validator: V) -> Self
-    where
-        N: Into<Cow<'static, str>>,
-        V: Validator + 'static,
-    {
-        self.table.insert(name.into(), Arc::new(validator));
-        self
-    }
-
-    /// Register a context-free `Fn(&str) -> bool` validator under
-    /// `name`. Convenience wrapper around [`Self::with`] for the
-    /// common case where the validator ignores
-    /// [`ValidationContext`].
-    #[must_use]
-    pub fn with_simple<N, F>(self, name: N, validator: F) -> Self
-    where
-        N: Into<Cow<'static, str>>,
-        F: Fn(&str) -> bool + Send + Sync + 'static,
-    {
-        self.with(name, move |s: &str, _: &ValidationContext| validator(s))
-    }
-
-    /// Look up a validator by name.
-    ///
-    /// Returns `None` when the name is unregistered; the
-    /// recognizer's build step surfaces that as a configuration
-    /// error.
-    #[must_use]
-    pub fn resolve(&self, name: &str) -> Option<Arc<dyn Validator>> {
-        self.table.get(name).cloned()
-    }
-}
-
-impl fmt::Debug for ValidatorRegistry {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        let names: Vec<&str> = self.table.keys().map(AsRef::as_ref).collect();
-        f.debug_struct("ValidatorRegistry")
-            .field("validators", &names)
-            .finish()
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/ng/mod.rs b/crates/nvisy-pattern/src/validators/ng/mod.rs
deleted file mode 100644
index dad8029a9..000000000
--- a/crates/nvisy-pattern/src/validators/ng/mod.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-//! Nigeria-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"ng.nin"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod nin;
-
-pub use self::nin::nin;
diff --git a/crates/nvisy-pattern/src/validators/ng/nin.rs b/crates/nvisy-pattern/src/validators/ng/nin.rs
deleted file mode 100644
index 50879c66f..000000000
--- a/crates/nvisy-pattern/src/validators/ng/nin.rs
+++ /dev/null
@@ -1,59 +0,0 @@
-//! Nigerian National Identification Number (NIN) validator.
-//!
-//! 11 digits issued by the National Identity Management
-//! Commission (NIMC). The last digit is a Verhoeff checksum
-//! over the preceding 10.
-
-use super::super::verhoeff::verhoeff;
-
-/// Return `true` when `value` is a valid 11-digit NIN.
-pub fn nin(value: &str) -> bool {
-    let digits: String = value.chars().filter(|c| c.is_ascii_digit()).collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 11 || extras > 0 {
-        return false;
-    }
-    verhoeff(&digits)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_nin() {
-        // Body 1234567890 → Verhoeff check 2.
-        assert!(nin("12345678902"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        assert!(nin("98765432102"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(nin("1234 5678 902"));
-        assert!(nin("123-456-78902"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!nin("12345678900"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!nin("1234567890"));
-        assert!(!nin("123456789033"));
-        assert!(!nin(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!nin("1234567890A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/phone.rs b/crates/nvisy-pattern/src/validators/phone.rs
deleted file mode 100644
index 19cd4bd44..000000000
--- a/crates/nvisy-pattern/src/validators/phone.rs
+++ /dev/null
@@ -1,39 +0,0 @@
-//! Region-aware phone-number validator backed by the
-//! `phonenumber` crate (Rust port of Google's libphonenumber).
-//!
-//! Two paths:
-//!
-//! 1. Inputs that parse as E.164 (carry their own `+CC` prefix)
-//!    validate directly, regardless of caller context.
-//! 2. Inputs in national format (no leading `+`) need a region
-//!    hint. When [`ValidationContext::country`] is set we use it;
-//!    otherwise we fail closed — region-less national-format
-//!    matching is genuinely ambiguous (a 13-digit run can be a
-//!    valid IL/IN phone *and* the leading 13 digits of a Visa
-//!    PAN), so without a country signal we'd rather miss a
-//!    handful of national-format numbers than mislabel card and
-//!    account numbers as phones.
-
-use std::str::FromStr;
-
-use phonenumber::country::Id;
-use phonenumber::parse;
-
-use super::ValidationContext;
-
-/// Return `true` when `value` parses as a valid phone number
-/// for the caller's jurisdiction (or as E.164 with an explicit
-/// `+CC` prefix).
-pub fn phone(value: &str, ctx: &ValidationContext) -> bool {
-    let trimmed = value.trim();
-
-    if parse(None, trimmed).map(|n| n.is_valid()).unwrap_or(false) {
-        return true;
-    }
-
-    ctx.country
-        .and_then(|c| Id::from_str(c.as_str()).ok())
-        .and_then(|region| parse(Some(region), trimmed).ok())
-        .map(|n| n.is_valid())
-        .unwrap_or(false)
-}
diff --git a/crates/nvisy-pattern/src/validators/pl/mod.rs b/crates/nvisy-pattern/src/validators/pl/mod.rs
deleted file mode 100644
index a34423de0..000000000
--- a/crates/nvisy-pattern/src/validators/pl/mod.rs
+++ /dev/null
@@ -1,14 +0,0 @@
-//! Poland-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"pl.pesel"`, `"pl.nip"`, `"pl.regon"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod nip;
-mod pesel;
-mod regon;
-
-pub use self::nip::nip;
-pub use self::pesel::pesel;
-pub use self::regon::regon;
diff --git a/crates/nvisy-pattern/src/validators/pl/nip.rs b/crates/nvisy-pattern/src/validators/pl/nip.rs
deleted file mode 100644
index c94d945d1..000000000
--- a/crates/nvisy-pattern/src/validators/pl/nip.rs
+++ /dev/null
@@ -1,65 +0,0 @@
-//! Polish NIP (Numer Identyfikacji Podatkowej) validator.
-//!
-//! 10 digits. Check digit per Ustawa z dnia 13 października 1995
-//! r. o zasadach ewidencji i identyfikacji podatników: weights
-//! `[6, 5, 7, 2, 3, 4, 5, 6, 7]` over the first 9 digits;
-//! check = `sum mod 11`. A computed value of 10 means the NIP is
-//! invalid (never assigned).
-
-/// Return `true` when `value` is a valid 10-digit NIP. Hyphen
-/// and space separators in the conventional `XXX-XXX-XX-XX` or
-/// `XXX-XX-XX-XXX` renderings are stripped before validation.
-pub fn nip(value: &str) -> bool {
-    let chars: Vec<char> = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .collect();
-    if chars.len() != 10 || !chars.iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    let weights = [6, 5, 7, 2, 3, 4, 5, 6, 7];
-    let sum: u32 = chars[..9]
-        .iter()
-        .zip(weights)
-        .map(|(c, w)| c.to_digit(10).unwrap() * w)
-        .sum();
-    let check = sum % 11;
-    if check == 10 {
-        return false;
-    }
-    check == chars[9].to_digit(10).unwrap()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_nip() {
-        // 1060000062 — Ministerstwo Finansów reference NIP.
-        assert!(nip("1060000062"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(nip("106-000-00-62"));
-        assert!(nip("106 000 00 62"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!nip("1060000060"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!nip("106000006"));
-        assert!(!nip("10600000622"));
-        assert!(!nip(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!nip("106000006A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/pl/pesel.rs b/crates/nvisy-pattern/src/validators/pl/pesel.rs
deleted file mode 100644
index 9ef957460..000000000
--- a/crates/nvisy-pattern/src/validators/pl/pesel.rs
+++ /dev/null
@@ -1,53 +0,0 @@
-//! Polish PESEL (Powszechny Elektroniczny System Ewidencji
-//! Ludności) checksum validator.
-//!
-//! 11 digits encoding birth date + sex + serial + check digit.
-//! Check digit per the Ministerstwo Cyfryzacji spec: weights
-//! `[1, 3, 7, 9, 1, 3, 7, 9, 1, 3]` over the first 10 digits;
-//! check = `(10 - sum mod 10) mod 10`.
-
-/// Return `true` when `value` is a valid 11-digit PESEL.
-pub fn pesel(value: &str) -> bool {
-    let chars: Vec<char> = value.chars().filter(|c| !c.is_ascii_whitespace()).collect();
-    if chars.len() != 11 || !chars.iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    let weights = [1, 3, 7, 9, 1, 3, 7, 9, 1, 3];
-    let sum: u32 = chars[..10]
-        .iter()
-        .zip(weights)
-        .map(|(c, w)| c.to_digit(10).unwrap() * w)
-        .sum();
-    let check = (10 - sum % 10) % 10;
-    check == chars[10].to_digit(10).unwrap()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_pesel() {
-        // 44051401359 — widely-quoted example from official PESEL
-        // documentation: born 1944-05-14, male, serial 0135,
-        // check 9.
-        assert!(pesel("44051401359"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!pesel("44051401350"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!pesel("4405140135"));
-        assert!(!pesel("440514013599"));
-        assert!(!pesel(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!pesel("4405140135A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/pl/regon.rs b/crates/nvisy-pattern/src/validators/pl/regon.rs
deleted file mode 100644
index 42bb2346c..000000000
--- a/crates/nvisy-pattern/src/validators/pl/regon.rs
+++ /dev/null
@@ -1,80 +0,0 @@
-//! Polish REGON (Rejestr Gospodarki Narodowej) validator.
-//!
-//! 9 digits (entity-level) or 14 digits (unit-level). Both use
-//! a weighted-sum mod-11 check, with the 9-digit weights
-//! `[8, 9, 2, 3, 4, 5, 6, 7]` and the 14-digit weights
-//! `[2, 4, 8, 5, 0, 9, 7, 3, 6, 1, 2, 4, 8]`. A computed value
-//! of 10 means the REGON is invalid (never assigned).
-
-const WEIGHTS_9: [u32; 8] = [8, 9, 2, 3, 4, 5, 6, 7];
-const WEIGHTS_14: [u32; 13] = [2, 4, 8, 5, 0, 9, 7, 3, 6, 1, 2, 4, 8];
-
-/// Return `true` when `value` is a valid REGON in either the
-/// 9-digit or 14-digit form. Hyphen and space separators are
-/// stripped before validation.
-pub fn regon(value: &str) -> bool {
-    let chars: Vec<char> = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .collect();
-    if !chars.iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    match chars.len() {
-        9 => valid_with(&chars, &WEIGHTS_9),
-        14 => valid_with(&chars[..9], &WEIGHTS_9) && valid_with(&chars, &WEIGHTS_14),
-        _ => false,
-    }
-}
-
-fn valid_with(chars: &[char], weights: &[u32]) -> bool {
-    let body = &chars[..weights.len()];
-    let check = chars[weights.len()];
-    let sum: u32 = body
-        .iter()
-        .zip(weights)
-        .map(|(c, w)| c.to_digit(10).unwrap() * w)
-        .sum();
-    let computed = sum % 11 % 10;
-    computed == check.to_digit(10).unwrap()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_9() {
-        // 123456785 — widely-quoted valid REGON.
-        assert!(regon("123456785"));
-    }
-
-    #[test]
-    fn accepts_canonical_14() {
-        // 12345678512347 — extends the 9-digit base with a valid
-        // 14-digit suffix check.
-        assert!(regon("12345678512347"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(regon("123-456-785"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum_9() {
-        assert!(!regon("123456789"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!regon("12345678"));
-        assert!(!regon("1234567890"));
-        assert!(!regon(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!regon("12345678A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/se/luhn.rs b/crates/nvisy-pattern/src/validators/se/luhn.rs
deleted file mode 100644
index 19d524498..000000000
--- a/crates/nvisy-pattern/src/validators/se/luhn.rs
+++ /dev/null
@@ -1,18 +0,0 @@
-//! Shared Luhn checksum for Swedish 10-digit identifiers.
-
-/// Return `true` when `digits` (exactly 10 decimal digits) pass
-/// the Luhn checksum used by both personnummer and
-/// organisationsnummer.
-pub(super) fn luhn10(digits: &[u32]) -> bool {
-    let mut sum = 0u32;
-    let check = digits[9];
-    for (i, d) in digits[..9].iter().rev().enumerate() {
-        if i.is_multiple_of(2) {
-            let m = d * 2;
-            sum += if m > 9 { m - 9 } else { m };
-        } else {
-            sum += d;
-        }
-    }
-    (sum + check).is_multiple_of(10)
-}
diff --git a/crates/nvisy-pattern/src/validators/se/mod.rs b/crates/nvisy-pattern/src/validators/se/mod.rs
deleted file mode 100644
index 67cb9f770..000000000
--- a/crates/nvisy-pattern/src/validators/se/mod.rs
+++ /dev/null
@@ -1,14 +0,0 @@
-//! Sweden-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"se.personnummer"`,
-//! `"se.organisationsnummer"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod luhn;
-mod organisationsnummer;
-mod personnummer;
-
-pub use self::organisationsnummer::organisationsnummer;
-pub use self::personnummer::personnummer;
diff --git a/crates/nvisy-pattern/src/validators/se/organisationsnummer.rs b/crates/nvisy-pattern/src/validators/se/organisationsnummer.rs
deleted file mode 100644
index 1315525fb..000000000
--- a/crates/nvisy-pattern/src/validators/se/organisationsnummer.rs
+++ /dev/null
@@ -1,62 +0,0 @@
-//! Swedish Organisationsnummer validator.
-//!
-//! 10 digits with the third digit ≥ 2 (Bolagsverket's rule that
-//! distinguishes organisationsnummer from personnummer) plus a
-//! Luhn checksum over all 10.
-
-use super::luhn::luhn10;
-
-/// Return `true` when `value` is a valid organisationsnummer.
-/// Hyphen separator is accepted; whitespace is ignored.
-pub fn organisationsnummer(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if extras > 0 || digits.len() != 10 {
-        return false;
-    }
-    if digits[2] < 2 {
-        return false;
-    }
-    luhn10(&digits)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_orgnr() {
-        // 556677-1233 — public Bolagsverket example shape.
-        assert!(organisationsnummer("5566771233"));
-        assert!(organisationsnummer("556677-1233"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        assert!(organisationsnummer("5522200004"));
-    }
-
-    #[test]
-    fn rejects_low_third_digit() {
-        // Third digit 1 → looks like a personnummer, not an org.
-        assert!(!organisationsnummer("5516771233"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!organisationsnummer("5566771230"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!organisationsnummer("556677123"));
-        assert!(!organisationsnummer(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/se/personnummer.rs b/crates/nvisy-pattern/src/validators/se/personnummer.rs
deleted file mode 100644
index 4f0320f3d..000000000
--- a/crates/nvisy-pattern/src/validators/se/personnummer.rs
+++ /dev/null
@@ -1,94 +0,0 @@
-//! Swedish Personnummer validator.
-//!
-//! Encoded as `YYMMDD-XXXX` (6-digit form) or `YYYYMMDD-XXXX`
-//! (8-digit form, post-2000 cohort). Samordningsnummer adds 60
-//! to the day field. The trailing 10 digits carry a Luhn
-//! checksum.
-
-use super::luhn::luhn10;
-
-/// Return `true` when `value` is a valid personnummer. Hyphen
-/// `-` and plus `+` separators (the latter marks a 100-year-old
-/// cohort) are accepted; whitespace is ignored.
-pub fn personnummer(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-' && *c != '+')
-        .count();
-    if extras > 0 || !matches!(digits.len(), 10 | 12) {
-        return false;
-    }
-    let pnr10: Vec<u32> = digits.into_iter().rev().take(10).rev().collect();
-
-    let month = pnr10[2] * 10 + pnr10[3];
-    let raw_day = pnr10[4] * 10 + pnr10[5];
-    let day = if raw_day >= 61 { raw_day - 60 } else { raw_day };
-    if !(1..=12).contains(&month) || !(1..=31).contains(&day) {
-        return false;
-    }
-
-    luhn10(&pnr10)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_personnummer() {
-        // 900101-1239 — 6-digit date form with valid Luhn.
-        assert!(personnummer("9001011239"));
-        assert!(personnummer("900101-1239"));
-    }
-
-    #[test]
-    fn accepts_8_digit_form() {
-        // 19900101-1239 — 12-digit form; Luhn applies to last 10.
-        assert!(personnummer("199001011239"));
-        assert!(personnummer("19900101-1239"));
-    }
-
-    #[test]
-    fn accepts_plus_separator() {
-        // `+` marks a 100-year-old cohort; structurally identical.
-        assert!(personnummer("900101+1239"));
-    }
-
-    #[test]
-    fn accepts_samordningsnummer() {
-        // Day 61 = samordningsnummer for day 1.
-        // 900161 + 1234 → compute matching Luhn.
-        let body = "900161123";
-        for last in 0..10 {
-            let s = format!("{body}{last}");
-            let digits: Vec<u32> = s.chars().map(|c| c.to_digit(10).unwrap()).collect();
-            if luhn10(&digits) {
-                assert!(personnummer(&s));
-                return;
-            }
-        }
-        panic!("no valid samordningsnummer found for body {body}");
-    }
-
-    #[test]
-    fn rejects_invalid_date() {
-        // Month 13.
-        assert!(!personnummer("9013011230"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!personnummer("9001011230"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!personnummer("900101123"));
-        assert!(!personnummer(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/sg/mod.rs b/crates/nvisy-pattern/src/validators/sg/mod.rs
deleted file mode 100644
index 40c726079..000000000
--- a/crates/nvisy-pattern/src/validators/sg/mod.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-//! Singapore-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"sg.nric"`, `"sg.uen"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod nric;
-mod uen;
-
-pub use self::nric::nric;
-pub use self::uen::uen;
diff --git a/crates/nvisy-pattern/src/validators/sg/nric.rs b/crates/nvisy-pattern/src/validators/sg/nric.rs
deleted file mode 100644
index 989f878f2..000000000
--- a/crates/nvisy-pattern/src/validators/sg/nric.rs
+++ /dev/null
@@ -1,115 +0,0 @@
-//! Singapore NRIC / FIN checksum validator.
-//!
-//! 9 characters: prefix letter + 7 digits + check letter.
-//!
-//! Prefix:
-//! - `S` — Singapore citizen / PR born before 2000
-//! - `T` — Singapore citizen / PR born 2000+
-//! - `F` — foreigner (Long-Term Pass) issued before 2000
-//! - `G` — foreigner issued 2000+
-//! - `M` — foreigner issued 2022+ (introduced after the original
-//!   NRIC spec; check digit uses a different table and offset).
-//!
-//! Algorithm: weighted sum of the 7 digits with
-//! `[2, 7, 6, 5, 4, 3, 2]`, plus an offset of 4 for `T`/`G`
-//! and 3 for `M`. The remainder mod 11 indexes into a
-//! prefix-specific letter alphabet; for `M`, the table is read
-//! at position `10 - r`.
-
-const ST_TABLE: &[u8; 11] = b"JZIHGFEDCBA";
-const FG_TABLE: &[u8; 11] = b"XWUTRQPNMLK";
-const M_TABLE: &[u8; 11] = b"KLJNPQRTUWX";
-const WEIGHTS: [u32; 7] = [2, 7, 6, 5, 4, 3, 2];
-
-/// Return `true` when `value` is a valid 9-character NRIC/FIN.
-pub fn nric(value: &str) -> bool {
-    let chars: Vec<char> = value
-        .trim()
-        .chars()
-        .map(|c| c.to_ascii_uppercase())
-        .collect();
-    if chars.len() != 9 {
-        return false;
-    }
-    let prefix = chars[0];
-    if !chars[1..8].iter().all(|c| c.is_ascii_digit()) || !chars[8].is_ascii_uppercase() {
-        return false;
-    }
-    let digits: String = chars[1..8].iter().collect();
-    let mut sum: u32 = digits
-        .chars()
-        .zip(WEIGHTS)
-        .map(|(c, w)| c.to_digit(10).unwrap() * w)
-        .sum();
-    sum += match prefix {
-        'T' | 'G' => 4,
-        'M' => 3,
-        'S' | 'F' => 0,
-        _ => return false,
-    };
-    let r = (sum % 11) as usize;
-    let (table, idx) = match prefix {
-        'S' | 'T' => (ST_TABLE, r),
-        'F' | 'G' => (FG_TABLE, r),
-        'M' => (M_TABLE, 10 - r),
-        _ => return false,
-    };
-    table[idx] as char == chars[8]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_s_prefix() {
-        assert!(nric("S2740116C"));
-    }
-
-    #[test]
-    fn accepts_t_prefix() {
-        assert!(nric("T1234567J"));
-    }
-
-    #[test]
-    fn accepts_f_prefix() {
-        assert!(nric("F1234567N"));
-    }
-
-    #[test]
-    fn accepts_g_prefix() {
-        assert!(nric("G1234567X"));
-    }
-
-    #[test]
-    fn accepts_m_prefix() {
-        assert!(nric("M1234567K"));
-    }
-
-    #[test]
-    fn accepts_lowercase_input() {
-        assert!(nric("s2740116c"));
-    }
-
-    #[test]
-    fn rejects_unknown_prefix() {
-        assert!(!nric("A2740116C"));
-    }
-
-    #[test]
-    fn rejects_wrong_check_letter() {
-        assert!(!nric("S2740116D"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!nric("S274011C"));
-        assert!(!nric("S27401166C"));
-        assert!(!nric(""));
-    }
-
-    #[test]
-    fn rejects_non_digit_body() {
-        assert!(!nric("S274A116C"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/sg/uen.rs b/crates/nvisy-pattern/src/validators/sg/uen.rs
deleted file mode 100644
index de743135a..000000000
--- a/crates/nvisy-pattern/src/validators/sg/uen.rs
+++ /dev/null
@@ -1,140 +0,0 @@
-//! Singapore Unique Entity Number (UEN) validator.
-//!
-//! Three formats issued by ACRA:
-//!
-//! - Format A: 9 chars — 8 digits + check letter.
-//! - Format B: 10 chars — 4-digit year of registration + 4
-//!   digits + check letter. The year cannot be in the future.
-//! - Format C: 10 chars — `T`/`S`/`R` prefix + 2 digits + 2-letter
-//!   entity-type code (from a fixed list) + 4 digits + check
-//!   letter.
-//!
-//! Each format uses its own weight vector and check-letter
-//! alphabet. Format C's modulo arithmetic subtracts 5 before
-//! taking `mod 11`.
-
-const FORMAT_A_WEIGHTS: [u32; 8] = [10, 4, 9, 3, 8, 2, 7, 1];
-const FORMAT_A_ALPHABET: &[u8; 11] = b"XMKECAWLJDB";
-const FORMAT_B_WEIGHTS: [u32; 9] = [10, 8, 6, 4, 9, 7, 5, 3, 1];
-const FORMAT_B_ALPHABET: &[u8; 11] = b"ZKCMDNERGWH";
-const FORMAT_C_WEIGHTS: [i64; 9] = [4, 3, 5, 3, 10, 2, 2, 5, 7];
-const FORMAT_C_ALPHABET: &str = "ABCDEFGHJKLMNPQRSTUVWX0123456789";
-
-const FORMAT_C_PREFIXES: &str = "TSR";
-const FORMAT_C_ENTITY_TYPES: &[&str] = &[
-    "LP", "LL", "FC", "PF", "RF", "MQ", "MM", "NB", "CC", "CS", "MB", "FM", "GS", "DP", "CP", "NR",
-    "CM", "CD", "MD", "HS", "VH", "CH", "MH", "CL", "XL", "CX", "HC", "RP", "TU", "TC", "FB", "FN",
-    "PA", "PB", "SS", "MC", "SM", "GA", "GB",
-];
-
-/// Return `true` when `value` is a valid UEN in any of the three
-/// ACRA-published formats.
-pub fn uen(value: &str) -> bool {
-    let normalized = value.trim().to_ascii_uppercase();
-    let chars: Vec<char> = normalized.chars().collect();
-    match chars.len() {
-        9 => validate_a(&chars),
-        10 if chars[0].is_ascii_alphabetic() => validate_c(&normalized, &chars),
-        10 => validate_b(&chars),
-        _ => false,
-    }
-}
-
-fn validate_a(chars: &[char]) -> bool {
-    if !chars[..8].iter().all(|c| c.is_ascii_digit()) || !chars[8].is_ascii_uppercase() {
-        return false;
-    }
-    let sum: u32 = chars[..8]
-        .iter()
-        .zip(FORMAT_A_WEIGHTS)
-        .map(|(c, w)| c.to_digit(10).unwrap() * w)
-        .sum();
-    FORMAT_A_ALPHABET[(sum % 11) as usize] as char == chars[8]
-}
-
-fn validate_b(chars: &[char]) -> bool {
-    if !chars[..9].iter().all(|c| c.is_ascii_digit()) || !chars[9].is_ascii_uppercase() {
-        return false;
-    }
-    let sum: u32 = chars[..9]
-        .iter()
-        .zip(FORMAT_B_WEIGHTS)
-        .map(|(c, w)| c.to_digit(10).unwrap() * w)
-        .sum();
-    FORMAT_B_ALPHABET[(sum % 11) as usize] as char == chars[9]
-}
-
-fn validate_c(normalized: &str, chars: &[char]) -> bool {
-    if !FORMAT_C_PREFIXES.contains(chars[0]) {
-        return false;
-    }
-    if !chars[1..3].iter().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    let entity_type = &normalized[3..5];
-    if !FORMAT_C_ENTITY_TYPES.contains(&entity_type) {
-        return false;
-    }
-    if !chars[5..9].iter().all(|c| c.is_ascii_digit()) || !chars[9].is_ascii_uppercase() {
-        return false;
-    }
-    let sum: i64 = chars[..9]
-        .iter()
-        .zip(FORMAT_C_WEIGHTS)
-        .map(|(c, w)| FORMAT_C_ALPHABET.find(*c).unwrap() as i64 * w)
-        .sum();
-    let idx = (sum - 5).rem_euclid(11) as usize;
-    FORMAT_C_ALPHABET.as_bytes()[idx] as char == chars[9]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_format_a() {
-        // Body 12345678 → check `M`.
-        assert!(uen("12345678M"));
-    }
-
-    #[test]
-    fn accepts_format_b() {
-        // Body 200512345 (year 2005) → check `R`.
-        assert!(uen("200512345R"));
-    }
-
-    #[test]
-    fn accepts_format_c() {
-        // T05LL1234 (limited liability partnership) → check `D`.
-        assert!(uen("T05LL1234D"));
-    }
-
-    #[test]
-    fn rejects_format_c_unknown_entity_type() {
-        // `ZZ` is not a valid entity type.
-        assert!(!uen("T05ZZ1234D"));
-    }
-
-    #[test]
-    fn rejects_format_a_wrong_checksum() {
-        assert!(!uen("12345678A"));
-    }
-
-    #[test]
-    fn rejects_format_b_wrong_checksum() {
-        assert!(!uen("200512345A"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!uen("1234567M"));
-        assert!(!uen("12345678MA"));
-        assert!(!uen(""));
-    }
-
-    #[test]
-    fn rejects_format_c_wrong_prefix() {
-        // `X` is not a valid format-C prefix.
-        assert!(!uen("X05LL1234D"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/th/mod.rs b/crates/nvisy-pattern/src/validators/th/mod.rs
deleted file mode 100644
index d5344175c..000000000
--- a/crates/nvisy-pattern/src/validators/th/mod.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-//! Thailand-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"th.national_id"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod national_id;
-
-pub use self::national_id::national_id;
diff --git a/crates/nvisy-pattern/src/validators/th/national_id.rs b/crates/nvisy-pattern/src/validators/th/national_id.rs
deleted file mode 100644
index 8235c2f50..000000000
--- a/crates/nvisy-pattern/src/validators/th/national_id.rs
+++ /dev/null
@@ -1,85 +0,0 @@
-//! Thai National Identification Number (เลขประจำตัวประชาชน)
-//! validator.
-//!
-//! 13 digits issued by the Department of Provincial
-//! Administration (กรมการปกครอง). The first digit encodes the
-//! citizenship category (1-8); 0 is reserved. The 13th digit is
-//! a weighted checksum with weights `[13, 12, 11, 10, 9, 8, 7,
-//! 6, 5, 4, 3, 2]` over the first 12 digits; check digit equals
-//! `(11 - sum mod 11) mod 10`.
-
-const WEIGHTS: [u32; 12] = [13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2];
-
-/// Return `true` when `value` is a valid 13-digit Thai NID.
-/// Whitespace and `-` separators in the rendering `X-XXXX-XXXXX-XX-X`
-/// are stripped before validation.
-pub fn national_id(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace() && *c != '-')
-        .count();
-    if digits.len() != 13 || extras > 0 {
-        return false;
-    }
-    if !(1..=8).contains(&digits[0]) {
-        return false;
-    }
-    let sum: u32 = digits[..12].iter().zip(WEIGHTS).map(|(d, w)| d * w).sum();
-    let check = (11 - (sum % 11)) % 10;
-    check == digits[12]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_nid() {
-        // Body 123456789012 → check 1.
-        assert!(national_id("1234567890121"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        assert!(national_id("1000000000009"));
-    }
-
-    #[test]
-    fn accepts_with_separators() {
-        assert!(national_id("1-2345-67890-12-1"));
-        assert!(national_id("1 2345 67890 12 1"));
-    }
-
-    #[test]
-    fn rejects_reserved_leading_zero() {
-        assert!(!national_id("0234567890121"));
-    }
-
-    #[test]
-    fn rejects_leading_nine() {
-        // Category 9 is not assigned.
-        assert!(!national_id("9234567890121"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!national_id("1234567890120"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!national_id("123456789012"));
-        assert!(!national_id("12345678901211"));
-        assert!(!national_id(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!national_id("123456789012A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/tr/mod.rs b/crates/nvisy-pattern/src/validators/tr/mod.rs
deleted file mode 100644
index fdd2e64cc..000000000
--- a/crates/nvisy-pattern/src/validators/tr/mod.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-//! Turkey-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"tr.tckn"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod tckn;
-
-pub use self::tckn::tckn;
diff --git a/crates/nvisy-pattern/src/validators/tr/tckn.rs b/crates/nvisy-pattern/src/validators/tr/tckn.rs
deleted file mode 100644
index 4b2791a38..000000000
--- a/crates/nvisy-pattern/src/validators/tr/tckn.rs
+++ /dev/null
@@ -1,86 +0,0 @@
-//! Turkish T.C. Kimlik No (TCKN) validator.
-//!
-//! 11-digit national identifier issued by Nüfus ve Vatandaşlık
-//! İşleri (NVI). The first digit cannot be zero. Two check
-//! digits:
-//!
-//! - 10th = `(sum_odd_positions * 7 - sum_even_positions) mod 10`
-//!   (over the first 9 digits, 1-indexed odd/even).
-//! - 11th = `sum_of_first_10 mod 10`.
-
-/// Return `true` when `value` is a valid 11-digit TCKN.
-pub fn tckn(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| c.is_ascii_digit())
-        .map(|c| c.to_digit(10).unwrap())
-        .collect();
-    let extras = value
-        .chars()
-        .filter(|c| !c.is_ascii_digit() && !c.is_ascii_whitespace())
-        .count();
-    if digits.len() != 11 || extras > 0 {
-        return false;
-    }
-    if digits[0] == 0 {
-        return false;
-    }
-
-    let odd_sum: i64 = (digits[0] + digits[2] + digits[4] + digits[6] + digits[8]) as i64;
-    let even_sum: i64 = (digits[1] + digits[3] + digits[5] + digits[7]) as i64;
-    let tenth = (odd_sum * 7 - even_sum).rem_euclid(10) as u32;
-    if tenth != digits[9] {
-        return false;
-    }
-
-    let eleventh: u32 = digits[..10].iter().sum::<u32>() % 10;
-    eleventh == digits[10]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_canonical_tckn() {
-        // Body 123456789 → 10th = 5, 11th = 0.
-        assert!(tckn("12345678950"));
-    }
-
-    #[test]
-    fn accepts_second_vector() {
-        assert!(tckn("10000000078"));
-    }
-
-    #[test]
-    fn accepts_third_vector() {
-        assert!(tckn("98765432150"));
-    }
-
-    #[test]
-    fn rejects_leading_zero() {
-        assert!(!tckn("02345678950"));
-    }
-
-    #[test]
-    fn rejects_wrong_tenth_digit() {
-        assert!(!tckn("12345678900"));
-    }
-
-    #[test]
-    fn rejects_wrong_eleventh_digit() {
-        assert!(!tckn("12345678955"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!tckn("1234567895"));
-        assert!(!tckn("123456789500"));
-        assert!(!tckn(""));
-    }
-
-    #[test]
-    fn rejects_non_digit() {
-        assert!(!tckn("1234567895A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/uk/driving_licence.rs b/crates/nvisy-pattern/src/validators/uk/driving_licence.rs
deleted file mode 100644
index a914a47af..000000000
--- a/crates/nvisy-pattern/src/validators/uk/driving_licence.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-//! UK Driving Licence (DVLA) structural validator.
-//!
-//! The 16-char DVLA number opens with a 5-char surname slot —
-//! letters padded on the right with `9`s when the surname is
-//! shorter than five characters. A licence whose surname slot
-//! is *all* `9`s, or that places a `9` before a letter (e.g.
-//! `9ABCD…`, `A9BCD…`), violates the padding rule and is
-//! structurally invalid.
-
-/// Return `true` when the leading 5-char surname slot of a
-/// 16-char DVLA driving licence number is structurally valid.
-///
-/// Rejects an all-`9` surname and any `9` that appears before a
-/// letter within the slot. Does not re-validate the rest of the
-/// regex-matched number — that is the regex's job.
-pub fn driving_licence(value: &str) -> bool {
-    let surname: Vec<char> = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace())
-        .take(5)
-        .collect();
-    if surname.len() != 5 {
-        return false;
-    }
-    if surname.iter().all(|c| *c == '9') {
-        return false;
-    }
-    let mut padding_started = false;
-    for c in &surname {
-        match c {
-            '9' => padding_started = true,
-            c if c.is_ascii_uppercase() => {
-                if padding_started {
-                    return false;
-                }
-            }
-            _ => return false,
-        }
-    }
-    true
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_padded_short_surname() {
-        // 4-letter surname `MORG` padged with one `9`.
-        assert!(driving_licence("MORG9753116SM9IJ"));
-        // 5-letter surname `MORGA`, no padding.
-        assert!(driving_licence("MORGA753116SM9IJ"));
-    }
-
-    #[test]
-    fn rejects_all_nine_surname() {
-        assert!(!driving_licence("99999753116SM9IJ"));
-    }
-
-    #[test]
-    fn rejects_padding_before_letter() {
-        // `9` precedes a letter in the surname slot.
-        assert!(!driving_licence("9MORG753116SM9IJ"));
-        assert!(!driving_licence("A9ORG753116SM9IJ"));
-    }
-
-    #[test]
-    fn rejects_non_alpha_padding_in_surname() {
-        assert!(!driving_licence("M0RGA753116SM9IJ"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/uk/mod.rs b/crates/nvisy-pattern/src/validators/uk/mod.rs
deleted file mode 100644
index 44f7c3529..000000000
--- a/crates/nvisy-pattern/src/validators/uk/mod.rs
+++ /dev/null
@@ -1,17 +0,0 @@
-//! UK-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"uk.nhs"`, `"uk.nino"`,
-//! `"uk.driving_licence"`, `"uk.vehicle_registration"`.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod driving_licence;
-mod nhs;
-mod nino;
-mod vehicle_registration;
-
-pub use self::driving_licence::driving_licence;
-pub use self::nhs::nhs;
-pub use self::nino::nino;
-pub use self::vehicle_registration::vehicle_registration;
diff --git a/crates/nvisy-pattern/src/validators/uk/nhs.rs b/crates/nvisy-pattern/src/validators/uk/nhs.rs
deleted file mode 100644
index da184202f..000000000
--- a/crates/nvisy-pattern/src/validators/uk/nhs.rs
+++ /dev/null
@@ -1,63 +0,0 @@
-//! UK NHS number checksum validator.
-//!
-//! See `assets/NOTICE.md` for third-party attribution.
-
-/// Return `true` if `value` is a valid 10-digit UK NHS number.
-///
-/// The NHS algorithm multiplies each of the 10 digits by descending
-/// weights `[10, 9, 8, …, 1]` and accepts the number when the sum
-/// is divisible by 11. Equivalent to checking that the last digit
-/// equals `(11 - (weighted_sum_of_first_9 % 11)) % 11`, rejecting
-/// the special case where the expected check digit would be 10.
-///
-/// Whitespace and `-` separators are stripped before validation,
-/// so `"943 476 5919"`, `"943-476-5919"`, and `"9434765919"` are
-/// all equivalent inputs.
-pub fn nhs(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .map(|c| c.to_digit(10))
-        .collect::<Option<Vec<_>>>()
-        .unwrap_or_default();
-    if digits.len() != 10 {
-        return false;
-    }
-    let total: u32 = digits.iter().zip((1..=10).rev()).map(|(d, w)| d * w).sum();
-    total.is_multiple_of(11)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn valid_known_numbers() {
-        // Test number commonly used in NHS sandboxes.
-        assert!(nhs("9434765919"));
-        // Spaces and dashes are stripped.
-        assert!(nhs("943 476 5919"));
-        assert!(nhs("943-476-5919"));
-    }
-
-    #[test]
-    fn invalid_check_digit() {
-        // Wrong final digit fails the mod-11 check.
-        assert!(!nhs("9434765918"));
-        assert!(!nhs("9434765910"));
-    }
-
-    #[test]
-    fn rejects_non_digit_payload() {
-        // Embedded letters can never become a 10-digit checksum
-        // input.
-        assert!(!nhs("ABC4765919"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!nhs("123"));
-        assert!(!nhs("12345678901"));
-        assert!(!nhs(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/uk/nino.rs b/crates/nvisy-pattern/src/validators/uk/nino.rs
deleted file mode 100644
index 4933c2b89..000000000
--- a/crates/nvisy-pattern/src/validators/uk/nino.rs
+++ /dev/null
@@ -1,71 +0,0 @@
-//! UK National Insurance Number prefix validator.
-//!
-//! Reserved-prefix exclusion lives here in the validator because
-//! Rust's `regex` crate does not support look-around.
-
-/// Return `true` when `value`'s leading two-letter prefix is not
-/// a reserved NINO prefix.
-///
-/// Reserved prefixes (case-insensitive):
-///
-/// - Whole pair: `BG`, `GB`, `NK`, `KN`, `NT`, `TN`, `ZZ`.
-/// - First letter `O` (HMRC reserved; not blocked by the regex
-///   character class, which spans `j-p`).
-///
-/// The check is structural only — it does not confirm the
-/// trailing suffix letter or any HMRC issuance state.
-pub fn nino(value: &str) -> bool {
-    let prefix: String = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace())
-        .take(2)
-        .collect();
-    if prefix.len() != 2 || !prefix.chars().all(|c| c.is_ascii_alphabetic()) {
-        return false;
-    }
-    let upper = prefix.to_ascii_uppercase();
-    if upper.starts_with('O') {
-        return false;
-    }
-    !matches!(
-        upper.as_str(),
-        "BG" | "GB" | "NK" | "KN" | "NT" | "TN" | "ZZ"
-    )
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_valid_prefix() {
-        assert!(nino("AB123456C"));
-        assert!(nino("JK 12 34 56 A"));
-    }
-
-    #[test]
-    fn rejects_reserved_prefixes() {
-        for reserved in ["BG", "GB", "NK", "KN", "NT", "TN", "ZZ"] {
-            let value = format!("{reserved}123456A");
-            assert!(!nino(&value), "{reserved} must be rejected");
-        }
-    }
-
-    #[test]
-    fn rejection_is_case_insensitive() {
-        assert!(!nino("bg123456A"));
-        assert!(!nino("Zz123456A"));
-    }
-
-    #[test]
-    fn rejects_non_alpha_prefix() {
-        assert!(!nino("12345678A"));
-        assert!(!nino(""));
-    }
-
-    #[test]
-    fn rejects_o_at_position_zero() {
-        assert!(!nino("OA123456A"));
-        assert!(!nino("oa123456A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/uk/vehicle_registration.rs b/crates/nvisy-pattern/src/validators/uk/vehicle_registration.rs
deleted file mode 100644
index 7ff755e53..000000000
--- a/crates/nvisy-pattern/src/validators/uk/vehicle_registration.rs
+++ /dev/null
@@ -1,69 +0,0 @@
-//! UK current-format Vehicle Registration Mark (VRM) age-ID
-//! validator.
-//!
-//! Current (2001+) plates encode the issuance half-year as a
-//! 2-digit "age identifier" at positions 3-4:
-//!
-//! - March issue: `02..=29` (March 2002 through March 2029)
-//! - September issue: `51..=79` (September 2001 through September
-//!   2029)
-//!
-//! The recognizer regex permits the broader range `01..=79`
-//! (cheap to express); this validator narrows it to the issued
-//! windows that the DVLA actually allocates.
-
-/// Return `true` when the 2-digit age identifier embedded in a
-/// 7-char current-format UK plate falls inside an issued range.
-///
-/// Strips whitespace and `-`, then reads characters at positions
-/// 2 and 3 of the canonicalized string.
-pub fn vehicle_registration(value: &str) -> bool {
-    let chars: Vec<char> = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .collect();
-    if chars.len() != 7 {
-        return false;
-    }
-    let age = match (chars[2].to_digit(10), chars[3].to_digit(10)) {
-        (Some(a), Some(b)) => a * 10 + b,
-        _ => return false,
-    };
-    matches!(age, 2..=29 | 51..=79)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_march_window() {
-        assert!(vehicle_registration("AB02ABC"));
-        assert!(vehicle_registration("AB29ABC"));
-        assert!(vehicle_registration("AB 15 ABC"));
-    }
-
-    #[test]
-    fn accepts_september_window() {
-        assert!(vehicle_registration("AB51ABC"));
-        assert!(vehicle_registration("AB79ABC"));
-        assert!(vehicle_registration("AB-65-ABC"));
-    }
-
-    #[test]
-    fn rejects_out_of_range() {
-        // 01 was used briefly in 2001 but is not in the modern
-        // issued range; presidio rejects it too.
-        assert!(!vehicle_registration("AB01ABC"));
-        assert!(!vehicle_registration("AB30ABC"));
-        assert!(!vehicle_registration("AB50ABC"));
-        assert!(!vehicle_registration("AB80ABC"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!vehicle_registration("AB51AB"));
-        assert!(!vehicle_registration("AB51ABCD"));
-        assert!(!vehicle_registration(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/us/aba_routing.rs b/crates/nvisy-pattern/src/validators/us/aba_routing.rs
deleted file mode 100644
index 9ea3042d4..000000000
--- a/crates/nvisy-pattern/src/validators/us/aba_routing.rs
+++ /dev/null
@@ -1,64 +0,0 @@
-//! US ABA routing number checksum validator.
-
-/// Return `true` if `value` is a valid 9-digit ABA RTN.
-///
-/// The ABA checksum sums the 9 digits with cyclic weights
-/// `[3, 7, 1]` and accepts the number when the total is
-/// divisible by 10.
-///
-/// Whitespace and `-` separators are stripped before validation,
-/// so `"121000358"`, `"1210-0035-8"`, and `"121 000 358"` are
-/// equivalent inputs.
-pub fn aba_routing(value: &str) -> bool {
-    let digits: Vec<u32> = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .map(|c| c.to_digit(10))
-        .collect::<Option<Vec<_>>>()
-        .unwrap_or_default();
-    if digits.len() != 9 {
-        return false;
-    }
-    let weights = [3, 7, 1, 3, 7, 1, 3, 7, 1];
-    let total: u32 = digits.iter().zip(weights).map(|(d, w)| d * w).sum();
-    total.is_multiple_of(10)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn valid_known_numbers() {
-        // Wells Fargo SF (verified test vector).
-        assert!(aba_routing("121000358"));
-        // JPMorgan Chase NY.
-        assert!(aba_routing("021000021"));
-        // Citibank NY.
-        assert!(aba_routing("021000089"));
-    }
-
-    #[test]
-    fn strips_separators() {
-        assert!(aba_routing("121-000-358"));
-        assert!(aba_routing("121 000 358"));
-    }
-
-    #[test]
-    fn rejects_wrong_checksum() {
-        assert!(!aba_routing("121000359"));
-        assert!(!aba_routing("000000001"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!aba_routing("12100035"));
-        assert!(!aba_routing("1210003580"));
-        assert!(!aba_routing(""));
-    }
-
-    #[test]
-    fn rejects_non_digit_payload() {
-        assert!(!aba_routing("12100035A"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/us/dea_number.rs b/crates/nvisy-pattern/src/validators/us/dea_number.rs
deleted file mode 100644
index 1e6e56e34..000000000
--- a/crates/nvisy-pattern/src/validators/us/dea_number.rs
+++ /dev/null
@@ -1,86 +0,0 @@
-//! US DEA (Drug Enforcement Administration) registration number
-//! checksum validator.
-//!
-//! See `assets/NOTICE.md` for third-party attribution.
-
-/// Return `true` if `value` is a valid DEA registration number.
-///
-/// DEA numbers are 9 characters: two letters (the registration
-/// type and the surname initial) followed by seven digits, where
-/// the last digit is a checksum.
-///
-/// The check takes the odd-position digits `d1, d3, d5` and the
-/// even-position digits `d2, d4, d6`, then verifies that
-/// `(sum(odd) + 2 * sum(even)) % 10 == d7`.
-///
-/// Whitespace and `-` separators are stripped before validation.
-pub fn dea_number(value: &str) -> bool {
-    let cleaned: String = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .collect();
-    if cleaned.len() != 9 {
-        return false;
-    }
-    let mut chars = cleaned.chars();
-    let first = chars.next().unwrap();
-    let second = chars.next().unwrap();
-    if !first.is_ascii_alphabetic() || !second.is_ascii_alphabetic() {
-        return false;
-    }
-    let digits: Vec<u32> = chars
-        .map(|c| c.to_digit(10))
-        .collect::<Option<Vec<_>>>()
-        .unwrap_or_default();
-    if digits.len() != 7 {
-        return false;
-    }
-    let sum_odd = digits[0] + digits[2] + digits[4];
-    let sum_even = digits[1] + digits[3] + digits[5];
-    let expected = (sum_odd + 2 * sum_even) % 10;
-    expected == digits[6]
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn valid_known_dea_numbers() {
-        // AB1234563: odd = 1+3+5 = 9, even = 2+4+6 = 12,
-        // (9 + 24) % 10 = 3 → matches d7 = 3.
-        assert!(dea_number("AB1234563"));
-        // BC9876562: odd = 9+7+5 = 21, even = 8+6+6 = 20,
-        // (21 + 40) % 10 = 1 → mismatch with d7 = 2. Let me pick
-        // a passing one. AF3456788: odd = 3+5+7 = 15, even =
-        // 4+6+8 = 18, (15 + 36) % 10 = 1 → mismatch d7 = 8.
-        // Easier: BB0000000 → odd = 0+0+0 = 0, even = 0+0+0 = 0,
-        // d7 = 0. Valid.
-        assert!(dea_number("BB0000000"));
-    }
-
-    #[test]
-    fn strips_separators() {
-        assert!(dea_number("AB-12-34563"));
-        assert!(dea_number("AB 12 34563"));
-    }
-
-    #[test]
-    fn rejects_wrong_check_digit() {
-        assert!(!dea_number("AB1234560"));
-        assert!(!dea_number("AB1234565"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!dea_number("AB123"));
-        assert!(!dea_number("AB12345630"));
-        assert!(!dea_number(""));
-    }
-
-    #[test]
-    fn rejects_non_letter_prefix() {
-        assert!(!dea_number("123456789"));
-        assert!(!dea_number("A21234563"));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/us/mod.rs b/crates/nvisy-pattern/src/validators/us/mod.rs
deleted file mode 100644
index 1a4d007c2..000000000
--- a/crates/nvisy-pattern/src/validators/us/mod.rs
+++ /dev/null
@@ -1,18 +0,0 @@
-//! US-specific post-match validators.
-//!
-//! Registered under the [`ValidatorRegistry::builtin`] set with
-//! dotted names — `"us.ssn"`, `"us.aba_routing"`, etc.
-//!
-//! [`ValidatorRegistry::builtin`]: super::ValidatorRegistry::builtin
-
-mod aba_routing;
-mod dea_number;
-mod npi;
-mod postal_code;
-mod ssn;
-
-pub use self::aba_routing::aba_routing;
-pub use self::dea_number::dea_number;
-pub use self::npi::npi;
-pub use self::postal_code::postal_code;
-pub use self::ssn::ssn;
diff --git a/crates/nvisy-pattern/src/validators/us/npi.rs b/crates/nvisy-pattern/src/validators/us/npi.rs
deleted file mode 100644
index 533a66085..000000000
--- a/crates/nvisy-pattern/src/validators/us/npi.rs
+++ /dev/null
@@ -1,66 +0,0 @@
-//! US National Provider Identifier (NPI) checksum validator.
-//!
-//! See `assets/NOTICE.md` for third-party attribution.
-
-use super::super::luhn::luhn;
-
-/// Return `true` if `value` is a valid 10-digit US NPI.
-///
-/// The CMS algorithm prepends the constant `"80840"` to the
-/// 10-digit identifier and runs the standard Luhn checksum on
-/// the resulting 15-digit string.
-///
-/// Whitespace and `-` separators are stripped before validation.
-pub fn npi(value: &str) -> bool {
-    let digits: String = value
-        .chars()
-        .filter(|c| !c.is_ascii_whitespace() && *c != '-')
-        .collect();
-    if digits.len() != 10 || !digits.chars().all(|c| c.is_ascii_digit()) {
-        return false;
-    }
-    // Reject all-same-digit bodies (e.g. `1111111111`); they pass
-    // Luhn but are not real provider numbers.
-    let body = &digits[..9];
-    if body.chars().all(|c| c == body.chars().next().unwrap()) {
-        return false;
-    }
-    let prefixed = format!("80840{digits}");
-    luhn(&prefixed)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn valid_known_npi_number() {
-        // Test vector validated against the CMS Luhn-on-80840 algorithm.
-        assert!(npi("1234567893"));
-    }
-
-    #[test]
-    fn strips_separators() {
-        assert!(npi("1234-567-893"));
-        assert!(npi("1234 567 893"));
-    }
-
-    #[test]
-    fn rejects_wrong_check_digit() {
-        assert!(!npi("1234567890"));
-        assert!(!npi("1234567899"));
-    }
-
-    #[test]
-    fn rejects_degenerate_all_same_digits() {
-        assert!(!npi("1111111111"));
-        assert!(!npi("2222222222"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!npi("123456789"));
-        assert!(!npi("12345678901"));
-        assert!(!npi(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/us/postal_code.rs b/crates/nvisy-pattern/src/validators/us/postal_code.rs
deleted file mode 100644
index 96a193a67..000000000
--- a/crates/nvisy-pattern/src/validators/us/postal_code.rs
+++ /dev/null
@@ -1,40 +0,0 @@
-//! US ZIP / ZIP+4 sanity validator.
-
-/// Return `true` if `value` is a plausible US ZIP code.
-///
-/// Accepts the 5-digit and 5-4 (`12345-1234`) forms; rejects the
-/// reserved all-zeros prefix (`00000`) which is not assigned by the
-/// USPS but is a frequent stand-in for "unknown".
-pub fn postal_code(value: &str) -> bool {
-    let digits: Vec<char> = value.chars().filter(char::is_ascii_digit).collect();
-    if digits.len() != 5 && digits.len() != 9 {
-        return false;
-    }
-    !digits[..5].iter().all(|c| *c == '0')
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn accepts_valid() {
-        assert!(postal_code("90210"));
-        assert!(postal_code("97477-1234"));
-        // USPS lowest assigned prefix is 00501 (Holtsville, NY).
-        assert!(postal_code("00501"));
-    }
-
-    #[test]
-    fn rejects_all_zero_prefix() {
-        assert!(!postal_code("00000"));
-        assert!(!postal_code("00000-1234"));
-    }
-
-    #[test]
-    fn rejects_wrong_length() {
-        assert!(!postal_code("1234"));
-        assert!(!postal_code("123456"));
-        assert!(!postal_code(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/us/ssn.rs b/crates/nvisy-pattern/src/validators/us/ssn.rs
deleted file mode 100644
index d31cc66ee..000000000
--- a/crates/nvisy-pattern/src/validators/us/ssn.rs
+++ /dev/null
@@ -1,84 +0,0 @@
-//! US Social Security Number format validator.
-
-/// Return `true` if `value` is a structurally valid US SSN in
-/// `AAA-GG-SSSS` format.
-///
-/// Validates the three parts as:
-///
-/// - **Area** (`AAA`): 001–899, excluding 666.
-/// - **Group** (`GG`): 01–99.
-/// - **Serial** (`SSSS`): 0001–9999.
-///
-/// This is a format check only — not a verification against SSA
-/// records.
-pub fn ssn(value: &str) -> bool {
-    let parts: Vec<&str> = value.split(['-', ' ', '.']).collect();
-    if parts.len() != 3 {
-        return false;
-    }
-    let area: u32 = match parts[0].parse() {
-        Ok(v) => v,
-        Err(_) => return false,
-    };
-    let group: u32 = match parts[1].parse() {
-        Ok(v) => v,
-        Err(_) => return false,
-    };
-    let serial: u32 = match parts[2].parse() {
-        Ok(v) => v,
-        Err(_) => return false,
-    };
-    area > 0 && area < 900 && area != 666 && group > 0 && serial > 0
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    #[test]
-    fn valid() {
-        assert!(ssn("123-45-6789"));
-        assert!(ssn("001-01-0001"));
-        assert!(ssn("899-99-9999"));
-    }
-
-    #[test]
-    fn accepts_space_and_dot_separators() {
-        assert!(ssn("123 45 6789"));
-        assert!(ssn("123.45.6789"));
-    }
-
-    #[test]
-    fn invalid_area_zero() {
-        assert!(!ssn("000-45-6789"));
-    }
-
-    #[test]
-    fn invalid_area_666() {
-        assert!(!ssn("666-45-6789"));
-    }
-
-    #[test]
-    fn invalid_area_900_plus() {
-        assert!(!ssn("900-45-6789"));
-        assert!(!ssn("999-45-6789"));
-    }
-
-    #[test]
-    fn invalid_group_zero() {
-        assert!(!ssn("123-00-6789"));
-    }
-
-    #[test]
-    fn invalid_serial_zero() {
-        assert!(!ssn("123-45-0000"));
-    }
-
-    #[test]
-    fn wrong_format() {
-        assert!(!ssn("12345-6789"));
-        assert!(!ssn("123456789"));
-        assert!(!ssn("abc-de-fghi"));
-        assert!(!ssn(""));
-    }
-}
diff --git a/crates/nvisy-pattern/src/validators/verhoeff.rs b/crates/nvisy-pattern/src/validators/verhoeff.rs
deleted file mode 100644
index b2de8dec9..000000000
--- a/crates/nvisy-pattern/src/validators/verhoeff.rs
+++ /dev/null
@@ -1,40 +0,0 @@
-//! Verhoeff checksum tables and helper.
-//!
-//! Verhoeff is a dihedral-group-based checksum that detects all
-//! single-digit and most transposition errors. Used by Indian
-//! Aadhaar and Nigerian NIN.
-
-const D: [[u32; 10]; 10] = [
-    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
-    [1, 2, 3, 4, 0, 6, 7, 8, 9, 5],
-    [2, 3, 4, 0, 1, 7, 8, 9, 5, 6],
-    [3, 4, 0, 1, 2, 8, 9, 5, 6, 7],
-    [4, 0, 1, 2, 3, 9, 5, 6, 7, 8],
-    [5, 9, 8, 7, 6, 0, 4, 3, 2, 1],
-    [6, 5, 9, 8, 7, 1, 0, 4, 3, 2],
-    [7, 6, 5, 9, 8, 2, 1, 0, 4, 3],
-    [8, 7, 6, 5, 9, 3, 2, 1, 0, 4],
-    [9, 8, 7, 6, 5, 4, 3, 2, 1, 0],
-];
-
-const P: [[u32; 10]; 8] = [
-    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
-    [1, 5, 7, 6, 2, 8, 3, 0, 9, 4],
-    [5, 8, 0, 3, 7, 9, 6, 1, 4, 2],
-    [8, 9, 1, 6, 0, 4, 3, 5, 2, 7],
-    [9, 4, 5, 3, 1, 2, 6, 8, 7, 0],
-    [4, 2, 8, 6, 5, 7, 3, 9, 0, 1],
-    [2, 7, 9, 3, 8, 0, 6, 4, 1, 5],
-    [7, 0, 4, 6, 9, 1, 3, 2, 5, 8],
-];
-
-/// Return `true` when `digits` (already validated as ASCII
-/// digits) passes the Verhoeff checksum.
-pub(super) fn verhoeff(digits: &str) -> bool {
-    let mut c = 0u32;
-    for (i, ch) in digits.chars().rev().enumerate() {
-        let d = ch.to_digit(10).unwrap() as usize;
-        c = D[c as usize][P[i % 8][d] as usize];
-    }
-    c == 0
-}
diff --git a/crates/nvisy-pattern/testdata/dictionaries/product_codes.csv b/crates/nvisy-pattern/testdata/dictionaries/product_codes.csv
deleted file mode 100644
index 15ab2d2be..000000000
--- a/crates/nvisy-pattern/testdata/dictionaries/product_codes.csv
+++ /dev/null
@@ -1,4 +0,0 @@
-WIDGET-100,Acme Standard Widget,std-widget
-WIDGET-200,Acme Premium Widget,premium-widget
-SPROCKET-50,Acme Industrial Sprocket,industrial-sprocket
-GADGET-X1,Acme Gadget X1,deluxe-gadget
diff --git a/crates/nvisy-pattern/testdata/dictionaries/product_codes.toml b/crates/nvisy-pattern/testdata/dictionaries/product_codes.toml
deleted file mode 100644
index 50c65d25c..000000000
--- a/crates/nvisy-pattern/testdata/dictionaries/product_codes.toml
+++ /dev/null
@@ -1,3 +0,0 @@
-name = "product_codes"
-label = "internal_id"
-score = 0.8
diff --git a/crates/nvisy-pattern/testdata/inputs/au/finance.txt b/crates/nvisy-pattern/testdata/inputs/au/finance.txt
deleted file mode 100644
index 548315426..000000000
--- a/crates/nvisy-pattern/testdata/inputs/au/finance.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-Company filing for the Australian Business Register:
-
-  Company name:                 Example Pty Ltd
-  Australian Business Number:   51 824 753 556
-  Australian Company Number:    123 456 780
-  Registered office:            Sydney NSW
-
-The ABN above must appear on all tax invoices issued.
diff --git a/crates/nvisy-pattern/testdata/inputs/au/health.txt b/crates/nvisy-pattern/testdata/inputs/au/health.txt
deleted file mode 100644
index 9d05266e9..000000000
--- a/crates/nvisy-pattern/testdata/inputs/au/health.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-Patient handover for Mr. Doe — refer to the Medicare card on
-file with Services Australia.
-
-Medicare number: 2228 12366 1
diff --git a/crates/nvisy-pattern/testdata/inputs/au/identity.txt b/crates/nvisy-pattern/testdata/inputs/au/identity.txt
deleted file mode 100644
index a5a907a36..000000000
--- a/crates/nvisy-pattern/testdata/inputs/au/identity.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-Client record for John Smith — see the Australian Tax File
-Number registered with the ATO.
-
-Tax File Number (TFN): 123 456 782
diff --git a/crates/nvisy-pattern/testdata/inputs/ca/contact.txt b/crates/nvisy-pattern/testdata/inputs/ca/contact.txt
deleted file mode 100644
index 6027def66..000000000
--- a/crates/nvisy-pattern/testdata/inputs/ca/contact.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Mailing address for Canada Post delivery:
-
-  150 Elgin Street
-  Ottawa, ON K1P 1A1
-  Canada
-
-Use the postal code above on the shipping label.
diff --git a/crates/nvisy-pattern/testdata/inputs/ca/identity.txt b/crates/nvisy-pattern/testdata/inputs/ca/identity.txt
deleted file mode 100644
index b016ddae6..000000000
--- a/crates/nvisy-pattern/testdata/inputs/ca/identity.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-Client record for John Doe — see Social Insurance Number on
-file with Service Canada.
-
-SIN: 123 456 782
diff --git a/crates/nvisy-pattern/testdata/inputs/contact.txt b/crates/nvisy-pattern/testdata/inputs/contact.txt
deleted file mode 100644
index fb6af992c..000000000
--- a/crates/nvisy-pattern/testdata/inputs/contact.txt
+++ /dev/null
@@ -1,10 +0,0 @@
-Hi team,
-
-Please reach out to alice.johnson@example.com if you have
-questions about the proposal. For urgent matters, call me at
-+1 (415) 555-0142 or my office line 415.555.0188. Background
-materials live at https://docs.example.com/proposal and the
-secondary mirror is http://backup.example.org/proposal-v2.
-
-Best,
-Bob
diff --git a/crates/nvisy-pattern/testdata/inputs/credentials.txt b/crates/nvisy-pattern/testdata/inputs/credentials.txt
deleted file mode 100644
index 94b8c8030..000000000
--- a/crates/nvisy-pattern/testdata/inputs/credentials.txt
+++ /dev/null
@@ -1,16 +0,0 @@
-# config.env (DO NOT COMMIT)
-#
-# Note: all credentials below are obvious-fake placeholders chosen
-# to exercise the credential patterns without tripping
-# push-protection scanners. AWS uses AWS's own documented example
-# key; Stripe uses `sk_test_` (test-mode prefix); GitHub uses an
-# EXAMPLE-suffixed token.
-
-AWS_ACCESS_KEY_ID=AKIAIOSFODNN7EXAMPLE
-GITHUB_TOKEN=ghp_EXAMPLE00000abcdefghijklmnopqrstuvwx
-STRIPE_API_KEY=sk_test_EXAMPLE0000abcdefghijklmnopqr
-api_key = "abcd1234efgh5678ijkl9012mnop"
-
------BEGIN RSA PRIVATE KEY-----
-MIIEogIBAAKCAQEAxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
------END RSA PRIVATE KEY-----
diff --git a/crates/nvisy-pattern/testdata/inputs/de/contact.txt b/crates/nvisy-pattern/testdata/inputs/de/contact.txt
deleted file mode 100644
index 1b72288fb..000000000
--- a/crates/nvisy-pattern/testdata/inputs/de/contact.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Postanschrift für den Versand der Rechnung:
-
-  Friedrichstraße 100
-  10117 Berlin
-  Deutschland
-
-Die Postleitzahl oben für das Versandlabel verwenden.
diff --git a/crates/nvisy-pattern/testdata/inputs/de/finance.txt b/crates/nvisy-pattern/testdata/inputs/de/finance.txt
deleted file mode 100644
index 7b595c63f..000000000
--- a/crates/nvisy-pattern/testdata/inputs/de/finance.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-Unternehmensdaten für die Gesellschaftsstammakte:
-
-  Firma:                   Beispiel GmbH
-  Handelsregister:          HRB 123456
-  Registergericht:          Amtsgericht Berlin-Charlottenburg
-  Sitz:                    Berlin
-
-Die Handelsregisternummer ist im Impressum aufzuführen.
diff --git a/crates/nvisy-pattern/testdata/inputs/de/health.txt b/crates/nvisy-pattern/testdata/inputs/de/health.txt
deleted file mode 100644
index d047e7f49..000000000
--- a/crates/nvisy-pattern/testdata/inputs/de/health.txt
+++ /dev/null
@@ -1,6 +0,0 @@
-Patientenakte aus der Praxis Dr. Schmidt — siehe BSNR, LANR
-und Krankenversichertennummer von der gesetzlichen Krankenkasse.
-
-Betriebsstättennummer (BSNR): 381789045
-Arztnummer (LANR):            123456601
-Krankenversichertennummer:    A000500015
diff --git a/crates/nvisy-pattern/testdata/inputs/de/identity.txt b/crates/nvisy-pattern/testdata/inputs/de/identity.txt
deleted file mode 100644
index 59e80639b..000000000
--- a/crates/nvisy-pattern/testdata/inputs/de/identity.txt
+++ /dev/null
@@ -1,9 +0,0 @@
-Personenakte für Herrn Max Mustermann — siehe Personalausweis,
-Reisepass, Steueridentifikationsnummer und USt-IdNr.
-
-Personalausweis (nPA):       L01X00T44
-Reisepassnummer:             C0J9H58P3
-Steueridentifikationsnummer: 65929970489
-Umsatzsteuer-IdNr:           DE129273398
-Führerscheinnummer:          B072RRE2I52
-Rentenversicherungsnummer:   15010685M016
diff --git a/crates/nvisy-pattern/testdata/inputs/de/vehicle.txt b/crates/nvisy-pattern/testdata/inputs/de/vehicle.txt
deleted file mode 100644
index 0d7e819e3..000000000
--- a/crates/nvisy-pattern/testdata/inputs/de/vehicle.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-Fahrzeug im Bestand:
-
-  Marke / Modell:          BMW 3er
-  Amtliches Kennzeichen:   B-AB 1234
-  Fahrzeugschein:          vorhanden
-  HU gültig:               ja
-
-Die Versicherung ist auf das oben eingetragene Kennzeichen zu buchen.
diff --git a/crates/nvisy-pattern/testdata/inputs/es/contact.txt b/crates/nvisy-pattern/testdata/inputs/es/contact.txt
deleted file mode 100644
index 1230e906e..000000000
--- a/crates/nvisy-pattern/testdata/inputs/es/contact.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Dirección postal para el envío de la factura:
-
-  Calle Gran Vía 28
-  28013 Madrid
-  España
-
-Utilizar el código postal anterior para el etiquetado.
diff --git a/crates/nvisy-pattern/testdata/inputs/es/finance.txt b/crates/nvisy-pattern/testdata/inputs/es/finance.txt
deleted file mode 100644
index 16da5b806..000000000
--- a/crates/nvisy-pattern/testdata/inputs/es/finance.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Datos fiscales de la sociedad:
-
-  Razón social:        Ejemplo SL
-  Código de identificación fiscal (CIF): A12345674
-  Domicilio fiscal:    Madrid
-
-Incluir el CIF anterior en cualquier factura emitida.
diff --git a/crates/nvisy-pattern/testdata/inputs/es/identity.txt b/crates/nvisy-pattern/testdata/inputs/es/identity.txt
deleted file mode 100644
index 6b4add85b..000000000
--- a/crates/nvisy-pattern/testdata/inputs/es/identity.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Expediente del cliente Juan García López — referenciar el DNI, el
-número de identidad de extranjero (NIE) del conviviente, y el
-pasaporte adjunto.
-
-DNI:                12345678Z
-NIE (conviviente):  X1234567L
-Pasaporte:          AAA123456
diff --git a/crates/nvisy-pattern/testdata/inputs/fi/identity.txt b/crates/nvisy-pattern/testdata/inputs/fi/identity.txt
deleted file mode 100644
index 0d8873c23..000000000
--- a/crates/nvisy-pattern/testdata/inputs/fi/identity.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-Asiakkaan tiedot — Matti Meikäläinen — katso
-henkilötunnus alta.
-
-Henkilötunnus: 010170-123F
diff --git a/crates/nvisy-pattern/testdata/inputs/finance.txt b/crates/nvisy-pattern/testdata/inputs/finance.txt
deleted file mode 100644
index 17d2d32d9..000000000
--- a/crates/nvisy-pattern/testdata/inputs/finance.txt
+++ /dev/null
@@ -1,13 +0,0 @@
-Wire transfer authorization
----------------------------
-
-Beneficiary:    Acme Industries Ltd.
-IBAN:           GB29 NWBK 6016 1331 9268 19
-SWIFT/BIC:      NWBKGB2L
-Charge card on file (backup): 4539 1488 0343 6467
-Settlement:     in US Dollar, optionally EUR.
-
-Crypto reimbursement options:
-- Bitcoin: 1A1zP1eP5QGefi2DMPTfTL5SLmv7DivfNa
-- Ethereum: 0x742d35Cc6634C0532925a3b844Bc9e7595f6E842
-- Accepted coins: Tether, USDC
diff --git a/crates/nvisy-pattern/testdata/inputs/in/finance.txt b/crates/nvisy-pattern/testdata/inputs/in/finance.txt
deleted file mode 100644
index 35b2be533..000000000
--- a/crates/nvisy-pattern/testdata/inputs/in/finance.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-GST registration for the entity:
-
-  Legal name:     Example Private Limited
-  GSTIN:          27AAAPL1234C1ZE
-  State:          Maharashtra
-  PAN:            AAAPL1234C
-
-The GSTIN above must be quoted on every tax invoice raised.
diff --git a/crates/nvisy-pattern/testdata/inputs/in/identity.txt b/crates/nvisy-pattern/testdata/inputs/in/identity.txt
deleted file mode 100644
index e85df843c..000000000
--- a/crates/nvisy-pattern/testdata/inputs/in/identity.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-Client record for Rohit Sharma — see the Aadhaar number issued
-by UIDAI, the PAN card on file with the Income Tax Department,
-and the Indian passport.
-
-Aadhaar:       2341 2341 2346
-PAN:           ABCPK1234E
-Passport:      M1234567
-Voter ID:      ABC1234567
diff --git a/crates/nvisy-pattern/testdata/inputs/in/vehicle.txt b/crates/nvisy-pattern/testdata/inputs/in/vehicle.txt
deleted file mode 100644
index bcf9bc795..000000000
--- a/crates/nvisy-pattern/testdata/inputs/in/vehicle.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Vehicle on file with the Regional Transport Office:
-
-  Make / model:                 Tata Nexon
-  Vehicle registration number:  MH12AB1234
-  RTO:                          Pune
-
-Insurance to be billed against the registration plate above.
diff --git a/crates/nvisy-pattern/testdata/inputs/internal.txt b/crates/nvisy-pattern/testdata/inputs/internal.txt
deleted file mode 100644
index 8a2e8f38a..000000000
--- a/crates/nvisy-pattern/testdata/inputs/internal.txt
+++ /dev/null
@@ -1,6 +0,0 @@
-Internal handoff note
----------------------
-
-Engineer EMP-12345 is taking over the WIDGET-200 line from EMP-67890.
-Older invoices reference it as premium-widget; the next batch ships
-as Acme Premium Widget. For legal questions contact counsel@example.com.
diff --git a/crates/nvisy-pattern/testdata/inputs/it/finance.txt b/crates/nvisy-pattern/testdata/inputs/it/finance.txt
deleted file mode 100644
index 70fe03e85..000000000
--- a/crates/nvisy-pattern/testdata/inputs/it/finance.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Dati fiscali della società:
-
-  Ragione sociale:    Esempio S.r.l.
-  Partita IVA:        00154980569
-  Sede legale:        Milano
-
-Indicare la P.IVA su tutte le fatture emesse.
diff --git a/crates/nvisy-pattern/testdata/inputs/it/identity.txt b/crates/nvisy-pattern/testdata/inputs/it/identity.txt
deleted file mode 100644
index 4aa5da78b..000000000
--- a/crates/nvisy-pattern/testdata/inputs/it/identity.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Pratica del cliente Mario Rossi — vedere il codice fiscale, la
-carta d'identità elettronica e il passaporto allegati.
-
-Codice fiscale:     RSSMRA70A01H501S
-Carta d'identità:   CA12345AB
-Passaporto:         YA1234567
-Patente di guida:   MI1234567A
diff --git a/crates/nvisy-pattern/testdata/inputs/kr/finance.txt b/crates/nvisy-pattern/testdata/inputs/kr/finance.txt
deleted file mode 100644
index 214f9ac97..000000000
--- a/crates/nvisy-pattern/testdata/inputs/kr/finance.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-사업자 정보:
-
-  법인명:           예시 주식회사
-  사업자등록번호:   123-45-67891
-  주소:             서울특별시 강남구
-
-위 사업자번호는 모든 세금계산서에 표기해야 합니다.
diff --git a/crates/nvisy-pattern/testdata/inputs/kr/identity.txt b/crates/nvisy-pattern/testdata/inputs/kr/identity.txt
deleted file mode 100644
index d66733323..000000000
--- a/crates/nvisy-pattern/testdata/inputs/kr/identity.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-고객 기록 — 김민수 — 주민등록번호와 한국 여권을
-참조하십시오.
-
-주민등록번호:   800101-1112343
-외국인등록번호: 900101-5112344
-여권번호:       M123N4567
-운전면허번호:   11-20-123456-78
diff --git a/crates/nvisy-pattern/testdata/inputs/network.txt b/crates/nvisy-pattern/testdata/inputs/network.txt
deleted file mode 100644
index 8588cb7cf..000000000
--- a/crates/nvisy-pattern/testdata/inputs/network.txt
+++ /dev/null
@@ -1,11 +0,0 @@
-Network audit excerpt
-=====================
-
-Inbound connections last hour:
-  192.168.1.42      -> web01 (HTTP)
-  10.0.0.7          -> db01  (PG, blocked)
-  203.0.113.55      -> api01 (HTTPS)
-
-IPv6 peer: 2001:0db8:85a3:0000:0000:8a2e:0370:7334
-Spoofed gateway MAC observed: 00:1A:2B:3C:4D:5E (matches a known
-test device 3C-22-FB-A1-B2-C3 reported last week).
diff --git a/crates/nvisy-pattern/testdata/inputs/ng/identity.txt b/crates/nvisy-pattern/testdata/inputs/ng/identity.txt
deleted file mode 100644
index 81fae4fde..000000000
--- a/crates/nvisy-pattern/testdata/inputs/ng/identity.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-Client record for Mr. Adekunle Okafor — see the National
-Identification Number issued by NIMC.
-
-NIN: 12345678902
diff --git a/crates/nvisy-pattern/testdata/inputs/ng/vehicle.txt b/crates/nvisy-pattern/testdata/inputs/ng/vehicle.txt
deleted file mode 100644
index 32f6d1158..000000000
--- a/crates/nvisy-pattern/testdata/inputs/ng/vehicle.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Vehicle registered with the Federal Road Safety Corps (FRSC):
-
-  Make / model:           Toyota Camry
-  Plate number:           ABC-123DE
-  LGA:                    Ikeja
-
-Insurance billed against the plate above.
diff --git a/crates/nvisy-pattern/testdata/inputs/personal.txt b/crates/nvisy-pattern/testdata/inputs/personal.txt
deleted file mode 100644
index 083e326fd..000000000
--- a/crates/nvisy-pattern/testdata/inputs/personal.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Background notes (interview prep)
----------------------------------
-
-Subject was born on 04/22/1979 and last visited the clinic on
-2024-06-15T09:30:00Z. They identify as Italian (dual citizenship,
-also Canadian), speak fluent English and Spanish, and observe
-Catholic traditions during major holidays.
diff --git a/crates/nvisy-pattern/testdata/inputs/pl/contact.txt b/crates/nvisy-pattern/testdata/inputs/pl/contact.txt
deleted file mode 100644
index a32c0184c..000000000
--- a/crates/nvisy-pattern/testdata/inputs/pl/contact.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Adres pocztowy do wysyłki faktury:
-
-  ul. Marszałkowska 100
-  00-001 Warszawa
-  Polska
-
-Użyj powyższego kodu pocztowego na etykiecie wysyłkowej.
diff --git a/crates/nvisy-pattern/testdata/inputs/pl/finance.txt b/crates/nvisy-pattern/testdata/inputs/pl/finance.txt
deleted file mode 100644
index 3a2088773..000000000
--- a/crates/nvisy-pattern/testdata/inputs/pl/finance.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-Dane podatnika spółki:
-
-  Nazwa firmy:           Przykład Sp. z o.o.
-  NIP:                   106-000-00-62
-  REGON:                 123456785
-  Siedziba:              Warszawa
-
-Numer NIP należy podawać na wszystkich fakturach.
diff --git a/crates/nvisy-pattern/testdata/inputs/pl/identity.txt b/crates/nvisy-pattern/testdata/inputs/pl/identity.txt
deleted file mode 100644
index b6dce254d..000000000
--- a/crates/nvisy-pattern/testdata/inputs/pl/identity.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-Akta klienta Jan Kowalski — sprawdź numer PESEL i numer ewidencyjny
-osoby na utrzymaniu.
-
-Numer PESEL:           44051401359
-Numer ewidencyjny:     78010112345
diff --git a/crates/nvisy-pattern/testdata/inputs/se/contact.txt b/crates/nvisy-pattern/testdata/inputs/se/contact.txt
deleted file mode 100644
index 042988215..000000000
--- a/crates/nvisy-pattern/testdata/inputs/se/contact.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Postadress för faktura:
-
-  Drottninggatan 100
-  111 60 Stockholm
-  Sverige
-
-Använd postnumret ovan på fraktetiketten.
diff --git a/crates/nvisy-pattern/testdata/inputs/se/finance.txt b/crates/nvisy-pattern/testdata/inputs/se/finance.txt
deleted file mode 100644
index 317e7bf54..000000000
--- a/crates/nvisy-pattern/testdata/inputs/se/finance.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Företagsuppgifter för bolagsregistret:
-
-  Företagsnamn:           Exempel AB
-  Organisationsnummer:    556677-1233
-  Säte:                   Stockholm
-
-Organisationsnumret ska anges på alla fakturor.
diff --git a/crates/nvisy-pattern/testdata/inputs/se/identity.txt b/crates/nvisy-pattern/testdata/inputs/se/identity.txt
deleted file mode 100644
index d44199406..000000000
--- a/crates/nvisy-pattern/testdata/inputs/se/identity.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-Klientuppgift för Anders Andersson — se personnummer enligt
-Skatteverkets registrering.
-
-Personnummer: 900101-1239
diff --git a/crates/nvisy-pattern/testdata/inputs/sg/contact.txt b/crates/nvisy-pattern/testdata/inputs/sg/contact.txt
deleted file mode 100644
index 354ee260b..000000000
--- a/crates/nvisy-pattern/testdata/inputs/sg/contact.txt
+++ /dev/null
@@ -1,6 +0,0 @@
-Mailing address for SingPost delivery:
-
-  1 Marina Bay Drive
-  Singapore 018989
-
-Use the postal code above on the shipping label.
diff --git a/crates/nvisy-pattern/testdata/inputs/sg/finance.txt b/crates/nvisy-pattern/testdata/inputs/sg/finance.txt
deleted file mode 100644
index b1c35662c..000000000
--- a/crates/nvisy-pattern/testdata/inputs/sg/finance.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Company filing with ACRA:
-
-  Entity name:                Example Pte Ltd
-  Unique Entity Number (UEN): 200512345R
-  Business address:           Singapore
-
-The UEN above must be quoted on tax invoices.
diff --git a/crates/nvisy-pattern/testdata/inputs/sg/identity.txt b/crates/nvisy-pattern/testdata/inputs/sg/identity.txt
deleted file mode 100644
index 0551fe5cd..000000000
--- a/crates/nvisy-pattern/testdata/inputs/sg/identity.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-Client record for Mr. Tan Wei Ming — see NRIC issued by ICA.
-
-NRIC: S2740116C
-FIN:  F1234567N
diff --git a/crates/nvisy-pattern/testdata/inputs/th/contact.txt b/crates/nvisy-pattern/testdata/inputs/th/contact.txt
deleted file mode 100644
index e3ce1f7c0..000000000
--- a/crates/nvisy-pattern/testdata/inputs/th/contact.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-ที่อยู่จัดส่งสำหรับไปรษณีย์ไทย:
-
-  ถนนสีลม 100
-  กรุงเทพมหานคร 10500
-  ประเทศไทย
-
-ใช้รหัสไปรษณีย์ด้านบนบนฉลากจัดส่ง
diff --git a/crates/nvisy-pattern/testdata/inputs/th/identity.txt b/crates/nvisy-pattern/testdata/inputs/th/identity.txt
deleted file mode 100644
index 1231cbc84..000000000
--- a/crates/nvisy-pattern/testdata/inputs/th/identity.txt
+++ /dev/null
@@ -1,3 +0,0 @@
-บันทึกลูกค้า — สมชาย ใจดี — ดูเลขประจำตัวประชาชนด้านล่าง
-
-เลขประจำตัวประชาชน: 1-2345-67890-12-1
diff --git a/crates/nvisy-pattern/testdata/inputs/tr/contact.txt b/crates/nvisy-pattern/testdata/inputs/tr/contact.txt
deleted file mode 100644
index ee964a14a..000000000
--- a/crates/nvisy-pattern/testdata/inputs/tr/contact.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-PTT teslimatı için posta adresi:
-
-  Atatürk Caddesi 100
-  34000 İstanbul
-  Türkiye
-
-Yukarıdaki posta kodunu kargo etiketinde kullanın.
diff --git a/crates/nvisy-pattern/testdata/inputs/tr/identity.txt b/crates/nvisy-pattern/testdata/inputs/tr/identity.txt
deleted file mode 100644
index 3b611ccc2..000000000
--- a/crates/nvisy-pattern/testdata/inputs/tr/identity.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-Müşteri kaydı — Ahmet Yılmaz — TCKN ve plaka aşağıdadır.
-
-TC Kimlik No: 12345678950
-Plaka:        34-ABC-1234
diff --git a/crates/nvisy-pattern/testdata/inputs/uk/contact.txt b/crates/nvisy-pattern/testdata/inputs/uk/contact.txt
deleted file mode 100644
index 9ce5dff97..000000000
--- a/crates/nvisy-pattern/testdata/inputs/uk/contact.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-Correspondence address for Royal Mail delivery:
-
-  10 Downing Street
-  London SW1A 2AA
-  United Kingdom
-
-Use the postcode above when filing the mailing label.
diff --git a/crates/nvisy-pattern/testdata/inputs/uk/identity.txt b/crates/nvisy-pattern/testdata/inputs/uk/identity.txt
deleted file mode 100644
index 52b36edb7..000000000
--- a/crates/nvisy-pattern/testdata/inputs/uk/identity.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-Patient handover for Mrs A. Patel — refer to the NHS patient
-file, the National Insurance reference, the DVLA driving licence,
-and her HM Passport Office passport.
-
-NHS number:             943 476 5919
-NI Number (NINO):       AB123456C
-Driving licence (DVLA): MORGA753116SM9IJ
-British passport:       AB1234567
diff --git a/crates/nvisy-pattern/testdata/inputs/uk/vehicle.txt b/crates/nvisy-pattern/testdata/inputs/uk/vehicle.txt
deleted file mode 100644
index b6d248c37..000000000
--- a/crates/nvisy-pattern/testdata/inputs/uk/vehicle.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-Vehicle on file:
-
-  Make / model:           BMW 3 Series
-  Registration plate:     AB51 ABC
-  DVLA V5C log book:      yes
-  MOT current:            yes
-
-Insurance to be billed against the registered vehicle above.
diff --git a/crates/nvisy-pattern/testdata/inputs/us/finance.txt b/crates/nvisy-pattern/testdata/inputs/us/finance.txt
deleted file mode 100644
index aa0619b92..000000000
--- a/crates/nvisy-pattern/testdata/inputs/us/finance.txt
+++ /dev/null
@@ -1,10 +0,0 @@
-ACH transfer instruction (US correspondent)
--------------------------------------------
-
-Beneficiary bank:    Wells Fargo Bank, N.A.
-Routing (ABA RTN):   121000358
-Checking account:    0123456789012
-Wire memo:           consultancy invoice 0042
-
-Use the checking account number above for the ACH debit and the
-routing number above for the bank deposit instruction.
diff --git a/crates/nvisy-pattern/testdata/inputs/us/health.txt b/crates/nvisy-pattern/testdata/inputs/us/health.txt
deleted file mode 100644
index 00ac7cb27..000000000
--- a/crates/nvisy-pattern/testdata/inputs/us/health.txt
+++ /dev/null
@@ -1,15 +0,0 @@
-Provider claim — Medicare crossover
------------------------------------
-
-Rendering provider:
-  Name:           Dr. Robert Hayes, MD
-  NPI:            1234567893
-  DEA license:    BB0000000
-  Taxonomy code:  207R00000X
-
-Beneficiary:
-  Name:           Henry Davies
-  Medicare MBI:   1EG4-TE5-MK73
-
-Claim submitted for evaluation under Medicare Part B; the DEA
-registration above authorizes prescribing of controlled substances.
diff --git a/crates/nvisy-pattern/testdata/inputs/us/identity.txt b/crates/nvisy-pattern/testdata/inputs/us/identity.txt
deleted file mode 100644
index ea9e0c77b..000000000
--- a/crates/nvisy-pattern/testdata/inputs/us/identity.txt
+++ /dev/null
@@ -1,12 +0,0 @@
-PATIENT INTAKE FORM
-
-Full name:        Jane Smith
-Date of birth:    1985-03-14
-Social security:  123-45-6789
-ITIN (taxpayer):  912-71-1234
-Driver license:   D123-4567-8901
-US passport:      A12345678
-Mailing address:  742 Evergreen Terrace
-                  Springfield, OR 97477-1234
-
-Confirm insurance card on file (see attached).
diff --git a/crates/nvisy-pattern/testdata/patterns/employee_id.toml b/crates/nvisy-pattern/testdata/patterns/employee_id.toml
deleted file mode 100644
index b959d33ee..000000000
--- a/crates/nvisy-pattern/testdata/patterns/employee_id.toml
+++ /dev/null
@@ -1,6 +0,0 @@
-name = "internal-employee-id"
-label = "internal_id"
-
-[[variants]]
-regex = "\\bEMP-\\d{5}\\b"
-score = 0.95
diff --git a/crates/nvisy-pattern/testdata/patterns/product_codes.toml b/crates/nvisy-pattern/testdata/patterns/product_codes.toml
deleted file mode 100644
index 87e7941be..000000000
--- a/crates/nvisy-pattern/testdata/patterns/product_codes.toml
+++ /dev/null
@@ -1,6 +0,0 @@
-name = "internal-product-code"
-label = "internal_id"
-
-[[variants]]
-regex = "\\b(?:WIDGET-\\d{3}|SPROCKET-\\d{2}|GADGET-X\\d)\\b"
-score = 0.9
diff --git a/crates/nvisy-pattern/tests/builtin.rs b/crates/nvisy-pattern/tests/builtin.rs
deleted file mode 100644
index 2025384d2..000000000
--- a/crates/nvisy-pattern/tests/builtin.rs
+++ /dev/null
@@ -1,143 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! cross-jurisdiction (`world`) fixtures.
-//!
-//! Each test scans one `testdata/inputs/<domain>.txt` fixture
-//! through a recognizer wired with every shipped pattern and
-//! dictionary, then asserts the entities a real document of that
-//! domain is expected to surface (substring + label, not
-//! byte-offset, so fixtures and regexes can evolve without
-//! brittle churn).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/contact.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::EMAIL_ADDRESS.label_ref(),
-        "alice.johnson@example.com",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::URL.label_ref(),
-        "https://docs.example.com/proposal",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::URL.label_ref(),
-        "http://backup.example.org/proposal-v2",
-    );
-    assert_label_present(&entities, builtins::PHONE_NUMBER.label_ref());
-}
-
-#[tokio::test]
-async fn builtin_credentials() {
-    let (_, entities) = scan(include_str!("../testdata/inputs/credentials.txt")).await;
-    assert_label_present(&entities, builtins::API_KEY.label_ref());
-    assert_label_present(&entities, builtins::PRIVATE_KEY.label_ref());
-    assert_label_present(&entities, builtins::AUTH_TOKEN.label_ref());
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::IBAN.label_ref(),
-        "GB29 NWBK 6016 1331 9268 19",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::SWIFT_CODE.label_ref(),
-        "NWBKGB2L",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::PAYMENT_CARD.label_ref(),
-        "4539 1488 0343 6467",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::CRYPTO_ADDRESS.label_ref(),
-        "1A1zP1eP5QGefi2DMPTfTL5SLmv7DivfNa",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::CRYPTO_ADDRESS.label_ref(),
-        "0x742d35Cc6634C0532925a3b844Bc9e7595f6E842",
-    );
-    // Currency dictionaries pick up `USD`, `EUR`, `Tether`, `USDC`.
-    assert_label_present(&entities, builtins::CURRENCY.label_ref());
-}
-
-#[tokio::test]
-async fn builtin_network() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/network.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::IP_ADDRESS.label_ref(),
-        "192.168.1.42",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::IP_ADDRESS.label_ref(),
-        "10.0.0.7",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::IP_ADDRESS.label_ref(),
-        "203.0.113.55",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::IP_ADDRESS.label_ref(),
-        "2001:0db8:85a3:0000:0000:8a2e:0370:7334",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::MAC_ADDRESS.label_ref(),
-        "00:1A:2B:3C:4D:5E",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::MAC_ADDRESS.label_ref(),
-        "3C-22-FB-A1-B2-C3",
-    );
-}
-
-#[tokio::test]
-async fn builtin_personal() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/personal.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::DATE_OF_BIRTH.label_ref(),
-        "04/22/1979",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::DATE_TIME.label_ref(),
-        "2024-06-15T09:30:00Z",
-    );
-    assert_label_present(&entities, builtins::NATIONALITY.label_ref());
-    assert_label_present(&entities, builtins::LANGUAGE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_au.rs b/crates/nvisy-pattern/tests/builtin_au.rs
deleted file mode 100644
index 2d79528cd..000000000
--- a/crates/nvisy-pattern/tests/builtin_au.rs
+++ /dev/null
@@ -1,52 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! AU-jurisdiction fixtures (`testdata/inputs/au/<domain>.txt`).
-//!
-//! Each test scans one AU fixture through a recognizer wired
-//! with every shipped pattern and dictionary, then asserts the
-//! entities a real Australian document of that domain is
-//! expected to surface (substring + label).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/au/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::TAX_ID.label_ref(),
-        "123 456 782",
-    );
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/au/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::COMPANY_ID.label_ref(),
-        "51 824 753 556",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::COMPANY_ID.label_ref(),
-        "123 456 780",
-    );
-    assert_label_present(&entities, builtins::COMPANY_ID.label_ref());
-}
-
-#[tokio::test]
-async fn builtin_health() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/au/health.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::INSURANCE_ID.label_ref(),
-        "2228 12366 1",
-    );
-}
diff --git a/crates/nvisy-pattern/tests/builtin_ca.rs b/crates/nvisy-pattern/tests/builtin_ca.rs
deleted file mode 100644
index 7860b3cbc..000000000
--- a/crates/nvisy-pattern/tests/builtin_ca.rs
+++ /dev/null
@@ -1,35 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! CA-jurisdiction fixtures (`testdata/inputs/ca/<domain>.txt`).
-//!
-//! Each test scans one CA fixture through a recognizer wired
-//! with every shipped pattern and dictionary, then asserts the
-//! entities a real Canadian document of that domain is expected
-//! to surface (substring + label).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/ca/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "123 456 782",
-    );
-}
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/ca/contact.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::POSTAL_CODE.label_ref(),
-        "K1P 1A1",
-    );
-    assert_label_present(&entities, builtins::POSTAL_CODE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_de.rs b/crates/nvisy-pattern/tests/builtin_de.rs
deleted file mode 100644
index 943367622..000000000
--- a/crates/nvisy-pattern/tests/builtin_de.rs
+++ /dev/null
@@ -1,114 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! DE-jurisdiction fixtures (`testdata/inputs/de/<domain>.txt`).
-//!
-//! Each test scans one DE fixture through a recognizer wired
-//! with every shipped pattern and dictionary, then asserts the
-//! entities a real German document of that domain is expected to
-//! surface (substring + label).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/de/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "L01X00T44",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::PASSPORT_NUMBER.label_ref(),
-        "C0J9H58P3",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::TAX_ID.label_ref(),
-        "65929970489",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::TAX_ID.label_ref(),
-        "DE129273398",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::DRIVERS_LICENSE.label_ref(),
-        "B072RRE2I52",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::NATIONAL_INSURANCE_NUMBER.label_ref(),
-        "15010685M016",
-    );
-}
-
-#[tokio::test]
-async fn builtin_health() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/de/health.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::MEDICAL_ID.label_ref(),
-        "381789045",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::MEDICAL_ID.label_ref(),
-        "123456601",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::INSURANCE_ID.label_ref(),
-        "A000500015",
-    );
-}
-
-#[tokio::test]
-async fn builtin_vehicle() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/de/vehicle.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::LICENSE_PLATE.label_ref(),
-        "B-AB 1234",
-    );
-}
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/de/contact.txt")).await;
-    assert_match(&text, &entities, builtins::POSTAL_CODE.label_ref(), "10117");
-    // English-language nationality dictionary stays silent on a
-    // German document — assert it didn't fire.
-    assert!(
-        !entities
-            .iter()
-            .any(|e| e.label == builtins::NATIONALITY.label_ref()),
-        "english-language NATIONALITY dictionary should not match on a DE fixture",
-    );
-    // Sanity: at least one PLZ entity surfaced.
-    assert_label_present(&entities, builtins::POSTAL_CODE.label_ref());
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/de/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::COMPANY_ID.label_ref(),
-        "HRB 123456",
-    );
-}
diff --git a/crates/nvisy-pattern/tests/builtin_es.rs b/crates/nvisy-pattern/tests/builtin_es.rs
deleted file mode 100644
index ecf671b5b..000000000
--- a/crates/nvisy-pattern/tests/builtin_es.rs
+++ /dev/null
@@ -1,61 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! ES-jurisdiction fixtures (`testdata/inputs/es/<domain>.txt`).
-//!
-//! Each test scans one ES fixture through a recognizer wired
-//! with every shipped pattern and dictionary, then asserts the
-//! entities a real Spanish document of that domain is expected
-//! to surface (substring + label).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/es/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "12345678Z",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "X1234567L",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::PASSPORT_NUMBER.label_ref(),
-        "AAA123456",
-    );
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/es/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::COMPANY_ID.label_ref(),
-        "A12345674",
-    );
-}
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/es/contact.txt")).await;
-    assert_match(&text, &entities, builtins::POSTAL_CODE.label_ref(), "28013");
-    // English-language nationality dictionary stays silent on a
-    // Spanish document — assert it didn't fire.
-    assert!(
-        !entities
-            .iter()
-            .any(|e| e.label == builtins::NATIONALITY.label_ref()),
-        "english-language NATIONALITY dictionary should not match on an ES fixture",
-    );
-    assert_label_present(&entities, builtins::POSTAL_CODE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_fi.rs b/crates/nvisy-pattern/tests/builtin_fi.rs
deleted file mode 100644
index 5488a5a3d..000000000
--- a/crates/nvisy-pattern/tests/builtin_fi.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! FI-jurisdiction fixtures (`testdata/inputs/fi/<domain>.txt`).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/fi/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "010170-123F",
-    );
-    assert_label_present(&entities, builtins::GOVERNMENT_ID.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_in.rs b/crates/nvisy-pattern/tests/builtin_in.rs
deleted file mode 100644
index fd7057f26..000000000
--- a/crates/nvisy-pattern/tests/builtin_in.rs
+++ /dev/null
@@ -1,60 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! IN-jurisdiction fixtures (`testdata/inputs/in/<domain>.txt`).
-//!
-//! Each test scans one IN fixture through a recognizer wired
-//! with every shipped pattern and dictionary, then asserts the
-//! entities a real Indian document of that domain is expected
-//! to surface (substring + label).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/in/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "2341 2341 2346",
-    );
-    assert_match(&text, &entities, builtins::TAX_ID.label_ref(), "ABCPK1234E");
-    assert_match(
-        &text,
-        &entities,
-        builtins::PASSPORT_NUMBER.label_ref(),
-        "M1234567",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "ABC1234567",
-    );
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/in/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::TAX_ID.label_ref(),
-        "27AAAPL1234C1ZE",
-    );
-    assert_match(&text, &entities, builtins::TAX_ID.label_ref(), "AAAPL1234C");
-}
-
-#[tokio::test]
-async fn builtin_vehicle() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/in/vehicle.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::LICENSE_PLATE.label_ref(),
-        "MH12AB1234",
-    );
-    assert_label_present(&entities, builtins::LICENSE_PLATE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_it.rs b/crates/nvisy-pattern/tests/builtin_it.rs
deleted file mode 100644
index eeb54c2b9..000000000
--- a/crates/nvisy-pattern/tests/builtin_it.rs
+++ /dev/null
@@ -1,61 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! IT-jurisdiction fixtures (`testdata/inputs/it/<domain>.txt`).
-//!
-//! Each test scans one IT fixture through a recognizer wired
-//! with every shipped pattern and dictionary, then asserts the
-//! entities a real Italian document of that domain is expected
-//! to surface (substring + label).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/it/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::TAX_ID.label_ref(),
-        "RSSMRA70A01H501S",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "CA12345AB",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::PASSPORT_NUMBER.label_ref(),
-        "YA1234567",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::DRIVERS_LICENSE.label_ref(),
-        "MI1234567A",
-    );
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/it/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::TAX_ID.label_ref(),
-        "00154980569",
-    );
-    // English-language nationality dictionary stays silent on an
-    // Italian document — assert it didn't fire.
-    assert!(
-        !entities
-            .iter()
-            .any(|e| e.label == builtins::NATIONALITY.label_ref()),
-        "english-language NATIONALITY dictionary should not match on an IT fixture",
-    );
-    assert_label_present(&entities, builtins::TAX_ID.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_kr.rs b/crates/nvisy-pattern/tests/builtin_kr.rs
deleted file mode 100644
index 919e4355e..000000000
--- a/crates/nvisy-pattern/tests/builtin_kr.rs
+++ /dev/null
@@ -1,48 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! KR-jurisdiction fixtures (`testdata/inputs/kr/<domain>.txt`).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/kr/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "800101-1112343",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "900101-5112344",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::PASSPORT_NUMBER.label_ref(),
-        "M123N4567",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::DRIVERS_LICENSE.label_ref(),
-        "11-20-123456-78",
-    );
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/kr/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::COMPANY_ID.label_ref(),
-        "123-45-67891",
-    );
-    assert_label_present(&entities, builtins::COMPANY_ID.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_ng.rs b/crates/nvisy-pattern/tests/builtin_ng.rs
deleted file mode 100644
index 088f7def7..000000000
--- a/crates/nvisy-pattern/tests/builtin_ng.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! NG-jurisdiction fixtures (`testdata/inputs/ng/<domain>.txt`).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/ng/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "12345678902",
-    );
-}
-
-#[tokio::test]
-async fn builtin_vehicle() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/ng/vehicle.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::LICENSE_PLATE.label_ref(),
-        "ABC-123DE",
-    );
-    assert_label_present(&entities, builtins::LICENSE_PLATE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_pl.rs b/crates/nvisy-pattern/tests/builtin_pl.rs
deleted file mode 100644
index b1489db1e..000000000
--- a/crates/nvisy-pattern/tests/builtin_pl.rs
+++ /dev/null
@@ -1,60 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! PL-jurisdiction fixtures (`testdata/inputs/pl/<domain>.txt`).
-//!
-//! Each test scans one PL fixture through a recognizer wired
-//! with every shipped pattern and dictionary, then asserts the
-//! entities a real Polish document of that domain is expected
-//! to surface (substring + label).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/pl/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "44051401359",
-    );
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/pl/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::TAX_ID.label_ref(),
-        "106-000-00-62",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::COMPANY_ID.label_ref(),
-        "123456785",
-    );
-}
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/pl/contact.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::POSTAL_CODE.label_ref(),
-        "00-001",
-    );
-    // English-language nationality dictionary stays silent on a
-    // Polish document — assert it didn't fire.
-    assert!(
-        !entities
-            .iter()
-            .any(|e| e.label == builtins::NATIONALITY.label_ref()),
-        "english-language NATIONALITY dictionary should not match on a PL fixture",
-    );
-    assert_label_present(&entities, builtins::POSTAL_CODE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_se.rs b/crates/nvisy-pattern/tests/builtin_se.rs
deleted file mode 100644
index 9cf741b4c..000000000
--- a/crates/nvisy-pattern/tests/builtin_se.rs
+++ /dev/null
@@ -1,41 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! SE-jurisdiction fixtures (`testdata/inputs/se/<domain>.txt`).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/se/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "900101-1239",
-    );
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/se/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::COMPANY_ID.label_ref(),
-        "556677-1233",
-    );
-}
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/se/contact.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::POSTAL_CODE.label_ref(),
-        "111 60",
-    );
-    assert_label_present(&entities, builtins::POSTAL_CODE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_sg.rs b/crates/nvisy-pattern/tests/builtin_sg.rs
deleted file mode 100644
index 83542c8dc..000000000
--- a/crates/nvisy-pattern/tests/builtin_sg.rs
+++ /dev/null
@@ -1,47 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! SG-jurisdiction fixtures (`testdata/inputs/sg/<domain>.txt`).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/sg/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "S2740116C",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "F1234567N",
-    );
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/sg/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::COMPANY_ID.label_ref(),
-        "200512345R",
-    );
-}
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/sg/contact.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::POSTAL_CODE.label_ref(),
-        "018989",
-    );
-    assert_label_present(&entities, builtins::POSTAL_CODE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_th.rs b/crates/nvisy-pattern/tests/builtin_th.rs
deleted file mode 100644
index 2ce212dfd..000000000
--- a/crates/nvisy-pattern/tests/builtin_th.rs
+++ /dev/null
@@ -1,25 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! TH-jurisdiction fixtures (`testdata/inputs/th/<domain>.txt`).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/th/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "1-2345-67890-12-1",
-    );
-}
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/th/contact.txt")).await;
-    assert_match(&text, &entities, builtins::POSTAL_CODE.label_ref(), "10500");
-    assert_label_present(&entities, builtins::POSTAL_CODE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_tr.rs b/crates/nvisy-pattern/tests/builtin_tr.rs
deleted file mode 100644
index 64b474008..000000000
--- a/crates/nvisy-pattern/tests/builtin_tr.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! TR-jurisdiction fixtures (`testdata/inputs/tr/<domain>.txt`).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/tr/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "12345678950",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::LICENSE_PLATE.label_ref(),
-        "34-ABC-1234",
-    );
-}
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/tr/contact.txt")).await;
-    assert_match(&text, &entities, builtins::POSTAL_CODE.label_ref(), "34000");
-    assert_label_present(&entities, builtins::POSTAL_CODE.label_ref());
-}
diff --git a/crates/nvisy-pattern/tests/builtin_uk.rs b/crates/nvisy-pattern/tests/builtin_uk.rs
deleted file mode 100644
index e166c73aa..000000000
--- a/crates/nvisy-pattern/tests/builtin_uk.rs
+++ /dev/null
@@ -1,65 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! UK-jurisdiction fixtures (`testdata/inputs/uk/<domain>.txt`).
-//!
-//! Each test scans one UK fixture through a recognizer wired
-//! with every shipped pattern and dictionary, then asserts the
-//! entities a real UK document of that domain is expected to
-//! surface (substring + label).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/uk/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::MEDICAL_ID.label_ref(),
-        "943 476 5919",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::NATIONAL_INSURANCE_NUMBER.label_ref(),
-        "AB123456C",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::DRIVERS_LICENSE.label_ref(),
-        "MORGA753116SM9IJ",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::PASSPORT_NUMBER.label_ref(),
-        "AB1234567",
-    );
-    // World nationality dictionary activates on UK text ("British").
-    assert_label_present(&entities, builtins::NATIONALITY.label_ref());
-}
-
-#[tokio::test]
-async fn builtin_contact() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/uk/contact.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::POSTAL_CODE.label_ref(),
-        "SW1A 2AA",
-    );
-}
-
-#[tokio::test]
-async fn builtin_vehicle() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/uk/vehicle.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::LICENSE_PLATE.label_ref(),
-        "AB51 ABC",
-    );
-}
diff --git a/crates/nvisy-pattern/tests/builtin_us.rs b/crates/nvisy-pattern/tests/builtin_us.rs
deleted file mode 100644
index 24b714cb4..000000000
--- a/crates/nvisy-pattern/tests/builtin_us.rs
+++ /dev/null
@@ -1,85 +0,0 @@
-//! End-to-end: shipped patterns + dictionaries against the
-//! US-jurisdiction fixtures (`testdata/inputs/us/<domain>.txt`).
-//!
-//! Each test scans one US fixture through a recognizer wired
-//! with every shipped pattern and dictionary, then asserts the
-//! entities a real US document of that domain is expected to
-//! surface (substring + label).
-
-mod fixtures;
-
-use fixtures::{assert_label_present, assert_match, scan};
-use nvisy_core::entity::builtins;
-
-#[tokio::test]
-async fn builtin_identity() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/us/identity.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::GOVERNMENT_ID.label_ref(),
-        "123-45-6789",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::TAX_ID.label_ref(),
-        "912-71-1234",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::DRIVERS_LICENSE.label_ref(),
-        "D123-4567-8901",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::PASSPORT_NUMBER.label_ref(),
-        "A12345678",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::POSTAL_CODE.label_ref(),
-        "97477-1234",
-    );
-}
-
-#[tokio::test]
-async fn builtin_finance() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/us/finance.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::BANK_ROUTING.label_ref(),
-        "121000358",
-    );
-    // bank_account is `\b\d{8,17}\b` with score 0.05 — it requires
-    // a context-keyword boost (e.g. `account`) to clear the
-    // confidence threshold. The fixture provides one.
-    assert_label_present(&entities, builtins::BANK_ACCOUNT.label_ref());
-}
-
-#[tokio::test]
-async fn builtin_health() {
-    let (text, entities) = scan(include_str!("../testdata/inputs/us/health.txt")).await;
-    assert_match(
-        &text,
-        &entities,
-        builtins::MEDICAL_ID.label_ref(),
-        "1234567893",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::MEDICAL_ID.label_ref(),
-        "1EG4-TE5-MK73",
-    );
-    assert_match(
-        &text,
-        &entities,
-        builtins::MEDICAL_ID.label_ref(),
-        "BB0000000",
-    );
-}
diff --git a/crates/nvisy-pattern/tests/enhancer_roundtrip.rs b/crates/nvisy-pattern/tests/enhancer_roundtrip.rs
deleted file mode 100644
index 057a57088..000000000
--- a/crates/nvisy-pattern/tests/enhancer_roundtrip.rs
+++ /dev/null
@@ -1,75 +0,0 @@
-//! End-to-end: feed real input through a [`Regex`] →
-//! [`PatternRecognizer`] (wrapped in [`ContextEnhanced`]) and verify
-//! that confidence is boosted, and a [`Refinement`] step is
-//! appended only for matches that had a nearby keyword.
-//!
-//! [`Refinement`]: nvisy_core::entity::TrailStepKind::Refinement
-//! [`ContextEnhanced`]: nvisy_context::ContextEnhanced
-
-use nvisy_core::entity::{TrailStepKind, builtins};
-use nvisy_core::modality::TextData;
-use nvisy_core::primitive::Confidence;
-use nvisy_core::recognition::{EntityRecognizer, RecognizerInput};
-use nvisy_pattern::{PatternRecognizer, Regex, Variant};
-
-#[tokio::test]
-async fn enhancer_boosts_matches_near_keyword_only() {
-    let variant = Variant::new(r"\b\d{3}-\d{2}-\d{4}\b")
-        .expect("ssn variant builds")
-        .with_score(Confidence::clamped(0.6));
-    let regex = Regex::builder()
-        .with_name("ssn")
-        .with_label(builtins::GOVERNMENT_ID.label_ref())
-        .with_context(vec!["ssn".to_owned(), "social security".to_owned()])
-        .with_variants(vec![variant])
-        .build()
-        .expect("ssn regex builds");
-
-    let recognizer = PatternRecognizer::builder()
-        .with_pattern(regex)
-        .build_context_enhanced()
-        .expect("recognizer builds");
-
-    // Two SSN-shaped numbers: one near the keyword, one not.
-    let text = "First SSN: 123-45-6789. Unrelated number 987-65-4329 elsewhere.";
-    let input = RecognizerInput::new(TextData::new(text.to_owned()));
-    let entities = recognizer
-        .recognize(&input)
-        .await
-        .expect("recognize")
-        .entities;
-    assert_eq!(entities.len(), 2, "two SSN matches expected");
-
-    // First match has `SSN:` within the default 5-word prefix/suffix
-    // window and gets boosted by the ContextEnhanced<PatternRecognizer> wrapper.
-    let near = entities
-        .iter()
-        .find(|e| &text[e.location.start..e.location.end] == "123-45-6789")
-        .expect("near match present");
-    assert!(
-        near.confidence.get() > 0.6,
-        "near-keyword match should be boosted",
-    );
-    assert!(
-        near.trail
-            .iter()
-            .any(|s| matches!(s.kind, TrailStepKind::Refinement)),
-        "near-keyword match should have a Refinement step",
-    );
-
-    // Second match is well outside the window → untouched.
-    let far = entities
-        .iter()
-        .find(|e| &text[e.location.start..e.location.end] == "987-65-4329")
-        .expect("far match present");
-    assert!(
-        (far.confidence.get() - 0.6).abs() < f64::EPSILON,
-        "far-from-keyword match should not be boosted",
-    );
-    assert!(
-        !far.trail
-            .iter()
-            .any(|s| matches!(s.kind, TrailStepKind::Refinement)),
-        "far-from-keyword match should have no Refinement step",
-    );
-}
diff --git a/crates/nvisy-pattern/tests/fixtures/mod.rs b/crates/nvisy-pattern/tests/fixtures/mod.rs
deleted file mode 100644
index 1e257e47f..000000000
--- a/crates/nvisy-pattern/tests/fixtures/mod.rs
+++ /dev/null
@@ -1,52 +0,0 @@
-//! Shared helpers for the `builtin_*` end-to-end test suites.
-//!
-//! Each per-region test file (`tests/builtin_world.rs`,
-//! `tests/builtin_us.rs`, `tests/builtin_uk.rs`) declares this
-//! module via `mod fixtures;` and calls [`scan`] + the
-//! `assert_*` helpers to express expectations against a single
-//! shared [`PatternRecognizer`] built from every shipped pattern
-//! and dictionary.
-
-use nvisy_core::entity::{Entity, EntityLabelRef};
-use nvisy_core::modality::{Text, TextData};
-use nvisy_core::recognition::{EntityRecognizer, RecognizerInput};
-use nvisy_pattern::PatternRecognizer;
-
-pub async fn scan(text: &str) -> (String, Vec<Entity<Text>>) {
-    let recognizer = PatternRecognizer::builder()
-        .with_builtin_patterns()
-        .with_builtin_dictionaries()
-        .build_context_enhanced()
-        .expect("shipped recognizer builds");
-    let input = RecognizerInput::new(TextData::new(text.to_owned()));
-    let entities = recognizer
-        .recognize(&input)
-        .await
-        .expect("shipped recognize")
-        .entities;
-    (text.to_owned(), entities)
-}
-
-#[track_caller]
-pub fn assert_match(text: &str, entities: &[Entity<Text>], label: EntityLabelRef, needle: &str) {
-    let hit = entities
-        .iter()
-        .any(|e| e.label == label && &text[e.location.start..e.location.end] == needle);
-    assert!(
-        hit,
-        "expected `{needle}` as {label:?}; got: {:?}",
-        entities
-            .iter()
-            .map(|e| (e.label.clone(), &text[e.location.start..e.location.end]))
-            .collect::<Vec<_>>()
-    );
-}
-
-#[track_caller]
-pub fn assert_label_present(entities: &[Entity<Text>], label: EntityLabelRef) {
-    assert!(
-        entities.iter().any(|e| e.label == label),
-        "expected at least one {label:?} entity; got labels: {:?}",
-        entities.iter().map(|e| e.label.clone()).collect::<Vec<_>>()
-    );
-}
diff --git a/crates/nvisy-pattern/tests/user_rules.rs b/crates/nvisy-pattern/tests/user_rules.rs
deleted file mode 100644
index a3b477fdc..000000000
--- a/crates/nvisy-pattern/tests/user_rules.rs
+++ /dev/null
@@ -1,91 +0,0 @@
-//! End-to-end: load user-supplied patterns from the on-disk wire
-//! shape (`testdata/patterns/*.toml`,
-//! `testdata/dictionaries/*.{toml,csv}`) through
-//! [`Regex::from_toml`], [`Dictionary::metadata_from_toml`], and
-//! [`Term::from_csv`], mix them with shipped patterns, and
-//! confirm a real internal-handoff document yields the custom
-//! entities.
-
-use nvisy_core::entity::builtins;
-use nvisy_core::modality::TextData;
-use nvisy_core::recognition::{EntityRecognizer, RecognizerInput};
-use nvisy_pattern::{Dictionary, PatternRecognizer, Regex, Term};
-
-#[tokio::test]
-async fn user_toml_rules_load_and_detect() {
-    let employee_id = Regex::from_toml(include_str!("../testdata/patterns/employee_id.toml"))
-        .expect("employee_id.toml parses");
-    let product_code_pattern =
-        Regex::from_toml(include_str!("../testdata/patterns/product_codes.toml"))
-            .expect("product_codes.toml parses");
-
-    let terms = Term::from_csv(include_str!("../testdata/dictionaries/product_codes.csv"))
-        .expect("product_codes.csv parses");
-    let product_code_dict =
-        Dictionary::metadata_from_toml(include_str!("../testdata/dictionaries/product_codes.toml"))
-            .expect("product_codes metadata parses")
-            .with_terms(terms)
-            .build()
-            .expect("dictionary builds");
-
-    // 4 rows × 3 columns; every non-empty cell becomes a term.
-    assert_eq!(product_code_dict.terms.len(), 12);
-
-    // Mix user patterns with shipped (so the input also sees email etc.).
-    let recognizer = PatternRecognizer::builder()
-        .with_pattern(employee_id)
-        .with_pattern(product_code_pattern)
-        .with_dictionary(product_code_dict)
-        .with_builtin_patterns()
-        .build_context_enhanced()
-        .expect("recognizer builds");
-
-    let text = include_str!("../testdata/inputs/internal.txt");
-    let input = RecognizerInput::new(TextData::new(text.to_owned()));
-    let entities = recognizer
-        .recognize(&input)
-        .await
-        .expect("recognize")
-        .entities;
-
-    // The custom regex finds both employee numbers.
-    let emp_hits: Vec<&str> = entities
-        .iter()
-        .filter(|e| e.label == builtins::INTERNAL_ID.label_ref())
-        .map(|e| &text[e.location.start..e.location.end])
-        .collect();
-    assert!(
-        emp_hits.contains(&"EMP-12345"),
-        "expected EMP-12345 among InternalId hits, got {emp_hits:?}"
-    );
-    assert!(
-        emp_hits.contains(&"EMP-67890"),
-        "expected EMP-67890 among InternalId hits, got {emp_hits:?}"
-    );
-
-    // Both the user regex and the user dictionary should fire on
-    // `WIDGET-200`: regex matches the code, dictionary matches the
-    // same code as a literal term.
-    assert!(
-        emp_hits.contains(&"WIDGET-200"),
-        "expected WIDGET-200 among InternalId hits, got {emp_hits:?}"
-    );
-
-    // Dictionary fires on the alias term `premium-widget` and the
-    // canonical full name `Acme Premium Widget` (substring of "as
-    // Acme Premium Widget."). Either is enough to prove the
-    // dictionary layer ran.
-    assert!(
-        emp_hits.contains(&"premium-widget") || emp_hits.contains(&"Acme Premium Widget"),
-        "expected dictionary alias/full-name hit, got {emp_hits:?}"
-    );
-
-    // Shipped email regex fires too — proves user + shipped coexist.
-    assert!(
-        entities
-            .iter()
-            .any(|e| e.label == builtins::EMAIL_ADDRESS.label_ref()
-                && &text[e.location.start..e.location.end] == "counsel@example.com"),
-        "expected shipped email regex to fire alongside user rules"
-    );
-}
diff --git a/crates/nvisy-server/Cargo.toml b/crates/nvisy-server/Cargo.toml
index eb2e7628b..df180cee9 100644
--- a/crates/nvisy-server/Cargo.toml
+++ b/crates/nvisy-server/Cargo.toml
@@ -19,7 +19,7 @@ homepage = { workspace = true }
 documentation = { workspace = true }
 
 [features]
-default = ["tabular", "image", "audio", "rich"]
+default = ["tabular", "image", "audio"]
 
 ## Tabular modality — csv, xlsx.
 tabular = ["nvisy-engine/tabular"]
@@ -27,19 +27,6 @@ tabular = ["nvisy-engine/tabular"]
 image = ["nvisy-engine/image"]
 ## Audio modality — wav, mp3.
 audio = ["nvisy-engine/audio"]
-## Rich-document modality — pdf, docx.
-rich = ["nvisy-engine/rich"]
-
-## Enable all OpenAI providers (GPT, Whisper STT).
-openai = ["nvisy-engine/openai"]
-## Enable Anthropic Claude completion provider.
-anthropic = ["nvisy-engine/anthropic"]
-## Enable Google Gemini.
-google = ["nvisy-engine/google"]
-
-## Enable the externalized BentoML backends (NER + OCR) — forwards
-## to `nvisy-engine/bento`.
-bento = ["nvisy-engine/bento"]
 
 [package.metadata.docs.rs]
 all-features = true
@@ -50,6 +37,12 @@ rustdoc-args = ["--cfg", "docsrs"]
 nvisy-core = { workspace = true, features = [] }
 nvisy-engine = { workspace = true, features = [] }
 
+# Upstream elide — server consumes modality + entity types
+# directly when rendering wire responses. Default features
+# off so server picks per-modality codec features explicitly
+# via its own `tabular` / `image` / `audio` flags.
+elide-core = { workspace = true, features = ["tabular", "image", "audio"] }
+
 # Serialization
 serde = { workspace = true, features = [] }
 schemars = { workspace = true, features = [] }
@@ -61,6 +54,7 @@ derive_more = { workspace = true, features = ["deref", "display"] }
 # Primitive datatypes
 uuid = { workspace = true, features = [] }
 jiff = { workspace = true, features = [] }
+semver = { workspace = true, features = ["serde"] }
 
 # Async runtime and parallelism
 futures = { workspace = true, features = [] }
diff --git a/crates/nvisy-server/src/handler/contexts.rs b/crates/nvisy-server/src/handler/contexts.rs
new file mode 100644
index 000000000..b10db47ab
--- /dev/null
+++ b/crates/nvisy-server/src/handler/contexts.rs
@@ -0,0 +1,136 @@
+//! Context resource handlers. Symmetric with
+//! [`super::policies`].
+
+use aide::axum::ApiRouter;
+use aide::axum::routing::{delete_with, get_with, post_with};
+use aide::transform::TransformOperation;
+use axum::extract::State;
+use axum::http::StatusCode;
+use nvisy_core::context::Context;
+use nvisy_engine::{ContextRegistry, EngineHandle};
+
+use super::error::Result;
+use super::request::{ContextIdPath, ContextVersionPath, NewContext};
+use super::response::ContextSummary;
+use crate::extract::{ActorId, Json, Path};
+use crate::middleware::{DEFAULT_READ_TIMEOUT, DEFAULT_WRITE_TIMEOUT, RouterTimeoutExt};
+use crate::service::ServiceState;
+
+const TARGET: &str = "nvisy_server::contexts";
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
+async fn put_context(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+    Json(NewContext(context)): Json<NewContext>,
+) -> Result<(StatusCode, Json<ContextSummary>)> {
+    engine.registry().put_context(actor_id, &context).await?;
+    Ok((
+        StatusCode::CREATED,
+        Json(ContextSummary {
+            id: context.id,
+            version: context.version,
+        }),
+    ))
+}
+
+fn put_context_docs(op: TransformOperation) -> TransformOperation {
+    op.id("putContext")
+        .tag("contexts")
+        .summary("Write a new context version")
+}
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%id, %version, %actor_id))]
+async fn get_context(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+    Path(ContextVersionPath { id, version }): Path<ContextVersionPath>,
+) -> Result<Json<Context>> {
+    let context = engine.registry().get_context(actor_id, id, version).await?;
+    Ok(Json(context))
+}
+
+fn get_context_docs(op: TransformOperation) -> TransformOperation {
+    op.id("getContext")
+        .tag("contexts")
+        .summary("Read a specific context version")
+}
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
+async fn get_latest_context(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+    Path(ContextIdPath { id }): Path<ContextIdPath>,
+) -> Result<Json<Context>> {
+    let context = engine.registry().latest_context(actor_id, id).await?;
+    Ok(Json(context))
+}
+
+fn get_latest_context_docs(op: TransformOperation) -> TransformOperation {
+    op.id("getLatestContext")
+        .tag("contexts")
+        .summary("Read the highest-version context for the id")
+}
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
+async fn list_contexts(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+) -> Result<Json<Vec<ContextSummary>>> {
+    let paged = engine.registry().list_contexts(actor_id).await?;
+    let summaries: Vec<ContextSummary> = paged
+        .items
+        .into_iter()
+        .map(|(id, version)| ContextSummary { id, version })
+        .collect();
+    Ok(Json(summaries))
+}
+
+fn list_contexts_docs(op: TransformOperation) -> TransformOperation {
+    op.id("listContexts")
+        .tag("contexts")
+        .summary("List every (context id, version) for the actor")
+}
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%id, %version, %actor_id))]
+async fn delete_context(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+    Path(ContextVersionPath { id, version }): Path<ContextVersionPath>,
+) -> Result<StatusCode> {
+    engine
+        .registry()
+        .delete_context(actor_id, id, version)
+        .await?;
+    Ok(StatusCode::NO_CONTENT)
+}
+
+fn delete_context_docs(op: TransformOperation) -> TransformOperation {
+    op.id("deleteContext")
+        .tag("contexts")
+        .summary("Remove one context version")
+}
+
+pub fn routes_v1() -> ApiRouter<ServiceState> {
+    let read = ApiRouter::new()
+        .api_route("/contexts", get_with(list_contexts, list_contexts_docs))
+        .api_route(
+            "/contexts/{id}/latest",
+            get_with(get_latest_context, get_latest_context_docs),
+        )
+        .api_route(
+            "/contexts/{id}/{version}",
+            get_with(get_context, get_context_docs),
+        )
+        .with_timeout(DEFAULT_READ_TIMEOUT);
+
+    let write = ApiRouter::new()
+        .api_route("/contexts", post_with(put_context, put_context_docs))
+        .api_route(
+            "/contexts/{id}/{version}",
+            delete_with(delete_context, delete_context_docs),
+        )
+        .with_timeout(DEFAULT_WRITE_TIMEOUT);
+
+    read.merge(write)
+}
diff --git a/crates/nvisy-server/src/handler/detections.rs b/crates/nvisy-server/src/handler/detections.rs
index d1eb7754c..61cf7b74c 100644
--- a/crates/nvisy-server/src/handler/detections.rs
+++ b/crates/nvisy-server/src/handler/detections.rs
@@ -1,25 +1,25 @@
-//! Detection-pass HTTP handlers.
+//! Detection-side handlers — the "find entities" half of the
+//! run lifecycle.
 //!
-//! | Method   | Path                       | Description                          |
-//! |----------|----------------------------|--------------------------------------|
-//! | `POST`   | `/detections`              | Submit a detection pass              |
-//! | `GET`    | `/detections`              | List detections (filterable)         |
-//! | `GET`    | `/detections/{id}`         | Get a detection snapshot             |
-//! | `POST`   | `/detections/{id}/cancel`  | Cancel an in-progress detection      |
-//! | `DELETE` | `/detections/{id}`         | Delete a finished detection          |
-//! | `DELETE` | `/detections`              | Delete all finished detections       |
+//! `POST /detections` starts a run from already-uploaded file
+//! ids. `GET /detections/{id}` returns the full run state
+//! (header + every per-document body inline); the response
+//! shape covers any [`RunState`] — clients filter on
+//! [`state`](super::response::RunStateDto) to render the
+//! detection vs redaction view of the same underlying run.
 
 use aide::axum::ApiRouter;
 use aide::axum::routing::{delete_with, get_with, post_with};
 use aide::transform::TransformOperation;
 use axum::extract::{Query, State};
 use axum::http::StatusCode;
-use nvisy_engine::detection::{DetectionEngine, DetectionFilter, DetectionSnapshot};
+use futures::future;
+use nvisy_engine::{EngineHandle, runs};
 
+use super::error::Result;
 use super::request::{DetectionPath, DetectionQuery, NewDetection};
-use super::response::{DetectionId, DetectionList, Page};
+use super::response::{DetectionId, Page, RunResponse};
 use crate::extract::{ActorId, Json, Path};
-use crate::handler::error::Result;
 use crate::middleware::{DEFAULT_READ_TIMEOUT, DEFAULT_WRITE_TIMEOUT, RouterTimeoutExt};
 use crate::service::ServiceState;
 
@@ -27,120 +27,104 @@ const TARGET: &str = "nvisy_server::detections";
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
 async fn create_detection(
-    State(engine): State<DetectionEngine>,
+    State(state): State<ServiceState>,
     ActorId(actor_id): ActorId,
     Json(req): Json<NewDetection>,
 ) -> Result<(StatusCode, Json<DetectionId>)> {
-    let input = req.into_engine_input(actor_id);
-    let id = engine.detect(input).await?;
-    tracing::info!(target: TARGET, %id, "detection pass submitted");
+    let batch = req.into_engine_input(state.analyzer_default());
+    let id = runs::start(state.engine(), actor_id, batch).await?;
+    tracing::info!(target: TARGET, %id, "detection started");
     Ok((StatusCode::ACCEPTED, Json(DetectionId { id })))
 }
 
 fn create_detection_docs(op: TransformOperation) -> TransformOperation {
     op.id("createDetection")
         .tag("detections")
-        .summary("Submit a detection pass")
+        .summary("Start a detection run")
         .description(
-            "Runs imports → extraction → recognition → deduplication → policy \
-             evaluation. Stops before applying any redaction; the result audit \
-             holds the policy chain's pending decisions for review before a \
-             matching redaction pass.",
+            "Runs the analyzer fan-out across the referenced files. Stops at \
+             `AwaitingReview`; the matching `POST /redactions` carries reviewer \
+             overrides and transitions to `Applied` / `PartiallyApplied`.",
         )
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
 async fn list_detections(
-    State(engine): State<DetectionEngine>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
     Query(query): Query<DetectionQuery>,
-) -> Result<Json<DetectionList>> {
-    let filter = DetectionFilter {
-        status: query.status,
-    };
-    let entries = engine.list_detections(actor_id, filter).await;
-    let page = Page::paginate(entries, &query.pagination);
-    Ok(Json(page))
+) -> Result<Json<Page<RunResponse>>> {
+    let runs_list = runs::list(&engine, actor_id).await;
+    let assembled = assemble_runs(&engine, actor_id, runs_list).await;
+    Ok(Json(Page::paginate(assembled, &query.pagination)))
 }
 
 fn list_detections_docs(op: TransformOperation) -> TransformOperation {
     op.id("listDetections")
         .tag("detections")
-        .summary("List detections")
-        .description("Returns detection passes for the caller, optionally filtered by status.")
+        .summary("List runs")
+        .description("Returns every run for the actor, with per-document bodies inlined.")
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
 async fn get_detection(
-    State(engine): State<DetectionEngine>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
     Path(DetectionPath { id }): Path<DetectionPath>,
-) -> Result<Json<DetectionSnapshot>> {
-    let snapshot = engine.get_detection(actor_id, id).await?;
-    Ok(Json(snapshot))
+) -> Result<Json<RunResponse>> {
+    let run = runs::get(&engine, actor_id, id).await?;
+    let docs = fetch_docs(&engine, actor_id, &run).await;
+    Ok(Json(RunResponse::assemble(run, docs)))
 }
 
 fn get_detection_docs(op: TransformOperation) -> TransformOperation {
     op.id("getDetection")
         .tag("detections")
-        .summary("Get a detection snapshot")
-        .description(
-            "Returns the detection pass snapshot. Once the pass reaches a \
-             terminal state with at least one audit, the snapshot's `result` \
-             field carries the immutable `DetectionResult` a redaction pass \
-             references.",
-        )
+        .summary("Get the full run state")
+        .description("Returns the run header + every per-document body inline.")
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
 async fn cancel_detection(
-    State(engine): State<DetectionEngine>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
     Path(DetectionPath { id }): Path<DetectionPath>,
 ) -> Result<StatusCode> {
-    engine.cancel_detection(actor_id, id).await?;
+    runs::cancel(&engine, actor_id, id).await?;
     Ok(StatusCode::ACCEPTED)
 }
 
 fn cancel_detection_docs(op: TransformOperation) -> TransformOperation {
     op.id("cancelDetection")
         .tag("detections")
-        .summary("Cancel an in-progress detection")
-        .description("Triggers cooperative cancellation at the next yield point.")
+        .summary("Mark a run as cancelled")
+        .description(
+            "Sets the run header to `Failed` with `reason = \"cancelled\"`. Only \
+             valid from `Analyzing` / `AwaitingReview`. In-flight per-doc fan-out \
+             tasks are not interrupted today; cooperative cancellation needs a \
+             `CancellationToken` plumbed through the pipeline.",
+        )
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
 async fn delete_detection(
-    State(engine): State<DetectionEngine>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
     Path(DetectionPath { id }): Path<DetectionPath>,
 ) -> Result<StatusCode> {
-    engine.delete_detection(actor_id, id).await?;
+    runs::delete(&engine, actor_id, id).await?;
     Ok(StatusCode::NO_CONTENT)
 }
 
 fn delete_detection_docs(op: TransformOperation) -> TransformOperation {
     op.id("deleteDetection")
         .tag("detections")
-        .summary("Delete a finished detection")
-        .description("Removes the in-memory record and cascades to the persisted result.")
-}
-
-#[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
-async fn delete_all_detections(
-    State(engine): State<DetectionEngine>,
-    ActorId(actor_id): ActorId,
-) -> Result<StatusCode> {
-    let deleted = engine.delete_all_detections(actor_id).await;
-    tracing::info!(target: TARGET, deleted, "all finished detections deleted");
-    Ok(StatusCode::NO_CONTENT)
-}
-
-fn delete_all_detections_docs(op: TransformOperation) -> TransformOperation {
-    op.id("deleteAllDetections")
-        .tag("detections")
-        .summary("Delete all finished detections")
-        .description("Active passes are preserved.")
+        .summary("Delete a run")
+        .description(
+            "Removes the run header + per-doc bodies. Does **not** cascade to \
+             input or redacted output files — files are first-class resources \
+             and survive their producing run.",
+        )
 }
 
 pub fn routes_v1() -> ApiRouter<ServiceState> {
@@ -158,8 +142,7 @@ pub fn routes_v1() -> ApiRouter<ServiceState> {
     let write = ApiRouter::new()
         .api_route(
             "/detections",
-            post_with(create_detection, create_detection_docs)
-                .delete_with(delete_all_detections, delete_all_detections_docs),
+            post_with(create_detection, create_detection_docs),
         )
         .api_route(
             "/detections/{id}",
@@ -173,3 +156,38 @@ pub fn routes_v1() -> ApiRouter<ServiceState> {
 
     read.merge(write)
 }
+
+/// Fetch every per-doc row for one run, concurrently. Failures
+/// on individual rows are dropped — the run header already
+/// records per-doc state, and a missing row means the run was
+/// concurrently deleted underneath us.
+pub(super) async fn fetch_docs(
+    engine: &EngineHandle,
+    actor_id: uuid::Uuid,
+    run: &nvisy_engine::runs::Run,
+) -> Vec<nvisy_engine::runs::RunDocument> {
+    let lookups = run
+        .document_ids
+        .iter()
+        .map(|doc_id| runs::get_doc(engine, actor_id, run.id, *doc_id));
+    future::join_all(lookups)
+        .await
+        .into_iter()
+        .filter_map(std::result::Result::ok)
+        .collect()
+}
+
+/// Assemble [`RunResponse`]s for many runs by fetching per-doc
+/// rows concurrently per run.
+async fn assemble_runs(
+    engine: &EngineHandle,
+    actor_id: uuid::Uuid,
+    runs_list: Vec<nvisy_engine::runs::Run>,
+) -> Vec<RunResponse> {
+    let mut out = Vec::with_capacity(runs_list.len());
+    for run in runs_list {
+        let docs = fetch_docs(engine, actor_id, &run).await;
+        out.push(RunResponse::assemble(run, docs));
+    }
+    out
+}
diff --git a/crates/nvisy-server/src/handler/files.rs b/crates/nvisy-server/src/handler/files.rs
index a0fdc9edf..b9f106bc3 100644
--- a/crates/nvisy-server/src/handler/files.rs
+++ b/crates/nvisy-server/src/handler/files.rs
@@ -1,33 +1,17 @@
-//! File upload, download, list, and deletion handlers.
-//!
-//! | Method   | Path                              | Description                                |
-//! |----------|-----------------------------------|--------------------------------------------|
-//! | `POST`   | `/files`                          | Upload raw bytes (octet-stream)            |
-//! | `GET`    | `/files`                          | List uploaded files (paginated metadata)   |
-//! | `GET`    | `/files/{id}`                     | Get file metadata (descriptor + digest)    |
-//! | `GET`    | `/files/{id}/content`             | Download raw bytes (octet-stream)          |
-//! | `GET`    | `/files/{id}/annotations`         | Read the annotation overlay                |
-//! | `PUT`    | `/files/{id}/annotations`         | Replace the annotation overlay             |
-//! | `DELETE` | `/files/{id}`                     | Delete a single file (cascades annotations)|
-//! | `DELETE` | `/files`                          | Delete all files                           |
-//!
-//! Paths are relative — the version prefix (e.g. `/api/v1`) is
-//! applied by the version module.
+//! File upload / download / list / delete handlers.
 
 use aide::axum::ApiRouter;
-use aide::axum::routing::{delete_with, get_with, post_with, put_with};
+use aide::axum::routing::{delete_with, get_with, post_with};
 use aide::transform::TransformOperation;
-use axum::body::{Body, Bytes};
+use axum::body::Bytes;
 use axum::extract::{Query, State};
-use axum::http::{HeaderMap, HeaderValue, StatusCode, header};
-use axum::response::{IntoResponse, Response};
-use nvisy_engine::document::AnyAnnotations;
-use nvisy_engine::registry::Registry;
-use nvisy_engine::{Content, ContentData, ContentDescriptor};
+use axum::http::{HeaderMap, StatusCode, header};
+use axum::response::Response;
+use nvisy_engine::{EngineHandle, FileDescriptor, FileRegistry};
 
-use super::error::Result;
-use super::request::{ContentPath, MAX_PAGE_LIMIT, Pagination};
-use super::response::{FileId, FileList, FileMetadata, Page};
+use super::error::{ErrorKind, Result};
+use super::request::{FilePath, FileQuery};
+use super::response::{FileId, FileMetadataResponse, Page};
 use crate::extract::{ActorId, Json, Path};
 use crate::middleware::{DEFAULT_READ_TIMEOUT, DEFAULT_WRITE_TIMEOUT, RouterTimeoutExt};
 use crate::service::ServiceState;
@@ -36,38 +20,44 @@ const TARGET: &str = "nvisy_server::files";
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
 async fn upload_file(
-    State(registry): State<Registry>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
     headers: HeaderMap,
     body: Bytes,
 ) -> Result<(StatusCode, Json<FileId>)> {
     let size = body.len();
-    let content_data = ContentData::from(body);
-
-    let mut descriptor = ContentDescriptor::new();
-    if let Some(mime) = headers
+    let filename = parse_filename(&headers);
+    let content_type = headers
         .get(header::CONTENT_TYPE)
         .and_then(|h| h.to_str().ok())
-    {
-        descriptor.content_type = Some(mime.to_owned());
-    }
-    if let Some(name) = headers
-        .get(header::CONTENT_DISPOSITION)
-        .and_then(|h| h.to_str().ok())
-        .and_then(parse_disposition_filename)
-    {
-        descriptor.filename = Some(name.into());
-    }
-
-    let content = Content::with_descriptor(content_data, descriptor);
-    let id = registry
-        .register_content(actor_id, content, None)
-        .await?
-        .content_source()
-        .as_uuid();
-
-    tracing::info!(target: TARGET, %id, size, "file uploaded");
-    Ok((StatusCode::CREATED, Json(FileId { id })))
+        .map(|s| s.to_owned().into());
+    let extension = filename
+        .as_deref()
+        .and_then(filename_extension)
+        .ok_or_else(|| {
+            ErrorKind::BadRequest.with_message(
+                "cannot derive file extension; supply a `Content-Disposition: \
+                 attachment; filename=\"…\"` header with an extension",
+            )
+        })?
+        .to_owned()
+        .into();
+
+    let descriptor = FileDescriptor {
+        filename: filename.map(Into::into),
+        content_type,
+        extension,
+        lineage: None,
+        descriptor_labels: Vec::new(),
+        descriptor_metadata: std::collections::HashMap::new(),
+    };
+
+    let metadata = engine
+        .registry()
+        .put_file(actor_id, descriptor, body)
+        .await?;
+    tracing::info!(target: TARGET, id = %metadata.id, size, "file uploaded");
+    Ok((StatusCode::CREATED, Json(FileId { id: metadata.id })))
 }
 
 fn upload_file_docs(op: TransformOperation) -> TransformOperation {
@@ -77,180 +67,110 @@ fn upload_file_docs(op: TransformOperation) -> TransformOperation {
         .description(
             "Accepts raw bytes as the request body. `Content-Type` carries the \
              caller-supplied MIME hint; `Content-Disposition` carries the original \
-             filename (e.g. `attachment; filename=\"scan.pdf\"`). Annotations are \
-             attached separately via `PUT /files/{id}/annotations`.",
+             filename (e.g. `attachment; filename=\"scan.pdf\"`). The extension \
+             the codec resolves on is derived from the filename.",
         )
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
-async fn get_file_metadata(
-    State(registry): State<Registry>,
+async fn get_file(
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
-    Path(ContentPath { id }): Path<ContentPath>,
-) -> Result<Json<FileMetadata>> {
-    let handle = registry.read_content(actor_id, id).await?;
-    let record = handle.record().await?;
-    Ok(Json(FileMetadata {
-        id,
-        descriptor: record.descriptor,
-        digest: record.digest,
-    }))
+    Path(FilePath { id }): Path<FilePath>,
+) -> Result<Json<FileMetadataResponse>> {
+    let metadata = engine.registry().get_file(actor_id, id).await?;
+    Ok(Json(metadata))
 }
 
-fn get_file_metadata_docs(op: TransformOperation) -> TransformOperation {
-    op.id("getFileMetadata")
-        .tag("files")
-        .summary("Get file metadata")
-        .description(
-            "Returns the file's caller-supplied descriptor + registry-computed \
-             digest. Annotations live at `GET /files/{id}/annotations`; bytes at \
-             `GET /files/{id}/content`.",
-        )
+fn get_file_docs(op: TransformOperation) -> TransformOperation {
+    op.id("getFile").tag("files").summary("Get file metadata")
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
-async fn download_file_content(
-    State(registry): State<Registry>,
+async fn get_file_content(
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
-    Path(ContentPath { id }): Path<ContentPath>,
+    Path(FilePath { id }): Path<FilePath>,
 ) -> Result<Response> {
-    let handle = registry.read_content(actor_id, id).await?;
-    let data = handle.content_data().await?;
-    let record = handle.record().await?;
-    let size = data.size();
-    tracing::debug!(target: TARGET, size, "file content downloaded");
-
-    let content_type = record
-        .content_type()
-        .and_then(|s| HeaderValue::from_str(s).ok())
-        .unwrap_or_else(|| HeaderValue::from_static("application/octet-stream"));
+    let metadata = engine.registry().get_file(actor_id, id).await?;
+    let bytes = engine.registry().get_file_bytes(actor_id, id).await?;
 
-    let mut response = (StatusCode::OK, Body::from(data.as_bytes().to_vec())).into_response();
-    response
-        .headers_mut()
-        .insert(header::CONTENT_TYPE, content_type);
-    if let Some(name) = record.filename_lossy()
-        && let Ok(disposition) = HeaderValue::from_str(&format!("attachment; filename=\"{name}\""))
+    let mut response = Response::new(axum::body::Body::from(bytes));
+    if let Some(ct) = metadata.content_type.as_ref()
+        && let Ok(value) = ct.as_str().parse()
     {
-        response
-            .headers_mut()
-            .insert(header::CONTENT_DISPOSITION, disposition);
+        response.headers_mut().insert(header::CONTENT_TYPE, value);
+    }
+    if let Some(name) = metadata.filename.as_ref() {
+        let disposition = format!("attachment; filename=\"{}\"", name.as_str());
+        if let Ok(value) = disposition.parse() {
+            response
+                .headers_mut()
+                .insert(header::CONTENT_DISPOSITION, value);
+        }
     }
     Ok(response)
 }
 
-fn download_file_content_docs(op: TransformOperation) -> TransformOperation {
-    op.id("downloadFileContent")
-        .tag("files")
-        .summary("Download raw file bytes")
-        .description(
-            "Returns the file's raw bytes. `Content-Type` carries the best-available \
-             MIME (caller-supplied, then sniffed); `Content-Disposition` carries the \
-             original filename when known.",
-        )
-}
-
-#[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
-async fn get_file_annotations(
-    State(registry): State<Registry>,
-    ActorId(actor_id): ActorId,
-    Path(ContentPath { id }): Path<ContentPath>,
-) -> Result<Json<AnyAnnotations>> {
-    let annotations = registry.load_annotations(actor_id, id).await?;
-    Ok(Json(annotations))
-}
-
-fn get_file_annotations_docs(op: TransformOperation) -> TransformOperation {
-    op.id("getFileAnnotations")
-        .tag("files")
-        .summary("Read the file's annotation overlay")
-        .description(
-            "Returns the per-modality annotation buckets (`text`, `tabular`, \
-             `image`, `audio`) plus document-level `labels`. Returns an empty \
-             overlay when nothing has been set.",
-        )
-}
-
-#[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
-async fn put_file_annotations(
-    State(registry): State<Registry>,
-    ActorId(actor_id): ActorId,
-    Path(ContentPath { id }): Path<ContentPath>,
-    Json(annotations): Json<AnyAnnotations>,
-) -> Result<StatusCode> {
-    registry
-        .store_annotations(actor_id, id, &annotations)
-        .await?;
-    tracing::info!(target: TARGET, "file annotations updated");
-    Ok(StatusCode::NO_CONTENT)
-}
-
-fn put_file_annotations_docs(op: TransformOperation) -> TransformOperation {
-    op.id("putFileAnnotations")
+fn get_file_content_docs(op: TransformOperation) -> TransformOperation {
+    op.id("getFileContent")
         .tag("files")
-        .summary("Replace the file's annotation overlay")
+        .summary("Download file bytes")
         .description(
-            "Replaces the stored annotations with the request body. Idempotent: \
-             PUT with an empty object clears all hints. Annotations attached \
-             here are picked up by every subsequent detection / redaction pass \
-             that references this content.",
+            "Returns the raw bytes of the file with the upload's `Content-Type` \
+             and a `Content-Disposition` carrying the original filename.",
         )
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
 async fn list_files(
-    State(registry): State<Registry>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
-    Query(pagination): Query<Pagination>,
-) -> Result<Json<FileList>> {
-    let limit = pagination.limit.min(MAX_PAGE_LIMIT);
-    let paged = registry
-        .list_content_with_record(actor_id, pagination.offset, limit)
-        .await?;
-    let page = Page::from_paged(paged, &pagination, |(id, record)| FileMetadata {
-        id,
-        descriptor: record.descriptor,
-        digest: record.digest,
-    });
-    tracing::debug!(target: TARGET, total = page.total, count = page.items.len(), "files listed");
-    Ok(Json(page))
+    Query(query): Query<FileQuery>,
+) -> Result<Json<Page<FileMetadataResponse>>> {
+    let items = engine.registry().list_files(actor_id).await?;
+    Ok(Json(Page::paginate(items, &query.pagination)))
 }
 
 fn list_files_docs(op: TransformOperation) -> TransformOperation {
     op.id("listFiles")
         .tag("files")
-        .summary("List stored files")
+        .summary("List files")
         .description(
-            "Paginated metadata listing. Each entry has the same shape as \
-             `GET /files/{id}`. Annotations are not included.",
+            "Returns metadata for every file the actor owns, including redacted \
+             outputs (recognisable by `lineage`). Bytes are not loaded; fetch via \
+             `GET /files/{id}/content` when needed.",
         )
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
 async fn delete_file(
-    State(registry): State<Registry>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
-    Path(ContentPath { id }): Path<ContentPath>,
+    Path(FilePath { id }): Path<FilePath>,
 ) -> Result<StatusCode> {
-    registry.unregister_content(actor_id, id).await?;
-    tracing::info!(target: TARGET, "file deleted");
+    engine.registry().delete_file(actor_id, id).await?;
     Ok(StatusCode::NO_CONTENT)
 }
 
 fn delete_file_docs(op: TransformOperation) -> TransformOperation {
     op.id("deleteFile")
         .tag("files")
-        .summary("Delete an uploaded file")
-        .description("Removes the file and cascades to its annotations.")
+        .summary("Delete a file")
+        .description(
+            "Removes the file's bytes + metadata. Does not cascade — runs that \
+             referenced the file remain (their `inputFileId` will resolve to a \
+             missing file on subsequent apply).",
+        )
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
 async fn delete_all_files(
-    State(registry): State<Registry>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
 ) -> Result<StatusCode> {
-    let deleted = registry.unregister_all_content(actor_id).await?;
-    tracing::info!(target: TARGET, deleted, "all files deleted");
+    let removed = engine.registry().delete_all_files(actor_id).await?;
+    tracing::info!(target: TARGET, removed, "all files deleted");
     Ok(StatusCode::NO_CONTENT)
 }
 
@@ -258,91 +178,52 @@ fn delete_all_files_docs(op: TransformOperation) -> TransformOperation {
     op.id("deleteAllFiles")
         .tag("files")
         .summary("Delete every file for the actor")
-        .description("Removes every file plus its annotations.")
 }
 
 pub fn routes_v1() -> ApiRouter<ServiceState> {
-    let read_routes = ApiRouter::new()
+    let read = ApiRouter::new()
         .api_route("/files", get_with(list_files, list_files_docs))
-        .api_route(
-            "/files/{id}",
-            get_with(get_file_metadata, get_file_metadata_docs),
-        )
+        .api_route("/files/{id}", get_with(get_file, get_file_docs))
         .api_route(
             "/files/{id}/content",
-            get_with(download_file_content, download_file_content_docs),
-        )
-        .api_route(
-            "/files/{id}/annotations",
-            get_with(get_file_annotations, get_file_annotations_docs),
+            get_with(get_file_content, get_file_content_docs),
         )
         .with_timeout(DEFAULT_READ_TIMEOUT);
 
-    let write_routes = ApiRouter::new()
+    let write = ApiRouter::new()
         .api_route(
             "/files",
             post_with(upload_file, upload_file_docs)
                 .delete_with(delete_all_files, delete_all_files_docs),
         )
         .api_route("/files/{id}", delete_with(delete_file, delete_file_docs))
-        .api_route(
-            "/files/{id}/annotations",
-            put_with(put_file_annotations, put_file_annotations_docs),
-        )
         .with_timeout(DEFAULT_WRITE_TIMEOUT);
 
-    read_routes.merge(write_routes)
+    read.merge(write)
 }
 
 /// Parse the `filename="..."` value out of a `Content-Disposition`
-/// header. Accepts both quoted and unquoted forms, returns `None`
-/// when the header doesn't carry one.
-fn parse_disposition_filename(header_value: &str) -> Option<String> {
-    for part in header_value.split(';').map(str::trim) {
-        if let Some(rest) = part
-            .strip_prefix("filename=")
-            .or_else(|| part.strip_prefix("filename*="))
-        {
+/// header. Accepts quoted and unquoted forms.
+fn parse_filename(headers: &HeaderMap) -> Option<String> {
+    let raw = headers
+        .get(header::CONTENT_DISPOSITION)
+        .and_then(|h| h.to_str().ok())?;
+    for part in raw.split(';') {
+        let part = part.trim();
+        if let Some(rest) = part.strip_prefix("filename=") {
             let unquoted = rest.trim_matches('"');
-            if !unquoted.is_empty() {
-                return Some(unquoted.to_owned());
-            }
+            return Some(unquoted.to_owned());
         }
     }
     None
 }
 
-#[cfg(test)]
-mod tests {
-    use super::parse_disposition_filename;
-
-    #[test]
-    fn quoted_filename() {
-        assert_eq!(
-            parse_disposition_filename("attachment; filename=\"scan.pdf\""),
-            Some("scan.pdf".to_owned())
-        );
-    }
-
-    #[test]
-    fn unquoted_filename() {
-        assert_eq!(
-            parse_disposition_filename("attachment; filename=scan.pdf"),
-            Some("scan.pdf".to_owned())
-        );
-    }
-
-    #[test]
-    fn no_filename() {
-        assert_eq!(parse_disposition_filename("attachment"), None);
-        assert_eq!(parse_disposition_filename("inline"), None);
-    }
-
-    #[test]
-    fn empty_filename() {
-        assert_eq!(
-            parse_disposition_filename("attachment; filename=\"\""),
-            None
-        );
+/// Extract the lowercase extension from a filename. `report.pdf`
+/// → `Some("pdf")`; `noext` → `None`.
+fn filename_extension(name: &str) -> Option<&str> {
+    let idx = name.rfind('.')?;
+    if idx == 0 || idx + 1 >= name.len() {
+        return None;
     }
+    Some(&name[idx + 1..])
 }
diff --git a/crates/nvisy-server/src/handler/infra.rs b/crates/nvisy-server/src/handler/infra.rs
index e28b354f7..ebf8f35b9 100644
--- a/crates/nvisy-server/src/handler/infra.rs
+++ b/crates/nvisy-server/src/handler/infra.rs
@@ -1,16 +1,14 @@
 //! Health handler.
 //!
-//! | Method | Path          | Description                          |
-//! |--------|---------------|--------------------------------------|
-//! | `GET`  | `/health`     | Liveness probe (unversioned)         |
+//! | Method | Path     | Description           |
+//! |--------|----------|-----------------------|
+//! | `GET`  | `/health`| Liveness probe        |
 
 use aide::axum::ApiRouter;
 use aide::axum::routing::get_with;
 use aide::transform::TransformOperation;
 use axum::extract::State;
-use nvisy_core::health::{ComponentCheck, Healthcheck, ServiceStatus};
-use nvisy_engine::core::probe_all;
-use nvisy_engine::detection::DetectionEngine;
+use nvisy_core::service::health::{ComponentCheck, ServiceStatus};
 
 use super::response::Health;
 use crate::extract::Json;
@@ -20,21 +18,17 @@ use crate::service::ServiceState;
 const TARGET: &str = "nvisy_server::infra";
 
 #[tracing::instrument(target = TARGET, skip_all)]
-async fn health_check(State(engine): State<DetectionEngine>) -> Json<Health> {
-    let fs_status = if engine.data_dir().is_dir() {
+async fn health_check(State(state): State<ServiceState>) -> Json<Health> {
+    let fs_status = if state.data_dir().is_dir() {
         ServiceStatus::Healthy
     } else {
         ServiceStatus::Unhealthy
     };
-
-    let mut checks = vec![ComponentCheck::new("filesystem", fs_status)];
-    checks.extend(probe_all([engine.registry() as &dyn Healthcheck]).await);
-
+    let checks = vec![ComponentCheck::new("filesystem", fs_status)];
     let overall = roll_up(&checks);
     if overall != ServiceStatus::Healthy {
         tracing::warn!(target: TARGET, ?overall, "health check degraded or unhealthy");
     }
-    tracing::debug!(target: TARGET, ?overall, "health check");
 
     Json(Health {
         status: overall,
@@ -60,9 +54,8 @@ fn health_docs(op: TransformOperation) -> TransformOperation {
         .tag("infra")
         .summary("Liveness probe")
         .description(
-            "Checks that the server is running, the data directory is accessible, \
-             and the registry is operational. Returns 200 with an overall status \
-             and per-component checks.",
+            "Checks that the server is running and the data directory is accessible. \
+             Returns 200 with an overall status and per-component checks.",
         )
 }
 
diff --git a/crates/nvisy-server/src/handler/mod.rs b/crates/nvisy-server/src/handler/mod.rs
index 7b40b45bc..372ffd5b0 100644
--- a/crates/nvisy-server/src/handler/mod.rs
+++ b/crates/nvisy-server/src/handler/mod.rs
@@ -4,18 +4,21 @@
 //! `routes_v1()` function that returns relative-path `ApiRouter`
 //! fragments. These are nested under `/api/v1` by [`routes`].
 //!
-//! The top-level [`routes()`] function assembles all versions plus
-//! unversioned routes (health, docs) into a single router.
+//! The top-level [`routes()`] function assembles all versions
+//! plus unversioned routes (health, docs) into a single router.
 //!
-//! Request and response types live in the `request` and `response`
-//! submodules. [`Error`], [`ErrorKind`], and [`Result`] are re-exported for
-//! use by middleware and extractors.
+//! Request and response types live in the `request` and
+//! `response` submodules. [`Error`], [`ErrorKind`], and
+//! [`Result`] are re-exported for use by middleware and
+//! extractors.
 
 pub mod error;
 
+mod contexts;
 mod detections;
 mod files;
 mod infra;
+mod policies;
 mod redactions;
 
 mod request;
@@ -27,26 +30,29 @@ use axum::http::Uri;
 pub use self::error::{Error, ErrorKind, Result};
 use crate::service::ServiceState;
 
-/// Build the complete route tree with versioned API and unversioned infra routes.
+/// Build the complete route tree.
 ///
 /// # Route structure
 ///
 /// ```text
-/// /health                                  (unversioned)
-/// /api/v1/files[/{id}]                     (DELETE /files removes every file)
-/// /api/v1/detections[/{id}[/cancel]]       (DELETE /detections removes every finished detection)
-/// /api/v1/redactions[/{id}[/cancel]]       (DELETE /redactions removes every finished redaction)
-/// /api/v1/openapi.json                     (added by OpenAPI middleware)
-/// /docs                                    (added by OpenAPI middleware)
+/// /health                                       (unversioned)
+/// /api/v1/files[/{id}[/content]]
+/// /api/v1/policies[/{id}[/{version}|latest]]
+/// /api/v1/contexts[/{id}[/{version}|latest]]
+/// /api/v1/detections[/{id}[/cancel]]
+/// /api/v1/redactions[/{id}]
+/// /api/v1/openapi.json                          (added by OpenAPI middleware)
+/// /docs                                         (added by OpenAPI middleware)
 /// ```
 ///
-/// Policies are not a persisted resource: the caller submits the
-/// full [`Policy`] bodies inline on `POST /api/v1/detections`.
-/// The audit refers to each rule by `PolicyDecisionRef` so the
-/// caller can correlate decisions against their own copy of the
-/// policy bytes when rendering.
+/// `/detections` and `/redactions` are filtered views of one
+/// underlying run keyspace: a redaction id equals the
+/// detection id it was applied from. Bytes (uploads and
+/// redacted outputs) flow through `/files`; redacted files
+/// carry [`FileLineage::RedactedFrom`] so the run that
+/// produced them is traceable.
 ///
-/// [`Policy`]: nvisy_engine::policy::Policy
+/// [`FileLineage::RedactedFrom`]: nvisy_core::FileLineage::RedactedFrom
 pub fn routes() -> ApiRouter<ServiceState> {
     ApiRouter::new()
         .merge(infra::health_routes())
@@ -55,20 +61,17 @@ pub fn routes() -> ApiRouter<ServiceState> {
 }
 
 /// All v1 API routes (nested under `/api/v1` by [`routes`]).
-///
-/// When a new API version is introduced, a parallel `v2_routes()`
-/// function can compose a different set of handlers under `/api/v2`.
 fn v1_routes() -> ApiRouter<ServiceState> {
     ApiRouter::new()
         .merge(files::routes_v1())
+        .merge(policies::routes_v1())
+        .merge(contexts::routes_v1())
         .merge(detections::routes_v1())
         .merge(redactions::routes_v1())
 }
 
-/// Catch-all for unmatched paths.
-///
-/// Returns 404 with guidance pointing to the current API version
-/// for `/api/*` paths.
+/// Catch-all for unmatched paths. Returns 404 with guidance
+/// pointing to the current API version for `/api/*` paths.
 async fn api_version_fallback(uri: Uri) -> Result<()> {
     let path = uri.path();
     if path.starts_with("/api/") {
diff --git a/crates/nvisy-server/src/handler/policies.rs b/crates/nvisy-server/src/handler/policies.rs
new file mode 100644
index 000000000..7bb1963a1
--- /dev/null
+++ b/crates/nvisy-server/src/handler/policies.rs
@@ -0,0 +1,139 @@
+//! Policy resource handlers.
+
+use aide::axum::ApiRouter;
+use aide::axum::routing::{delete_with, get_with, post_with};
+use aide::transform::TransformOperation;
+use axum::extract::State;
+use axum::http::StatusCode;
+use nvisy_core::policy::Policy;
+use nvisy_engine::{EngineHandle, PolicyRegistry};
+
+use super::error::Result;
+use super::request::{NewPolicy, PolicyIdPath, PolicyVersionPath};
+use super::response::PolicySummary;
+use crate::extract::{ActorId, Json, Path};
+use crate::middleware::{DEFAULT_READ_TIMEOUT, DEFAULT_WRITE_TIMEOUT, RouterTimeoutExt};
+use crate::service::ServiceState;
+
+const TARGET: &str = "nvisy_server::policies";
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
+async fn put_policy(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+    Json(NewPolicy(policy)): Json<NewPolicy>,
+) -> Result<(StatusCode, Json<PolicySummary>)> {
+    engine.registry().put_policy(actor_id, &policy).await?;
+    Ok((
+        StatusCode::CREATED,
+        Json(PolicySummary {
+            id: policy.id,
+            version: policy.version,
+        }),
+    ))
+}
+
+fn put_policy_docs(op: TransformOperation) -> TransformOperation {
+    op.id("putPolicy")
+        .tag("policies")
+        .summary("Write a new policy version")
+        .description(
+            "Stores the full policy under `(actor_id, policy.id, policy.version)`. \
+             Returns `Conflict` if the same `(id, version)` already exists.",
+        )
+}
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%id, %version, %actor_id))]
+async fn get_policy(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+    Path(PolicyVersionPath { id, version }): Path<PolicyVersionPath>,
+) -> Result<Json<Policy>> {
+    let policy = engine.registry().get_policy(actor_id, id, version).await?;
+    Ok(Json(policy))
+}
+
+fn get_policy_docs(op: TransformOperation) -> TransformOperation {
+    op.id("getPolicy")
+        .tag("policies")
+        .summary("Read a specific policy version")
+}
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
+async fn get_latest_policy(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+    Path(PolicyIdPath { id }): Path<PolicyIdPath>,
+) -> Result<Json<Policy>> {
+    let policy = engine.registry().latest_policy(actor_id, id).await?;
+    Ok(Json(policy))
+}
+
+fn get_latest_policy_docs(op: TransformOperation) -> TransformOperation {
+    op.id("getLatestPolicy")
+        .tag("policies")
+        .summary("Read the highest-version policy for the id")
+}
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
+async fn list_policies(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+) -> Result<Json<Vec<PolicySummary>>> {
+    let paged = engine.registry().list_policies(actor_id).await?;
+    let summaries: Vec<PolicySummary> = paged
+        .items
+        .into_iter()
+        .map(|(id, version)| PolicySummary { id, version })
+        .collect();
+    Ok(Json(summaries))
+}
+
+fn list_policies_docs(op: TransformOperation) -> TransformOperation {
+    op.id("listPolicies")
+        .tag("policies")
+        .summary("List every (policy id, version) for the actor")
+}
+
+#[tracing::instrument(target = TARGET, skip_all, fields(%id, %version, %actor_id))]
+async fn delete_policy(
+    State(engine): State<EngineHandle>,
+    ActorId(actor_id): ActorId,
+    Path(PolicyVersionPath { id, version }): Path<PolicyVersionPath>,
+) -> Result<StatusCode> {
+    engine
+        .registry()
+        .delete_policy(actor_id, id, version)
+        .await?;
+    Ok(StatusCode::NO_CONTENT)
+}
+
+fn delete_policy_docs(op: TransformOperation) -> TransformOperation {
+    op.id("deletePolicy")
+        .tag("policies")
+        .summary("Remove one policy version")
+}
+
+pub fn routes_v1() -> ApiRouter<ServiceState> {
+    let read = ApiRouter::new()
+        .api_route("/policies", get_with(list_policies, list_policies_docs))
+        .api_route(
+            "/policies/{id}/latest",
+            get_with(get_latest_policy, get_latest_policy_docs),
+        )
+        .api_route(
+            "/policies/{id}/{version}",
+            get_with(get_policy, get_policy_docs),
+        )
+        .with_timeout(DEFAULT_READ_TIMEOUT);
+
+    let write = ApiRouter::new()
+        .api_route("/policies", post_with(put_policy, put_policy_docs))
+        .api_route(
+            "/policies/{id}/{version}",
+            delete_with(delete_policy, delete_policy_docs),
+        )
+        .with_timeout(DEFAULT_WRITE_TIMEOUT);
+
+    read.merge(write)
+}
diff --git a/crates/nvisy-server/src/handler/redactions.rs b/crates/nvisy-server/src/handler/redactions.rs
index c6aab7ae9..2b42d49bc 100644
--- a/crates/nvisy-server/src/handler/redactions.rs
+++ b/crates/nvisy-server/src/handler/redactions.rs
@@ -1,25 +1,25 @@
-//! Redaction-pass HTTP handlers.
+//! Redaction-side handlers — the "apply the redaction" half of
+//! the run lifecycle.
 //!
-//! | Method   | Path                       | Description                          |
-//! |----------|----------------------------|--------------------------------------|
-//! | `POST`   | `/redactions`              | Submit a redaction pass              |
-//! | `GET`    | `/redactions`              | List redactions (filterable)         |
-//! | `GET`    | `/redactions/{id}`         | Get a redaction snapshot             |
-//! | `POST`   | `/redactions/{id}/cancel`  | Cancel an in-progress redaction      |
-//! | `DELETE` | `/redactions/{id}`         | Delete a finished redaction          |
-//! | `DELETE` | `/redactions`              | Delete all finished redactions       |
+//! `POST /redactions` carries reviewer overrides and triggers
+//! the apply transition on the referenced detection. The
+//! response carries the per-doc apply outcomes (one redacted
+//! file id per successful input). `GET /redactions/{id}`
+//! returns the same full run state as `GET /detections/{id}` —
+//! they're views on the same underlying run.
 
 use aide::axum::ApiRouter;
-use aide::axum::routing::{delete_with, get_with, post_with};
+use aide::axum::routing::{get_with, post_with};
 use aide::transform::TransformOperation;
-use axum::extract::{Query, State};
+use axum::extract::State;
 use axum::http::StatusCode;
-use nvisy_engine::redaction::{RedactionEngine, RedactionFilter, RedactionSnapshot};
+use nvisy_engine::{EngineHandle, runs};
 
-use super::request::{NewRedaction, RedactionPath, RedactionQuery};
-use super::response::{Page, RedactionId, RedactionList};
+use super::detections::fetch_docs;
+use super::error::Result;
+use super::request::{NewRedaction, RedactionPath};
+use super::response::{RedactionOutput, RedactionResult, RunResponse};
 use crate::extract::{ActorId, Json, Path};
-use crate::handler::error::Result;
 use crate::middleware::{DEFAULT_READ_TIMEOUT, DEFAULT_WRITE_TIMEOUT, RouterTimeoutExt};
 use crate::service::ServiceState;
 
@@ -27,132 +27,95 @@ const TARGET: &str = "nvisy_server::redactions";
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
 async fn create_redaction(
-    State(engine): State<RedactionEngine>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
     Json(req): Json<NewRedaction>,
-) -> Result<(StatusCode, Json<RedactionId>)> {
-    let input = req.into_engine_input(actor_id);
-    let id = engine.redact(input).await?;
-    tracing::info!(target: TARGET, %id, "redaction pass submitted");
-    Ok((StatusCode::ACCEPTED, Json(RedactionId { id })))
+) -> Result<(StatusCode, Json<RedactionResult>)> {
+    let detection_id = req.detection_id;
+
+    // Apply every reviewer override first. Each is its own
+    // engine call so a missing entity surfaces a clean error
+    // before we kick off the more-expensive apply fan-out.
+    for o in &req.overrides {
+        runs::override_entity(
+            &engine,
+            actor_id,
+            detection_id,
+            o.doc_id,
+            o.entity_id,
+            o.action.clone(),
+        )
+        .await?;
+    }
+
+    runs::apply(&engine, actor_id, detection_id).await?;
+
+    // Re-read the run + per-doc rows to render the result. The
+    // header + rows now carry the Applied / PartiallyApplied /
+    // Failed state per doc plus the output file ids.
+    let run = runs::get(&engine, actor_id, detection_id).await?;
+    let docs = fetch_docs(&engine, actor_id, &run).await;
+    let outputs: Vec<RedactionOutput> = docs
+        .into_iter()
+        .map(|d| {
+            let dto: super::response::runs::RunDocumentDto = d.into();
+            RedactionOutput {
+                doc_id: dto.id,
+                input_file_id: dto.input_file_id,
+                output_file_id: dto.output_file_id,
+                state: dto.state,
+                failure_reason: dto.failure_reason,
+            }
+        })
+        .collect();
+
+    tracing::info!(
+        target: TARGET,
+        run_id = %detection_id,
+        outputs = outputs.len(),
+        "redaction applied"
+    );
+    Ok((
+        StatusCode::ACCEPTED,
+        Json(RedactionResult {
+            id: detection_id,
+            outputs,
+        }),
+    ))
 }
 
 fn create_redaction_docs(op: TransformOperation) -> TransformOperation {
     op.id("createRedaction")
         .tag("redactions")
-        .summary("Submit a redaction pass")
+        .summary("Apply a detection")
         .description(
-            "References a previously completed detection by id and optionally \
-             carries per-entity overrides (accept / reject / replace / add). \
-             Re-imports the original content, applies overrides to the audit, \
-             runs redaction + validation, and writes to the configured exports.",
-        )
-}
-
-#[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
-async fn list_redactions(
-    State(engine): State<RedactionEngine>,
-    ActorId(actor_id): ActorId,
-    Query(query): Query<RedactionQuery>,
-) -> Result<Json<RedactionList>> {
-    let filter = RedactionFilter {
-        status: query.status,
-        detection_id: query.detection_id,
-    };
-    let entries = engine.list_redactions(actor_id, filter).await;
-    let page = Page::paginate(entries, &query.pagination);
-    Ok(Json(page))
-}
-
-fn list_redactions_docs(op: TransformOperation) -> TransformOperation {
-    op.id("listRedactions")
-        .tag("redactions")
-        .summary("List redactions")
-        .description(
-            "Returns redaction passes for the caller, optionally filtered by \
-             status and / or detection id.",
+            "Applies reviewer overrides + transitions the referenced detection \
+             into the redaction phase. Returns the produced file ids; download \
+             via `GET /files/{id}/content`. Re-applying the same detection \
+             returns `Conflict`.",
         )
 }
 
 #[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
 async fn get_redaction(
-    State(engine): State<RedactionEngine>,
+    State(engine): State<EngineHandle>,
     ActorId(actor_id): ActorId,
     Path(RedactionPath { id }): Path<RedactionPath>,
-) -> Result<Json<RedactionSnapshot>> {
-    let snapshot = engine.get_redaction(actor_id, id).await?;
-    Ok(Json(snapshot))
+) -> Result<Json<RunResponse>> {
+    let run = runs::get(&engine, actor_id, id).await?;
+    let docs = fetch_docs(&engine, actor_id, &run).await;
+    Ok(Json(RunResponse::assemble(run, docs)))
 }
 
 fn get_redaction_docs(op: TransformOperation) -> TransformOperation {
     op.id("getRedaction")
         .tag("redactions")
-        .summary("Get a redaction snapshot")
-        .description(
-            "Returns the redaction pass snapshot. Once terminal with at least \
-             one audit, the snapshot's `result` field carries the final audit \
-             with `Execution` populated and `RedactionDecision` provenance on \
-             every override-touched entry.",
-        )
-}
-
-#[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
-async fn cancel_redaction(
-    State(engine): State<RedactionEngine>,
-    ActorId(actor_id): ActorId,
-    Path(RedactionPath { id }): Path<RedactionPath>,
-) -> Result<StatusCode> {
-    engine.cancel_redaction(actor_id, id).await?;
-    Ok(StatusCode::ACCEPTED)
-}
-
-fn cancel_redaction_docs(op: TransformOperation) -> TransformOperation {
-    op.id("cancelRedaction")
-        .tag("redactions")
-        .summary("Cancel an in-progress redaction")
-        .description("Triggers cooperative cancellation at the next yield point.")
-}
-
-#[tracing::instrument(target = TARGET, skip_all, fields(%id, %actor_id))]
-async fn delete_redaction(
-    State(engine): State<RedactionEngine>,
-    ActorId(actor_id): ActorId,
-    Path(RedactionPath { id }): Path<RedactionPath>,
-) -> Result<StatusCode> {
-    engine.delete_redaction(actor_id, id).await?;
-    Ok(StatusCode::NO_CONTENT)
-}
-
-fn delete_redaction_docs(op: TransformOperation) -> TransformOperation {
-    op.id("deleteRedaction")
-        .tag("redactions")
-        .summary("Delete a finished redaction")
-        .description("Removes the in-memory record and cascades to the persisted result.")
-}
-
-#[tracing::instrument(target = TARGET, skip_all, fields(%actor_id))]
-async fn delete_all_redactions(
-    State(engine): State<RedactionEngine>,
-    ActorId(actor_id): ActorId,
-) -> Result<StatusCode> {
-    let deleted = engine.delete_all_redactions(actor_id).await;
-    tracing::info!(target: TARGET, deleted, "all finished redactions deleted");
-    Ok(StatusCode::NO_CONTENT)
-}
-
-fn delete_all_redactions_docs(op: TransformOperation) -> TransformOperation {
-    op.id("deleteAllRedactions")
-        .tag("redactions")
-        .summary("Delete all finished redactions")
-        .description("Active passes are preserved.")
+        .summary("Get the run state for a redaction")
+        .description("Same shape as `GET /detections/{id}` — they're views on the same run.")
 }
 
 pub fn routes_v1() -> ApiRouter<ServiceState> {
     let read = ApiRouter::new()
-        .api_route(
-            "/redactions",
-            get_with(list_redactions, list_redactions_docs),
-        )
         .api_route(
             "/redactions/{id}",
             get_with(get_redaction, get_redaction_docs),
@@ -162,16 +125,7 @@ pub fn routes_v1() -> ApiRouter<ServiceState> {
     let write = ApiRouter::new()
         .api_route(
             "/redactions",
-            post_with(create_redaction, create_redaction_docs)
-                .delete_with(delete_all_redactions, delete_all_redactions_docs),
-        )
-        .api_route(
-            "/redactions/{id}",
-            delete_with(delete_redaction, delete_redaction_docs),
-        )
-        .api_route(
-            "/redactions/{id}/cancel",
-            post_with(cancel_redaction, cancel_redaction_docs),
+            post_with(create_redaction, create_redaction_docs),
         )
         .with_timeout(DEFAULT_WRITE_TIMEOUT);
 
diff --git a/crates/nvisy-server/src/handler/request/analyzer.rs b/crates/nvisy-server/src/handler/request/analyzer.rs
new file mode 100644
index 000000000..5ba1b25fe
--- /dev/null
+++ b/crates/nvisy-server/src/handler/request/analyzer.rs
@@ -0,0 +1,285 @@
+//! Per-request analyzer params: layered overrides on top of the
+//! deployment's [`AnalyzerParams`] default.
+//!
+//! Every override field defaults to "inherit"; clients say only
+//! what they want different from the deployment default. The
+//! resolution method folds the two into a final
+//! [`AnalyzerParams`] handed to the engine. Total — no failure
+//! mode; the engine compile validates the resolved spec.
+//!
+//! ## Override kinds
+//!
+//! - [`ScalarOverride<T>`] wraps a single-value field. Three
+//!   variants: `Inherit` (use default), `Replace { value }`
+//!   (set to value), `Remove` (clear an `Option` slot, even if
+//!   the default had it set).
+//! - [`CollectionOverride<T, S>`] wraps a `Vec<T>` field. Three
+//!   variants: `Inherit`, `Replace { values }`, and
+//!   `Patch { extend, remove }` — filter the default by removing
+//!   matching items via [`S`] selectors, then append `extend`.
+//!
+//! ## Wire shape per field
+//!
+//! - `recognizers` ([`RecognizerOverrides`]): nested struct
+//!   carrying one override per kind. `pattern` is at-most-one
+//!   (scalar); `ner` and `llm` are lists with selector-based
+//!   patch semantics.
+//! - `enrichers` ([`EnricherOverrides`]): nested struct, each
+//!   slot at-most-one (scalar). Slots: `language`, `ocr`, `stt`.
+//! - `deduplication`, `scope`: scalars.
+//! - `labelCatalog`: collection keyed by label name.
+
+use nvisy_core::plan::{
+    AnalyzerParams, DeduplicationParams, EnricherParams, LanguageEnricherParams,
+    LlmRecognizerParams, NerRecognizerParams, OcrEnricherParams, PatternRecognizerParams,
+    RecognizerParams, ScopeParams, SttEnricherParams,
+};
+use nvisy_core::schema::LabelSchema;
+use schemars::JsonSchema;
+use serde::Deserialize;
+
+/// Per-request analyzer params. Every field defaults to
+/// [`ScalarOverride::Inherit`] / [`CollectionOverride::Inherit`];
+/// the request may set any subset.
+#[derive(Debug, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct AnalyzerOverrides {
+    /// Per-kind recognizer overrides.
+    #[serde(default)]
+    pub recognizers: RecognizerOverrides,
+    /// Per-kind enricher overrides.
+    #[serde(default)]
+    pub enrichers: EnricherOverrides,
+    /// Deduplication pipeline. Scalar — replace or inherit.
+    #[serde(default)]
+    pub deduplication: ScalarOverride<DeduplicationParams>,
+    /// Caller-asserted scope. Scalar — replace or inherit.
+    #[serde(default)]
+    pub scope: ScalarOverride<ScopeParams>,
+    /// Per-request label catalog.
+    #[serde(default)]
+    pub label_catalog: CollectionOverride<LabelSchema, LabelSelector>,
+}
+
+impl Default for AnalyzerOverrides {
+    fn default() -> Self {
+        Self {
+            recognizers: RecognizerOverrides::default(),
+            enrichers: EnricherOverrides::default(),
+            deduplication: ScalarOverride::Inherit,
+            scope: ScalarOverride::Inherit,
+            label_catalog: CollectionOverride::Inherit,
+        }
+    }
+}
+
+impl AnalyzerOverrides {
+    /// Fold the request's overrides into the deployment default
+    /// to produce the final params handed to the engine.
+    pub fn resolve(self, default: &AnalyzerParams) -> AnalyzerParams {
+        AnalyzerParams {
+            recognizers: self.recognizers.resolve(&default.recognizers),
+            enrichers: self.enrichers.resolve(&default.enrichers),
+            deduplication: self.deduplication.resolve(&default.deduplication),
+            scope: self.scope.resolve(&default.scope),
+            label_catalog: self
+                .label_catalog
+                .resolve(&default.label_catalog, label_matches),
+        }
+    }
+}
+
+/// Per-kind overrides on the recognizer slots of
+/// [`RecognizerParams`]. Pattern is scalar (at-most-one); NER
+/// and LLM are collections.
+#[derive(Debug, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct RecognizerOverrides {
+    /// Pattern recognizer slot. Scalar.
+    #[serde(default)]
+    pub pattern: ScalarOverride<PatternRecognizerParams>,
+    /// NER recognizer list. Selectors match by `name`.
+    #[serde(default)]
+    pub ner: CollectionOverride<NerRecognizerParams, NerSelector>,
+    /// LLM recognizer list. Selectors match by `name`.
+    #[serde(default)]
+    pub llm: CollectionOverride<LlmRecognizerParams, LlmSelector>,
+}
+
+impl Default for RecognizerOverrides {
+    fn default() -> Self {
+        Self {
+            pattern: ScalarOverride::Inherit,
+            ner: CollectionOverride::Inherit,
+            llm: CollectionOverride::Inherit,
+        }
+    }
+}
+
+impl RecognizerOverrides {
+    fn resolve(self, default: &RecognizerParams) -> RecognizerParams {
+        RecognizerParams {
+            pattern: self.pattern.resolve_optional(default.pattern.as_ref()),
+            ner: self.ner.resolve(&default.ner, ner_matches),
+            llm: self.llm.resolve(&default.llm, llm_matches),
+        }
+    }
+}
+
+/// Per-kind overrides on the enricher slots of
+/// [`EnricherParams`]. Every kind is scalar (at-most-one).
+#[derive(Debug, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct EnricherOverrides {
+    /// Language enricher slot.
+    #[serde(default)]
+    pub language: ScalarOverride<LanguageEnricherParams>,
+    /// OCR enricher slot (image modality only).
+    #[serde(default)]
+    pub ocr: ScalarOverride<OcrEnricherParams>,
+    /// STT enricher slot (audio modality only).
+    #[serde(default)]
+    pub stt: ScalarOverride<SttEnricherParams>,
+}
+
+impl Default for EnricherOverrides {
+    fn default() -> Self {
+        Self {
+            language: ScalarOverride::Inherit,
+            ocr: ScalarOverride::Inherit,
+            stt: ScalarOverride::Inherit,
+        }
+    }
+}
+
+impl EnricherOverrides {
+    fn resolve(self, default: &EnricherParams) -> EnricherParams {
+        EnricherParams {
+            language: self.language.resolve_optional(default.language.as_ref()),
+            ocr: self.ocr.resolve_optional(default.ocr.as_ref()),
+            stt: self.stt.resolve_optional(default.stt.as_ref()),
+        }
+    }
+}
+
+/// Per-field override for a scalar (single-value) field.
+#[derive(Debug, Default, Deserialize, JsonSchema)]
+#[serde(tag = "mode", rename_all = "snake_case")]
+pub enum ScalarOverride<T> {
+    /// Use the server default's value for this field.
+    #[default]
+    Inherit,
+    /// Replace the server default's value entirely.
+    Replace {
+        /// The value to use.
+        value: T,
+    },
+    /// Clear an `Option` slot, even if the default had it set.
+    /// On non-`Option` slots (`deduplication`, `scope`) `Remove`
+    /// is equivalent to `Inherit` — there's nothing to clear.
+    Remove,
+}
+
+impl<T: Clone> ScalarOverride<T> {
+    /// Resolve against a *required* default. `Remove` falls back
+    /// to `Inherit` semantics (the slot is not optional, so
+    /// clearing it isn't meaningful).
+    fn resolve(self, default: &T) -> T {
+        match self {
+            Self::Inherit | Self::Remove => default.clone(),
+            Self::Replace { value } => value,
+        }
+    }
+
+    /// Resolve against an *optional* slot.
+    fn resolve_optional(self, default: Option<&T>) -> Option<T> {
+        match self {
+            Self::Inherit => default.cloned(),
+            Self::Replace { value } => Some(value),
+            Self::Remove => None,
+        }
+    }
+}
+
+/// Per-field override for a collection (`Vec<T>`) field.
+#[derive(Debug, Default, Deserialize, JsonSchema)]
+#[serde(tag = "mode", rename_all = "snake_case")]
+pub enum CollectionOverride<T, S> {
+    /// Use the server default's collection.
+    #[default]
+    Inherit,
+    /// Replace the server default's collection entirely.
+    Replace {
+        /// The values to use.
+        values: Vec<T>,
+    },
+    /// Filter the server default by removing matching entries
+    /// (selectors), then append `extend`. Both arrays are
+    /// required on the wire (use `[]` for "no items"); the
+    /// derive can't supply a default-on-omit without forcing
+    /// `Default` bounds onto `T` / `S` we don't want.
+    Patch {
+        /// Items to add after applying `remove`.
+        extend: Vec<T>,
+        /// Selectors that drop matching items from the
+        /// inherited collection.
+        remove: Vec<S>,
+    },
+}
+
+impl<T: Clone, S> CollectionOverride<T, S> {
+    fn resolve(self, default: &[T], matches: impl Fn(&S, &T) -> bool) -> Vec<T> {
+        match self {
+            Self::Inherit => default.to_vec(),
+            Self::Replace { values } => values,
+            Self::Patch { extend, remove } => {
+                let mut out: Vec<T> = default
+                    .iter()
+                    .filter(|item| !remove.iter().any(|sel| matches(sel, item)))
+                    .cloned()
+                    .collect();
+                out.extend(extend);
+                out
+            }
+        }
+    }
+}
+
+/// Selector for [`CollectionOverride::Patch::remove`] on the
+/// `ner` list. NER recognizers are keyed by `name`.
+#[derive(Debug, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct NerSelector {
+    /// NER recognizer name to match.
+    pub name: String,
+}
+
+/// Selector for [`CollectionOverride::Patch::remove`] on the
+/// `llm` list. LLM recognizers are keyed by `name`.
+#[derive(Debug, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct LlmSelector {
+    /// LLM recognizer name to match.
+    pub name: String,
+}
+
+/// Selector for [`CollectionOverride::Patch::remove`] on the
+/// `labelCatalog` field. Labels are keyed by `name`.
+#[derive(Debug, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct LabelSelector {
+    /// Label name to match.
+    pub name: String,
+}
+
+fn ner_matches(sel: &NerSelector, spec: &NerRecognizerParams) -> bool {
+    sel.name == spec.name
+}
+
+fn llm_matches(sel: &LlmSelector, spec: &LlmRecognizerParams) -> bool {
+    sel.name == spec.name
+}
+
+fn label_matches(sel: &LabelSelector, label: &LabelSchema) -> bool {
+    sel.name == label.name
+}
diff --git a/crates/nvisy-server/src/handler/request/contexts.rs b/crates/nvisy-server/src/handler/request/contexts.rs
new file mode 100644
index 000000000..4582c5899
--- /dev/null
+++ b/crates/nvisy-server/src/handler/request/contexts.rs
@@ -0,0 +1,12 @@
+//! Request bodies for `/contexts` endpoints.
+
+use nvisy_core::context::Context;
+use schemars::JsonSchema;
+use serde::Deserialize;
+
+/// Body for `POST /contexts`. Same shape as `NewPolicy` —
+/// caller-supplied id + version; the server keys storage by
+/// `(actor_id, context.id, context.version)`.
+#[derive(Debug, Deserialize, JsonSchema)]
+#[serde(transparent)]
+pub struct NewContext(pub Context);
diff --git a/crates/nvisy-server/src/handler/request/detections.rs b/crates/nvisy-server/src/handler/request/detections.rs
index 379c3f049..687681291 100644
--- a/crates/nvisy-server/src/handler/request/detections.rs
+++ b/crates/nvisy-server/src/handler/request/detections.rs
@@ -1,36 +1,60 @@
 //! Request bodies for `/detections` endpoints.
 
-use nvisy_engine::core::ingestion::ImportFile;
-use nvisy_engine::detection::{DetectionInput, DetectionPlan, DetectionStatus};
-use nvisy_engine::policy::Policy;
+use std::collections::HashMap;
+
+use nvisy_core::plan::AnalyzerParams;
+use nvisy_engine::runs::{DocumentInput, StartBatch};
 use schemars::JsonSchema;
 use serde::Deserialize;
 use uuid::Uuid;
 
-use crate::handler::request::pagination::Pagination;
+use super::analyzer::AnalyzerOverrides;
+use super::pagination::Pagination;
+use super::refs::ResourceRef;
 
-/// Body for `POST /detections`.
-#[derive(Debug, Deserialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
+/// Body for `POST /detections`. Documents are file ids
+/// previously uploaded via `POST /files`; the server resolves
+/// them at start time.
+///
+/// The `analyzer` field carries per-request *overrides* on top
+/// of the deployment's default [`AnalyzerParams`] — clients that
+/// omit it inherit the default. See [`AnalyzerOverrides`] for the
+/// per-field merge semantics.
+#[derive(Debug, Default, Deserialize, JsonSchema)]
+#[serde(default, rename_all = "camelCase")]
 pub struct NewDetection {
-    /// Policies to apply, in precedence order
-    /// (index 0 is highest precedence).
-    pub policies: Vec<Policy>,
-    /// Content sources to ingest at the start of the pass.
-    pub imports: Vec<ImportFile>,
-    /// Per-phase behaviour knobs.
-    #[serde(default)]
-    pub plan: DetectionPlan,
+    /// Policies to apply, in precedence order.
+    pub policy_refs: Vec<ResourceRef>,
+    /// Contexts to apply.
+    pub context_refs: Vec<ResourceRef>,
+    /// File ids to analyze. Must exist for the calling actor.
+    pub documents: Vec<Uuid>,
+    /// Per-request metadata merged with each document's
+    /// descriptor at policy-evaluation time.
+    pub metadata: HashMap<String, String>,
+    /// Analyzer overrides. Field-by-field merge into the
+    /// deployment default; omitted fields inherit the default.
+    pub analyzer: AnalyzerOverrides,
+    /// Per-doc concurrency cap. `None` for the engine default.
+    pub concurrency: Option<usize>,
 }
 
 impl NewDetection {
-    /// Project the request into the typed engine input.
-    pub fn into_engine_input(self, actor_id: Uuid) -> DetectionInput {
-        DetectionInput {
-            actor_id,
-            policies: self.policies,
-            imports: self.imports,
-            plan: self.plan,
+    /// Project the request into the typed engine input,
+    /// resolving the analyzer params against the server's
+    /// configured default.
+    pub fn into_engine_input(self, analyzer_default: &AnalyzerParams) -> StartBatch {
+        StartBatch {
+            policy_refs: self.policy_refs.into_iter().map(Into::into).collect(),
+            context_refs: self.context_refs.into_iter().map(Into::into).collect(),
+            documents: self
+                .documents
+                .into_iter()
+                .map(|file_id| DocumentInput { file_id })
+                .collect(),
+            metadata: self.metadata,
+            analyzer: self.analyzer.resolve(analyzer_default),
+            concurrency: self.concurrency,
         }
     }
 }
@@ -39,9 +63,6 @@ impl NewDetection {
 #[derive(Debug, Deserialize, JsonSchema)]
 #[serde(rename_all = "camelCase")]
 pub struct DetectionQuery {
-    /// Optional status filter.
-    #[serde(default)]
-    pub status: Option<DetectionStatus>,
     /// Pagination knobs.
     #[serde(flatten)]
     pub pagination: Pagination,
diff --git a/crates/nvisy-server/src/handler/request/files.rs b/crates/nvisy-server/src/handler/request/files.rs
new file mode 100644
index 000000000..954de19b7
--- /dev/null
+++ b/crates/nvisy-server/src/handler/request/files.rs
@@ -0,0 +1,17 @@
+//! Query parameters for `/files` endpoints. (Upload body is
+//! raw bytes; descriptor fields come via HTTP headers
+//! `Content-Type` + `Content-Disposition`.)
+
+use schemars::JsonSchema;
+use serde::Deserialize;
+
+use super::pagination::Pagination;
+
+/// Query parameters for `GET /files`.
+#[derive(Debug, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct FileQuery {
+    /// Pagination knobs.
+    #[serde(flatten)]
+    pub pagination: Pagination,
+}
diff --git a/crates/nvisy-server/src/handler/request/mod.rs b/crates/nvisy-server/src/handler/request/mod.rs
index 592912535..71ff9ee01 100644
--- a/crates/nvisy-server/src/handler/request/mod.rs
+++ b/crates/nvisy-server/src/handler/request/mod.rs
@@ -6,12 +6,23 @@
 //! [`Deserialize`]: serde::Deserialize
 //! [`JsonSchema`]: schemars::JsonSchema
 
+pub mod analyzer;
+mod contexts;
 mod detections;
+mod files;
 mod pagination;
 mod path;
+mod policies;
 mod redactions;
+mod refs;
 
+pub use self::contexts::NewContext;
 pub use self::detections::{DetectionQuery, NewDetection};
+pub use self::files::FileQuery;
 pub use self::pagination::{MAX_PAGE_LIMIT, Pagination};
-pub use self::path::{ContentPath, DetectionPath, RedactionPath};
-pub use self::redactions::{NewRedaction, RedactionQuery};
+pub use self::path::{
+    ContextIdPath, ContextVersionPath, DetectionPath, FilePath, PolicyIdPath, PolicyVersionPath,
+    RedactionPath,
+};
+pub use self::policies::NewPolicy;
+pub use self::redactions::NewRedaction;
diff --git a/crates/nvisy-server/src/handler/request/path.rs b/crates/nvisy-server/src/handler/request/path.rs
index 51c490f40..8d22c1660 100644
--- a/crates/nvisy-server/src/handler/request/path.rs
+++ b/crates/nvisy-server/src/handler/request/path.rs
@@ -1,30 +1,68 @@
 //! Typed path parameters for API endpoints.
 //!
 //! One newtype per resource so the OpenAPI schema documents each
-//! distinct parameter shape (aide collapses type aliases). Every
-//! one carries a single UUID `id`.
+//! distinct parameter shape (aide collapses type aliases).
 
 use schemars::JsonSchema;
+use semver::Version;
 use serde::Deserialize;
 use uuid::Uuid;
 
 /// Path parameter for `/files/{id}` endpoints.
 #[derive(Debug, Deserialize, JsonSchema)]
-pub struct ContentPath {
-    /// Content identifier.
+pub struct FilePath {
+    /// File identifier.
     pub id: Uuid,
 }
 
-/// Path parameter for `/detections/{id}` endpoints.
+/// Path parameter for `/detections/{id}` endpoints. Path id is
+/// the underlying run id — `/detections` and `/redactions` are
+/// filtered views of the same `runs` keyspace.
 #[derive(Debug, Deserialize, JsonSchema)]
 pub struct DetectionPath {
-    /// Detection identifier.
+    /// Run identifier.
     pub id: Uuid,
 }
 
 /// Path parameter for `/redactions/{id}` endpoints.
 #[derive(Debug, Deserialize, JsonSchema)]
 pub struct RedactionPath {
-    /// Redaction identifier.
+    /// Run identifier.
+    pub id: Uuid,
+}
+
+/// Path parameter for `/policies/{id}/{version}` endpoints.
+#[derive(Debug, Deserialize, JsonSchema)]
+pub struct PolicyVersionPath {
+    /// Policy identifier.
+    pub id: Uuid,
+    /// Semver version.
+    #[schemars(with = "String")]
+    pub version: Version,
+}
+
+/// Path parameter for `/policies/{id}/latest` and
+/// `/policies/{id}` listing-by-id endpoints.
+#[derive(Debug, Deserialize, JsonSchema)]
+pub struct PolicyIdPath {
+    /// Policy identifier.
+    pub id: Uuid,
+}
+
+/// Path parameter for `/contexts/{id}/{version}` endpoints.
+#[derive(Debug, Deserialize, JsonSchema)]
+pub struct ContextVersionPath {
+    /// Context identifier.
+    pub id: Uuid,
+    /// Semver version.
+    #[schemars(with = "String")]
+    pub version: Version,
+}
+
+/// Path parameter for `/contexts/{id}/latest` and
+/// `/contexts/{id}` listing-by-id endpoints.
+#[derive(Debug, Deserialize, JsonSchema)]
+pub struct ContextIdPath {
+    /// Context identifier.
     pub id: Uuid,
 }
diff --git a/crates/nvisy-server/src/handler/request/policies.rs b/crates/nvisy-server/src/handler/request/policies.rs
new file mode 100644
index 000000000..5bf334c43
--- /dev/null
+++ b/crates/nvisy-server/src/handler/request/policies.rs
@@ -0,0 +1,13 @@
+//! Request bodies for `/policies` endpoints.
+
+use nvisy_core::policy::Policy;
+use schemars::JsonSchema;
+use serde::Deserialize;
+
+/// Body for `POST /policies`. The full [`Policy`] inline — id +
+/// version are caller-supplied (UUIDv7 + semver). Server stores
+/// it under `(actor_id, policy.id, policy.version)`; re-posting
+/// the same `(id, version)` returns `Conflict`.
+#[derive(Debug, Deserialize, JsonSchema)]
+#[serde(transparent)]
+pub struct NewPolicy(pub Policy);
diff --git a/crates/nvisy-server/src/handler/request/redactions.rs b/crates/nvisy-server/src/handler/request/redactions.rs
index bd99fa289..2f28605be 100644
--- a/crates/nvisy-server/src/handler/request/redactions.rs
+++ b/crates/nvisy-server/src/handler/request/redactions.rs
@@ -1,53 +1,38 @@
 //! Request bodies for `/redactions` endpoints.
 
-use nvisy_engine::core::ingestion::ExportFile;
-use nvisy_engine::redaction::{RedactionInput, RedactionOverride, RedactionPlan, RedactionStatus};
+use nvisy_core::policy::RuleAction;
 use schemars::JsonSchema;
 use serde::Deserialize;
 use uuid::Uuid;
 
-use crate::handler::request::pagination::Pagination;
-
-/// Body for `POST /redactions`.
+/// Body for `POST /redactions`. References a prior detection
+/// (the same run id) and optionally carries reviewer overrides
+/// to apply per-entity before the redaction runs.
 #[derive(Debug, Deserialize, JsonSchema)]
 #[serde(rename_all = "camelCase")]
 pub struct NewRedaction {
-    /// Detection pass this redaction targets.
+    /// Detection (run) to apply. The redaction reuses the same
+    /// underlying run; the resulting redaction id equals
+    /// [`detection_id`](Self::detection_id).
     pub detection_id: Uuid,
-    /// Per-entity decision overrides.
-    #[serde(default)]
-    pub overrides: Vec<RedactionOverride>,
-    /// Per-phase behaviour knobs (validation thresholds etc.).
+    /// Per-entity decision overrides. Applied in order before
+    /// the redaction fan-out; an override on an entity id that
+    /// doesn't exist in the run returns `NotFound`.
     #[serde(default)]
-    pub plan: RedactionPlan,
-    /// Sinks to write redacted content to.
-    #[serde(default)]
-    pub exports: Vec<ExportFile>,
-}
-
-impl NewRedaction {
-    pub fn into_engine_input(self, actor_id: Uuid) -> RedactionInput {
-        RedactionInput {
-            actor_id,
-            detection_id: self.detection_id,
-            overrides: self.overrides,
-            plan: self.plan,
-            exports: self.exports,
-        }
-    }
+    pub overrides: Vec<NewOverride>,
 }
 
-/// Query parameters for `GET /redactions`.
-#[derive(Debug, Deserialize, JsonSchema)]
+/// One reviewer override: a per-entity action decided before
+/// the redaction runs. The server loops these into
+/// [`nvisy_engine::runs::override_entity`] before the apply
+/// transition.
+#[derive(Debug, Clone, Deserialize, JsonSchema)]
 #[serde(rename_all = "camelCase")]
-pub struct RedactionQuery {
-    /// Optional status filter.
-    #[serde(default)]
-    pub status: Option<RedactionStatus>,
-    /// Optional filter: only redactions for this detection.
-    #[serde(default)]
-    pub detection_id: Option<Uuid>,
-    /// Pagination knobs.
-    #[serde(flatten)]
-    pub pagination: Pagination,
+pub struct NewOverride {
+    /// Document id within the run.
+    pub doc_id: Uuid,
+    /// Entity id within the document.
+    pub entity_id: Uuid,
+    /// The action to apply.
+    pub action: RuleAction,
 }
diff --git a/crates/nvisy-server/src/handler/request/refs.rs b/crates/nvisy-server/src/handler/request/refs.rs
new file mode 100644
index 000000000..8060fb15b
--- /dev/null
+++ b/crates/nvisy-server/src/handler/request/refs.rs
@@ -0,0 +1,31 @@
+//! `(id, version)` reference shape used by `POST /detections`
+//! for `policyRefs` and `contextRefs`. JsonSchema-wrapped mirror
+//! of [`nvisy_engine::runs::ResourceRef`] (engine type doesn't
+//! derive `JsonSchema`).
+
+use nvisy_engine::runs::ResourceRef as EngineRef;
+use schemars::JsonSchema;
+use semver::Version;
+use serde::Deserialize;
+use uuid::Uuid;
+
+/// Reference to a stored policy or context resource by
+/// `(id, version)`.
+#[derive(Debug, Clone, Deserialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct ResourceRef {
+    /// Resource UUID.
+    pub id: Uuid,
+    /// Resource version.
+    #[schemars(with = "String")]
+    pub version: Version,
+}
+
+impl From<ResourceRef> for EngineRef {
+    fn from(r: ResourceRef) -> Self {
+        EngineRef {
+            id: r.id,
+            version: r.version,
+        }
+    }
+}
diff --git a/crates/nvisy-server/src/handler/response/check.rs b/crates/nvisy-server/src/handler/response/check.rs
index 8b8a37d0d..85e46dbf9 100644
--- a/crates/nvisy-server/src/handler/response/check.rs
+++ b/crates/nvisy-server/src/handler/response/check.rs
@@ -1,7 +1,7 @@
 //! Health response body.
 
 use jiff::Timestamp;
-use nvisy_core::health::{ComponentCheck, ServiceStatus};
+use nvisy_core::service::health::{ComponentCheck, ServiceStatus};
 use schemars::JsonSchema;
 use serde::Serialize;
 
diff --git a/crates/nvisy-server/src/handler/response/contexts.rs b/crates/nvisy-server/src/handler/response/contexts.rs
new file mode 100644
index 000000000..a1b0d6011
--- /dev/null
+++ b/crates/nvisy-server/src/handler/response/contexts.rs
@@ -0,0 +1,18 @@
+//! Context response shapes. Symmetric with [`super::policies`].
+
+use schemars::JsonSchema;
+use semver::Version;
+use serde::Serialize;
+use uuid::Uuid;
+
+/// Summary returned by `POST /contexts` and on each entry of
+/// `GET /contexts` list responses.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct ContextSummary {
+    /// Context id.
+    pub id: Uuid,
+    /// Context version.
+    #[schemars(with = "String")]
+    pub version: Version,
+}
diff --git a/crates/nvisy-server/src/handler/response/detections.rs b/crates/nvisy-server/src/handler/response/detections.rs
index e756186cf..9f2269402 100644
--- a/crates/nvisy-server/src/handler/response/detections.rs
+++ b/crates/nvisy-server/src/handler/response/detections.rs
@@ -1,19 +1,19 @@
-//! Detection response types.
+//! Detection response shapes.
+//!
+//! Most `/detections` reads return the shared [`RunResponse`]
+//! (header + per-doc bodies); see [`super::runs`]. This module
+//! holds the only detection-specific shape, the just-created
+//! id returned by `POST /detections`.
 
-use nvisy_engine::detection::DetectionEntry;
 use schemars::JsonSchema;
 use serde::Serialize;
 use uuid::Uuid;
 
-use super::page::Page;
-
-/// Response body for `GET /detections`.
-pub type DetectionList = Page<DetectionEntry>;
-
-/// Response body for `POST /detections`.
+/// Returned by `POST /detections`. The id is the new run id —
+/// the same id the matching `POST /redactions` will reference.
 #[derive(Debug, Serialize, JsonSchema)]
 #[serde(rename_all = "camelCase")]
 pub struct DetectionId {
-    /// Identifier assigned to the submitted detection pass.
+    /// Engine-minted UUIDv7.
     pub id: Uuid,
 }
diff --git a/crates/nvisy-server/src/handler/response/files.rs b/crates/nvisy-server/src/handler/response/files.rs
index ef4057ab5..b5ceccfd2 100644
--- a/crates/nvisy-server/src/handler/response/files.rs
+++ b/crates/nvisy-server/src/handler/response/files.rs
@@ -1,38 +1,20 @@
-//! File response types.
+//! File response shapes.
 
-use nvisy_engine::{ContentDescriptor, ContentDigest};
+use nvisy_core::FileMetadata;
 use schemars::JsonSchema;
 use serde::Serialize;
 use uuid::Uuid;
 
-use super::page::Page;
-
-/// Response body for `GET /files`.
-pub type FileList = Page<FileMetadata>;
-
-/// Metadata for a stored file. Returned by `GET /files/{id}` and
-/// per-item by `GET /files`.
-///
-/// Annotations live at the separate `GET /files/{id}/annotations`
-/// subresource and are not embedded here — bundling them would
-/// couple the cacheability of the immutable descriptor/digest to
-/// every annotation edit.
-#[derive(Debug, Serialize, JsonSchema)]
-#[serde(rename_all = "camelCase")]
-pub struct FileMetadata {
-    /// Identifier of the file.
-    pub id: Uuid,
-    /// Caller-supplied descriptor (filename, MIME hint, policy
-    /// metadata).
-    pub descriptor: ContentDescriptor,
-    /// Byte-derived digest (size, sha256, sniffed MIME).
-    pub digest: ContentDigest,
-}
-
-/// Response body for `POST /files`.
+/// Just-uploaded id; returned by `POST /files` so clients can
+/// reference the file in subsequent calls without parsing the
+/// metadata.
 #[derive(Debug, Serialize, JsonSchema)]
 #[serde(rename_all = "camelCase")]
 pub struct FileId {
-    /// Identifier assigned to the uploaded file.
+    /// Engine-minted UUIDv7.
     pub id: Uuid,
 }
+
+/// Re-export of [`FileMetadata`] as the response shape (already
+/// has `Serialize + JsonSchema` via nvisy-core).
+pub type FileMetadataResponse = FileMetadata;
diff --git a/crates/nvisy-server/src/handler/response/mod.rs b/crates/nvisy-server/src/handler/response/mod.rs
index 8f02e6002..e7d2b49fd 100644
--- a/crates/nvisy-server/src/handler/response/mod.rs
+++ b/crates/nvisy-server/src/handler/response/mod.rs
@@ -1,22 +1,33 @@
 //! Typed response bodies for API endpoints.
 //!
-//! Each struct derives [`Serialize`] and
-//! [`JsonSchema`] for automatic OpenAPI schema
-//! generation via aide.
+//! Each struct derives [`Serialize`] and [`JsonSchema`] for
+//! automatic OpenAPI schema generation via aide.
+//!
+//! Per-run wrappers (the per-modality entity records, the
+//! [`runs::RunResponse`] assembler, etc.) live in
+//! [`runs`] — a public submodule so the CLI and other
+//! external consumers can name them on their own request /
+//! response types without going through `pub use`.
 //!
 //! [`Serialize`]: serde::Serialize
 //! [`JsonSchema`]: schemars::JsonSchema
 
 mod check;
+mod contexts;
 mod detections;
 mod error;
 mod files;
 mod page;
+mod policies;
 mod redactions;
+pub mod runs;
 
 pub use self::check::Health;
-pub use self::detections::{DetectionId, DetectionList};
+pub use self::contexts::ContextSummary;
+pub use self::detections::DetectionId;
 pub use self::error::ErrorResponse;
-pub use self::files::{FileId, FileList, FileMetadata};
+pub use self::files::{FileId, FileMetadataResponse};
 pub use self::page::Page;
-pub use self::redactions::{RedactionId, RedactionList};
+pub use self::policies::PolicySummary;
+pub use self::redactions::{RedactionOutput, RedactionResult};
+pub use self::runs::RunResponse;
diff --git a/crates/nvisy-server/src/handler/response/page.rs b/crates/nvisy-server/src/handler/response/page.rs
index 8e5c14070..0ee7a8f28 100644
--- a/crates/nvisy-server/src/handler/response/page.rs
+++ b/crates/nvisy-server/src/handler/response/page.rs
@@ -1,20 +1,11 @@
 //! Paginated response wrapper.
 
-use nvisy_engine::registry::PagedResult;
 use schemars::JsonSchema;
 use serde::Serialize;
 
 use crate::handler::request::Pagination;
 
 /// Paginated response wrapper used by every list endpoint.
-///
-/// Two construction paths depending on where the data comes from:
-///
-/// - [`Page::paginate`] — for in-memory collections that have
-///   already been materialised + sorted. Slices the full `Vec`.
-/// - [`Page::from_paged`] — for registry-backed lists where the
-///   storage layer did the windowing. Wraps a [`PagedResult`] +
-///   computes `has_more` from `offset + items.len() < total`.
 #[derive(Debug, Serialize, JsonSchema)]
 #[serde(rename_all = "camelCase")]
 pub struct Page<T: Serialize + JsonSchema> {
@@ -31,11 +22,6 @@ impl<T: Serialize + JsonSchema> Page<T> {
     /// a flat collection. `limit` is clamped to [`MAX_PAGE_LIMIT`]
     /// to bound memory use.
     ///
-    /// Use [`Self::from_paged`] when the storage layer already did
-    /// the windowing; this method is for in-memory collections that
-    /// must be materialised in full (typically because they require
-    /// sorting before slicing).
-    ///
     /// [`MAX_PAGE_LIMIT`]: crate::handler::request::MAX_PAGE_LIMIT
     pub fn paginate(items: Vec<T>, pagination: &Pagination) -> Self {
         let limit = pagination
@@ -53,20 +39,4 @@ impl<T: Serialize + JsonSchema> Page<T> {
             items,
         }
     }
-
-    /// Wrap a registry-side [`PagedResult`]. Maps each item through
-    /// `f` (typically projecting a storage row to a wire summary)
-    /// and computes `has_more` from the window position.
-    pub fn from_paged<U>(
-        paged: PagedResult<U>,
-        pagination: &Pagination,
-        f: impl FnMut(U) -> T,
-    ) -> Self {
-        let items: Vec<T> = paged.items.into_iter().map(f).collect();
-        Self {
-            has_more: pagination.offset + items.len() < paged.total,
-            total: paged.total,
-            items,
-        }
-    }
 }
diff --git a/crates/nvisy-server/src/handler/response/policies.rs b/crates/nvisy-server/src/handler/response/policies.rs
new file mode 100644
index 000000000..0c256d366
--- /dev/null
+++ b/crates/nvisy-server/src/handler/response/policies.rs
@@ -0,0 +1,23 @@
+//! Policy response shapes.
+//!
+//! Reads return the [`nvisy_core::policy::Policy`] type
+//! directly (already `Serialize + JsonSchema`). `POST /policies`
+//! returns the `(id, version)` summary so clients can reference
+//! the policy without parsing the full body.
+
+use schemars::JsonSchema;
+use semver::Version;
+use serde::Serialize;
+use uuid::Uuid;
+
+/// Summary returned by `POST /policies` and on each entry of
+/// `GET /policies` list responses.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct PolicySummary {
+    /// Policy id.
+    pub id: Uuid,
+    /// Policy version.
+    #[schemars(with = "String")]
+    pub version: Version,
+}
diff --git a/crates/nvisy-server/src/handler/response/redactions.rs b/crates/nvisy-server/src/handler/response/redactions.rs
index 5fc9ee1b5..3a62eef1e 100644
--- a/crates/nvisy-server/src/handler/response/redactions.rs
+++ b/crates/nvisy-server/src/handler/response/redactions.rs
@@ -1,19 +1,41 @@
-//! Redaction response types.
+//! Redaction response shapes.
 
-use nvisy_engine::redaction::RedactionEntry;
 use schemars::JsonSchema;
 use serde::Serialize;
 use uuid::Uuid;
 
-use super::page::Page;
+use super::runs::RunDocStateDto;
 
-/// Response body for `GET /redactions`.
-pub type RedactionList = Page<RedactionEntry>;
-
-/// Response body for `POST /redactions`.
+/// Returned by `POST /redactions`. The id is the same run id
+/// as the matching detection. `outputs` carries one entry per
+/// input file in the run, so clients see which input produced
+/// which redacted file and which per-doc applies failed.
 #[derive(Debug, Serialize, JsonSchema)]
 #[serde(rename_all = "camelCase")]
-pub struct RedactionId {
-    /// Identifier assigned to the submitted redaction pass.
+pub struct RedactionResult {
+    /// Run id (= detection id).
     pub id: Uuid,
+    /// One per input file. Position matches input order.
+    pub outputs: Vec<RedactionOutput>,
+}
+
+/// One per-doc apply outcome.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct RedactionOutput {
+    /// Doc id within the run.
+    pub doc_id: Uuid,
+    /// Input file the apply read.
+    pub input_file_id: Uuid,
+    /// Redacted output file, when this doc applied successfully.
+    /// `None` when the doc failed — see
+    /// [`state`](Self::state) + [`failure_reason`](Self::failure_reason).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub output_file_id: Option<Uuid>,
+    /// Per-doc lifecycle state after apply.
+    pub state: RunDocStateDto,
+    /// Detail when [`state`](Self::state) is
+    /// [`RunDocStateDto::Failed`].
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub failure_reason: Option<String>,
 }
diff --git a/crates/nvisy-server/src/handler/response/runs.rs b/crates/nvisy-server/src/handler/response/runs.rs
new file mode 100644
index 000000000..36af99885
--- /dev/null
+++ b/crates/nvisy-server/src/handler/response/runs.rs
@@ -0,0 +1,520 @@
+//! Run + per-document response wrappers.
+//!
+//! Engine types ([`nvisy_engine::runs::Run`],
+//! [`RunDocument`], [`DocBody`], [`EntityRecord<M>`]) don't
+//! derive `JsonSchema`. The server defines flat per-modality
+//! mirror DTOs so the OpenAPI spec fully describes the wire
+//! shape, with `From` conversions from the engine types.
+//!
+//! Provenance (the per-recognizer audit trail) and
+//! `recognized_range` are intentionally dropped from the wire —
+//! they're audit-internal and not on the review UI's critical
+//! path. Add them back when there's a consumer.
+
+use std::collections::HashMap;
+
+use elide_core::entity::Entity;
+use elide_core::modality::audio::{Audio, AudioLocation};
+use elide_core::modality::image::{Image, ImageLocation};
+use elide_core::modality::tabular::{Tabular, TabularLocation};
+use elide_core::modality::text::{Text, TextLocation};
+use jiff::Timestamp;
+use nvisy_core::policy::RuleAction;
+use nvisy_engine::runs::{
+    DocBody, EntityRecord, ModalityKind, Run, RunDocState, RunDocument, RunState,
+};
+use schemars::JsonSchema;
+use semver::Version;
+use serde::Serialize;
+use uuid::Uuid;
+
+/// Run header + every per-document body, packaged as one
+/// response. `GET /detections/{id}` and `GET /redactions/{id}`
+/// both render through this — the caller filters by
+/// [`state`](RunResponse::state) to know which view they're in.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct RunResponse {
+    /// Run id (same as the detection id and the redaction id —
+    /// detections and redactions are filtered views of the same
+    /// underlying run).
+    pub id: Uuid,
+    /// Top-level run state.
+    pub state: RunStateDto,
+    /// Detail when [`state`](Self::state) is
+    /// [`RunStateDto::Failed`].
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub failure_reason: Option<String>,
+    /// UUIDv7 timestamp the run was started.
+    #[schemars(with = "String")]
+    pub started_at: Timestamp,
+    /// UUIDv7 timestamp of the most recent state transition.
+    #[schemars(with = "String")]
+    pub updated_at: Timestamp,
+    /// Policies the caller submitted.
+    pub policy_refs: Vec<ResourceRefDto>,
+    /// Contexts the caller submitted.
+    #[serde(skip_serializing_if = "Vec::is_empty")]
+    pub context_refs: Vec<ResourceRefDto>,
+    /// Per-request metadata.
+    #[serde(skip_serializing_if = "HashMap::is_empty")]
+    pub metadata: HashMap<String, String>,
+    /// Per-document state. One entry per input file in the run.
+    pub documents: Vec<RunDocumentDto>,
+}
+
+/// Wire-format mirror of [`RunState`].
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum RunStateDto {
+    /// Analyze in flight or queued.
+    Analyzing,
+    /// Analyze finished; awaiting reviewer + apply.
+    AwaitingReview,
+    /// Apply ran; every document succeeded.
+    Applied,
+    /// Apply ran; some documents succeeded, others failed.
+    PartiallyApplied,
+    /// Run failed before producing per-doc state worth
+    /// reviewing; the [`failure_reason`](RunResponse::failure_reason)
+    /// field carries the detail.
+    Failed,
+}
+
+/// Wire-format mirror of [`nvisy_engine::runs::ResourceRef`].
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct ResourceRefDto {
+    /// Resource UUID.
+    pub id: Uuid,
+    /// Resource version.
+    #[schemars(with = "String")]
+    pub version: Version,
+}
+
+/// Wire-format mirror of [`RunDocument`]. Inlines the
+/// per-modality body so the response is one flat array of docs
+/// rather than a header + indirected fetches.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct RunDocumentDto {
+    /// Doc id within the run.
+    pub id: Uuid,
+    /// Input file the doc analyzed.
+    pub input_file_id: Uuid,
+    /// Redacted output file, when apply succeeded for this doc.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub output_file_id: Option<Uuid>,
+    /// Per-doc lifecycle state.
+    pub state: RunDocStateDto,
+    /// Detail when [`state`](Self::state) is
+    /// [`RunDocStateDto::Failed`].
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub failure_reason: Option<String>,
+    /// Modality the codec resolved this doc to.
+    pub modality: ModalityDto,
+    /// Recognized entities + reviewer overrides. Variant
+    /// matches [`modality`](Self::modality).
+    pub body: DocBodyDto,
+}
+
+/// Wire-format mirror of [`RunDocState`].
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum RunDocStateDto {
+    /// Awaiting its turn on the analyze semaphore.
+    Queued,
+    /// Analyze in flight.
+    Analyzing,
+    /// Analyze finished; awaiting reviewer + apply.
+    AwaitingReview,
+    /// Apply ran for this doc; the redacted output is in
+    /// [`output_file_id`](RunDocumentDto::output_file_id).
+    Applied,
+    /// Analyze or apply errored.
+    Failed,
+    /// Per-doc timeout fired.
+    TimedOut,
+}
+
+/// Wire-format mirror of [`ModalityKind`].
+#[derive(Debug, Clone, Copy, Serialize, JsonSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum ModalityDto {
+    /// Text.
+    Text,
+    /// Tabular.
+    Tabular,
+    /// Image.
+    Image,
+    /// Audio.
+    Audio,
+}
+
+/// Per-modality body of recognized entities + reviewer
+/// overrides. Tagged by `modality`; variant matches
+/// [`RunDocumentDto::modality`].
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(tag = "modality", rename_all = "snake_case")]
+pub enum DocBodyDto {
+    /// Text entities.
+    Text {
+        /// Recognized entities, in source-coordinate order.
+        entities: Vec<TextEntityRecordDto>,
+    },
+    /// Tabular entities.
+    Tabular {
+        /// Recognized entities.
+        entities: Vec<TabularEntityRecordDto>,
+    },
+    /// Image entities.
+    Image {
+        /// Recognized entities.
+        entities: Vec<ImageEntityRecordDto>,
+    },
+    /// Audio entities.
+    Audio {
+        /// Recognized entities.
+        entities: Vec<AudioEntityRecordDto>,
+    },
+}
+
+/// Shared scalar fields every per-modality entity record
+/// carries.
+fn entity_common<M>(entity: &Entity<M>) -> EntityCommon
+where
+    M: elide_core::modality::Modality,
+{
+    EntityCommon {
+        id: entity.id,
+        label: entity.label.as_str().to_owned(),
+        confidence: f32::from(entity.confidence),
+        coref: entity.coref.as_ref().map(|c| c.as_str().to_owned()),
+        language: entity.language.as_ref().map(|l| l.as_str().to_owned()),
+    }
+}
+
+/// Scalar-only fields shared across the four
+/// per-modality entity DTOs. Inlined into each via
+/// `#[serde(flatten)]`.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct EntityCommon {
+    /// Stable per-entity id.
+    pub id: Uuid,
+    /// Label name (e.g. `"person_name"`, `"email_address"`).
+    pub label: String,
+    /// Confidence in `[0.0, 1.0]`.
+    pub confidence: f32,
+    /// Coreference cluster id, when resolved.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub coref: Option<String>,
+    /// Detected language, when resolved.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub language: Option<String>,
+}
+
+/// Wire-format mirror of [`EntityRecord<Text>`].
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TextEntityRecordDto {
+    /// Shared scalars.
+    #[serde(flatten)]
+    pub common: EntityCommon,
+    /// Byte range in the source text.
+    pub location: TextLocationDto,
+    /// Reviewer override action, when set.
+    #[serde(skip_serializing_if = "Option::is_none", rename = "override")]
+    pub r#override: Option<RuleAction>,
+}
+
+/// Wire-format mirror of [`TextLocation`].
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TextLocationDto {
+    /// Byte offset where the range starts.
+    pub start: usize,
+    /// Byte offset where the range ends (exclusive).
+    pub end: usize,
+    /// 1-based page number, when known.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub page: Option<u32>,
+}
+
+/// Wire-format mirror of [`EntityRecord<Tabular>`].
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TabularEntityRecordDto {
+    /// Shared scalars.
+    #[serde(flatten)]
+    pub common: EntityCommon,
+    /// Cell coordinates.
+    pub location: TabularLocationDto,
+    /// Reviewer override action, when set.
+    #[serde(skip_serializing_if = "Option::is_none", rename = "override")]
+    pub r#override: Option<RuleAction>,
+}
+
+/// Wire-format mirror of [`TabularLocation`].
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct TabularLocationDto {
+    /// Zero-based row index.
+    pub row_index: u32,
+    /// Zero-based column index.
+    pub column_index: u32,
+    /// Byte offset within the cell where the entity starts.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub start_offset: Option<usize>,
+    /// Byte offset within the cell where the entity ends.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub end_offset: Option<usize>,
+    /// Column name from the header row, when known.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub column_name: Option<String>,
+    /// Sheet name, for multi-sheet sources.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub sheet_name: Option<String>,
+}
+
+/// Wire-format mirror of [`EntityRecord<Image>`]. Image
+/// location is the entity's bounding box; the polygon variant
+/// from elide is reduced to its bounding box in the wire format.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct ImageEntityRecordDto {
+    /// Shared scalars.
+    #[serde(flatten)]
+    pub common: EntityCommon,
+    /// Axis-aligned bounding box in pixel coordinates.
+    pub location: ImageLocationDto,
+    /// Reviewer override action, when set.
+    #[serde(skip_serializing_if = "Option::is_none", rename = "override")]
+    pub r#override: Option<RuleAction>,
+}
+
+/// Wire-format mirror of the image bounding box. Coordinates
+/// are floats matching elide's internal precision.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct ImageLocationDto {
+    /// X coordinate of the box's top-left corner.
+    pub x: f64,
+    /// Y coordinate of the box's top-left corner.
+    pub y: f64,
+    /// Box width.
+    pub width: f64,
+    /// Box height.
+    pub height: f64,
+    /// 1-based page number for multi-page documents.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub page: Option<u32>,
+}
+
+/// Wire-format mirror of [`EntityRecord<Audio>`]. Audio
+/// location is a time span over the source stream.
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct AudioEntityRecordDto {
+    /// Shared scalars.
+    #[serde(flatten)]
+    pub common: EntityCommon,
+    /// Time span on the source stream.
+    pub location: AudioLocationDto,
+    /// Reviewer override action, when set.
+    #[serde(skip_serializing_if = "Option::is_none", rename = "override")]
+    pub r#override: Option<RuleAction>,
+}
+
+/// Wire-format mirror of [`AudioLocation`].
+#[derive(Debug, Serialize, JsonSchema)]
+#[serde(rename_all = "camelCase")]
+pub struct AudioLocationDto {
+    /// Start of the span, in milliseconds from the stream
+    /// origin.
+    pub start_ms: u64,
+    /// End of the span, in milliseconds from the stream origin.
+    pub end_ms: u64,
+    /// Diarization label of the speaker, when a diarizer
+    /// assigned one.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub speaker: Option<String>,
+}
+
+// ---- Conversions ----
+
+impl RunResponse {
+    /// Assemble a response from the run header + the per-doc
+    /// rows the caller fetched (one fjall round-trip per id).
+    pub fn assemble(run: Run, documents: Vec<RunDocument>) -> Self {
+        let (state, failure_reason) = match run.state {
+            RunState::Analyzing => (RunStateDto::Analyzing, None),
+            RunState::AwaitingReview => (RunStateDto::AwaitingReview, None),
+            RunState::Applied => (RunStateDto::Applied, None),
+            RunState::PartiallyApplied => (RunStateDto::PartiallyApplied, None),
+            RunState::Failed { reason } => (RunStateDto::Failed, Some(reason)),
+        };
+        RunResponse {
+            id: run.id,
+            state,
+            failure_reason,
+            started_at: run.started_at,
+            updated_at: run.updated_at,
+            policy_refs: run.policy_refs.into_iter().map(Into::into).collect(),
+            context_refs: run.context_refs.into_iter().map(Into::into).collect(),
+            metadata: run.metadata,
+            documents: documents.into_iter().map(Into::into).collect(),
+        }
+    }
+}
+
+impl From<nvisy_engine::runs::ResourceRef> for ResourceRefDto {
+    fn from(r: nvisy_engine::runs::ResourceRef) -> Self {
+        ResourceRefDto {
+            id: r.id,
+            version: r.version,
+        }
+    }
+}
+
+impl From<RunDocument> for RunDocumentDto {
+    fn from(doc: RunDocument) -> Self {
+        let (state, failure_reason) = match doc.state {
+            RunDocState::Queued => (RunDocStateDto::Queued, None),
+            RunDocState::Analyzing => (RunDocStateDto::Analyzing, None),
+            RunDocState::AwaitingReview => (RunDocStateDto::AwaitingReview, None),
+            RunDocState::Applied => (RunDocStateDto::Applied, None),
+            RunDocState::Failed { reason } => (RunDocStateDto::Failed, Some(reason)),
+            RunDocState::TimedOut => (RunDocStateDto::TimedOut, None),
+        };
+        RunDocumentDto {
+            id: doc.id,
+            input_file_id: doc.input_file_id,
+            output_file_id: doc.output_file_id,
+            state,
+            failure_reason,
+            modality: doc.modality.into(),
+            body: doc.body.into(),
+        }
+    }
+}
+
+impl From<ModalityKind> for ModalityDto {
+    fn from(m: ModalityKind) -> Self {
+        match m {
+            ModalityKind::Text => ModalityDto::Text,
+            ModalityKind::Tabular => ModalityDto::Tabular,
+            ModalityKind::Image => ModalityDto::Image,
+            ModalityKind::Audio => ModalityDto::Audio,
+        }
+    }
+}
+
+impl From<DocBody> for DocBodyDto {
+    fn from(body: DocBody) -> Self {
+        match body {
+            DocBody::Text { entities } => DocBodyDto::Text {
+                entities: entities.into_iter().map(Into::into).collect(),
+            },
+            DocBody::Tabular { entities } => DocBodyDto::Tabular {
+                entities: entities.into_iter().map(Into::into).collect(),
+            },
+            DocBody::Image { entities } => DocBodyDto::Image {
+                entities: entities.into_iter().map(Into::into).collect(),
+            },
+            DocBody::Audio { entities } => DocBodyDto::Audio {
+                entities: entities.into_iter().map(Into::into).collect(),
+            },
+        }
+    }
+}
+
+impl From<EntityRecord<Text>> for TextEntityRecordDto {
+    fn from(record: EntityRecord<Text>) -> Self {
+        let common = entity_common(&record.entity);
+        TextEntityRecordDto {
+            common,
+            location: record.entity.location.into(),
+            r#override: record.r#override,
+        }
+    }
+}
+
+impl From<TextLocation> for TextLocationDto {
+    fn from(loc: TextLocation) -> Self {
+        TextLocationDto {
+            start: loc.start,
+            end: loc.end,
+            page: loc.page,
+        }
+    }
+}
+
+impl From<EntityRecord<Tabular>> for TabularEntityRecordDto {
+    fn from(record: EntityRecord<Tabular>) -> Self {
+        let common = entity_common(&record.entity);
+        TabularEntityRecordDto {
+            common,
+            location: record.entity.location.into(),
+            r#override: record.r#override,
+        }
+    }
+}
+
+impl From<TabularLocation> for TabularLocationDto {
+    fn from(loc: TabularLocation) -> Self {
+        TabularLocationDto {
+            row_index: loc.row_index,
+            column_index: loc.column_index,
+            start_offset: loc.start_offset,
+            end_offset: loc.end_offset,
+            column_name: loc.column_name.map(|s| s.as_str().to_owned()),
+            sheet_name: loc.sheet_name.map(|s| s.as_str().to_owned()),
+        }
+    }
+}
+
+impl From<EntityRecord<Image>> for ImageEntityRecordDto {
+    fn from(record: EntityRecord<Image>) -> Self {
+        let common = entity_common(&record.entity);
+        ImageEntityRecordDto {
+            common,
+            location: record.entity.location.into(),
+            r#override: record.r#override,
+        }
+    }
+}
+
+impl From<ImageLocation> for ImageLocationDto {
+    fn from(loc: ImageLocation) -> Self {
+        let bb = loc.bounding_box;
+        ImageLocationDto {
+            x: bb.min.x,
+            y: bb.min.y,
+            width: bb.width(),
+            height: bb.height(),
+            page: loc.page,
+        }
+    }
+}
+
+impl From<EntityRecord<Audio>> for AudioEntityRecordDto {
+    fn from(record: EntityRecord<Audio>) -> Self {
+        let common = entity_common(&record.entity);
+        AudioEntityRecordDto {
+            common,
+            location: record.entity.location.into(),
+            r#override: record.r#override,
+        }
+    }
+}
+
+impl From<AudioLocation> for AudioLocationDto {
+    fn from(loc: AudioLocation) -> Self {
+        AudioLocationDto {
+            start_ms: loc.span.start_millis(),
+            end_ms: loc.span.end_millis(),
+            speaker: loc.speaker_id.map(|s| s.as_str().to_owned()),
+        }
+    }
+}
diff --git a/crates/nvisy-server/src/service/mod.rs b/crates/nvisy-server/src/service/mod.rs
index b303c01a1..8c1b60e96 100644
--- a/crates/nvisy-server/src/service/mod.rs
+++ b/crates/nvisy-server/src/service/mod.rs
@@ -1,73 +1,63 @@
 //! Application state and dependency injection.
 //!
-//! [`ServiceState`] holds both [`DetectionEngine`] and
-//! [`RedactionEngine`]. The redaction engine is constructed from
-//! the detection engine via
-//! [`RedactionEngine::from_detection`][rfd] so they share the
-//! same registry, runtime config, optional key provider, and an
-//! in-memory `DetectionState` read-handle for the
-//! detect→redact handoff.
-//!
-//! Individual handlers extract whichever engine they need through
-//! `FromRef`.
-//!
-//! [rfd]: nvisy_engine::redaction::RedactionEngine::from_detection
+//! [`ServiceState`] wraps the single [`EngineHandle`] every
+//! handler needs plus the deployment's default
+//! [`AnalyzerParams`]. Cheaply cloneable (the engine handle is
+//! `Arc`-backed and the analyzer spec is shared via `Arc`), so
+//! per-HTTP-request handlers get a clone via `FromRef` without
+//! coordinating lifetime.
 
 use std::path::{Path, PathBuf};
+use std::sync::Arc;
 
 use nvisy_core::Result;
-use nvisy_engine::core::RuntimeConfig;
-use nvisy_engine::detection::DetectionEngine;
-use nvisy_engine::redaction::RedactionEngine;
-use nvisy_engine::registry::Registry;
+use nvisy_core::plan::AnalyzerParams;
+use nvisy_engine::EngineHandle;
 
-/// Shared application state threaded through all handlers.
+/// Shared application state threaded through every handler.
 #[must_use = "state does nothing unless you use it"]
 #[derive(Clone)]
 pub struct ServiceState {
-    detection: DetectionEngine,
-    redaction: RedactionEngine,
+    engine: EngineHandle,
+    data_dir: PathBuf,
+    analyzer_default: Arc<AnalyzerParams>,
 }
 
 impl ServiceState {
-    /// Creates a new service state from a resolved [`RuntimeConfig`]
-    /// and data directory.
-    ///
-    /// Async because engine construction may download model
-    /// artifacts on first use. The redaction engine is constructed
-    /// from the detection engine so they share the underlying
-    /// registry, runtime config, and key provider.
-    ///
-    /// # Errors
-    ///
-    /// Returns an error if the registry database cannot be opened.
-    pub async fn new(config: RuntimeConfig, data_dir: PathBuf) -> Result<Self> {
-        let detection = DetectionEngine::open(data_dir, config).await?;
-        let redaction = RedactionEngine::from_detection(&detection);
+    /// Open the engine database under `data_dir`. The deployment's
+    /// default analyzer is whatever the caller passes — typically
+    /// the `[analyzer]` section of `Nvisy.toml`, or
+    /// [`AnalyzerParams::default()`] when the section is absent
+    /// (degenerate: no recognizers, no enrichers — runs that omit
+    /// `analyzer` overrides will detect nothing).
+    pub async fn new(data_dir: PathBuf, analyzer_default: AnalyzerParams) -> Result<Self> {
+        let engine = EngineHandle::open(&data_dir)?;
         Ok(Self {
-            detection,
-            redaction,
+            engine,
+            data_dir,
+            analyzer_default: Arc::new(analyzer_default),
         })
     }
 
-    /// Returns the data directory path.
+    /// The engine handle every handler reaches for.
+    pub fn engine(&self) -> &EngineHandle {
+        &self.engine
+    }
+
+    /// The data directory the engine database lives under.
     pub fn data_dir(&self) -> &Path {
-        self.detection.data_dir()
+        &self.data_dir
     }
-}
 
-macro_rules! impl_di {
-    ($($extract:expr => $t:ty),+ $(,)?) => {$(
-        impl axum::extract::FromRef<ServiceState> for $t {
-            fn from_ref(state: &ServiceState) -> Self {
-                $extract(state)
-            }
-        }
-    )+};
+    /// The deployment's default [`AnalyzerParams`]. Requests that
+    /// inherit any analyzer field resolve against this.
+    pub fn analyzer_default(&self) -> &AnalyzerParams {
+        &self.analyzer_default
+    }
 }
 
-impl_di!(
-    |s: &ServiceState| s.detection.clone() => DetectionEngine,
-    |s: &ServiceState| s.redaction.clone() => RedactionEngine,
-    |s: &ServiceState| s.detection.registry().clone() => Registry,
-);
+impl axum::extract::FromRef<ServiceState> for EngineHandle {
+    fn from_ref(state: &ServiceState) -> Self {
+        state.engine.clone()
+    }
+}
diff --git a/crates/nvisy-stt/Cargo.toml b/crates/nvisy-stt/Cargo.toml
deleted file mode 100644
index 6890e7897..000000000
--- a/crates/nvisy-stt/Cargo.toml
+++ /dev/null
@@ -1,42 +0,0 @@
-# https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[package]
-name = "nvisy-stt"
-description = "Speech-to-text extractor (transcription + timestamps + diarization) for the Nvisy runtime"
-keywords = ["nvisy", "stt", "transcription", "diarization"]
-categories = ["multimedia::audio", "api-bindings"]
-readme = "README.md"
-
-version = { workspace = true }
-rust-version = { workspace = true }
-edition = { workspace = true }
-license = { workspace = true }
-publish = { workspace = true }
-
-authors = { workspace = true }
-repository = { workspace = true }
-homepage = { workspace = true }
-documentation = { workspace = true }
-
-[features]
-default = []
-
-[package.metadata.docs.rs]
-all-features = true
-rustdoc-args = ["--cfg", "docsrs"]
-
-[dependencies]
-# Internal crates
-nvisy-core = { workspace = true, features = [] }
-
-# Primitive datatypes
-uuid = { workspace = true, features = [] }
-
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-
-# Observability
-tracing = { workspace = true, features = [] }
-
-[dev-dependencies]
-tokio = { workspace = true, features = ["macros", "rt"] }
diff --git a/crates/nvisy-stt/README.md b/crates/nvisy-stt/README.md
deleted file mode 100644
index 0a49351f1..000000000
--- a/crates/nvisy-stt/README.md
+++ /dev/null
@@ -1,34 +0,0 @@
-# nvisy-stt
-
-Speech-to-text extractor for the Nvisy runtime. Mirrors the layout of
-[`nvisy-ner`][ner] and [`nvisy-ocr`][ocr]: a per-call
-[`backend::SttBackend`] trait, shipped implementations behind that
-trait, and an [`SttExtractor`] that type-erases the backend and plugs
-into the toolkit pipeline as [`Extractor<Audio>`][extractor].
-
-## Backends
-
-| Backend         | Status              |
-|-----------------|---------------------|
-| `NoopBackend`   | shipped             |
-
-[`backend::NoopBackend`] returns an empty [`Transcription`] for every
-request and acts as both the default when STT isn't configured and the
-test stub. Real provider backends (OpenAI Whisper, Deepgram,
-AssemblyAI) plug in later as feature-gated siblings — the trait is the
-integration point.
-
-## Transcription shape
-
-The extractor produces a [`Transcription`] with an ordered list of
-[`backend::TranscribedSegment`]s. Each segment carries a `TimeSpan`,
-the recognised text, and optional fields backends may or may not
-populate (`speaker_id` for diarized providers, `language` for
-code-switching providers, `confidence`, word-level breakdown).
-Backends without diarization leave `speaker_id = None` — the field
-exists so diarization-capable providers can plug in without a shape
-change.
-
-[ner]: https://docs.rs/nvisy-ner
-[ocr]: https://docs.rs/nvisy-ocr
-[extractor]: https://docs.rs/nvisy-core/latest/nvisy_core/extraction/trait.Extractor.html
diff --git a/crates/nvisy-stt/src/backend/mod.rs b/crates/nvisy-stt/src/backend/mod.rs
deleted file mode 100644
index 8772d14f0..000000000
--- a/crates/nvisy-stt/src/backend/mod.rs
+++ /dev/null
@@ -1,16 +0,0 @@
-//! Backend layer: the [`SttBackend`] trait and its shipped impls.
-//!
-//! One trait covers every flavour of provider — hosted APIs that emit
-//! a single full-clip segment (OpenAI Whisper), hosted APIs that emit
-//! diarized multi-speaker segments (Deepgram, AssemblyAI), and
-//! local/self-hosted inference services. Today only [`NoopBackend`]
-//! (returns no segments) ships; real providers will land as
-//! feature-gated siblings.
-
-mod noop_backend;
-mod stt_backend;
-mod transcribed_segment;
-
-pub use self::noop_backend::NoopBackend;
-pub use self::stt_backend::{SttBackend, SttRequest, SttResponse};
-pub use self::transcribed_segment::{TranscribedSegment, TranscribedWord};
diff --git a/crates/nvisy-stt/src/backend/noop_backend.rs b/crates/nvisy-stt/src/backend/noop_backend.rs
deleted file mode 100644
index 1dacf5c61..000000000
--- a/crates/nvisy-stt/src/backend/noop_backend.rs
+++ /dev/null
@@ -1,24 +0,0 @@
-//! [`NoopBackend`]: no-op [`SttBackend`] for tests and as the default
-//! when no real STT provider is configured.
-
-use nvisy_core::Result;
-use nvisy_core::entity::ModelProvenance;
-
-use super::stt_backend::{SttBackend, SttRequest, SttResponse};
-
-const MODEL_NAME: &str = "stt-noop";
-
-/// No-op STT backend: every call returns an empty response.
-#[derive(Debug, Default, Clone, Copy)]
-pub struct NoopBackend;
-
-#[async_trait::async_trait]
-impl SttBackend for NoopBackend {
-    fn provenance(&self) -> ModelProvenance {
-        ModelProvenance::new(MODEL_NAME.to_owned())
-    }
-
-    async fn transcribe(&self, _request: SttRequest<'_>) -> Result<SttResponse> {
-        Ok(SttResponse::default())
-    }
-}
diff --git a/crates/nvisy-stt/src/backend/stt_backend.rs b/crates/nvisy-stt/src/backend/stt_backend.rs
deleted file mode 100644
index f002c1b9c..000000000
--- a/crates/nvisy-stt/src/backend/stt_backend.rs
+++ /dev/null
@@ -1,77 +0,0 @@
-//! [`SttBackend`]: the unified per-call speech-to-text backend trait.
-//!
-//! One trait covers every flavour of provider — hosted APIs that emit
-//! a single full-clip segment (OpenAI Whisper), hosted APIs that emit
-//! diarized multi-speaker segments (Deepgram, AssemblyAI), and
-//! local/self-hosted inference services. Each backend turns a
-//! request (audio bytes + optional language hint) into a response
-//! (an ordered list of [`TranscribedSegment`]).
-//!
-//! Object-safe: extractors hold `Arc<dyn SttBackend>` and dispatch
-//! per call.
-
-use nvisy_core::Result;
-use nvisy_core::entity::ModelProvenance;
-use nvisy_core::primitive::LanguageTag;
-use uuid::Uuid;
-
-use super::transcribed_segment::TranscribedSegment;
-
-/// One per-call STT request handed to an [`SttBackend`].
-#[derive(Debug, Clone)]
-pub struct SttRequest<'a> {
-    /// Raw audio bytes (MP3, WAV, FLAC, etc.). The backend is
-    /// responsible for honouring whatever container/codec it accepts.
-    pub audio: &'a [u8],
-    /// Original filename. Some backends use the extension for
-    /// MIME-type detection on multipart uploads.
-    pub filename: &'a str,
-    /// Caller-asserted language. Backends that support per-call
-    /// language hinting use this; others ignore it.
-    pub language: Option<&'a LanguageTag>,
-    /// Correlation UUID for tracing.
-    pub correlation_id: Option<Uuid>,
-}
-
-/// One per-call STT response from an [`SttBackend`].
-///
-/// Wraps the segments the backend produced in their original order.
-#[derive(Debug, Clone, Default)]
-pub struct SttResponse {
-    /// Segments predicted for the request, in backend order.
-    pub segments: Vec<TranscribedSegment>,
-}
-
-impl SttResponse {
-    /// Construct a response from segments.
-    #[must_use]
-    pub fn new(segments: Vec<TranscribedSegment>) -> Self {
-        Self { segments }
-    }
-}
-
-/// Per-call speech-to-text backend.
-///
-/// Implemented by everything that turns `(audio, language?)` into
-/// transcribed segments — hosted provider clients (OpenAI Whisper,
-/// Deepgram, AssemblyAI), local model wrappers, and the in-process
-/// no-op test stub.
-#[async_trait::async_trait]
-pub trait SttBackend: Send + Sync + 'static {
-    /// Backend identity (model / service name + provenance kind).
-    ///
-    /// The document-side extraction phase reads this after STT runs and
-    /// stamps it into [`AudioExtraction::Transcription`] on the
-    /// document's metadata, so the audit records *which* STT pass
-    /// produced the document.
-    ///
-    /// [`AudioExtraction::Transcription`]: nvisy_core::modality::AudioExtraction::Transcription
-    fn provenance(&self) -> ModelProvenance;
-
-    /// Transcribe `request`.
-    ///
-    /// # Errors
-    ///
-    /// Returns the underlying transport / parse / inference error.
-    async fn transcribe(&self, request: SttRequest<'_>) -> Result<SttResponse>;
-}
diff --git a/crates/nvisy-stt/src/backend/transcribed_segment.rs b/crates/nvisy-stt/src/backend/transcribed_segment.rs
deleted file mode 100644
index 08f23543b..000000000
--- a/crates/nvisy-stt/src/backend/transcribed_segment.rs
+++ /dev/null
@@ -1,47 +0,0 @@
-//! [`TranscribedSegment`] + [`TranscribedWord`]: the per-segment
-//! transcription unit produced by an [`SttBackend`].
-//!
-//! [`SttBackend`]: super::SttBackend
-
-use nvisy_core::primitive::{Confidence, LanguageTag, TimeSpan};
-
-/// One transcription segment from an [`SttBackend`].
-///
-/// Carries a time interval, the recognised text, and a handful of
-/// optional fields that providers may or may not fill in.
-/// `speaker_id` is populated only by providers with diarization
-/// (Deepgram, AssemblyAI, etc.) — OpenAI Whisper leaves it `None`.
-/// `language` is populated by providers that emit per-segment language
-/// detection (handy for code-switching audio).
-///
-/// [`SttBackend`]: super::SttBackend
-#[derive(Debug, Clone, PartialEq)]
-pub struct TranscribedSegment {
-    /// Time interval (microsecond precision) within the source clip.
-    pub time_span: TimeSpan,
-    /// Speaker label, if the backend performed diarization.
-    pub speaker_id: Option<String>,
-    /// Detected language for this segment, if reported.
-    pub language: Option<LanguageTag>,
-    /// Recognised text for this segment.
-    pub text: String,
-    /// Backend confidence in the segment, if reported.
-    pub confidence: Option<Confidence>,
-    /// Word-level breakdown, if the backend emitted one. Empty
-    /// otherwise.
-    pub words: Vec<TranscribedWord>,
-}
-
-/// One word inside a [`TranscribedSegment`].
-///
-/// Populated by backends that emit word-level timestamps (OpenAI
-/// Whisper with `timestamp_granularities=["word"]`, Deepgram, etc.).
-#[derive(Debug, Clone, PartialEq)]
-pub struct TranscribedWord {
-    /// Time interval covering this word.
-    pub time_span: TimeSpan,
-    /// The word as the backend transcribed it.
-    pub text: String,
-    /// Per-word confidence, if reported.
-    pub confidence: Option<Confidence>,
-}
diff --git a/crates/nvisy-stt/src/extraction/mod.rs b/crates/nvisy-stt/src/extraction/mod.rs
deleted file mode 100644
index d7621b4ff..000000000
--- a/crates/nvisy-stt/src/extraction/mod.rs
+++ /dev/null
@@ -1,82 +0,0 @@
-//! Type-erased STT extractor wrapping any [`SttBackend`].
-//!
-//! [`SttBackend`]: crate::backend::SttBackend
-
-mod transcription;
-
-use std::fmt;
-use std::sync::Arc;
-
-use nvisy_core::Result;
-use nvisy_core::entity::ModelProvenance;
-use nvisy_core::extraction::{Extractor, ExtractorOutput, Span};
-use nvisy_core::modality::{Audio, AudioExtraction};
-
-pub use self::transcription::Transcription;
-use crate::backend::{SttBackend, SttRequest};
-
-const TARGET: &str = "nvisy_stt::extraction";
-
-/// Type-erased STT extractor wrapping any [`SttBackend`].
-///
-/// Owns an `Arc<dyn SttBackend>` and forwards transcription requests
-/// to it, providing a concrete, object-safe entry point without
-/// generics at every call site. Cloning shares the backend.
-#[derive(Clone)]
-pub struct SttExtractor {
-    backend: Arc<dyn SttBackend>,
-}
-
-impl fmt::Debug for SttExtractor {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("SttExtractor").finish_non_exhaustive()
-    }
-}
-
-impl SttExtractor {
-    /// Wrap any [`SttBackend`] implementation.
-    pub fn new(backend: impl SttBackend) -> Self {
-        Self {
-            backend: Arc::new(backend),
-        }
-    }
-
-    /// Provenance of the wrapped backend, forwarded from
-    /// [`SttBackend::provenance`].
-    pub fn provenance(&self) -> ModelProvenance {
-        self.backend.provenance()
-    }
-}
-
-#[async_trait::async_trait]
-impl Extractor<Audio> for SttExtractor {
-    type Output = Transcription;
-
-    #[tracing::instrument(target = TARGET, skip_all, fields(audio_bytes = span.data.bytes.len()))]
-    async fn extract(&self, span: &Span<Audio>) -> Result<ExtractorOutput<Audio, Self::Output>> {
-        let synthesized;
-        let filename = match span.data.filename.as_deref() {
-            Some(name) => name,
-            None => {
-                synthesized = format!("audio.{}", span.data.extension());
-                synthesized.as_str()
-            }
-        };
-        let request = SttRequest {
-            audio: span.data.bytes.as_ref(),
-            filename,
-            language: span.language.as_ref(),
-            correlation_id: span.correlation_id,
-        };
-        let response = self.backend.transcribe(request).await?;
-        tracing::debug!(
-            target: TARGET,
-            segments = response.segments.len(),
-            "transcription complete",
-        );
-        Ok(ExtractorOutput::new(
-            Transcription::new(response.segments),
-            AudioExtraction::Transcription(self.provenance()),
-        ))
-    }
-}
diff --git a/crates/nvisy-stt/src/extraction/transcription.rs b/crates/nvisy-stt/src/extraction/transcription.rs
deleted file mode 100644
index a2da982c2..000000000
--- a/crates/nvisy-stt/src/extraction/transcription.rs
+++ /dev/null
@@ -1,39 +0,0 @@
-//! [`Transcription`]: the value emitted by an [`SttExtractor`].
-//!
-//! [`SttExtractor`]: super::SttExtractor
-
-use crate::backend::TranscribedSegment;
-
-/// Output of an STT extraction pass: an ordered list of
-/// [`TranscribedSegment`]s covering the source audio.
-///
-/// Empty if the backend returned nothing (e.g. silence or the no-op
-/// backend).
-#[derive(Debug, Clone, Default, PartialEq)]
-pub struct Transcription {
-    /// Segments in source order.
-    pub segments: Vec<TranscribedSegment>,
-}
-
-impl Transcription {
-    /// Build a transcription from segments.
-    #[must_use]
-    pub fn new(segments: Vec<TranscribedSegment>) -> Self {
-        Self { segments }
-    }
-
-    /// Concatenated text of every segment, joined with a single space.
-    ///
-    /// Convenience for downstream callers (NER recognizers) that want
-    /// the recognised text as one string. Segment-level timestamps are
-    /// preserved in [`segments`] for callers that need them.
-    ///
-    /// [`segments`]: Self::segments
-    pub fn full_text(&self) -> String {
-        self.segments
-            .iter()
-            .map(|s| s.text.as_str())
-            .collect::<Vec<_>>()
-            .join(" ")
-    }
-}
diff --git a/crates/nvisy-stt/src/lib.rs b/crates/nvisy-stt/src/lib.rs
deleted file mode 100644
index 58964b350..000000000
--- a/crates/nvisy-stt/src/lib.rs
+++ /dev/null
@@ -1,8 +0,0 @@
-#![forbid(unsafe_code)]
-#![cfg_attr(docsrs, feature(doc_cfg))]
-#![doc = include_str!("../README.md")]
-
-pub mod backend;
-mod extraction;
-
-pub use self::extraction::{SttExtractor, Transcription};
diff --git a/crates/nvisy-toolkit/Cargo.toml b/crates/nvisy-toolkit/Cargo.toml
deleted file mode 100644
index f16707d0b..000000000
--- a/crates/nvisy-toolkit/Cargo.toml
+++ /dev/null
@@ -1,102 +0,0 @@
-# https://doc.rust-lang.org/cargo/reference/manifest.html
-
-[package]
-name = "nvisy-toolkit"
-description = "Composable component library for Nvisy pipelines: recognizers, dedup layers, checks, redaction strategies"
-keywords = ["nvisy", "toolkit", "components", "pipeline"]
-categories = ["data-structures"]
-readme = "README.md"
-
-version = { workspace = true }
-rust-version = { workspace = true }
-edition = { workspace = true }
-license = { workspace = true }
-publish = true
-
-authors = { workspace = true }
-repository = { workspace = true }
-homepage = { workspace = true }
-documentation = { workspace = true }
-
-[features]
-default = ["text", "tabular", "image", "audio", "rich"]
-
-## Re-export `DecodedBuffer<Text>` from `nvisy-codec` and pull in
-## every shipped text-format codec (txt/json/markdown/html).
-text = ["nvisy-codec/text"]
-## Re-export `DecodedBuffer<Tabular>` and pull in csv/xlsx codecs.
-tabular = ["nvisy-codec/tabular"]
-## Re-export `DecodedBuffer<Image>` and pull in png/jpeg/tiff codecs.
-image = ["nvisy-codec/image"]
-## Re-export `DecodedBuffer<Audio>` and pull in wav/mp3 codecs.
-audio = ["nvisy-codec/audio"]
-## Pull in rich-document codecs (pdf/docx). Implies `image` since
-## PDF rendering produces image children.
-rich = ["image", "nvisy-codec/rich"]
-
-## Reversible AES-GCM encrypt/decrypt redaction operators.
-encrypt = ["dep:aes-gcm"]
-
-## Externalised BentoML backends (NER + OCR).
-bento = ["nvisy-ner/bento", "nvisy-ocr/bento"]
-
-[package.metadata.docs.rs]
-all-features = true
-rustdoc-args = ["--cfg", "docsrs"]
-
-[dependencies]
-# Internal crates
-nvisy-codec = { workspace = true, features = [] }
-nvisy-core = { workspace = true, features = [] }
-nvisy-llm = { workspace = true, features = [] }
-nvisy-ner = { workspace = true, features = [] }
-nvisy-ocr = { workspace = true, features = [] }
-nvisy-pattern = { workspace = true, features = [] }
-nvisy-stt = { workspace = true, features = [] }
-
-# Serialization
-serde = { workspace = true, features = [] }
-schemars = { workspace = true, features = [] }
-
-# Primitive datatypes
-uuid = { workspace = true, features = [] }
-type-map = { workspace = true, features = [] }
-
-# Encoding and hashing
-base64 = { workspace = true, features = [] }
-sha2 = { workspace = true, features = [] }
-aes-gcm = { workspace = true, features = ["aes", "alloc"], optional = true }
-
-# Async runtime and parallelism
-async-trait = { workspace = true, features = [] }
-tokio = { workspace = true, features = ["rt", "sync", "time", "macros"] }
-
-# Observability
-tracing = { workspace = true, features = [] }
-
-# Text processing (unicode-aware folding for leak detection)
-unicode-normalization = { workspace = true, features = [] }
-regex = { workspace = true, features = [] }
-
-[dev-dependencies]
-# Internal test utilities (Entity::test_builder, …).
-nvisy-core = { workspace = true, features = ["test-utils"] }
-# ContextEnhanced<R> wrapper type returned by PatternRecognizer::build().
-nvisy-context = { workspace = true, features = [] }
-# Codec front door for E2E pipeline tests. Production builds pull
-# nvisy-codec via the toolkit's per-modality features; the dev
-# entry pins txt/csv/json so the pipeline_*.rs tests compile.
-nvisy-codec = { workspace = true, features = ["txt", "csv", "json"] }
-# Pattern recognizer used by the end-to-end example and pipeline tests.
-nvisy-pattern = { workspace = true }
-# Fake-data anonymizer wrapped through Vault in the integration test.
-nvisy-fake = { workspace = true }
-# JSON parsing for asserting redacted JSON still parses cleanly.
-serde_json = { workspace = true, features = [] }
-# Async-trait + tokio runtime for the example.
-async-trait = { workspace = true }
-tokio = { workspace = true, features = ["rt-multi-thread", "macros"] }
-
-[[example]]
-name = "pipeline"
-path = "examples/pipeline.rs"
diff --git a/crates/nvisy-toolkit/README.md b/crates/nvisy-toolkit/README.md
deleted file mode 100644
index 5cada4474..000000000
--- a/crates/nvisy-toolkit/README.md
+++ /dev/null
@@ -1,40 +0,0 @@
-# nvisy-toolkit
-
-[![Build](https://img.shields.io/github/actions/workflow/status/nvisycom/runtime/build.yml?branch=main&label=build%20%26%20test&style=flat-square)](https://github.com/nvisycom/runtime/actions/workflows/build.yml)
-
-Composable component library for Nvisy pipelines — the registries
-and policies a consumer plugs into their own document-processing
-flow.
-
-## Overview
-
-Hosts the per-stage component machinery that the document
-orchestrator drives: `RecognizerRegistry` + `Detection` plan (which
-recognizers run per modality), `ExtractorRegistry` + `Extraction`
-plan (OCR / STT), `LayerPipeline` for post-detection deduplication,
-`CheckPipeline` for validation, and `RedactionRegistry<M>` for
-custom anonymizers.
-
-Also ships the `MemoryBuffer<M>` ingestion helper that wraps bytes
-in a `DataAt<M>` (and `RedactAt<Text>`) buffer suitable for direct
-consumption by recognizers. Sits one level above `nvisy-core`:
-toolkit owns reusable pieces; the orchestration that strings them
-into a full pipeline lives one layer up.
-
-## Documentation
-
-See [`docs/`](../../docs/) for architecture, security, and API documentation.
-
-## Changelog
-
-See [CHANGELOG.md](../../CHANGELOG.md) for release notes and version history.
-
-## License
-
-Apache 2.0 License, see [LICENSE.txt](../../LICENSE.txt)
-
-## Support
-
-- **Documentation**: [docs.nvisy.com](https://docs.nvisy.com)
-- **Issues**: [GitHub Issues](https://github.com/nvisycom/runtime/issues)
-- **Email**: [support@nvisy.com](mailto:support@nvisy.com)
diff --git a/crates/nvisy-toolkit/examples/pipeline.rs b/crates/nvisy-toolkit/examples/pipeline.rs
deleted file mode 100644
index 66c6c908f..000000000
--- a/crates/nvisy-toolkit/examples/pipeline.rs
+++ /dev/null
@@ -1,144 +0,0 @@
-//! End-to-end toolkit pipeline against a synthetic text input,
-//! using the real codec read/write surface
-//! ([`CodecRegistry`] → [`DocumentHandle`]): **detection →
-//! deduplication → redaction**, all in one process, no external
-//! services required.
-//!
-//! Mirrors the producer side of what the engine runs per text
-//! block. Each section below corresponds to one phase the engine
-//! orchestrates.
-//!
-//! Run:
-//!
-//! ```text
-//! cargo run -p nvisy-toolkit --example pipeline
-//! ```
-//!
-//! [`CodecRegistry`]: nvisy_codec::CodecRegistry
-//! [`DocumentHandle`]: nvisy_codec::DocumentHandle
-
-use std::str::from_utf8;
-
-use nvisy_codec::{CodecRegistry, DocumentHandle};
-use nvisy_core::Result;
-use nvisy_core::entity::builtins;
-use nvisy_core::modality::{Text, TextData};
-use nvisy_core::primitive::ConfidenceThreshold;
-use nvisy_core::recognition::RecognizerInput;
-use nvisy_core::redaction::RedactAt;
-use nvisy_pattern::PatternRecognizer;
-use nvisy_toolkit::deduplication::{LayerContext, LayerParams, LayerPipeline};
-use nvisy_toolkit::detection::RecognizerRegistry;
-use nvisy_toolkit::redaction::RedactionRegistry;
-use nvisy_toolkit::redaction::anonymizer::{Mask, Redact, Replace};
-
-const SAMPLE: &str = "Email alice@example.test or call +1 415 555 0100. \
-                      Card 4111111111111111 expires 12/27.";
-
-#[tokio::main]
-async fn main() -> Result<()> {
-    // ── Phase 0: ingestion ────────────────────────────────────────
-    // Decode the sample through the codec registry — the same front
-    // door the engine uses — by resolving the `.txt` extension to a
-    // loader and pulling the resulting handle back into a typed
-    // text DocumentHandle.
-    let registry = CodecRegistry::with_builtin();
-    let untyped = registry.decode(SAMPLE, "txt").await?;
-    let mut source = untyped
-        .into_text()
-        .expect("txt extension resolves to text modality");
-    println!("source = {SAMPLE}\n");
-
-    // ── Phase 1: detection ────────────────────────────────────────
-    // Pattern-only registry so the example needs no external
-    // services. Add NER / LLM recognizers with extra
-    // `.with_recognizer(...)` calls.
-    let pattern = PatternRecognizer::builder()
-        .with_builtin_patterns()
-        .with_builtin_dictionaries()
-        .build_context_enhanced()?;
-    let detection = RecognizerRegistry::new().with_recognizer(pattern);
-
-    let input = RecognizerInput::new(TextData::new(SAMPLE.to_owned()));
-    let entities = detection.run::<Text>(input).await?;
-
-    println!(
-        "detection: {} entit{}",
-        entities.len(),
-        plural(entities.len())
-    );
-    for entity in &entities {
-        let matched = &SAMPLE[entity.location.start..entity.location.end];
-        println!(
-            "  - {:?} {:?} at {}..{} (confidence {:.2})",
-            entity.label,
-            matched,
-            entity.location.start,
-            entity.location.end,
-            entity.confidence.get(),
-        );
-    }
-
-    // ── Phase 2: deduplication ────────────────────────────────────
-    // Canonical four-layer pipeline: calibrate → filter → fuse →
-    // resolve. Drops overlapping / low-confidence detections so the
-    // redaction phase sees a conflict-free entity set.
-    //
-    // The 0.5 threshold filters out the ISO-639 short-code matches
-    // (`or`, `am`, ...) from the languages dictionary, which load
-    // at 0.30 because they collide with common English words. See
-    // `assets/dictionaries/general/languages.toml`'s `column_scores`.
-    let params = LayerParams {
-        confidence_threshold: Some(ConfidenceThreshold::new(0.5).unwrap()),
-        ..LayerParams::default()
-    };
-    let ctx = LayerContext::<Text, DocumentHandle<Text>>::new(&source);
-    let dedup = LayerPipeline::<Text, DocumentHandle<Text>>::from_params(&params)?;
-
-    let before = entities.len();
-    let entities = dedup.run(entities, &ctx).await;
-    println!(
-        "\ndeduplication: kept {} of {} (dropped {})",
-        entities.len(),
-        before,
-        before - entities.len(),
-    );
-
-    // ── Phase 3: redaction ────────────────────────────────────────
-    // Register an operator per kind plus a catch-all fallback.
-    // `apply_all` runs the per-kind resolver across every entity and
-    // returns a `Redactions` batch; `redact_at` flushes the batch
-    // back into the codec handler in place.
-    let redaction = RedactionRegistry::<Text>::new()
-        .insert_label(
-            builtins::EMAIL_ADDRESS.label_ref(),
-            Replace::new("[{label}]"),
-        )
-        .insert_label(
-            builtins::PHONE_NUMBER.label_ref(),
-            Replace::new("[{label}]"),
-        )
-        .insert_label(builtins::PAYMENT_CARD.label_ref(), Mask::stars())
-        .with_fallback(Redact);
-
-    // `source` is the codec-backed handle; it satisfies `DataAt<Text>`,
-    // so apply_all pulls only the per-entity byte range and hands
-    // *that* (not the whole document) to each anonymizer.
-    let redactions = redaction.apply_all(&entities, &source).await?;
-    println!("\nredaction: produced {} replacement(s)", redactions.len());
-    source.redact_at(redactions).await?;
-
-    // Serialize the post-redaction handler back through the codec.
-    // `Handler::encode` is what the engine's export phase calls to
-    // produce output bytes; it knows the format's reassembly rules
-    // (line terminators for txt, JSON envelope for json, …) so the
-    // example doesn't have to.
-    let encoded = source.handler().encode()?;
-    let redacted = from_utf8(encoded.as_bytes()).expect("txt encode produces UTF-8");
-    println!("\nredacted = {redacted}");
-    Ok(())
-}
-
-fn plural(n: usize) -> &'static str {
-    if n == 1 { "y" } else { "ies" }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/calibrate/mod.rs b/crates/nvisy-toolkit/src/deduplication/calibrate/mod.rs
deleted file mode 100644
index 01799bbd8..000000000
--- a/crates/nvisy-toolkit/src/deduplication/calibrate/mod.rs
+++ /dev/null
@@ -1,246 +0,0 @@
-//! [`CalibrateLayer`]: scale entity confidences by per-recognizer
-//! multipliers.
-//!
-//! Compensates for score-distribution differences between
-//! detectors — regex always returns `1.0` while NER returns
-//! `0.3–0.9`, so a multiplier of `0.8` on `"pattern"` brings them
-//! into the same range before deduplication runs.
-//!
-//! Keys in [`CalibrationMap`] are recognizer source names from the
-//! entity's [`Recognition`] trail step (typically `"pattern"`,
-//! `"ner"`, `"llm-ner"`).
-//!
-//! [`Recognition`]: nvisy_core::entity::TrailStepKind::Recognition
-
-use std::borrow::Cow;
-use std::collections::HashMap;
-
-use nvisy_core::entity::{Entity, TrailStep};
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::Modality;
-use nvisy_core::primitive::Confidence;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::layer::{Layer, LayerContext};
-
-const TARGET: &str = "nvisy_toolkit::deduplication::calibrate";
-
-/// Per-recognizer confidence multipliers applied before
-/// deduplication.
-///
-/// Maps a recognizer source name to a scaling factor. Recognizers
-/// not present in the map are left unchanged (implicit multiplier
-/// `1.0`). Both built-in recognizer names (`"pattern"`, `"ner"`)
-/// and runtime-configured custom names are accepted as keys.
-#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(transparent)]
-pub struct CalibrationMap(HashMap<Cow<'static, str>, f64>);
-
-impl CalibrationMap {
-    /// Empty calibration map.
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Insert a multiplier for a recognizer name.
-    pub fn insert(&mut self, recognizer: impl Into<Cow<'static, str>>, multiplier: f64) {
-        self.0.insert(recognizer.into(), multiplier);
-    }
-
-    /// Look up the multiplier for a recognizer name, or `None`.
-    pub fn get(&self, recognizer: &str) -> Option<f64> {
-        self.0.get(recognizer).copied()
-    }
-
-    /// True when no multipliers are registered.
-    pub fn is_empty(&self) -> bool {
-        self.0.is_empty()
-    }
-
-    /// Number of registered multipliers.
-    pub fn len(&self) -> usize {
-        self.0.len()
-    }
-}
-
-impl<K, V> FromIterator<(K, V)> for CalibrationMap
-where
-    K: Into<Cow<'static, str>>,
-    V: Into<f64>,
-{
-    fn from_iter<I: IntoIterator<Item = (K, V)>>(iter: I) -> Self {
-        Self(
-            iter.into_iter()
-                .map(|(k, v)| (k.into(), v.into()))
-                .collect(),
-        )
-    }
-}
-
-/// [`Layer`] that scales per-entity confidence by per-recognizer
-/// calibration multipliers in place.
-///
-/// For each entity, looks up the multiplier for the *originating*
-/// recognizer (the first `Recognition` trail step) and scales the
-/// confidence. Refinement / fusion sources appended later in the
-/// trail don't override the originating recognizer's calibration —
-/// calibration is a per-detector statement about score shape, and
-/// the score belongs to whichever detector produced it.
-///
-/// Entities whose originating recognizer is absent from the map
-/// are left unchanged (implicit multiplier `1.0`). Results are
-/// clamped to `[0.0, 1.0]`. Adjusted entities receive a
-/// [`Calibration`] step on their trail.
-///
-/// Drops nothing — returns an empty vec from [`Layer::apply`].
-///
-/// [`Calibration`]: nvisy_core::entity::TrailStepKind::Calibration
-pub struct CalibrateLayer {
-    calibration: CalibrationMap,
-}
-
-impl CalibrateLayer {
-    /// Construct a calibration layer from a map.
-    pub fn new(calibration: CalibrationMap) -> Self {
-        Self { calibration }
-    }
-}
-
-#[async_trait::async_trait]
-impl<M: Modality, R: TextAt<M> + ?Sized> Layer<M, R> for CalibrateLayer {
-    async fn apply(
-        &self,
-        entities: &mut Vec<Entity<M>>,
-        _ctx: &LayerContext<'_, M, R>,
-    ) -> Vec<Entity<M>> {
-        if self.calibration.is_empty() {
-            return Vec::new();
-        }
-
-        let mut adjusted = 0usize;
-
-        for entity in entities.iter_mut() {
-            let multiplier = entity
-                .recognizers()
-                .next()
-                .and_then(|name| self.calibration.get(name));
-
-            if let Some(m) = multiplier {
-                let before = entity.confidence;
-                let after_raw = (before.get() * m).clamp(0.0, 1.0);
-                let after = Confidence::new(after_raw).expect("clamped to [0,1]");
-                entity.confidence = after;
-                entity.trail.push(TrailStep::calibration(
-                    before,
-                    after,
-                    format!("scaled by {m:.3}"),
-                ));
-                adjusted += 1;
-
-                tracing::trace!(
-                    target: TARGET,
-                    entity_id = %entity.id,
-                    multiplier = m,
-                    before = before.get(),
-                    after = after.get(),
-                    "calibrated confidence",
-                );
-            }
-        }
-
-        tracing::debug!(
-            target: TARGET,
-            total = entities.len(),
-            adjusted,
-            "calibration pass complete",
-        );
-
-        Vec::new()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{Entity, ModelProvenance, TrailProvenance, TrailStep, TrailStepKind};
-    use nvisy_core::modality::Text;
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-    use crate::deduplication::test_resolver;
-
-    fn conf(v: f64) -> Confidence {
-        Confidence::new(v).expect("confidence in [0,1]")
-    }
-
-    fn trail_for(source: &str, confidence: Confidence) -> Vec<TrailStep> {
-        vec![TrailStep::recognition(
-            source,
-            confidence,
-            TrailProvenance::Model(ModelProvenance::new(source)),
-            "",
-        )]
-    }
-
-    async fn apply<M: Modality>(layer: &CalibrateLayer, entities: &mut Vec<Entity<M>>) {
-        let resolver = test_resolver::<M>();
-        let ctx = LayerContext::new(&*resolver);
-        let dropped = layer.apply(entities, &ctx).await;
-        assert!(dropped.is_empty(), "calibrate never drops");
-    }
-
-    #[tokio::test]
-    async fn scales_confidence_and_appends_calibration_step() {
-        let mut calibration = CalibrationMap::new();
-        calibration.insert("pattern", 0.5);
-        let layer = CalibrateLayer::new(calibration);
-        let mut entities: Vec<Entity<Text>> = vec![
-            Entity::test_builder(0, 4)
-                .with_trail(trail_for("pattern", conf(0.8)))
-                .with_confidence(conf(0.8))
-                .test_build(),
-        ];
-        apply(&layer, &mut entities).await;
-        assert!((entities[0].confidence.get() - 0.4).abs() < f64::EPSILON);
-        assert!(
-            entities[0]
-                .trail
-                .iter()
-                .any(|s| matches!(s.kind, TrailStepKind::Calibration))
-        );
-    }
-
-    #[tokio::test]
-    async fn clamps_to_one() {
-        let mut calibration = CalibrationMap::new();
-        calibration.insert("pattern", 2.0);
-        let layer = CalibrateLayer::new(calibration);
-        let mut entities: Vec<Entity<Text>> = vec![
-            Entity::test_builder(0, 4)
-                .with_trail(trail_for("pattern", conf(0.8)))
-                .with_confidence(conf(0.8))
-                .test_build(),
-        ];
-        apply(&layer, &mut entities).await;
-        assert!((entities[0].confidence.get() - 1.0).abs() < f64::EPSILON);
-    }
-
-    #[tokio::test]
-    async fn picks_originating_recognizer_multiplier() {
-        let mut calibration = CalibrationMap::new();
-        calibration.insert("pattern", 0.5);
-        calibration.insert("ner", 0.8);
-        let layer = CalibrateLayer::new(calibration);
-        let mut trail = trail_for("pattern", conf(1.0));
-        trail.extend(trail_for("ner", conf(1.0)));
-        let mut entities: Vec<Entity<Text>> = vec![
-            Entity::test_builder(0, 4)
-                .with_trail(trail)
-                .with_confidence(conf(1.0))
-                .test_build(),
-        ];
-        apply(&layer, &mut entities).await;
-        // First recognition step is "pattern" -> multiplier 0.5.
-        assert!((entities[0].confidence.get() - 0.5).abs() < f64::EPSILON);
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/filter/mod.rs b/crates/nvisy-toolkit/src/deduplication/filter/mod.rs
deleted file mode 100644
index 044ab5a1a..000000000
--- a/crates/nvisy-toolkit/src/deduplication/filter/mod.rs
+++ /dev/null
@@ -1,169 +0,0 @@
-//! [`FilterLayer`]: drop entities outside the allowed labels or
-//! below the confidence floor.
-//!
-//! Dropped entities are returned from [`Layer::apply`] so the
-//! pipeline can attribute them in its drop-reason roll-up.
-//!
-//! [`Layer::apply`]: super::layer::Layer::apply
-
-use nvisy_core::entity::{Entity, EntityLabelRef};
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::Modality;
-use nvisy_core::primitive::ConfidenceThreshold;
-
-use super::layer::{Layer, LayerContext};
-
-/// [`Layer`] that drops entities outside the allowed labels or
-/// below the confidence floor. Returns the dropped entities from
-/// [`Layer::apply`].
-///
-/// Construct empty with [`FilterLayer::new`] (default = pass
-/// everything) and configure via [`with_allowed_labels`] /
-/// [`with_confidence_threshold`].
-///
-/// [`with_allowed_labels`]: Self::with_allowed_labels
-/// [`with_confidence_threshold`]: Self::with_confidence_threshold
-#[derive(Debug, Clone, Default)]
-pub struct FilterLayer {
-    allowed_labels: Option<Vec<EntityLabelRef>>,
-    confidence_threshold: Option<ConfidenceThreshold>,
-}
-
-impl FilterLayer {
-    /// Empty filter: keeps every entity.
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Drop entities whose `label` is outside this set.
-    /// `None` keeps every label (same as not calling this).
-    #[must_use]
-    pub fn with_allowed_labels(mut self, labels: Option<Vec<EntityLabelRef>>) -> Self {
-        self.allowed_labels = labels;
-        self
-    }
-
-    /// Drop entities whose calibrated `confidence` is below this
-    /// floor. `None` keeps every confidence level (same as not
-    /// calling this).
-    #[must_use]
-    pub fn with_confidence_threshold(mut self, threshold: Option<ConfidenceThreshold>) -> Self {
-        self.confidence_threshold = threshold;
-        self
-    }
-
-    /// Whether `entity` clears every configured filter knob.
-    pub fn passes<M: Modality>(&self, entity: &Entity<M>) -> bool {
-        if let Some(ref labels) = self.allowed_labels
-            && !labels.contains(&entity.label)
-        {
-            return false;
-        }
-        if let Some(threshold) = self.confidence_threshold
-            && !threshold.admits(entity.confidence)
-        {
-            return false;
-        }
-        true
-    }
-}
-
-#[async_trait::async_trait]
-impl<M: Modality, R: TextAt<M> + ?Sized> Layer<M, R> for FilterLayer {
-    async fn apply(
-        &self,
-        entities: &mut Vec<Entity<M>>,
-        _ctx: &LayerContext<'_, M, R>,
-    ) -> Vec<Entity<M>> {
-        let mut dropped = Vec::new();
-        entities.retain(|e| {
-            let keep = self.passes(e);
-            if !keep {
-                dropped.push(e.clone());
-            }
-            keep
-        });
-        dropped
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{Entity, EntityLabelRef, builtins};
-    use nvisy_core::modality::Text;
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-    use crate::deduplication::test_resolver;
-
-    fn ent(label: EntityLabelRef, conf: f64) -> Entity<Text> {
-        Entity::test_builder(0, 4)
-            .with_label(label)
-            .with_confidence(Confidence::new(conf).expect("in range"))
-            .test_build()
-    }
-
-    async fn apply<M: Modality>(
-        layer: FilterLayer,
-        entities: &mut Vec<Entity<M>>,
-    ) -> Vec<Entity<M>> {
-        let resolver = test_resolver::<M>();
-        let ctx = LayerContext::new(&*resolver);
-        layer.apply(entities, &ctx).await
-    }
-
-    #[tokio::test]
-    async fn default_layer_keeps_everything() {
-        let mut entities: Vec<Entity<Text>> = vec![
-            ent(builtins::PERSON_NAME.label_ref(), 0.9),
-            ent(builtins::EMAIL_ADDRESS.label_ref(), 0.4),
-        ];
-        let dropped = apply(FilterLayer::new(), &mut entities).await;
-        assert_eq!(entities.len(), 2);
-        assert!(dropped.is_empty());
-    }
-
-    #[tokio::test]
-    async fn allowed_labels_drops_outsiders() {
-        let mut entities: Vec<Entity<Text>> = vec![
-            ent(builtins::PERSON_NAME.label_ref(), 0.9),
-            ent(builtins::EMAIL_ADDRESS.label_ref(), 0.9),
-        ];
-        let layer =
-            FilterLayer::new().with_allowed_labels(Some(vec![builtins::PERSON_NAME.label_ref()]));
-        let dropped = apply(layer, &mut entities).await;
-        assert_eq!(entities.len(), 1);
-        assert_eq!(entities[0].label, builtins::PERSON_NAME.label_ref());
-        assert_eq!(dropped.len(), 1);
-        assert_eq!(dropped[0].label, builtins::EMAIL_ADDRESS.label_ref());
-    }
-
-    #[tokio::test]
-    async fn confidence_threshold_drops_below() {
-        let mut entities: Vec<Entity<Text>> = vec![
-            ent(builtins::PERSON_NAME.label_ref(), 0.95),
-            ent(builtins::PERSON_NAME.label_ref(), 0.40),
-        ];
-        let layer =
-            FilterLayer::new().with_confidence_threshold(Some(ConfidenceThreshold::clamped(0.5)));
-        let dropped = apply(layer, &mut entities).await;
-        assert_eq!(entities.len(), 1);
-        assert!(entities[0].confidence.get() >= 0.5);
-        assert_eq!(dropped.len(), 1);
-    }
-
-    #[tokio::test]
-    async fn labels_and_threshold_compose() {
-        let mut entities: Vec<Entity<Text>> = vec![
-            ent(builtins::PERSON_NAME.label_ref(), 0.95),   // keep
-            ent(builtins::PERSON_NAME.label_ref(), 0.40),   // drop: threshold
-            ent(builtins::EMAIL_ADDRESS.label_ref(), 0.95), // drop: label
-        ];
-        let layer = FilterLayer::new()
-            .with_allowed_labels(Some(vec![builtins::PERSON_NAME.label_ref()]))
-            .with_confidence_threshold(Some(ConfidenceThreshold::clamped(0.5)));
-        let dropped = apply(layer, &mut entities).await;
-        assert_eq!(entities.len(), 1);
-        assert_eq!(dropped.len(), 2);
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/fuse/group.rs b/crates/nvisy-toolkit/src/deduplication/fuse/group.rs
deleted file mode 100644
index 8d8e7f420..000000000
--- a/crates/nvisy-toolkit/src/deduplication/fuse/group.rs
+++ /dev/null
@@ -1,214 +0,0 @@
-//! [`GroupingCriteria`]: how entity values and locations are
-//! matched when fusing co-referent detections.
-//!
-//! Grouping runs in two phases:
-//!
-//! 1. **Hash phase**: bucket entities by `(kind, value)` into a
-//!    [`HashMap`] for O(n) partitioning.
-//! 2. **Overlap phase**: within each bucket, sub-group by location
-//!    overlap (O(k²) per bucket, where k is typically small). For
-//!    substring criteria, a cross-bucket merge pass handles value
-//!    containment across different hash keys.
-
-use std::collections::{HashMap, HashSet};
-use std::mem;
-
-use nvisy_core::entity::{Entity, EntityLabelRef};
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::{Modality, Overlap};
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::key::GroupKey;
-
-const TARGET: &str = "nvisy_toolkit::deduplication::fuse::group";
-
-/// How entity values and locations are matched when grouping.
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum GroupingCriteria {
-    /// Exact value match + overlapping location.
-    #[default]
-    Strict,
-    /// Case-insensitive, trimmed value match + overlapping location.
-    Normalized,
-    /// Substring containment (shorter value is prefix/substring of longer)
-    /// + overlapping location. Groups "John" with "John Smith".
-    Narrowing,
-    /// Same as narrowing but ignores location — groups the same entity
-    /// across non-overlapping regions (e.g. cross-chunk deduplication).
-    Widening,
-}
-
-impl GroupingCriteria {
-    /// Whether two values match under this criteria.
-    pub fn values_match(self, a: &str, b: &str) -> bool {
-        match self {
-            Self::Strict => a == b,
-            Self::Normalized => a.trim().eq_ignore_ascii_case(b.trim()),
-            Self::Narrowing | Self::Widening => {
-                let (short, long) = if a.len() <= b.len() { (a, b) } else { (b, a) };
-                long.contains(short)
-            }
-        }
-    }
-
-    /// Whether this criteria requires overlapping locations for grouping.
-    pub fn requires_location_overlap(self) -> bool {
-        !matches!(self, Self::Widening)
-    }
-
-    /// Whether this criteria uses substring containment for value matching.
-    pub fn is_substring(self) -> bool {
-        matches!(self, Self::Narrowing | Self::Widening)
-    }
-
-    /// Normalise a value for HashMap bucketing under this criteria.
-    pub fn bucket_value(self, value: &str) -> String {
-        match self {
-            Self::Normalized => value.trim().to_lowercase(),
-            _ => value.to_owned(),
-        }
-    }
-}
-
-/// Extension trait that groups entities for deduplication.
-pub(super) trait GroupEntities<M: Modality> {
-    /// Partition entities into groups of candidates that should be fused.
-    ///
-    /// Each returned `Vec<Entity>` contains entities that share the same
-    /// kind/value (per the criteria) and overlap in location (unless the
-    /// criteria ignores location).
-    fn group<V: TextAt<M> + ?Sized>(
-        self,
-        criteria: GroupingCriteria,
-        view: &V,
-    ) -> impl Future<Output = Vec<Vec<Entity<M>>>> + Send;
-}
-
-impl<M> GroupEntities<M> for Vec<Entity<M>>
-where
-    M: Modality,
-    M::Location: Overlap,
-{
-    async fn group<V: TextAt<M> + ?Sized>(
-        self,
-        criteria: GroupingCriteria,
-        view: &V,
-    ) -> Vec<Vec<Entity<M>>> {
-        let check_overlap = criteria.requires_location_overlap();
-        let is_substring = criteria.is_substring();
-        let entity_count = self.len();
-
-        // Phase 1: bucket by (kind, value).
-        let mut buckets: HashMap<GroupKey, Vec<Entity<M>>> = HashMap::new();
-        for entity in self {
-            let key = GroupKey::new(&entity, criteria, view).await;
-            buckets.entry(key).or_default().push(entity);
-        }
-
-        tracing::trace!(
-            target: TARGET,
-            entities = entity_count,
-            buckets = buckets.len(),
-            criteria = ?criteria,
-            "hash phase complete",
-        );
-
-        // Phase 2a: within each bucket, sub-group by location overlap.
-        let mut groups: Vec<Vec<Entity<M>>> = Vec::new();
-        let mut label_groups: HashMap<EntityLabelRef, Vec<usize>> = HashMap::new();
-
-        for (_key, bucket) in buckets {
-            let label = bucket[0].label.clone();
-            let mut sub_groups: Vec<Vec<Entity<M>>> = Vec::new();
-
-            for entity in bucket {
-                let bucket_match = sub_groups.iter_mut().find(|g| {
-                    !check_overlap
-                        || g.iter()
-                            .any(|member| member.location.overlaps(&entity.location))
-                });
-                match bucket_match {
-                    Some(g) => g.push(entity),
-                    None => sub_groups.push(vec![entity]),
-                }
-            }
-
-            for sg in sub_groups {
-                let idx = groups.len();
-                groups.push(sg);
-                if is_substring {
-                    label_groups.entry(label.clone()).or_default().push(idx);
-                }
-            }
-        }
-
-        // Phase 2b: for substring criteria, merge groups within the same
-        // kind whose values have a containment relationship.
-        if is_substring {
-            let mut total_merges = 0usize;
-            for indices in label_groups.values() {
-                let mut merged_into: HashSet<usize> = HashSet::new();
-                for i in 0..indices.len() {
-                    if merged_into.contains(&indices[i]) {
-                        continue;
-                    }
-                    for j in (i + 1)..indices.len() {
-                        if merged_into.contains(&indices[j]) {
-                            continue;
-                        }
-                        let mut any_value_match = false;
-                        for a in &groups[indices[i]] {
-                            for b in &groups[indices[j]] {
-                                let va = view.text_at(&a.location).await;
-                                let vb = view.text_at(&b.location).await;
-                                if let (Some(va), Some(vb)) = (va.as_deref(), vb.as_deref())
-                                    && criteria.values_match(va, vb)
-                                {
-                                    any_value_match = true;
-                                    break;
-                                }
-                            }
-                            if any_value_match {
-                                break;
-                            }
-                        }
-                        let location_ok = !check_overlap
-                            || groups[indices[i]].iter().any(|a| {
-                                groups[indices[j]]
-                                    .iter()
-                                    .any(|b| a.location.overlaps(&b.location))
-                            });
-                        if location_ok && any_value_match {
-                            let donor = mem::take(&mut groups[indices[j]]);
-                            groups[indices[i]].extend(donor);
-                            merged_into.insert(indices[j]);
-                            total_merges += 1;
-                        }
-                    }
-                }
-            }
-            groups.retain(|g| !g.is_empty());
-
-            if total_merges > 0 {
-                tracing::trace!(
-                    target: TARGET,
-                    total_merges,
-                    "cross-bucket substring merge complete",
-                );
-            }
-        }
-
-        tracing::debug!(
-            target: TARGET,
-            entities = entity_count,
-            groups = groups.len(),
-            criteria = ?criteria,
-            "grouping complete",
-        );
-
-        groups
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/fuse/key.rs b/crates/nvisy-toolkit/src/deduplication/fuse/key.rs
deleted file mode 100644
index 7b45803be..000000000
--- a/crates/nvisy-toolkit/src/deduplication/fuse/key.rs
+++ /dev/null
@@ -1,43 +0,0 @@
-//! [`GroupKey`]: hash key for the first grouping phase.
-
-use nvisy_core::entity::{Entity, EntityLabelRef};
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::Modality;
-
-use super::group::GroupingCriteria;
-
-/// Hash key for the first grouping phase.
-///
-/// For [`Strict`], [`Narrowing`], and [`Widening`] this stores the
-/// exact value; for [`Normalized`] it stores the lowercased,
-/// trimmed form.
-///
-/// [`Strict`]: GroupingCriteria::Strict
-/// [`Narrowing`]: GroupingCriteria::Narrowing
-/// [`Widening`]: GroupingCriteria::Widening
-/// [`Normalized`]: GroupingCriteria::Normalized
-#[derive(Hash, PartialEq, Eq)]
-pub(super) struct GroupKey {
-    pub(super) label: EntityLabelRef,
-    pub(super) value: String,
-}
-
-impl GroupKey {
-    pub(super) async fn new<M: Modality, V: TextAt<M> + ?Sized>(
-        entity: &Entity<M>,
-        criteria: GroupingCriteria,
-        view: &V,
-    ) -> Self {
-        // Entities without a text value (e.g. image bounding boxes)
-        // get a unique sentinel so they don't all bucket together.
-        // They will still be grouped by location overlap in phase 2.
-        let value = match view.text_at(&entity.location).await {
-            Some(v) => criteria.bucket_value(&v),
-            None => entity.id.to_string(),
-        };
-        Self {
-            label: entity.label.clone(),
-            value,
-        }
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/fuse/mod.rs b/crates/nvisy-toolkit/src/deduplication/fuse/mod.rs
deleted file mode 100644
index be661e92f..000000000
--- a/crates/nvisy-toolkit/src/deduplication/fuse/mod.rs
+++ /dev/null
@@ -1,172 +0,0 @@
-//! [`FuseLayer`]: collapse co-referent entities into one.
-//!
-//! [`GroupingCriteria`] decides *which* entities are candidates
-//! for fusion; [`DeduplicationStrategy`] decides *how* their
-//! confidences combine into the survivor's score.
-
-mod group;
-mod key;
-mod strategy;
-
-use std::cmp::Ordering;
-use std::collections::HashSet;
-
-use nvisy_core::entity::{Entity, TrailStep};
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::{Modality, Overlap};
-use nvisy_core::primitive::Confidence;
-
-use self::group::GroupEntities;
-pub use self::group::GroupingCriteria;
-pub use self::strategy::{DeduplicationStrategy, FusionWeights};
-use super::layer::{Layer, LayerContext};
-use super::span_size::SpanSize;
-
-const TARGET: &str = "nvisy_toolkit::deduplication::fuse";
-
-/// [`Layer`] that groups co-referent entities and merges each group
-/// into a single entity.
-///
-/// Drops nothing — but reshapes the collection (each group of N
-/// entities becomes one). Returns an empty vec from [`Layer::apply`].
-pub struct FuseLayer {
-    strategy: DeduplicationStrategy,
-    criteria: GroupingCriteria,
-}
-
-impl FuseLayer {
-    /// Construct a fuse layer from a strategy + criteria.
-    pub fn new(strategy: DeduplicationStrategy, criteria: GroupingCriteria) -> Self {
-        Self { strategy, criteria }
-    }
-}
-
-#[async_trait::async_trait]
-impl<M, R> Layer<M, R> for FuseLayer
-where
-    M: Modality,
-    M::Location: Overlap + SpanSize,
-    R: TextAt<M> + ?Sized,
-{
-    async fn apply(
-        &self,
-        entities: &mut Vec<Entity<M>>,
-        ctx: &LayerContext<'_, M, R>,
-    ) -> Vec<Entity<M>> {
-        if entities.len() <= 1 {
-            return Vec::new();
-        }
-
-        let owned = std::mem::take(entities);
-        let groups = owned.group(self.criteria, ctx.resolver).await;
-
-        tracing::debug!(
-            target: TARGET,
-            strategy = ?self.strategy,
-            groups = groups.len(),
-            "fusing entity groups",
-        );
-
-        for group in groups {
-            entities.push(fuse_group(&self.strategy, group, ctx.resolver).await);
-        }
-
-        Vec::new()
-    }
-}
-
-/// Fuse a group of co-referent entities into one.
-async fn fuse_group<M, V: TextAt<M> + ?Sized>(
-    strategy: &DeduplicationStrategy,
-    mut group: Vec<Entity<M>>,
-    view: &V,
-) -> Entity<M>
-where
-    M: Modality,
-    M::Location: SpanSize,
-{
-    debug_assert!(!group.is_empty());
-
-    if group.len() == 1 {
-        return group
-            .into_iter()
-            .next()
-            .expect("group.len() == 1 by guard above");
-    }
-
-    let fused_confidence = strategy.compute_confidence(&group);
-
-    // Classify: if every entity in the group has the same set of
-    // recognizer source names, this is plain deduplication (same
-    // detector produced duplicates); otherwise it's ensemble fusion.
-    let label = classify_fusion(&group);
-
-    // Sort by descending confidence: highest-confidence entity
-    // becomes the base since it carries the most trusted metadata.
-    group.sort_by(|a, b| {
-        b.confidence
-            .partial_cmp(&a.confidence)
-            .unwrap_or(Ordering::Equal)
-    });
-    let mut result = group.remove(0);
-    let rest = group;
-
-    // Prefer the largest span: for text, the longer match is more
-    // specific (e.g. "John Smith" over "John"); for images, the
-    // larger bounding box. Adopt the winner's location so the
-    // span stays consistent.
-    for e in &rest {
-        if e.location
-            .span_cmp(&result.location)
-            .unwrap_or(Ordering::Less)
-            == Ordering::Greater
-        {
-            result.location.clone_from(&e.location);
-        }
-    }
-
-    // Merge trails (order-preserving union of every step in every
-    // entity, deduplicated by exact equality).
-    for e in &rest {
-        for step in &e.trail {
-            if !result.trail.contains(step) {
-                result.trail.push(step.clone());
-            }
-        }
-    }
-
-    let before = result.confidence;
-    let after = Confidence::new(fused_confidence.clamp(0.0, 1.0)).expect("clamped to [0,1]");
-    result.confidence = after;
-    let reason = format!("{label} of {} entities", rest.len() + 1);
-    result.trail.push(TrailStep::fusion(before, after, reason));
-
-    let value = view.text_at(&result.location).await.unwrap_or_default();
-    tracing::trace!(
-        target: TARGET,
-        entity_id = %result.id,
-        fused_from = rest.len() + 1,
-        confidence = fused_confidence,
-        label,
-        value,
-        "fused entity group",
-    );
-
-    result
-}
-
-/// Label the group as plain deduplication or ensemble fusion. Used
-/// only for the trail step's `reason` text — no behaviour depends on
-/// this discriminant.
-fn classify_fusion<M: Modality>(group: &[Entity<M>]) -> &'static str {
-    let first: HashSet<&str> = group[0].recognizers().collect();
-    let all_same = group
-        .iter()
-        .skip(1)
-        .all(|e| e.recognizers().collect::<HashSet<_>>() == first);
-    if all_same {
-        "deduplication"
-    } else {
-        "ensemble fusion"
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/fuse/strategy.rs b/crates/nvisy-toolkit/src/deduplication/fuse/strategy.rs
deleted file mode 100644
index 6ced08ad8..000000000
--- a/crates/nvisy-toolkit/src/deduplication/fuse/strategy.rs
+++ /dev/null
@@ -1,184 +0,0 @@
-//! [`DeduplicationStrategy`]: how to combine confidence scores when
-//! fusing a group of co-referent entities.
-
-use std::borrow::Cow;
-use std::collections::HashMap;
-
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::Modality;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Per-recognizer evidence weights used by [`DeduplicationStrategy`]
-/// when combining confidences across a fused group.
-///
-/// Maps a recognizer source name to a weight. Recognizers not
-/// present in the map contribute the neutral weight `1.0`. Both
-/// built-in recognizer names (`"pattern"`, `"ner"`) and
-/// runtime-configured custom names are accepted as keys.
-///
-/// Semantics differ per strategy:
-///
-/// - In [`DeduplicationStrategy::WeightedAverage`], a weight `w`
-///   means "this detector counts `w` times more than an unweighted
-///   one" in the average.
-/// - In [`DeduplicationStrategy::NoisyOr`], a weight `w` means
-///   "this detector contributes `w` units of evidence" — fractional
-///   weights dampen correlated detectors.
-#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(transparent)]
-pub struct FusionWeights(HashMap<Cow<'static, str>, f64>);
-
-impl FusionWeights {
-    /// Empty weight map.
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Insert a weight for a recognizer name.
-    pub fn insert(&mut self, recognizer: impl Into<Cow<'static, str>>, weight: f64) {
-        self.0.insert(recognizer.into(), weight);
-    }
-
-    /// Look up the weight for a recognizer name, or `None`.
-    pub fn get(&self, recognizer: &str) -> Option<f64> {
-        self.0.get(recognizer).copied()
-    }
-
-    /// True when no weights are registered.
-    pub fn is_empty(&self) -> bool {
-        self.0.is_empty()
-    }
-
-    /// Number of registered weights.
-    pub fn len(&self) -> usize {
-        self.0.len()
-    }
-}
-
-impl<K, V> FromIterator<(K, V)> for FusionWeights
-where
-    K: Into<Cow<'static, str>>,
-    V: Into<f64>,
-{
-    fn from_iter<I: IntoIterator<Item = (K, V)>>(iter: I) -> Self {
-        Self(
-            iter.into_iter()
-                .map(|(k, v)| (k.into(), v.into()))
-                .collect(),
-        )
-    }
-}
-
-/// Strategy for combining confidence scores from multiple detectors.
-#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
-#[serde(tag = "kind", rename_all = "snake_case")]
-pub enum DeduplicationStrategy {
-    /// Take the maximum confidence across all detectors.
-    #[default]
-    MaxConfidence,
-    /// Weighted average across the group: each entity's confidence
-    /// contributes proportionally to its weight; the result is
-    /// `Σ(wᵢ·pᵢ) / Σ(wᵢ)`.
-    ///
-    /// Per-entity weight comes from the entity's originating
-    /// recognizer — the source of its first `Recognition` trail
-    /// step. (Each input entity at fuse time has its own recognition
-    /// step(s) but hasn't been merged with siblings yet, so the
-    /// first one is unambiguously what produced this entity.) If
-    /// that recognizer isn't in `weights`, the per-entity weight
-    /// defaults to `1.0`. `WeightedAverage { weights: empty }`
-    /// reduces to a plain unweighted average.
-    ///
-    /// # Non-monotonicity
-    ///
-    /// Unlike [`MaxConfidence`] or [`NoisyOr`], the fused
-    /// confidence can be *lower* than the highest input. A
-    /// high-confidence hit from a low-weight detector gets pulled
-    /// down by a lower-confidence hit from a high-weight detector.
-    /// This is the price of expressing per-detector trust. If you
-    /// want "more detectors agreeing should only increase
-    /// confidence," use [`NoisyOr`] instead.
-    ///
-    /// [`MaxConfidence`]: Self::MaxConfidence
-    /// [`NoisyOr`]: Self::NoisyOr
-    WeightedAverage {
-        /// Per-recognizer weight.
-        #[serde(default)]
-        weights: FusionWeights,
-    },
-    /// Weighted noisy-OR: `P = 1 − ∏(1 − pᵢ)^wᵢ`.
-    ///
-    /// Unweighted (`weights: empty`) reduces to standard noisy-OR
-    /// `P = 1 − ∏(1 − pᵢ)`, which treats each detector as an
-    /// independent witness — each contributes its own evidence and
-    /// the joint probability that *all* are wrong is `∏(1 − pᵢ)`.
-    ///
-    /// # Correlation: use weights
-    ///
-    /// Independence is rarely literally true. Two regex patterns
-    /// matching the same SSN format are ~100% correlated, not
-    /// independent; standard noisy-OR will double-count and
-    /// inflate. Set fractional weights on correlated detectors so
-    /// each contributes only as much fresh evidence as it carries
-    /// — two redundant patterns at `weights: { "pattern-a": 0.5,
-    /// "pattern-b": 0.5 }` together count as one full witness.
-    ///
-    /// Per-entity weight uses the entity's originating recognizer
-    /// (first `Recognition` trail step), matching [`WeightedAverage`].
-    /// Missing recognizers contribute the standard unit weight
-    /// `1.0` (vanilla noisy-OR contribution).
-    ///
-    /// [`WeightedAverage`]: Self::WeightedAverage
-    NoisyOr {
-        /// Per-recognizer evidence weight.
-        #[serde(default)]
-        weights: FusionWeights,
-    },
-}
-
-impl DeduplicationStrategy {
-    /// Combine confidences across `group` per this strategy.
-    pub(super) fn compute_confidence<M: Modality>(&self, group: &[Entity<M>]) -> f64 {
-        debug_assert!(
-            !group.is_empty(),
-            "fuse never calls compute_confidence with an empty group",
-        );
-        match self {
-            Self::MaxConfidence => group
-                .iter()
-                .map(|e| e.confidence.get())
-                .fold(f64::NEG_INFINITY, f64::max),
-
-            Self::NoisyOr { weights } => {
-                1.0 - group
-                    .iter()
-                    .map(|e| {
-                        let w = entity_weight(e, weights);
-                        (1.0 - e.confidence.get()).powf(w)
-                    })
-                    .product::<f64>()
-            }
-
-            Self::WeightedAverage { weights } => {
-                let (wsum, total_w) = group.iter().fold((0.0_f64, 0.0_f64), |(s, t), e| {
-                    let w = entity_weight(e, weights);
-                    (s + e.confidence.get() * w, t + w)
-                });
-                wsum / total_w
-            }
-        }
-    }
-}
-
-/// Look up the weight for an entity's originating recognizer (the
-/// source of its first `Recognition` trail step). Missing-from-map
-/// and missing-trail-step both fall through to the neutral weight
-/// `1.0`.
-fn entity_weight<M: Modality>(entity: &Entity<M>, weights: &FusionWeights) -> f64 {
-    entity
-        .recognizers()
-        .next()
-        .and_then(|name| weights.get(name))
-        .unwrap_or(1.0)
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/layer.rs b/crates/nvisy-toolkit/src/deduplication/layer.rs
deleted file mode 100644
index 54b578ec9..000000000
--- a/crates/nvisy-toolkit/src/deduplication/layer.rs
+++ /dev/null
@@ -1,99 +0,0 @@
-//! [`Layer`]: one stage of an entity-processing pipeline.
-//!
-//! A layer takes a mutable entity collection plus a read-only
-//! [`LayerContext`], mutates the entities in place, and returns
-//! whichever entities it dropped so the pipeline can attribute
-//! them in its drop-reason roll-up. Layers that only *add* to or
-//! *reshape* the collection return an empty `Vec`.
-//!
-//! The four built-in layers ([`CalibrateLayer`], [`FilterLayer`],
-//! [`FuseLayer`], [`ResolveConflictsLayer`]) cover the canonical
-//! recipe. Operators compose them — or their own custom layers —
-//! through [`LayerPipeline`].
-//!
-//! # Generic over the resolver
-//!
-//! Both `Layer<M, R>` and [`LayerContext<'_, M, R>`] are
-//! parameterised by `R: TextAt<M>` so the value-resolver call
-//! (only [`FuseLayer`] needs it among
-//! the built-ins) is monomorphised. Object safety still holds:
-//! [`LayerPipeline`] stores `Box<dyn Layer<M, R>>` for a specific
-//! `R` chosen at construction.
-//!
-//! [`CalibrateLayer`]: super::calibrate::CalibrateLayer
-//! [`FilterLayer`]: super::filter::FilterLayer
-//! [`FuseLayer`]: super::fuse::FuseLayer
-//! [`ResolveConflictsLayer`]: super::resolve::ResolveConflictsLayer
-//! [`LayerPipeline`]: super::pipeline::LayerPipeline
-//! [`fuse`]: super::fuse::FuseLayer
-
-use std::marker::PhantomData;
-
-use nvisy_core::entity::Entity;
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::Modality;
-use uuid::Uuid;
-
-/// Read-only context every [`Layer::apply`] call receives.
-///
-/// Carries the value resolver (for layers that need to read source
-/// text at a location — fuse, today) and an optional correlation id
-/// (for tracing — see [`RecognizerInput::correlation_id`]).
-///
-/// Built with [`Self::new`] then optionally extended through the
-/// `with_*` setters:
-///
-/// ```ignore
-/// let ctx = LayerContext::new(&resolver).with_correlation_id(run_id);
-/// pipeline.run(entities, &ctx).await;
-/// ```
-///
-/// [`RecognizerInput::correlation_id`]: nvisy_core::recognition::RecognizerInput::correlation_id
-pub struct LayerContext<'a, M: Modality, R: TextAt<M> + ?Sized> {
-    /// Resolver for "what value sits at this location?". Layers that
-    /// only inspect entity metadata can ignore this field.
-    pub resolver: &'a R,
-    /// Optional correlation id used to stitch tracing spans across
-    /// the run. `None` means the layer (or pipeline) emits its spans
-    /// without a correlation id, which is fine for ad-hoc / test
-    /// usage.
-    pub correlation_id: Option<Uuid>,
-    /// Phantom binding `M` into the lifetime/type so the trait bound
-    /// on `R` carries through without an unused-param error.
-    _marker: PhantomData<&'a M>,
-}
-
-impl<'a, M: Modality, R: TextAt<M> + ?Sized> LayerContext<'a, M, R> {
-    /// Construct a context with just a value resolver; correlation
-    /// id defaults to `None`.
-    pub fn new(resolver: &'a R) -> Self {
-        Self {
-            resolver,
-            correlation_id: None,
-            _marker: PhantomData,
-        }
-    }
-
-    /// Attach a correlation id (typically a run id).
-    pub fn with_correlation_id(mut self, correlation_id: Uuid) -> Self {
-        self.correlation_id = Some(correlation_id);
-        self
-    }
-}
-
-/// One stage of an entity-processing pipeline.
-///
-/// Each layer mutates `entities` in place; the returned `Vec` is the
-/// list of entities the layer **dropped**. Layers that don't drop
-/// anything (calibrate, fuse) return an empty vec; layers that filter
-/// or resolve conflicts return the discarded entities so the pipeline
-/// can roll up drop counts.
-#[async_trait::async_trait]
-pub trait Layer<M: Modality, R: TextAt<M> + ?Sized>: Send + Sync {
-    /// Run this layer against `entities`. Returns the dropped vec.
-    async fn apply(
-        &self,
-        entities: &mut Vec<Entity<M>>,
-        ctx: &LayerContext<'_, M, R>,
-    ) -> Vec<Entity<M>>;
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/mod.rs b/crates/nvisy-toolkit/src/deduplication/mod.rs
deleted file mode 100644
index 26b6784fb..000000000
--- a/crates/nvisy-toolkit/src/deduplication/mod.rs
+++ /dev/null
@@ -1,77 +0,0 @@
-//! Entity deduplication: composable layers run through a
-//! [`LayerPipeline`].
-//!
-//! # Layer submodules
-//!
-//! Each of the four canonical step kinds is its own public
-//! submodule, named for what it does to the entity set:
-//!
-//! - [`calibrate`] — per-recognizer confidence scaling.
-//! - [`filter`] — drop entities outside the allowed kinds or below
-//!   a confidence floor.
-//! - [`fuse`] — group + combine co-referent entities.
-//! - [`resolve`] — break cross-kind span overlaps.
-//!
-//! # Plumbing (re-exported at the root)
-//!
-//! - [`Layer`] / [`LayerContext`] — the trait every step implements
-//!   plus its per-call context.
-//! - [`LayerPipeline`] — the orchestrator that runs a stack of
-//!   layers in order.
-//! - [`LayerParams`] — the per-call knob bag callers set.
-//! - [`SpanSize`] — helper for span-length tiebreaks (used by
-//!   `fuse` and `resolve` internally; exposed for custom layers).
-//!
-//! # Canonical recipe
-//!
-//! [`LayerPipeline::from_params`] assembles the canonical four-step
-//! recipe from a [`LayerParams`]:
-//!
-//! 1. **Calibrate** raw confidence scores per-recognizer.
-//! 2. **Filter** by allowed kinds + confidence floor.
-//! 3. **Fuse** co-referent entities into one (group + combine).
-//! 4. **Suppress** entities whose matched text is on a
-//!    caller-supplied allow list.
-//! 5. **Resolve conflicts** between different kinds on the same span.
-//!
-//! Operators can swap steps, drop steps, or insert their own custom
-//! [`Layer`] impls by building the pipeline manually with
-//! [`LayerPipeline::new`] + [`LayerPipeline::with_layer`].
-
-pub mod calibrate;
-pub mod filter;
-pub mod fuse;
-pub mod resolve;
-pub mod suppress;
-
-mod layer;
-mod params;
-mod pipeline;
-mod span_size;
-
-#[cfg(test)]
-use std::marker::PhantomData;
-
-#[cfg(test)]
-use nvisy_core::extraction::TextAt;
-#[cfg(test)]
-use nvisy_core::modality::Modality;
-
-pub use self::layer::{Layer, LayerContext};
-pub use self::params::LayerParams;
-pub use self::pipeline::LayerPipeline;
-pub use self::span_size::SpanSize;
-
-#[cfg(test)]
-pub(crate) fn test_resolver<M: Modality>() -> Box<dyn TextAt<M>> {
-    struct Noop<M>(PhantomData<M>);
-
-    #[async_trait::async_trait]
-    impl<M: Modality> TextAt<M> for Noop<M> {
-        async fn text_at(&self, _location: &M::Location) -> Option<String> {
-            None
-        }
-    }
-
-    Box::new(Noop::<M>(PhantomData))
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/params.rs b/crates/nvisy-toolkit/src/deduplication/params.rs
deleted file mode 100644
index 8f73b6ae2..000000000
--- a/crates/nvisy-toolkit/src/deduplication/params.rs
+++ /dev/null
@@ -1,70 +0,0 @@
-//! [`LayerParams`]: the per-call knob bag that drives the
-//! canonical deduplication recipe.
-//!
-//! Bundles every per-layer setting the five-step recipe needs
-//! ([`CalibrationMap`], filtering thresholds + allowed kinds,
-//! [`DeduplicationStrategy`], [`GroupingCriteria`],
-//! [`SuppressionParams`], [`ConflictResolution`]) into a single
-//! deserialisable shape callers set once per request.
-//! [`LayerPipeline::from_params`] reads it and assembles the
-//! five-step pipeline.
-//!
-//! [`CalibrationMap`]: super::calibrate::CalibrationMap
-//! [`DeduplicationStrategy`]: super::fuse::DeduplicationStrategy
-//! [`GroupingCriteria`]: super::fuse::GroupingCriteria
-//! [`SuppressionParams`]: super::suppress::SuppressionParams
-//! [`ConflictResolution`]: super::resolve::ConflictResolution
-//! [`LayerPipeline::from_params`]: super::pipeline::LayerPipeline::from_params
-
-use nvisy_core::entity::EntityLabelRef;
-use nvisy_core::primitive::ConfidenceThreshold;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::calibrate::CalibrationMap;
-use super::fuse::{DeduplicationStrategy, GroupingCriteria};
-use super::resolve::ConflictResolution;
-use super::suppress::SuppressionParams;
-
-/// Configuration for the deduplication pipeline's five-step recipe.
-///
-/// Owns the sole confidence threshold in the pipeline: detection
-/// layers and recognizers do not filter on confidence themselves —
-/// per-method skew is folded in via [`calibration`], and the
-/// resulting calibrated score is checked against
-/// [`confidence_threshold`] here.
-///
-/// [`calibration`]: Self::calibration
-/// [`confidence_threshold`]: Self::confidence_threshold
-#[derive(Debug, Clone, Default, PartialEq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct LayerParams {
-    /// Per-recognizer confidence scaling applied first.
-    #[serde(default, skip_serializing_if = "CalibrationMap::is_empty")]
-    pub calibration: CalibrationMap,
-    /// Drop entities whose `label` is outside this set. `None`
-    /// keeps every label.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub allowed_labels: Option<Vec<EntityLabelRef>>,
-    /// Minimum calibrated confidence an entity must clear to survive
-    /// the filter step. `None` keeps every confidence level.
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub confidence_threshold: Option<ConfidenceThreshold>,
-    /// How to match entity values and locations when grouping for
-    /// fusion.
-    #[serde(default)]
-    pub grouping: GroupingCriteria,
-    /// Strategy for combining confidence scores within a fused
-    /// group.
-    #[serde(default)]
-    pub strategy: DeduplicationStrategy,
-    /// Allow-list inputs consumed by [`SuppressionLayer`].
-    ///
-    /// [`SuppressionLayer`]: super::suppress::SuppressionLayer
-    #[serde(default, skip_serializing_if = "SuppressionParams::is_empty")]
-    pub suppression: SuppressionParams,
-    /// How to resolve conflicts when different entity kinds overlap
-    /// the same span.
-    #[serde(default)]
-    pub conflict_resolution: ConflictResolution,
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/pipeline.rs b/crates/nvisy-toolkit/src/deduplication/pipeline.rs
deleted file mode 100644
index 35bc7ea58..000000000
--- a/crates/nvisy-toolkit/src/deduplication/pipeline.rs
+++ /dev/null
@@ -1,196 +0,0 @@
-//! [`LayerPipeline`]: ordered stack of [`Layer`]s run head-to-tail.
-//!
-//! Build a pipeline either from canonical defaults
-//! ([`LayerPipeline::from_params`]) or from individual layers
-//! ([`LayerPipeline::new`] + [`LayerPipeline::with_layer`]). Run it
-//! against an entity list with [`LayerPipeline::run`].
-
-use std::marker::PhantomData;
-
-use nvisy_core::Error;
-use nvisy_core::entity::Entity;
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::{Modality, Overlap};
-
-use super::calibrate::CalibrateLayer;
-use super::filter::FilterLayer;
-use super::fuse::FuseLayer;
-use super::layer::{Layer, LayerContext};
-use super::params::LayerParams;
-use super::resolve::ResolveConflictsLayer;
-use super::span_size::SpanSize;
-use super::suppress::SuppressionLayer;
-
-const TARGET: &str = "nvisy_toolkit::deduplication";
-
-/// Ordered stack of layers, run head-to-tail against an entity list.
-///
-/// Construction is open: callers can compose any sequence of
-/// [`Layer<M, R>`] impls (built-in or custom). The canonical
-/// four-layer dedup recipe is available through [`Self::from_params`].
-///
-/// `R` is the resolver type passed in through the [`LayerContext`] at
-/// `run` time. It's a type parameter on the pipeline (rather than a
-/// `dyn TextAt<M>`) so each layer's `text_at` call is
-/// monomorphised. Layers that don't touch the resolver compile
-/// against any `R: TextAt<M>` uniformly.
-pub struct LayerPipeline<M: Modality, R: TextAt<M> + ?Sized> {
-    layers: Vec<Box<dyn Layer<M, R>>>,
-    _marker: PhantomData<fn(&M, &R)>,
-}
-
-impl<M: Modality, R: TextAt<M> + ?Sized> LayerPipeline<M, R> {
-    /// Empty pipeline. Use [`Self::with_layer`] to append layers.
-    pub fn new() -> Self {
-        Self {
-            layers: Vec::new(),
-            _marker: PhantomData,
-        }
-    }
-
-    /// Append a layer.
-    pub fn with_layer<L: Layer<M, R> + 'static>(mut self, layer: L) -> Self {
-        self.layers.push(Box::new(layer));
-        self
-    }
-
-    /// Run every layer in registration order against `entities`.
-    ///
-    /// The pipeline aggregates each layer's dropped count into a
-    /// single info log line per run. Layers themselves emit per-layer
-    /// debug/trace spans.
-    pub async fn run(
-        &self,
-        mut entities: Vec<Entity<M>>,
-        ctx: &LayerContext<'_, M, R>,
-    ) -> Vec<Entity<M>> {
-        if entities.is_empty() {
-            return entities;
-        }
-        let before = entities.len();
-
-        let mut dropped_total = 0usize;
-        for layer in &self.layers {
-            let dropped = layer.apply(&mut entities, ctx).await;
-            dropped_total += dropped.len();
-        }
-
-        tracing::info!(
-            target: TARGET,
-            before,
-            after = entities.len(),
-            reduced = before.saturating_sub(entities.len()),
-            dropped = dropped_total,
-            "layer pipeline complete",
-        );
-
-        entities
-    }
-}
-
-impl<M, R> LayerPipeline<M, R>
-where
-    M: Modality,
-    M::Location: Overlap + SpanSize,
-    R: TextAt<M> + ?Sized,
-{
-    /// Build the canonical five-layer recipe: calibrate → filter →
-    /// fuse → suppress → resolve. Every layer's config is read
-    /// from `params`.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when any
-    /// `params.allow_values_regex` entry fails to compile.
-    pub fn from_params(params: &LayerParams) -> Result<Self, Error> {
-        let filter = FilterLayer::new()
-            .with_allowed_labels(params.allowed_labels.clone())
-            .with_confidence_threshold(params.confidence_threshold);
-        let suppress = SuppressionLayer::from_params(&params.suppression)?;
-        Ok(Self::new()
-            .with_layer(CalibrateLayer::new(params.calibration.clone()))
-            .with_layer(filter)
-            .with_layer(FuseLayer::new(params.strategy.clone(), params.grouping))
-            .with_layer(suppress)
-            .with_layer(ResolveConflictsLayer::new(params.conflict_resolution)))
-    }
-}
-
-impl<M: Modality, R: TextAt<M> + ?Sized> Default for LayerPipeline<M, R> {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::sync::Arc;
-
-    use async_trait::async_trait;
-    use nvisy_core::entity::{Entity, builtins};
-    use nvisy_core::modality::{Text, TextLocation};
-
-    use super::*;
-    use crate::deduplication::suppress::SuppressionParams;
-
-    struct TextSliceResolver(Arc<String>);
-
-    #[async_trait]
-    impl TextAt<Text> for TextSliceResolver {
-        async fn text_at(&self, location: &TextLocation) -> Option<String> {
-            self.0.get(location.start..location.end).map(String::from)
-        }
-    }
-
-    fn email(start: usize, end: usize) -> Entity<Text> {
-        Entity::test_builder(start, end)
-            .with_label(builtins::EMAIL_ADDRESS.label_ref())
-            .test_build()
-    }
-
-    fn url(start: usize, end: usize) -> Entity<Text> {
-        Entity::test_builder(start, end)
-            .with_label(builtins::URL.label_ref())
-            .test_build()
-    }
-
-    /// Pipeline-order contract: fuse collapses same-kind duplicates
-    /// before suppress sees them; suppress drops allowlisted
-    /// entities before resolve adjudicates cross-kind conflicts.
-    ///
-    /// Inputs: two PERSON_NAME hits at the same span (duplicates
-    /// of an allowlisted email-like value) plus one URL hit at an
-    /// overlapping span. After the pipeline, only the URL should
-    /// survive.
-    ///
-    /// Without `fuse → suppress` ordering, the duplicate would
-    /// survive (suppress only drops one of the two). Without
-    /// `suppress → resolve` ordering, the resolve step would
-    /// pick a winner between EMAIL and URL — possibly the EMAIL —
-    /// before the allow-list could remove it.
-    #[tokio::test]
-    async fn fuse_then_suppress_then_resolve() {
-        let source = "noreply@foo.com /docs";
-        let resolver = TextSliceResolver(Arc::new(source.to_owned()));
-
-        let params = LayerParams {
-            suppression: SuppressionParams::new()
-                .with_allow_values(vec!["noreply@foo.com".to_owned()]),
-            ..Default::default()
-        };
-
-        let pipeline: LayerPipeline<Text, _> =
-            LayerPipeline::from_params(&params).expect("pipeline builds");
-        let ctx = LayerContext::new(&resolver);
-
-        // Two EMAIL hits at [0, 15) — same kind, same span, fuse
-        // collapses to one. The collapsed entity matches the
-        // allow-list; suppress drops it. A URL at [0, 21) remains
-        // for resolve to leave untouched.
-        let entities = vec![email(0, 15), email(0, 15), url(0, 21)];
-
-        let survivors = pipeline.run(entities, &ctx).await;
-        assert_eq!(survivors.len(), 1, "only the URL should survive");
-        assert_eq!(survivors[0].label, builtins::URL.label_ref());
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/resolve/mod.rs b/crates/nvisy-toolkit/src/deduplication/resolve/mod.rs
deleted file mode 100644
index d4ed14abf..000000000
--- a/crates/nvisy-toolkit/src/deduplication/resolve/mod.rs
+++ /dev/null
@@ -1,188 +0,0 @@
-//! [`ResolveConflictsLayer`]: drop the loser when different kinds
-//! overlap the same span.
-//!
-//! When entities of different kinds claim the same text (e.g.
-//! `"555-1234"` matches both `PhoneNumber` and `EmailAddress`),
-//! this layer keeps only the winner per the configured
-//! [`ConflictResolution`] strategy.
-
-mod strategy;
-
-use nvisy_core::entity::Entity;
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::{Modality, Overlap};
-
-pub use self::strategy::ConflictResolution;
-use super::layer::{Layer, LayerContext};
-use super::span_size::SpanSize;
-
-const TARGET: &str = "nvisy_toolkit::deduplication::resolve";
-
-/// [`Layer`] that resolves cross-kind span overlaps in place per the
-/// given [`ConflictResolution`] strategy.
-///
-/// Returns the dropped (losing) entities from [`Layer::apply`].
-pub struct ResolveConflictsLayer {
-    strategy: ConflictResolution,
-}
-
-impl ResolveConflictsLayer {
-    /// Construct a resolve layer from a strategy.
-    pub fn new(strategy: ConflictResolution) -> Self {
-        Self { strategy }
-    }
-}
-
-#[async_trait::async_trait]
-impl<M, R> Layer<M, R> for ResolveConflictsLayer
-where
-    M: Modality,
-    M::Location: Overlap + SpanSize,
-    R: TextAt<M> + ?Sized,
-{
-    async fn apply(
-        &self,
-        entities: &mut Vec<Entity<M>>,
-        _ctx: &LayerContext<'_, M, R>,
-    ) -> Vec<Entity<M>> {
-        if entities.len() <= 1 {
-            return Vec::new();
-        }
-
-        let len = entities.len();
-        let mut losers = vec![false; len];
-        let mut resolved = 0usize;
-
-        for i in 0..len {
-            if losers[i] {
-                continue;
-            }
-            for j in (i + 1)..len {
-                if losers[j] {
-                    continue;
-                }
-                if entities[i].label == entities[j].label {
-                    continue;
-                }
-                if !entities[i].location.overlaps(&entities[j].location) {
-                    continue;
-                }
-
-                if self.strategy.keeps_first(&entities[i], &entities[j]) {
-                    losers[j] = true;
-                } else {
-                    losers[i] = true;
-                }
-                resolved += 1;
-            }
-        }
-
-        if resolved > 0 {
-            tracing::debug!(
-                target: TARGET,
-                resolved,
-                strategy = ?self.strategy,
-                "span conflicts resolved",
-            );
-        }
-
-        let mut dropped = Vec::new();
-        let mut idx = 0usize;
-        entities.retain(|entity| {
-            let lost = losers[idx];
-            idx += 1;
-            if lost {
-                dropped.push(entity.clone());
-            }
-            !lost
-        });
-        dropped
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{Entity, builtins};
-    use nvisy_core::modality::Text;
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-    use crate::deduplication::test_resolver;
-
-    fn conf(v: f64) -> Confidence {
-        Confidence::new(v).expect("confidence in [0,1]")
-    }
-
-    async fn apply<M>(strategy: ConflictResolution, entities: &mut Vec<Entity<M>>) -> Vec<Entity<M>>
-    where
-        M: Modality,
-        M::Location: Overlap + SpanSize,
-    {
-        let resolver = test_resolver::<M>();
-        let ctx = LayerContext::new(&*resolver);
-        let layer = ResolveConflictsLayer::new(strategy);
-        layer.apply(entities, &ctx).await
-    }
-
-    #[tokio::test]
-    async fn highest_confidence_keeps_winner() {
-        let mut entities: Vec<Entity<Text>> = vec![
-            Entity::test_builder(0, 8)
-                .with_label(builtins::PHONE_NUMBER.label_ref())
-                .test_build(),
-            Entity::test_builder(0, 8)
-                .with_label(builtins::EMAIL_ADDRESS.label_ref())
-                .with_confidence(conf(0.8))
-                .test_build(),
-        ];
-        let _ = apply(ConflictResolution::HighestConfidence, &mut entities).await;
-        assert_eq!(entities.len(), 1);
-        assert_eq!(entities[0].label, builtins::PHONE_NUMBER.label_ref());
-    }
-
-    #[tokio::test]
-    async fn non_overlapping_not_resolved() {
-        let mut entities: Vec<Entity<Text>> = vec![
-            Entity::test_builder(0, 8)
-                .with_label(builtins::PHONE_NUMBER.label_ref())
-                .test_build(),
-            Entity::test_builder(20, 24)
-                .with_label(builtins::EMAIL_ADDRESS.label_ref())
-                .with_confidence(conf(0.8))
-                .test_build(),
-        ];
-        let _ = apply(ConflictResolution::HighestConfidence, &mut entities).await;
-        assert_eq!(entities.len(), 2);
-    }
-
-    #[tokio::test]
-    async fn same_kind_not_resolved() {
-        let mut entities: Vec<Entity<Text>> = vec![
-            Entity::test_builder(0, 8)
-                .with_label(builtins::PHONE_NUMBER.label_ref())
-                .test_build(),
-            Entity::test_builder(0, 8)
-                .with_label(builtins::PHONE_NUMBER.label_ref())
-                .with_confidence(conf(0.8))
-                .test_build(),
-        ];
-        let _ = apply(ConflictResolution::HighestConfidence, &mut entities).await;
-        assert_eq!(entities.len(), 2);
-    }
-
-    #[tokio::test]
-    async fn longest_span_keeps_longer() {
-        let mut entities: Vec<Entity<Text>> = vec![
-            Entity::test_builder(0, 3)
-                .with_label(builtins::PHONE_NUMBER.label_ref())
-                .test_build(),
-            Entity::test_builder(0, 8)
-                .with_label(builtins::EMAIL_ADDRESS.label_ref())
-                .with_confidence(conf(0.7))
-                .test_build(),
-        ];
-        let _ = apply(ConflictResolution::LongestSpan, &mut entities).await;
-        assert_eq!(entities.len(), 1);
-        assert_eq!(entities[0].label, builtins::EMAIL_ADDRESS.label_ref());
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/resolve/strategy.rs b/crates/nvisy-toolkit/src/deduplication/resolve/strategy.rs
deleted file mode 100644
index 5b7ccbc0f..000000000
--- a/crates/nvisy-toolkit/src/deduplication/resolve/strategy.rs
+++ /dev/null
@@ -1,40 +0,0 @@
-//! [`ConflictResolution`]: how to resolve cross-kind span overlaps.
-
-use std::cmp::Ordering;
-
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::Modality;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use crate::deduplication::span_size::SpanSize;
-
-/// How to resolve conflicts when different entity kinds overlap the
-/// same text span.
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum ConflictResolution {
-    /// Keep only the entity with the highest confidence score.
-    #[default]
-    HighestConfidence,
-    /// Keep only the entity with the longest span (most specific match).
-    LongestSpan,
-}
-
-impl ConflictResolution {
-    /// `true` when `a` wins the overlap against `b` under this
-    /// strategy.
-    pub(super) fn keeps_first<M>(&self, a: &Entity<M>, b: &Entity<M>) -> bool
-    where
-        M: Modality,
-        M::Location: SpanSize,
-    {
-        match self {
-            Self::HighestConfidence => a.confidence.get() >= b.confidence.get(),
-            Self::LongestSpan => {
-                a.location.span_cmp(&b.location).unwrap_or(Ordering::Equal) != Ordering::Less
-            }
-        }
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/span_size.rs b/crates/nvisy-toolkit/src/deduplication/span_size.rs
deleted file mode 100644
index a9a38c16c..000000000
--- a/crates/nvisy-toolkit/src/deduplication/span_size.rs
+++ /dev/null
@@ -1,117 +0,0 @@
-//! [`SpanSize`]: compare the spatial / temporal extent of two
-//! same-modality locations.
-//!
-//! Used by [`FuseLayer`] and [`ResolveConflictsLayer`] to pick the
-//! most representative span when merging or breaking ties between
-//! entities.
-//!
-//! [`FuseLayer`]: super::fuse::FuseLayer
-//! [`ResolveConflictsLayer`]: super::resolve::ResolveConflictsLayer
-
-use std::cmp::Ordering;
-
-use nvisy_core::modality::{AudioLocation, ImageLocation, TabularLocation, TextLocation};
-
-/// Extension trait for comparing the spatial/temporal extent of two
-/// locations of the same modality.
-///
-/// Used during deduplication to select the most representative span
-/// when merging a group of duplicate entities (e.g. prefer "John Smith"
-/// over "John", or the larger bounding box).
-///
-/// Size metric per modality:
-/// - **Text**: byte length (`end - start`).
-/// - **Image**: bounding box area (`width * height`).
-/// - **Audio**: time span duration.
-/// - **Tabular**: cell text length.
-pub trait SpanSize {
-    /// Compare two locations by their modality's notion of "size".
-    /// Returns `None` when the modality offers no meaningful size
-    /// (e.g. tabular cells without intra-cell offsets).
-    fn span_cmp(&self, other: &Self) -> Option<Ordering>;
-}
-
-impl SpanSize for TextLocation {
-    fn span_cmp(&self, other: &Self) -> Option<Ordering> {
-        Some(self.len().cmp(&other.len()))
-    }
-}
-
-impl SpanSize for ImageLocation {
-    fn span_cmp(&self, other: &Self) -> Option<Ordering> {
-        self.area().partial_cmp(&other.area())
-    }
-}
-
-impl SpanSize for AudioLocation {
-    fn span_cmp(&self, other: &Self) -> Option<Ordering> {
-        Some(
-            self.time_span
-                .duration_us()
-                .cmp(&other.time_span.duration_us()),
-        )
-    }
-}
-
-impl SpanSize for TabularLocation {
-    fn span_cmp(&self, other: &Self) -> Option<Ordering> {
-        let len_a = self
-            .end_offset
-            .unwrap_or(0)
-            .saturating_sub(self.start_offset.unwrap_or(0));
-        let len_b = other
-            .end_offset
-            .unwrap_or(0)
-            .saturating_sub(other.start_offset.unwrap_or(0));
-        Some(len_a.cmp(&len_b))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::cmp::Ordering;
-
-    use nvisy_core::primitive::{BoundingBox, TimeSpan};
-
-    use super::*;
-
-    #[test]
-    fn text_larger_span_wins() {
-        let a = TextLocation::new(0, 10);
-        let b = TextLocation::new(0, 5);
-        assert_eq!(a.span_cmp(&b), Some(Ordering::Greater));
-        assert_eq!(b.span_cmp(&a), Some(Ordering::Less));
-    }
-
-    #[test]
-    fn text_equal_spans() {
-        let a = TextLocation::new(0, 5);
-        let b = TextLocation::new(3, 8);
-        assert_eq!(a.span_cmp(&b), Some(Ordering::Equal));
-    }
-
-    #[test]
-    fn image_larger_area_wins() {
-        let a = ImageLocation::new(BoundingBox {
-            x: 0.0,
-            y: 0.0,
-            width: 100.0,
-            height: 100.0,
-        });
-        let b = ImageLocation::new(BoundingBox {
-            x: 0.0,
-            y: 0.0,
-            width: 50.0,
-            height: 50.0,
-        });
-        assert_eq!(a.span_cmp(&b), Some(Ordering::Greater));
-        assert_eq!(b.span_cmp(&a), Some(Ordering::Less));
-    }
-
-    #[test]
-    fn audio_longer_duration_wins() {
-        let a = AudioLocation::new(TimeSpan::new(0, 10_000_000));
-        let b = AudioLocation::new(TimeSpan::new(0, 5_000_000));
-        assert_eq!(a.span_cmp(&b), Some(Ordering::Greater));
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/suppress/mod.rs b/crates/nvisy-toolkit/src/deduplication/suppress/mod.rs
deleted file mode 100644
index 71b46a263..000000000
--- a/crates/nvisy-toolkit/src/deduplication/suppress/mod.rs
+++ /dev/null
@@ -1,366 +0,0 @@
-//! [`SuppressionLayer`]: drop entities whose matched text is on a
-//! caller-supplied allow list.
-//!
-//! See [`SuppressionParams`] for the three allow-list shapes.
-//!
-//! All three operate on the **entity's resolved text** (sliced from
-//! the source via [`TextAt::text_at`]), not the surrounding
-//! document. When the resolver returns `None` (e.g. malformed
-//! location), the entity is kept — better to surface a false
-//! positive than silently drop something we can't verify.
-//!
-//! Returns dropped entities from [`Layer::apply`] so the pipeline
-//! can attribute them in its drop-reason roll-up.
-//!
-//! [`Layer::apply`]: super::layer::Layer::apply
-//! [`TextAt::text_at`]: nvisy_core::extraction::TextAt::text_at
-
-mod params;
-
-use nvisy_core::Error;
-use nvisy_core::entity::Entity;
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::Modality;
-use regex::Regex;
-
-pub use self::params::SuppressionParams;
-use super::layer::{Layer, LayerContext};
-
-const TARGET: &str = "nvisy_toolkit::deduplication::suppress";
-
-/// [`Layer`] that drops entities whose resolved text is on a
-/// caller-supplied allow list.
-///
-/// Construct via [`SuppressionLayer::new`] (empty, fast no-op) or
-/// [`SuppressionLayer::from_params`] (pre-validates the regex
-/// inputs). An empty layer short-circuits in
-/// [`Layer::apply`] without touching the resolver.
-#[derive(Debug, Clone, Default)]
-pub struct SuppressionLayer {
-    /// Pre-lowercased exact-match values.
-    allow_values: Vec<String>,
-    /// Pre-lowercased substring values.
-    allow_values_substring: Vec<String>,
-    /// Pre-compiled regex patterns.
-    allow_values_regex: Vec<Regex>,
-}
-
-impl SuppressionLayer {
-    /// Empty layer: passes every entity through unchanged.
-    #[must_use]
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Construct from a [`SuppressionParams`]. Each regex source
-    /// is compiled once here.
-    ///
-    /// Empty strings are silently dropped at construction from
-    /// all three lists: each would match every entity (or every
-    /// position) and is virtually never what the author meant.
-    /// Treating them as configuration mistakes and ignoring them
-    /// is safer than wiping every result.
-    ///
-    /// # Errors
-    ///
-    /// Returns a validation error when any non-empty entry in
-    /// [`SuppressionParams::allow_values_regex`] is not a valid
-    /// regular expression.
-    pub fn from_params(params: &SuppressionParams) -> Result<Self, Error> {
-        let allow_values = params
-            .allow_values
-            .iter()
-            .filter(|v| !v.is_empty())
-            .map(|v| v.to_ascii_lowercase())
-            .collect();
-        let allow_values_substring = params
-            .allow_values_substring
-            .iter()
-            .filter(|v| !v.is_empty())
-            .map(|v| v.to_ascii_lowercase())
-            .collect();
-        let allow_values_regex = params
-            .allow_values_regex
-            .iter()
-            .filter(|src| !src.is_empty())
-            .map(|src| {
-                Regex::new(src).map_err(|e| {
-                    Error::validation(
-                        format!("invalid allow_values_regex `{src}`: {e}"),
-                        "nvisy-toolkit",
-                    )
-                })
-            })
-            .collect::<Result<Vec<_>, _>>()?;
-        Ok(Self {
-            allow_values,
-            allow_values_substring,
-            allow_values_regex,
-        })
-    }
-
-    /// Return `true` when no allow-list values are configured.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.allow_values.is_empty()
-            && self.allow_values_substring.is_empty()
-            && self.allow_values_regex.is_empty()
-    }
-
-    /// Return `true` when `text` matches any configured allow-list
-    /// entry under exact / substring / regex semantics.
-    #[must_use]
-    pub fn suppresses(&self, text: &str) -> bool {
-        let lowered = text.to_ascii_lowercase();
-        if self.allow_values.iter().any(|v| v == &lowered) {
-            return true;
-        }
-        if self
-            .allow_values_substring
-            .iter()
-            .any(|v| lowered.contains(v.as_str()))
-        {
-            return true;
-        }
-        if self.allow_values_regex.iter().any(|r| r.is_match(text)) {
-            return true;
-        }
-        false
-    }
-}
-
-#[async_trait::async_trait]
-impl<M, R> Layer<M, R> for SuppressionLayer
-where
-    M: Modality,
-    R: TextAt<M> + ?Sized,
-{
-    async fn apply(
-        &self,
-        entities: &mut Vec<Entity<M>>,
-        ctx: &LayerContext<'_, M, R>,
-    ) -> Vec<Entity<M>> {
-        if self.is_empty() || entities.is_empty() {
-            return Vec::new();
-        }
-
-        let mut suppressed_flags = Vec::with_capacity(entities.len());
-        for entity in entities.iter() {
-            let suppress = match ctx.resolver.text_at(&entity.location).await {
-                Some(text) => self.suppresses(&text),
-                None => false,
-            };
-            suppressed_flags.push(suppress);
-        }
-
-        let mut suppressed_count = 0usize;
-        let mut dropped = Vec::new();
-        let mut idx = 0usize;
-        entities.retain(|entity| {
-            let drop = suppressed_flags[idx];
-            idx += 1;
-            if drop {
-                suppressed_count += 1;
-                dropped.push(entity.clone());
-            }
-            !drop
-        });
-
-        if suppressed_count > 0 {
-            tracing::debug!(
-                target: TARGET,
-                suppressed = suppressed_count,
-                "entities suppressed by allow list",
-            );
-        }
-        dropped
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use std::sync::Arc;
-
-    use nvisy_core::entity::{Entity, builtins};
-    use nvisy_core::modality::{Text, TextLocation};
-
-    use super::*;
-
-    /// Test resolver that resolves locations to a slice of a known
-    /// string. The Noop test_resolver in the parent module returns
-    /// `None`, which is fine for layers that don't touch text but
-    /// useless here.
-    struct TextSliceResolver {
-        text: Arc<String>,
-    }
-
-    #[async_trait::async_trait]
-    impl TextAt<Text> for TextSliceResolver {
-        async fn text_at(&self, location: &TextLocation) -> Option<String> {
-            self.text
-                .get(location.start..location.end)
-                .map(String::from)
-        }
-    }
-
-    fn entity(start: usize, end: usize) -> Entity<Text> {
-        Entity::test_builder(start, end)
-            .with_label(builtins::EMAIL_ADDRESS.label_ref())
-            .test_build()
-    }
-
-    fn params(values: &[&str], substrings: &[&str], regexes: &[&str]) -> SuppressionParams {
-        SuppressionParams {
-            allow_values: values.iter().map(|s| (*s).to_owned()).collect(),
-            allow_values_substring: substrings.iter().map(|s| (*s).to_owned()).collect(),
-            allow_values_regex: regexes.iter().map(|s| (*s).to_owned()).collect(),
-        }
-    }
-
-    async fn apply_to(
-        layer: &SuppressionLayer,
-        source: &str,
-        mut entities: Vec<Entity<Text>>,
-    ) -> (Vec<Entity<Text>>, Vec<Entity<Text>>) {
-        let resolver = TextSliceResolver {
-            text: Arc::new(source.to_owned()),
-        };
-        let ctx = LayerContext::new(&resolver);
-        let dropped = layer.apply(&mut entities, &ctx).await;
-        (entities, dropped)
-    }
-
-    #[tokio::test]
-    async fn empty_layer_is_noop() {
-        let layer = SuppressionLayer::new();
-        let source = "noreply@foo.com matters";
-        let (kept, dropped) = apply_to(&layer, source, vec![entity(0, 15)]).await;
-        assert_eq!(kept.len(), 1);
-        assert!(dropped.is_empty());
-    }
-
-    #[tokio::test]
-    async fn exact_match_drops_entity() {
-        let layer = SuppressionLayer::from_params(&params(&["noreply@foo.com"], &[], &[]))
-            .expect("layer builds");
-        let source = "noreply@foo.com matters";
-        let (kept, dropped) = apply_to(&layer, source, vec![entity(0, 15)]).await;
-        assert!(kept.is_empty());
-        assert_eq!(dropped.len(), 1);
-    }
-
-    #[tokio::test]
-    async fn exact_match_is_case_insensitive() {
-        let layer = SuppressionLayer::from_params(&params(&["NoReply@Foo.com"], &[], &[]))
-            .expect("layer builds");
-        let source = "noreply@foo.com matters";
-        let (kept, _) = apply_to(&layer, source, vec![entity(0, 15)]).await;
-        assert!(kept.is_empty(), "case-insensitive allow-list should drop");
-    }
-
-    #[tokio::test]
-    async fn exact_match_does_not_drop_partial_overlap() {
-        // Allow value is a substring of the entity, but not an
-        // exact equal — exact mode keeps it.
-        let layer = SuppressionLayer::from_params(&params(&["noreply@foo.com"], &[], &[]))
-            .expect("layer builds");
-        let source = "noreply@foo.com support team";
-        let (kept, _) = apply_to(&layer, source, vec![entity(0, 28)]).await;
-        assert_eq!(kept.len(), 1, "exact mode must not drop on partial overlap");
-    }
-
-    #[tokio::test]
-    async fn substring_match_drops_partial_overlap() {
-        let layer = SuppressionLayer::from_params(&params(&[], &["noreply@foo.com"], &[]))
-            .expect("layer builds");
-        let source = "noreply@foo.com support team";
-        let (kept, dropped) = apply_to(&layer, source, vec![entity(0, 28)]).await;
-        assert!(kept.is_empty());
-        assert_eq!(dropped.len(), 1);
-    }
-
-    #[tokio::test]
-    async fn regex_match_drops_entity() {
-        let layer = SuppressionLayer::from_params(&params(&[], &[], &[r"^test-.*@foo\.com$"]))
-            .expect("layer builds");
-        let source = "test-1234@foo.com";
-        let (kept, dropped) = apply_to(&layer, source, vec![entity(0, source.len())]).await;
-        assert!(kept.is_empty());
-        assert_eq!(dropped.len(), 1);
-    }
-
-    #[tokio::test]
-    async fn invalid_regex_at_construction_errors() {
-        let result = SuppressionLayer::from_params(&params(&[], &[], &["["]));
-        assert!(result.is_err(), "invalid regex must error at construction");
-    }
-
-    #[tokio::test]
-    async fn unresolved_text_keeps_entity() {
-        // Pass an entity with a location outside the source text.
-        // text_at returns None, the layer falls open and keeps the
-        // entity rather than silently dropping it.
-        let layer = SuppressionLayer::from_params(&params(&["noreply@foo.com"], &[], &[]))
-            .expect("layer builds");
-        let source = "short";
-        let (kept, dropped) = apply_to(&layer, source, vec![entity(100, 200)]).await;
-        assert_eq!(kept.len(), 1);
-        assert!(dropped.is_empty());
-    }
-
-    #[tokio::test]
-    async fn empty_substring_entry_does_not_suppress_everything() {
-        // `str::contains("")` is always true; without the
-        // construction-time filter, an empty entry would wipe
-        // every match. Confirm the filter holds.
-        let layer = SuppressionLayer::from_params(&params(&[], &[""], &[])).expect("layer builds");
-        let source = "noreply@foo.com matters";
-        let (kept, dropped) = apply_to(&layer, source, vec![entity(0, 15)]).await;
-        assert_eq!(kept.len(), 1, "empty substring must not drop");
-        assert!(dropped.is_empty());
-    }
-
-    #[tokio::test]
-    async fn empty_exact_entry_is_ignored() {
-        // An empty exact entry could only match an empty entity,
-        // which recognizers don't emit. Filtering it costs
-        // nothing and keeps the lookup short.
-        let layer = SuppressionLayer::from_params(&params(&[""], &[], &[])).expect("layer builds");
-        let source = "noreply@foo.com matters";
-        let (kept, dropped) = apply_to(&layer, source, vec![entity(0, 15)]).await;
-        assert_eq!(kept.len(), 1);
-        assert!(dropped.is_empty());
-    }
-
-    #[tokio::test]
-    async fn empty_regex_entry_is_ignored() {
-        // An empty regex matches at every position. Same
-        // catastrophe as empty substring; filter at construction.
-        let layer = SuppressionLayer::from_params(&params(&[], &[], &[""])).expect("layer builds");
-        let source = "noreply@foo.com matters";
-        let (kept, dropped) = apply_to(&layer, source, vec![entity(0, 15)]).await;
-        assert_eq!(kept.len(), 1, "empty regex must not drop");
-        assert!(dropped.is_empty());
-    }
-
-    #[tokio::test]
-    async fn union_across_modes() {
-        // Three allow-list shapes, three entities. Each entity is
-        // suppressed by exactly one mode; all three drop.
-        let layer = SuppressionLayer::from_params(&params(
-            &["alpha@x.com"],
-            &["bravo"],
-            &[r"^charlie-\d+$"],
-        ))
-        .expect("layer builds");
-        let source = "alpha@x.com bravo-team-12 charlie-99";
-        let entities = vec![
-            entity(0, 11),  // exact
-            entity(12, 25), // substring
-            entity(26, 36), // regex
-        ];
-        let (kept, dropped) = apply_to(&layer, source, entities).await;
-        assert!(kept.is_empty(), "all three should be suppressed");
-        assert_eq!(dropped.len(), 3);
-    }
-}
diff --git a/crates/nvisy-toolkit/src/deduplication/suppress/params.rs b/crates/nvisy-toolkit/src/deduplication/suppress/params.rs
deleted file mode 100644
index d229572db..000000000
--- a/crates/nvisy-toolkit/src/deduplication/suppress/params.rs
+++ /dev/null
@@ -1,82 +0,0 @@
-//! [`SuppressionParams`]: caller-supplied allow lists consumed by
-//! [`SuppressionLayer`].
-//!
-//! Three independent allow-list shapes apply by union — an entity
-//! is dropped when **any** of them fires:
-//!
-//! - exact ASCII case-insensitive equality
-//! - substring containment
-//! - regex match
-//!
-//! All three operate on the entity's resolved text (sliced from
-//! the source via [`TextAt::text_at`]), not the surrounding
-//! document.
-//!
-//! [`SuppressionLayer`]: super::SuppressionLayer
-//! [`TextAt::text_at`]: nvisy_core::extraction::TextAt::text_at
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-/// Caller-supplied allow lists consumed by [`SuppressionLayer`].
-///
-/// All three lists default to empty; the layer short-circuits as
-/// a fast no-op when every list is empty.
-///
-/// [`SuppressionLayer`]: super::SuppressionLayer
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
-#[derive(Serialize, Deserialize, JsonSchema)]
-pub struct SuppressionParams {
-    /// Drop entities whose matched text equals one of these values
-    /// (ASCII case-insensitive). Use for known false-positive
-    /// values like `noreply@yourcompany.com`.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub allow_values: Vec<String>,
-    /// Drop entities whose matched text contains one of these
-    /// values as a substring (ASCII case-insensitive). Use when an
-    /// over-matching recognizer surrounds a known false-positive
-    /// value with extra text.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub allow_values_substring: Vec<String>,
-    /// Drop entities whose matched text matches one of these
-    /// regular expressions. Compiled once at layer construction.
-    #[serde(default, skip_serializing_if = "Vec::is_empty")]
-    pub allow_values_regex: Vec<String>,
-}
-
-impl SuppressionParams {
-    /// Empty params: every allow list defaults to empty.
-    #[must_use]
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Set the exact-match list.
-    #[must_use]
-    pub fn with_allow_values(mut self, values: Vec<String>) -> Self {
-        self.allow_values = values;
-        self
-    }
-
-    /// Set the substring-match list.
-    #[must_use]
-    pub fn with_allow_values_substring(mut self, values: Vec<String>) -> Self {
-        self.allow_values_substring = values;
-        self
-    }
-
-    /// Set the regex-match list.
-    #[must_use]
-    pub fn with_allow_values_regex(mut self, values: Vec<String>) -> Self {
-        self.allow_values_regex = values;
-        self
-    }
-
-    /// Return `true` when no allow-list values are configured.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.allow_values.is_empty()
-            && self.allow_values_substring.is_empty()
-            && self.allow_values_regex.is_empty()
-    }
-}
diff --git a/crates/nvisy-toolkit/src/detection/chunks.rs b/crates/nvisy-toolkit/src/detection/chunks.rs
deleted file mode 100644
index e16f71b82..000000000
--- a/crates/nvisy-toolkit/src/detection/chunks.rs
+++ /dev/null
@@ -1,137 +0,0 @@
-//! Chunk-driven detection: walk an [`Handler<M>`]'s chunks,
-//! run the [`RecognizerRegistry`] against each, and lift every
-//! entity's offset back to source-coordinate `M::Location` via
-//! [`Handler::lift_chunk`] before returning the merged list.
-//!
-//! Exposed as an extension trait
-//! ([`RecognizerRegistryExt::detect`]) on [`RecognizerRegistry`]:
-//! callers `use nvisy_toolkit::detection::RecognizerRegistryExt;`
-//! and write `registry.detect(&mut handler).await?`. The handler's
-//! modality determines the output entity type.
-//!
-//! Cell-shaped tabular values, JSON-escaped strings, and other
-//! formats whose chunk payload doesn't match raw source bytes work
-//! out of the box because the codec's `lift_chunk` knows how to
-//! translate each one. The modality-to-modality reshape (turning
-//! the recognizer's `Entity<Text>` into the handler's `Entity<M>`)
-//! is handled by [`LiftedFromText`].
-//!
-//! Today recognizers are all `Text`-only — they scan a `&str` and
-//! emit byte offsets — so detection always runs as `Text` against
-//! the chunk payload and then reshapes upward. If/when a real
-//! per-modality recognizer ships (e.g. image-only OCR-driven
-//! recognition), it would skip this helper entirely and call
-//! [`RecognizerRegistry::run`] for its own modality directly.
-//!
-//! [`Handler<M>`]: nvisy_codec::Handler
-//! [`Handler::lift_chunk`]: nvisy_codec::Handler::lift_chunk
-
-use nvisy_codec::Handler;
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Modality, Tabular, TabularLocation, Text, TextData, TextLocation};
-use nvisy_core::recognition::RecognizerInput;
-
-use super::RecognizerRegistry;
-
-/// Extension trait adding chunk-driven detection to
-/// [`RecognizerRegistry`]. Each registered `Text` recognizer runs
-/// against every chunk yielded by `handler`; matches are lifted to
-/// source-coordinate `M::Location` via
-/// [`Handler::lift_chunk`] and reshaped into `Entity<M>` via
-/// [`LiftedFromText`].
-///
-/// Callers `use nvisy_toolkit::detection::RecognizerRegistryExt;`
-/// to bring the method into scope.
-///
-/// [`Handler::lift_chunk`]: nvisy_codec::Handler::lift_chunk
-#[async_trait::async_trait]
-pub trait RecognizerRegistryExt {
-    /// Walk `handler`'s chunks, run every registered `Text`
-    /// recognizer against each chunk's data, lift each match to
-    /// the handler's source-coordinate `M::Location`, and reshape
-    /// into `Entity<M>`. Entities whose offsets fall outside the
-    /// chunk's value (e.g. a match landing inside a JSON `\"`
-    /// escape pair) are dropped: the lift has no source pre-image
-    /// for them.
-    async fn detect<M, H>(&self, handler: &mut H) -> Result<Vec<Entity<M>>>
-    where
-        M: Modality + LiftedFromText,
-        M::Data: Into<TextData>,
-        H: Handler<M> + ?Sized;
-}
-
-#[async_trait::async_trait]
-impl RecognizerRegistryExt for RecognizerRegistry {
-    async fn detect<M, H>(&self, handler: &mut H) -> Result<Vec<Entity<M>>>
-    where
-        M: Modality + LiftedFromText,
-        M::Data: Into<TextData>,
-        H: Handler<M> + ?Sized,
-    {
-        let mut out = Vec::new();
-        while let Some(chunk) = handler.next_chunk().await? {
-            let input = RecognizerInput::new(chunk.data.clone().into())
-                .with_context_hints(chunk.hints.clone());
-            let text_entities = self.run::<Text>(input).await?;
-            for text_entity in text_entities {
-                let Some(loc) = handler
-                    .lift_chunk(&chunk, text_entity.location.start..text_entity.location.end)
-                else {
-                    continue;
-                };
-                out.push(M::from_text(text_entity, loc));
-            }
-        }
-        Ok(out)
-    }
-}
-
-/// Modality marker that knows how to absorb a `Text`-modality
-/// finding into its own coordinate system.
-///
-/// Implementations describe the entity-level reshape only:
-/// taking an `Entity<Text>` produced by a recognizer (with
-/// chunk-local offsets) and a pre-lifted `Self::Location`,
-/// produce an `Entity<Self>` with identity-relevant fields
-/// preserved.
-///
-/// Today's impls:
-///
-/// - [`Text`] — identity: the recognizer's entity already carries
-///   `Entity<Text>`; just install the lifted location.
-/// - [`Tabular`] — rebuild with the lifted [`TabularLocation`];
-///   preserve id, entity_id, kind, confidence, trail, language.
-///
-/// Future modalities (Audio, Image) would add impls here.
-pub trait LiftedFromText: Modality + Sized {
-    /// Take a [`Text`]-modality entity emitted by a recognizer
-    /// against the source bytes of a chunk, plus the pre-lifted
-    /// location, and produce a `Self`-modality entity.
-    fn from_text(text_entity: Entity<Text>, location: Self::Location) -> Entity<Self>;
-}
-
-impl LiftedFromText for Text {
-    fn from_text(mut entity: Entity<Text>, location: TextLocation) -> Entity<Text> {
-        entity.location = location;
-        entity
-    }
-}
-
-impl LiftedFromText for Tabular {
-    fn from_text(text_entity: Entity<Text>, location: TabularLocation) -> Entity<Tabular> {
-        let mut builder = Entity::<Tabular>::builder()
-            .with_id(text_entity.id)
-            .with_label(text_entity.label)
-            .with_location(location)
-            .with_confidence(text_entity.confidence)
-            .with_trail(text_entity.trail);
-        if let Some(entity_id) = text_entity.entity_id {
-            builder = builder.with_entity_id(entity_id);
-        }
-        if let Some(language) = text_entity.language {
-            builder = builder.with_language(language);
-        }
-        builder.build().expect("entity reshape")
-    }
-}
diff --git a/crates/nvisy-toolkit/src/detection/llm.rs b/crates/nvisy-toolkit/src/detection/llm.rs
deleted file mode 100644
index 4a390d011..000000000
--- a/crates/nvisy-toolkit/src/detection/llm.rs
+++ /dev/null
@@ -1,13 +0,0 @@
-//! Re-export the [`nvisy_llm`] surface as
-//! `nvisy_toolkit::detection::llm`.
-//!
-//! A consumer that wants the shipped LLM-driven recognizer
-//! ([`LlmRecognizer<M>`], implementing [`EntityRecognizer<M>`]) only
-//! needs the `nvisy-toolkit` dep. The recognizer registers straight
-//! into the shared [`RecognizerRegistry`] like any other recognizer.
-//!
-//! [`LlmRecognizer<M>`]: nvisy_llm::LlmRecognizer
-//! [`EntityRecognizer<M>`]: nvisy_core::recognition::EntityRecognizer
-//! [`RecognizerRegistry`]: super::RecognizerRegistry
-
-pub use nvisy_llm::*;
diff --git a/crates/nvisy-toolkit/src/detection/mod.rs b/crates/nvisy-toolkit/src/detection/mod.rs
deleted file mode 100644
index 77aefafe0..000000000
--- a/crates/nvisy-toolkit/src/detection/mod.rs
+++ /dev/null
@@ -1,33 +0,0 @@
-//! Detection: the [`RecognizerRegistry`] plus per-backend re-exports.
-//!
-//! Toolkit owns the registry side of detection — typed lists of
-//! `Arc<dyn EntityRecognizer<M>>` per modality. The trait the registry
-//! holds ([`EntityRecognizer`]) lives in `nvisy-core` so backend crates
-//! can implement it without depending on toolkit. Each backend crate
-//! is re-exported here under its own submodule so a consumer that
-//! wants the shipped recognizers only needs to depend on
-//! `nvisy-toolkit`.
-//!
-//! Three backend submodules ship today:
-//!
-//! - [`pattern`] — `nvisy_pattern`: regex + dictionary rules.
-//! - [`ner`] — `nvisy_ner`: zero-shot and adapter NER recognizers.
-//! - [`llm`] — `nvisy_llm`: LLM-driven recognizers.
-//!
-//! TOML-deserialisable backend selectors and config bundles live in
-//! a higher pipeline layer that turns config into concrete
-//! `Arc<dyn EntityRecognizer<M>>` instances inserted into
-//! [`RecognizerRegistry`].
-//!
-//! [`EntityRecognizer`]: nvisy_core::recognition::EntityRecognizer
-
-mod chunks;
-pub mod llm;
-pub mod ner;
-pub mod pattern;
-mod registry;
-
-pub use nvisy_core::recognition::EntityRecognizer;
-
-pub use self::chunks::{LiftedFromText, RecognizerRegistryExt};
-pub use self::registry::RecognizerRegistry;
diff --git a/crates/nvisy-toolkit/src/detection/ner.rs b/crates/nvisy-toolkit/src/detection/ner.rs
deleted file mode 100644
index 41022861b..000000000
--- a/crates/nvisy-toolkit/src/detection/ner.rs
+++ /dev/null
@@ -1,9 +0,0 @@
-//! Re-export the [`nvisy_ner`] backend surface as
-//! `nvisy_toolkit::detection::ner`.
-//!
-//! A consumer that wants the shipped NER backends only needs the
-//! `nvisy-toolkit` dep — `nvisy_toolkit::detection::ner::backend::Backend`,
-//! `nvisy_toolkit::detection::ner::backend::NoopBackend`, etc. are
-//! all reachable here.
-
-pub use nvisy_ner::*;
diff --git a/crates/nvisy-toolkit/src/detection/pattern.rs b/crates/nvisy-toolkit/src/detection/pattern.rs
deleted file mode 100644
index e9f249444..000000000
--- a/crates/nvisy-toolkit/src/detection/pattern.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-//! Re-export the [`nvisy_pattern`] backend surface as
-//! `nvisy_toolkit::detection::pattern`.
-//!
-//! A consumer that wants the shipped pattern recognizers only needs
-//! the `nvisy-toolkit` dep —
-//! `nvisy_toolkit::detection::pattern::PatternRecognizer`,
-//! `nvisy_toolkit::detection::pattern::patterns::all()`, etc. are all
-//! reachable here.
-
-pub use nvisy_pattern::*;
diff --git a/crates/nvisy-toolkit/src/detection/registry.rs b/crates/nvisy-toolkit/src/detection/registry.rs
deleted file mode 100644
index 17cefa8dd..000000000
--- a/crates/nvisy-toolkit/src/detection/registry.rs
+++ /dev/null
@@ -1,170 +0,0 @@
-//! [`RecognizerRegistry`]: open per-modality recognizer container.
-//!
-//! One [`TypeMap`] slot per [`Modality`], keyed by the marker
-//! type. Each slot is `Vec<Arc<dyn EntityRecognizer<M>>>`; iteration
-//! order at dispatch matches registration order. Populate via
-//! [`with_recognizer`] after constructing each backend.
-//!
-//! Scope: this type knows nothing about [`Document`]. It only owns
-//! recognizers and runs them against a [`RecognizerInput`]. Walking a
-//! document, lifting block-local spans to modality coordinates, and
-//! per-modality node dispatch live in the engine's detection phase.
-//!
-//! Failure is fail-fast within a modality: on the first task error
-//! every other in-flight task in that modality is aborted and the
-//! error is returned.
-//!
-//! [`with_recognizer`]: RecognizerRegistry::with_recognizer
-//! [`Modality`]: nvisy_core::modality::Modality
-//! [`RecognizerInput`]: nvisy_core::recognition::RecognizerInput
-//! [`TypeMap`]: type_map::concurrent::TypeMap
-
-use std::fmt;
-use std::sync::Arc;
-
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::Modality;
-use nvisy_core::recognition::{EntityRecognizer, RecognizerInput, RecognizerOutput};
-use nvisy_core::{Error, Result};
-use tokio::task::JoinSet;
-use tracing::Instrument;
-use type_map::concurrent::TypeMap;
-
-const TARGET: &str = "nvisy_toolkit::detection";
-
-/// Open per-modality recognizer container.
-///
-/// Holds a [`TypeMap`] keyed by the modality marker type `M`.
-/// Each slot is a `Vec<Arc<dyn EntityRecognizer<M>>>`. Adding a
-/// new modality requires zero edits to the registry — the modality
-/// crate's `impl Modality` is all that's needed.
-///
-/// [`TypeMap`]: type_map::concurrent::TypeMap
-#[derive(Default)]
-pub struct RecognizerRegistry {
-    slots: TypeMap,
-}
-
-impl RecognizerRegistry {
-    /// Empty registry. Populate with [`with_recognizer`].
-    ///
-    /// [`with_recognizer`]: Self::with_recognizer
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Register a recognizer for modality `M`. Appended to that
-    /// modality's slot; iteration order at dispatch matches
-    /// registration order. Chainable.
-    #[must_use]
-    pub fn with_recognizer<M: Modality>(
-        mut self,
-        recognizer: impl EntityRecognizer<M> + 'static,
-    ) -> Self {
-        self.slot_mut::<M>().push(Arc::new(recognizer));
-        self
-    }
-
-    /// Number of recognizers registered for modality `M`.
-    #[must_use]
-    pub fn count<M: Modality>(&self) -> usize {
-        self.slots
-            .get::<Slot<M>>()
-            .map_or(0, |slot| slot.recognizers.len())
-    }
-
-    /// Run every registered recognizer for modality `M` against
-    /// `input` in parallel and return the combined entity set.
-    ///
-    /// Returns `Ok(Vec::new())` when no recognizers are registered
-    /// for `M`.
-    pub async fn run<M>(&self, input: RecognizerInput<M>) -> Result<Vec<Entity<M>>>
-    where
-        M: Modality,
-        M::Data: fmt::Debug,
-    {
-        let Some(slot) = self.slots.get::<Slot<M>>() else {
-            return Ok(Vec::new());
-        };
-        if slot.recognizers.is_empty() {
-            return Ok(Vec::new());
-        }
-
-        let span = tracing::debug_span!(
-            target: TARGET,
-            "detect",
-            modality = M::NAME,
-            input = ?input.data,
-            correlation_id = input.correlation_id.as_ref().map(|id| id.to_string()),
-        );
-
-        let input = Arc::new(input);
-        let mut set: JoinSet<Result<RecognizerOutput<M>>> = JoinSet::new();
-        for recognizer in &slot.recognizers {
-            let recognizer = Arc::clone(recognizer);
-            let input = Arc::clone(&input);
-            set.spawn(async move { recognizer.recognize(&input).await });
-        }
-
-        async move { collect_join_set::<M>(set).await }
-            .instrument(span)
-            .await
-    }
-
-    fn slot_mut<M: Modality>(&mut self) -> &mut Vec<Arc<dyn EntityRecognizer<M>>> {
-        let slot = self.slots.entry::<Slot<M>>();
-        &mut slot.or_insert_with(Slot::default).recognizers
-    }
-}
-
-/// `TypeMap` entry for one modality. Wrapping `Vec` in a named
-/// struct keeps the entry shape `Default + Send + Sync` without
-/// needing those bounds in the `EntityRecognizer<M>` super-traits.
-struct Slot<M: Modality> {
-    recognizers: Vec<Arc<dyn EntityRecognizer<M>>>,
-}
-
-impl<M: Modality> Default for Slot<M> {
-    fn default() -> Self {
-        Self {
-            recognizers: Vec::new(),
-        }
-    }
-}
-
-async fn collect_join_set<M: Modality>(
-    mut set: JoinSet<Result<RecognizerOutput<M>>>,
-) -> Result<Vec<Entity<M>>> {
-    let mut all = Vec::new();
-    while let Some(joined) = set.join_next().await {
-        match joined {
-            Ok(Ok(output)) => {
-                tracing::debug!(
-                    target: TARGET,
-                    detected = output.entities.len(),
-                    "recognizer produced entities",
-                );
-                all.extend(output.entities);
-            }
-            Ok(Err(e)) => {
-                set.abort_all();
-                return Err(e);
-            }
-            Err(join_err) => {
-                set.abort_all();
-                return Err(Error::runtime(
-                    format!("recognizer task panicked or was cancelled: {join_err}"),
-                    "recognizer-registry",
-                    false,
-                ));
-            }
-        }
-    }
-    Ok(all)
-}
-
-impl fmt::Debug for RecognizerRegistry {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("RecognizerRegistry").finish_non_exhaustive()
-    }
-}
diff --git a/crates/nvisy-toolkit/src/extraction/mod.rs b/crates/nvisy-toolkit/src/extraction/mod.rs
deleted file mode 100644
index 87c3d57f3..000000000
--- a/crates/nvisy-toolkit/src/extraction/mod.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! Extraction: the [`ExtractorRegistry`] plus per-backend re-exports.
-//!
-//! Toolkit owns the registry side of extraction — typed slots for one
-//! per-modality extractor each. The trait the registry holds
-//! ([`Extractor`]) lives in `nvisy-core` so backend crates can
-//! implement it without depending on toolkit. Each backend crate is
-//! re-exported here under its own submodule ([`ocr`] and [`stt`]) so a
-//! consumer that wants the shipped backends only needs to depend on
-//! `nvisy-toolkit`.
-//!
-//! TOML-deserialisable backend selectors and config bundles live in
-//! a higher pipeline layer that turns config into concrete
-//! `Arc<dyn Extractor<M>>` instances inserted into
-//! [`ExtractorRegistry`].
-//!
-//! [`Extractor`]: nvisy_core::extraction::Extractor
-
-#[cfg(feature = "image")]
-pub mod ocr;
-#[cfg(feature = "audio")]
-pub mod stt;
-
-mod registry;
-
-pub use nvisy_core::extraction::Extractor;
-
-#[cfg(feature = "image")]
-pub use self::ocr::ImageExtractorOutput;
-pub use self::registry::ExtractorRegistry;
-#[cfg(feature = "audio")]
-pub use self::stt::AudioExtractorOutput;
diff --git a/crates/nvisy-toolkit/src/extraction/ocr.rs b/crates/nvisy-toolkit/src/extraction/ocr.rs
deleted file mode 100644
index 57c125af0..000000000
--- a/crates/nvisy-toolkit/src/extraction/ocr.rs
+++ /dev/null
@@ -1,21 +0,0 @@
-//! Re-export the [`nvisy_ocr`] backend surface as
-//! `nvisy_toolkit::extraction::ocr`.
-//!
-//! A consumer that wants the shipped OCR backends only needs the
-//! `nvisy-toolkit` dep — `nvisy_toolkit::extraction::ocr::OcrExtractor`,
-//! `nvisy_toolkit::extraction::ocr::backend::OcrBackend`,
-//! `nvisy_toolkit::extraction::ocr::backend::NoopBackend`, etc. are
-//! all reachable here.
-
-use nvisy_ocr::backend::OcrResponse;
-pub use nvisy_ocr::*;
-
-/// Output shape produced by every image-modality extractor.
-///
-/// Pinned to [`OcrResponse`] today. Widening the type later (to
-/// accommodate scene-text / layout-detection backends with
-/// different shapes) is a breaking change worth taking deliberately
-/// rather than hiding behind generics.
-///
-/// [`OcrResponse`]: nvisy_ocr::backend::OcrResponse
-pub type ImageExtractorOutput = OcrResponse;
diff --git a/crates/nvisy-toolkit/src/extraction/registry.rs b/crates/nvisy-toolkit/src/extraction/registry.rs
deleted file mode 100644
index 69fa983e2..000000000
--- a/crates/nvisy-toolkit/src/extraction/registry.rs
+++ /dev/null
@@ -1,93 +0,0 @@
-//! [`ExtractorRegistry`]: per-modality container of pre-built
-//! extractors keyed by modality.
-//!
-//! At most one extractor per modality today (OCR for [`Image`], STT
-//! for [`Audio`]). The output shape is pinned per modality so the
-//! document-side glue knows how to map the extractor's `Output` into
-//! `Block<M>` values — OCR backends produce `Vec<OcrOutput>`, STT
-//! backends produce `Transcription`. Custom backends slot in by
-//! implementing the matching [`Extractor`] trait shape.
-//!
-//! Toolkit holds no `from_config` builder — config-driven
-//! construction is the consumer's concern. A higher pipeline layer
-//! deserialises the `[extractor.*]` sections and inserts the chosen
-//! [`Arc<dyn Extractor<M>>`] into the registry.
-//!
-//! [`Extractor`]: nvisy_core::extraction::Extractor
-
-use std::sync::Arc;
-
-use nvisy_core::extraction::Extractor;
-use nvisy_core::modality::{Audio, Image};
-
-#[cfg(feature = "image")]
-use super::ocr::ImageExtractorOutput;
-#[cfg(feature = "audio")]
-use super::stt::AudioExtractorOutput;
-
-/// Per-modality container of pre-built extractors.
-///
-/// Built once at engine startup, cloned per run (the inner `Arc`s
-/// keep the actual extractors shared without copying).
-#[derive(Default, Clone)]
-pub struct ExtractorRegistry {
-    /// Image-modality extractor slot. `None` opts the technique out
-    /// entirely.
-    #[cfg(feature = "image")]
-    pub image: Option<Arc<dyn Extractor<Image, Output = ImageExtractorOutput>>>,
-    /// Audio-modality extractor slot. `None` opts the technique out
-    /// entirely.
-    #[cfg(feature = "audio")]
-    pub audio: Option<Arc<dyn Extractor<Audio, Output = AudioExtractorOutput>>>,
-}
-
-impl ExtractorRegistry {
-    /// Build an empty registry. Useful for tests; production callers
-    /// populate the slots with [`with_image_extractor`] /
-    /// [`with_audio_extractor`] after constructing the chosen
-    /// backend.
-    ///
-    /// [`with_image_extractor`]: Self::with_image_extractor
-    /// [`with_audio_extractor`]: Self::with_audio_extractor
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Install the image-modality extractor. Takes ownership of the
-    /// extractor and wraps it in `Arc` internally.
-    #[cfg(feature = "image")]
-    #[must_use]
-    pub fn with_image_extractor<E>(mut self, extractor: E) -> Self
-    where
-        E: Extractor<Image, Output = ImageExtractorOutput> + 'static,
-    {
-        self.image = Some(Arc::new(extractor));
-        self
-    }
-
-    /// Install the audio-modality extractor. Takes ownership of the
-    /// extractor and wraps it in `Arc` internally.
-    #[cfg(feature = "audio")]
-    #[must_use]
-    pub fn with_audio_extractor<E>(mut self, extractor: E) -> Self
-    where
-        E: Extractor<Audio, Output = AudioExtractorOutput> + 'static,
-    {
-        self.audio = Some(Arc::new(extractor));
-        self
-    }
-
-    /// `true` when no extractor is installed.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        #[cfg(feature = "image")]
-        let image_empty = self.image.is_none();
-        #[cfg(not(feature = "image"))]
-        let image_empty = true;
-        #[cfg(feature = "audio")]
-        let audio_empty = self.audio.is_none();
-        #[cfg(not(feature = "audio"))]
-        let audio_empty = true;
-        image_empty && audio_empty
-    }
-}
diff --git a/crates/nvisy-toolkit/src/extraction/stt.rs b/crates/nvisy-toolkit/src/extraction/stt.rs
deleted file mode 100644
index 2b48da7ce..000000000
--- a/crates/nvisy-toolkit/src/extraction/stt.rs
+++ /dev/null
@@ -1,12 +0,0 @@
-//! Re-export the [`nvisy_stt`] surface as
-//! `nvisy_toolkit::extraction::stt`.
-//!
-//! A consumer that wants the shipped STT backends only needs the
-//! `nvisy-toolkit` dep — `nvisy_toolkit::extraction::stt::SttExtractor`,
-//! `nvisy_toolkit::extraction::stt::backend::NoopBackend`, etc. are
-//! all reachable here.
-
-pub use nvisy_stt::{backend, *};
-
-/// Output shape produced by every audio-modality extractor.
-pub type AudioExtractorOutput = nvisy_stt::Transcription;
diff --git a/crates/nvisy-toolkit/src/ingestion/mod.rs b/crates/nvisy-toolkit/src/ingestion/mod.rs
deleted file mode 100644
index 235bee10f..000000000
--- a/crates/nvisy-toolkit/src/ingestion/mod.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! Ingestion: the read-side edge of a toolkit pipeline.
-//!
-//! Re-exports the codec front door so a toolkit-only consumer reaches
-//! the canonical "load bytes → typed handle" pipeline without
-//! pulling `nvisy-codec` into its own dep list:
-//!
-//! - [`CodecRegistry`] is the entry point — call
-//!   [`CodecRegistry::with_builtin`] to get every shipped format the
-//!   active features compile in, then [`decode`] to
-//!   commit bytes to a typed handle.
-//! - [`DocumentHandle<M>`] is the typed handle. It implements the
-//!   resolver traits ([`TextAt`], [`DataAt`]) plus [`RedactAt`] that
-//!   detection / deduplication / redaction phases bound on, so the
-//!   handle itself is the buffer the pipeline operates on.
-//!
-//! The per-modality `*At` impls on `DocumentHandle<M>` in
-//! `nvisy-codec` are cfg-gated on `internal_text` / `internal_tabular`
-//! / `internal_image` / `internal_audio`. Each toolkit feature
-//! (`text` / `tabular` / `image` / `audio` / `rich`) forwards to the
-//! matching codec umbrella, which in turn activates the right
-//! `internal_*` flags.
-//!
-//! [`CodecRegistry`]: nvisy_codec::CodecRegistry
-//! [`CodecRegistry::with_builtin`]: nvisy_codec::CodecRegistry::with_builtin
-//! [`decode`]: nvisy_codec::CodecRegistry::decode
-//! [`DocumentHandle<M>`]: nvisy_codec::DocumentHandle
-//! [`TextAt`]: nvisy_core::extraction::TextAt
-//! [`DataAt`]: nvisy_core::extraction::DataAt
-//! [`RedactAt`]: nvisy_core::redaction::RedactAt
-
-pub use nvisy_codec::{CodecRegistry, DocumentHandle};
diff --git a/crates/nvisy-toolkit/src/lib.rs b/crates/nvisy-toolkit/src/lib.rs
deleted file mode 100644
index 0d4a7eee9..000000000
--- a/crates/nvisy-toolkit/src/lib.rs
+++ /dev/null
@@ -1,10 +0,0 @@
-#![forbid(unsafe_code)]
-#![cfg_attr(docsrs, feature(doc_cfg))]
-#![doc = include_str!("../README.md")]
-
-pub mod deduplication;
-pub mod detection;
-pub mod extraction;
-pub mod ingestion;
-pub mod redaction;
-pub mod validation;
diff --git a/crates/nvisy-toolkit/src/redaction/anonymizer/encrypt.rs b/crates/nvisy-toolkit/src/redaction/anonymizer/encrypt.rs
deleted file mode 100644
index 5303ee7b0..000000000
--- a/crates/nvisy-toolkit/src/redaction/anonymizer/encrypt.rs
+++ /dev/null
@@ -1,134 +0,0 @@
-//! [`Encrypt`]: reversible AES-256-GCM redaction.
-//!
-//! Output wire shape: `base64(nonce || ciphertext || auth_tag)`,
-//! where `nonce` is a fresh 12-byte random value per call.
-//!
-//! The nonce is prepended so [`Decrypt`] can parse it without
-//! out-of-band metadata; the auth tag is appended by `aes-gcm` and
-//! verified on decryption. The whole blob is base64-encoded to keep
-//! the substituted payload printable.
-//!
-//! Inverse lives in [`Decrypt`] (sibling module).
-//!
-//! [`Decrypt`]: crate::redaction::deanonymizer::Decrypt
-
-use aes_gcm::aead::rand_core::RngCore;
-use aes_gcm::aead::{Aead, KeyInit, OsRng};
-use aes_gcm::{Aes256Gcm, Key, Nonce};
-use base64::Engine;
-use base64::engine::general_purpose::STANDARD;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Tabular, Text, TextData};
-use nvisy_core::{Error, Result};
-
-use crate::redaction::{Anonymizer, LeakProfile, TabularReplacement, TextReplacement};
-
-/// Nonce length in bytes (AES-GCM standard). Shared with
-/// [`crate::redaction::deanonymizer::Decrypt`] so both sides of
-/// the AES-256-GCM envelope agree on the nonce prefix.
-pub(crate) const NONCE_LEN: usize = 12;
-
-const TARGET: &str = "nvisy_toolkit::redaction::encrypt";
-
-/// AES-256-GCM encrypt operator.
-///
-/// Holds a 32-byte key. Use [`Decrypt`] with the same key to recover
-/// the originals from an audit trail.
-///
-/// [`Decrypt`]: crate::redaction::deanonymizer::Decrypt
-#[derive(Clone)]
-pub struct Encrypt {
-    key: Key<Aes256Gcm>,
-}
-
-impl Encrypt {
-    /// Construct from a 32-byte key.
-    pub fn from_key(key: [u8; 32]) -> Self {
-        Self { key: key.into() }
-    }
-
-    /// Encrypt `value` and return the base64-encoded
-    /// `nonce || ciphertext || auth_tag` blob.
-    fn encrypt_value(&self, value: &str) -> Result<String> {
-        let cipher = Aes256Gcm::new(&self.key);
-
-        let mut nonce_bytes = [0u8; NONCE_LEN];
-        OsRng.fill_bytes(&mut nonce_bytes);
-        let nonce = Nonce::from_slice(&nonce_bytes);
-
-        let ciphertext = cipher
-            .encrypt(nonce, value.as_bytes())
-            .map_err(|e| Error::runtime(format!("encryption failed: {e}"), TARGET, false))?;
-
-        let mut blob = Vec::with_capacity(NONCE_LEN + ciphertext.len());
-        blob.extend_from_slice(&nonce_bytes);
-        blob.extend_from_slice(&ciphertext);
-
-        Ok(STANDARD.encode(&blob))
-    }
-}
-
-#[async_trait::async_trait]
-impl Anonymizer<Text> for Encrypt {
-    fn leak_profile(&self) -> LeakProfile {
-        LeakProfile::Recoverable
-    }
-
-    async fn apply(&self, _entity: &Entity<Text>, source: &TextData) -> Result<TextReplacement> {
-        Ok(TextReplacement::substituted(
-            self.encrypt_value(source.text.as_str())?,
-        ))
-    }
-}
-
-#[async_trait::async_trait]
-impl Anonymizer<Tabular> for Encrypt {
-    fn leak_profile(&self) -> LeakProfile {
-        LeakProfile::Recoverable
-    }
-
-    async fn apply(
-        &self,
-        _entity: &Entity<Tabular>,
-        source: &TextData,
-    ) -> Result<TabularReplacement> {
-        Ok(TabularReplacement::substituted(
-            self.encrypt_value(source.text.as_str())?,
-        ))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{TrailStep, builtins};
-    use nvisy_core::modality::TextLocation;
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-
-    pub(super) const KEY: [u8; 32] = [
-        0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f,
-        0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e,
-        0x1f, 0x20,
-    ];
-
-    fn entity(start: usize, end: usize) -> Entity<Text> {
-        Entity::builder()
-            .with_label(builtins::EMAIL_ADDRESS.label_ref())
-            .with_location(TextLocation::new(start, end))
-            .with_confidence(Confidence::new(1.0).unwrap())
-            .with_trail(Vec::<TrailStep>::new())
-            .build()
-            .unwrap()
-    }
-
-    #[tokio::test]
-    async fn ciphertext_differs_each_call() {
-        let enc = Encrypt::from_key(KEY);
-        let source = TextData::new("alice@example.test");
-        let entity = entity(0, 18);
-        let a = enc.apply(&entity, &source).await.unwrap();
-        let b = enc.apply(&entity, &source).await.unwrap();
-        assert_ne!(a, b, "fresh nonce must produce different ciphertext");
-    }
-}
diff --git a/crates/nvisy-toolkit/src/redaction/anonymizer/hash.rs b/crates/nvisy-toolkit/src/redaction/anonymizer/hash.rs
deleted file mode 100644
index c4f569257..000000000
--- a/crates/nvisy-toolkit/src/redaction/anonymizer/hash.rs
+++ /dev/null
@@ -1,158 +0,0 @@
-//! [`Hash`]: replace the matched value with a one-way SHA-2 hash.
-//!
-//! Configurable along two axes:
-//!
-//! - `algorithm` — [`HashAlgorithm::Sha256`] (default) or
-//!   [`HashAlgorithm::Sha512`].
-//! - `salt` — optional bytes prepended to the value before hashing.
-//!   Salting blocks pre-computed rainbow attacks and makes equal
-//!   plaintext values hash to different ciphertext across deployments.
-//!
-//! The output is the lowercase hex representation of the digest.
-
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Text, TextData};
-use sha2::{Digest, Sha256, Sha512};
-
-use crate::redaction::{Anonymizer, LeakProfile, TextReplacement};
-
-/// Which SHA-2 variant to use.
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default)]
-#[derive(serde::Serialize, serde::Deserialize, schemars::JsonSchema)]
-#[serde(rename_all = "snake_case")]
-pub enum HashAlgorithm {
-    /// SHA-256 — 32-byte digest, 64-char hex.
-    #[default]
-    Sha256,
-    /// SHA-512 — 64-byte digest, 128-char hex.
-    Sha512,
-}
-
-/// One-way SHA-2 hash operator.
-#[derive(Debug, Clone, Default)]
-pub struct Hash {
-    algorithm: HashAlgorithm,
-    salt: Vec<u8>,
-}
-
-impl Hash {
-    /// Build a hash operator with the given algorithm and no salt.
-    pub fn new(algorithm: HashAlgorithm) -> Self {
-        Self {
-            algorithm,
-            salt: Vec::new(),
-        }
-    }
-
-    /// SHA-256 with no salt — the most common default.
-    pub fn sha256() -> Self {
-        Self::new(HashAlgorithm::Sha256)
-    }
-
-    /// SHA-512 with no salt.
-    pub fn sha512() -> Self {
-        Self::new(HashAlgorithm::Sha512)
-    }
-
-    /// Attach a salt that's prepended to the value before hashing.
-    #[must_use]
-    pub fn with_salt(mut self, salt: impl Into<Vec<u8>>) -> Self {
-        self.salt = salt.into();
-        self
-    }
-}
-
-#[async_trait::async_trait]
-impl Anonymizer<Text> for Hash {
-    fn leak_profile(&self) -> LeakProfile {
-        // A hash is recoverable when the candidate plaintext space
-        // is small enough to brute-force or rainbow-table; salting
-        // raises the bar but doesn't remove recoverability in the
-        // information-theoretic sense.
-        LeakProfile::Recoverable
-    }
-
-    async fn apply(&self, _entity: &Entity<Text>, source: &TextData) -> Result<TextReplacement> {
-        let digest = match self.algorithm {
-            HashAlgorithm::Sha256 => hex(Sha256::new()
-                .chain_update(&self.salt)
-                .chain_update(source.text.as_str())
-                .finalize()
-                .as_slice()),
-            HashAlgorithm::Sha512 => hex(Sha512::new()
-                .chain_update(&self.salt)
-                .chain_update(source.text.as_str())
-                .finalize()
-                .as_slice()),
-        };
-        Ok(TextReplacement::substituted(digest))
-    }
-}
-
-fn hex(bytes: &[u8]) -> String {
-    const TABLE: &[u8; 16] = b"0123456789abcdef";
-    let mut out = String::with_capacity(bytes.len() * 2);
-    for &b in bytes {
-        out.push(TABLE[(b >> 4) as usize] as char);
-        out.push(TABLE[(b & 0xf) as usize] as char);
-    }
-    out
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{TrailStep, builtins};
-    use nvisy_core::modality::TextLocation;
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-
-    fn entity(start: usize, end: usize) -> Entity<Text> {
-        Entity::builder()
-            .with_label(builtins::EMAIL_ADDRESS.label_ref())
-            .with_location(TextLocation::new(start, end))
-            .with_confidence(Confidence::new(1.0).unwrap())
-            .with_trail(Vec::<TrailStep>::new())
-            .build()
-            .unwrap()
-    }
-
-    #[tokio::test]
-    async fn sha256_matches_known_vector() {
-        // SHA-256("alice") = "2bd806c97f0e00af1a1fc3328fa763a9269723c8db8fac4f93af71db186d6e90"
-        let op = Hash::sha256();
-        let source = TextData::new("alice");
-        let entity = entity(0, 5);
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(
-            out,
-            TextReplacement::substituted(
-                "2bd806c97f0e00af1a1fc3328fa763a9269723c8db8fac4f93af71db186d6e90"
-            )
-        );
-    }
-
-    #[tokio::test]
-    async fn salt_changes_digest() {
-        let plain = Hash::sha256();
-        let salted = Hash::sha256().with_salt("pepper");
-        let source = TextData::new("alice");
-        let entity = entity(0, 5);
-        let a = plain.apply(&entity, &source).await.unwrap();
-        let b = salted.apply(&entity, &source).await.unwrap();
-        assert_ne!(a, b, "salt must change the digest");
-    }
-
-    #[tokio::test]
-    async fn sha512_emits_128_hex_chars() {
-        let op = Hash::sha512();
-        let source = TextData::new("alice");
-        let entity = entity(0, 5);
-        let TextReplacement::Substituted { value } = op.apply(&entity, &source).await.unwrap()
-        else {
-            panic!("Hash always substitutes");
-        };
-        assert_eq!(value.len(), 128);
-    }
-}
diff --git a/crates/nvisy-toolkit/src/redaction/anonymizer/keep.rs b/crates/nvisy-toolkit/src/redaction/anonymizer/keep.rs
deleted file mode 100644
index 5b74002bd..000000000
--- a/crates/nvisy-toolkit/src/redaction/anonymizer/keep.rs
+++ /dev/null
@@ -1,31 +0,0 @@
-//! [`Keep`]: pass the matched span through unchanged.
-//!
-//! Useful in mixed policies — e.g. mask every kind by default but
-//! keep `builtins::CURRENCY.label_ref()` so prices remain readable. The
-//! replacement records the original value verbatim so the audit
-//! trail still has a row.
-
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Text, TextData};
-
-use crate::redaction::{Anonymizer, LeakProfile, TextReplacement};
-
-/// Pass the matched span through unchanged.
-#[derive(Debug, Clone, Copy, Default)]
-pub struct Keep;
-
-#[async_trait::async_trait]
-impl Anonymizer<Text> for Keep {
-    fn leak_profile(&self) -> LeakProfile {
-        // The original value is unchanged: strictly more leaky than
-        // every other operator. The `Recoverable` ordering matches
-        // "the original is trivially derivable from the output."
-        LeakProfile::Recoverable
-    }
-
-    async fn apply(&self, _entity: &Entity<Text>, source: &TextData) -> Result<TextReplacement> {
-        let value = source.text.as_str();
-        Ok(TextReplacement::substituted(value))
-    }
-}
diff --git a/crates/nvisy-toolkit/src/redaction/anonymizer/mask.rs b/crates/nvisy-toolkit/src/redaction/anonymizer/mask.rs
deleted file mode 100644
index 3609933a6..000000000
--- a/crates/nvisy-toolkit/src/redaction/anonymizer/mask.rs
+++ /dev/null
@@ -1,198 +0,0 @@
-//! [`Mask`]: replace characters of the matched value with a fixed
-//! mask character, optionally leaving a prefix and/or suffix of
-//! the original value unmasked.
-//!
-//! Configurable along three axes:
-//!
-//! - `mask_char` — the character substituted in (default `'*'`).
-//! - `keep_prefix` — number of leading characters to leave visible
-//!   (default `0`). Use [`Mask::with_keep_prefix`].
-//! - `keep_suffix` — number of trailing characters to leave
-//!   visible (default `0`). Use [`Mask::with_keep_suffix`].
-//!
-//! Counts are character-based, not byte-based. When the combined
-//! `keep_prefix + keep_suffix` exceeds the value's length the
-//! whole value passes through unmasked. The output length always
-//! equals the input length.
-//!
-//! Common patterns:
-//!
-//! - Mask everything: `Mask::stars()`.
-//! - Show last 4 of a card: `Mask::stars().with_keep_suffix(4)`.
-//! - Show BIN and last 4: `Mask::stars().with_keep_prefix(4).with_keep_suffix(4)`.
-
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Tabular, Text, TextData};
-
-use crate::redaction::{Anonymizer, LeakProfile, TabularReplacement, TextReplacement};
-
-/// Character-replacement masking operator.
-#[derive(Debug, Clone, Copy)]
-pub struct Mask {
-    mask_char: char,
-    keep_prefix: usize,
-    keep_suffix: usize,
-}
-
-impl Mask {
-    /// Build a `Mask` operator with the given mask character and
-    /// no preserved prefix / suffix — i.e. mask the whole value.
-    pub fn new(mask_char: char) -> Self {
-        Self {
-            mask_char,
-            keep_prefix: 0,
-            keep_suffix: 0,
-        }
-    }
-
-    /// Mask every character with `'*'`.
-    pub fn stars() -> Self {
-        Self::new('*')
-    }
-
-    /// Leave the first `n` characters of the value unmasked. Useful
-    /// for "show BIN" card-display patterns.
-    #[must_use]
-    pub fn with_keep_prefix(mut self, n: usize) -> Self {
-        self.keep_prefix = n;
-        self
-    }
-
-    /// Leave the last `n` characters of the value unmasked. Useful
-    /// for "show last 4" card-display patterns.
-    #[must_use]
-    pub fn with_keep_suffix(mut self, n: usize) -> Self {
-        self.keep_suffix = n;
-        self
-    }
-
-    /// Render `value` under this mask: keep `keep_prefix` leading
-    /// and `keep_suffix` trailing characters verbatim, replace the
-    /// rest with `mask_char`. When the kept regions cover (or
-    /// overlap) the whole value the input is returned unchanged.
-    /// Character-based, not byte-based, so multi-byte codepoints
-    /// stay intact.
-    pub fn render(&self, value: &str) -> String {
-        let total = value.chars().count();
-        // Clamp so prefix + suffix never overlap: when they sum past
-        // `total`, the whole value passes through unmasked.
-        if self.keep_prefix.saturating_add(self.keep_suffix) >= total {
-            return value.to_string();
-        }
-        let suffix_start = total - self.keep_suffix;
-        value
-            .chars()
-            .enumerate()
-            .map(|(i, c)| {
-                if i < self.keep_prefix || i >= suffix_start {
-                    c
-                } else {
-                    self.mask_char
-                }
-            })
-            .collect()
-    }
-}
-
-impl Default for Mask {
-    fn default() -> Self {
-        Self::stars()
-    }
-}
-
-#[async_trait::async_trait]
-impl Anonymizer<Text> for Mask {
-    fn leak_profile(&self) -> LeakProfile {
-        LeakProfile::Partial
-    }
-
-    async fn apply(&self, _entity: &Entity<Text>, source: &TextData) -> Result<TextReplacement> {
-        Ok(TextReplacement::substituted(
-            self.render(source.text.as_str()),
-        ))
-    }
-}
-
-#[async_trait::async_trait]
-impl Anonymizer<Tabular> for Mask {
-    fn leak_profile(&self) -> LeakProfile {
-        LeakProfile::Partial
-    }
-
-    async fn apply(
-        &self,
-        _entity: &Entity<Tabular>,
-        source: &TextData,
-    ) -> Result<TabularReplacement> {
-        Ok(TabularReplacement::substituted(
-            self.render(source.text.as_str()),
-        ))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{TrailStep, builtins};
-    use nvisy_core::modality::TextLocation;
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-
-    fn entity(start: usize, end: usize) -> Entity<Text> {
-        Entity::builder()
-            .with_label(builtins::PAYMENT_CARD.label_ref())
-            .with_location(TextLocation::new(start, end))
-            .with_confidence(Confidence::new(1.0).unwrap())
-            .with_trail(Vec::<TrailStep>::new())
-            .build()
-            .unwrap()
-    }
-
-    #[tokio::test]
-    async fn mask_all_with_stars() {
-        let op = Mask::stars();
-        let source = TextData::new("4111111111111111");
-        let entity = entity(0, source.text.len());
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(out, TextReplacement::substituted("****************"));
-    }
-
-    #[tokio::test]
-    async fn keep_suffix_shows_last_n() {
-        let op = Mask::new('#').with_keep_suffix(4);
-        let source = TextData::new("4111111111111111");
-        let entity = entity(0, source.text.len());
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(out, TextReplacement::substituted("############1111"));
-    }
-
-    #[tokio::test]
-    async fn keep_prefix_shows_first_n() {
-        let op = Mask::new('#').with_keep_prefix(4);
-        let source = TextData::new("4111111111111111");
-        let entity = entity(0, source.text.len());
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(out, TextReplacement::substituted("4111############"));
-    }
-
-    #[tokio::test]
-    async fn keep_prefix_and_suffix_show_both_ends() {
-        let op = Mask::new('*').with_keep_prefix(4).with_keep_suffix(4);
-        let source = TextData::new("4111111111111111");
-        let entity = entity(0, source.text.len());
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(out, TextReplacement::substituted("4111********1111"));
-    }
-
-    #[tokio::test]
-    async fn keep_exceeding_length_passes_value_through() {
-        // 16-char value with prefix+suffix asking for 20 visible
-        // chars: nothing left to mask, return verbatim.
-        let op = Mask::new('*').with_keep_prefix(10).with_keep_suffix(10);
-        let source = TextData::new("4111111111111111");
-        let entity = entity(0, source.text.len());
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(out, TextReplacement::substituted("4111111111111111"));
-    }
-}
diff --git a/crates/nvisy-toolkit/src/redaction/anonymizer/mod.rs b/crates/nvisy-toolkit/src/redaction/anonymizer/mod.rs
deleted file mode 100644
index 8fe59172e..000000000
--- a/crates/nvisy-toolkit/src/redaction/anonymizer/mod.rs
+++ /dev/null
@@ -1,27 +0,0 @@
-//! Built-in [`Anonymizer<M>`] implementations shipped with the
-//! toolkit.
-//!
-//! Each operator is a typed Rust struct; consumers construct it
-//! with the parameters they want and register the instance
-//! against the [`EntityLabelRef`]s it should run for.
-//!
-//! [`Anonymizer<M>`]: crate::redaction::Anonymizer
-//! [`EntityLabelRef`]: nvisy_core::entity::EntityLabelRef
-
-#[cfg(feature = "encrypt")]
-#[cfg_attr(docsrs, doc(cfg(feature = "encrypt")))]
-pub(crate) mod encrypt;
-mod hash;
-mod keep;
-mod mask;
-mod redact;
-mod replace;
-
-#[cfg(feature = "encrypt")]
-#[cfg_attr(docsrs, doc(cfg(feature = "encrypt")))]
-pub use self::encrypt::Encrypt;
-pub use self::hash::{Hash, HashAlgorithm};
-pub use self::keep::Keep;
-pub use self::mask::Mask;
-pub use self::redact::Redact;
-pub use self::replace::Replace;
diff --git a/crates/nvisy-toolkit/src/redaction/anonymizer/redact.rs b/crates/nvisy-toolkit/src/redaction/anonymizer/redact.rs
deleted file mode 100644
index a2ddb31a8..000000000
--- a/crates/nvisy-toolkit/src/redaction/anonymizer/redact.rs
+++ /dev/null
@@ -1,26 +0,0 @@
-//! [`Redact`]: delete the matched span entirely.
-//!
-//! The codec writes nothing back at the entity's location; the span
-//! disappears from the output. This is the strongest text operator —
-//! no trace of the original value or its shape remains.
-
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Text, TextData};
-
-use crate::redaction::{Anonymizer, LeakProfile, TextReplacement};
-
-/// Delete the matched span entirely.
-#[derive(Debug, Clone, Copy, Default)]
-pub struct Redact;
-
-#[async_trait::async_trait]
-impl Anonymizer<Text> for Redact {
-    fn leak_profile(&self) -> LeakProfile {
-        LeakProfile::Irrecoverable
-    }
-
-    async fn apply(&self, _entity: &Entity<Text>, _source: &TextData) -> Result<TextReplacement> {
-        Ok(TextReplacement::Removed)
-    }
-}
diff --git a/crates/nvisy-toolkit/src/redaction/anonymizer/replace.rs b/crates/nvisy-toolkit/src/redaction/anonymizer/replace.rs
deleted file mode 100644
index e968fd9c8..000000000
--- a/crates/nvisy-toolkit/src/redaction/anonymizer/replace.rs
+++ /dev/null
@@ -1,127 +0,0 @@
-//! [`Replace`]: substitute the matched span with a fixed template
-//! string.
-//!
-//! Templates support two placeholders that are expanded at apply
-//! time:
-//!
-//! - `{label}` — the entity's label name (e.g. `person_name`).
-//! - `{value}` — the original matched substring.
-//!
-//! The default template is `[{label}]`.
-
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Tabular, Text, TextData};
-
-use crate::redaction::{Anonymizer, LeakProfile, TabularReplacement, TextReplacement};
-
-/// Substitute the matched span with a template string.
-#[derive(Debug, Clone)]
-pub struct Replace {
-    template: String,
-}
-
-impl Replace {
-    /// Build a `Replace` operator with the given template. See the
-    /// module docs for placeholder syntax.
-    pub fn new(template: impl Into<String>) -> Self {
-        Self {
-            template: template.into(),
-        }
-    }
-}
-
-impl Default for Replace {
-    /// Default template is `[{label}]` so users who don't
-    /// configure a template still get a visible label-tagged
-    /// marker.
-    fn default() -> Self {
-        Self::new("[{label}]")
-    }
-}
-
-#[async_trait::async_trait]
-impl Anonymizer<Text> for Replace {
-    fn leak_profile(&self) -> LeakProfile {
-        // Position and length of the rewritten span are still
-        // observable in the output even though the original value is
-        // gone.
-        LeakProfile::Partial
-    }
-
-    async fn apply(&self, entity: &Entity<Text>, source: &TextData) -> Result<TextReplacement> {
-        let value = source.text.as_str();
-        let rendered = render(&self.template, entity.label.as_str(), value);
-        Ok(TextReplacement::substituted(rendered))
-    }
-}
-
-#[async_trait::async_trait]
-impl Anonymizer<Tabular> for Replace {
-    fn leak_profile(&self) -> LeakProfile {
-        LeakProfile::Partial
-    }
-
-    async fn apply(
-        &self,
-        entity: &Entity<Tabular>,
-        source: &TextData,
-    ) -> Result<TabularReplacement> {
-        let value = source.text.as_str();
-        let rendered = render(&self.template, entity.label.as_str(), value);
-        Ok(TabularReplacement::substituted(rendered))
-    }
-}
-
-fn render(template: &str, label: &str, value: &str) -> String {
-    template.replace("{label}", label).replace("{value}", value)
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{EntityLabelRef, TrailStep, builtins};
-    use nvisy_core::modality::TextLocation;
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-
-    fn entity(label: EntityLabelRef, start: usize, end: usize) -> Entity<Text> {
-        Entity::builder()
-            .with_label(label)
-            .with_location(TextLocation::new(start, end))
-            .with_confidence(Confidence::new(1.0).unwrap())
-            .with_trail(Vec::<TrailStep>::new())
-            .build()
-            .unwrap()
-    }
-
-    #[tokio::test]
-    async fn default_template_emits_bracketed_label() {
-        let op = Replace::default();
-        let source = TextData::new("alice@example.test");
-        let entity = entity(builtins::EMAIL_ADDRESS.label_ref(), 0, 18);
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(out, TextReplacement::substituted("[email_address]"));
-    }
-
-    #[tokio::test]
-    async fn template_with_value_placeholder() {
-        let op = Replace::new("<<{value}::{label}>>");
-        let source = TextData::new("alice");
-        let entity = entity(builtins::EMAIL_ADDRESS.label_ref(), 0, source.text.len());
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(
-            out,
-            TextReplacement::substituted("<<alice::email_address>>")
-        );
-    }
-
-    #[tokio::test]
-    async fn empty_source_yields_empty_value_placeholder() {
-        let op = Replace::new("[{value}]");
-        let source = TextData::new("");
-        let entity = entity(builtins::PERSON_NAME.label_ref(), 0, 0);
-        let out = op.apply(&entity, &source).await.unwrap();
-        assert_eq!(out, TextReplacement::substituted("[]"));
-    }
-}
diff --git a/crates/nvisy-toolkit/src/redaction/deanonymizer/decrypt.rs b/crates/nvisy-toolkit/src/redaction/deanonymizer/decrypt.rs
deleted file mode 100644
index da6a3eae5..000000000
--- a/crates/nvisy-toolkit/src/redaction/deanonymizer/decrypt.rs
+++ /dev/null
@@ -1,207 +0,0 @@
-//! [`Decrypt`]: inverse of [`Encrypt`].
-//!
-//! Implements [`Deanonymizer<M>`] as a **self-contained** recovery
-//! operator: the ciphertext blob in the replacement carries
-//! everything needed to decode it, so the [`Entity`] argument is
-//! ignored. See the [`Deanonymizer`] module docs for the audit-
-//! keyed vs. self-contained distinction.
-//!
-//! Wire shape matches what [`Encrypt`] produces:
-//! `base64(nonce || ciphertext || auth_tag)` with a 12-byte nonce.
-//!
-//! [`Encrypt`]: crate::redaction::anonymizer::Encrypt
-//! [`Deanonymizer`]: nvisy_core::redaction::Deanonymizer
-//! [`Deanonymizer<M>`]: nvisy_core::redaction::Deanonymizer
-
-use aes_gcm::aead::{Aead, KeyInit};
-use aes_gcm::{Aes256Gcm, Key, Nonce};
-use base64::Engine;
-use base64::engine::general_purpose::STANDARD;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Tabular, Text, TextData};
-use nvisy_core::redaction::{Deanonymizer, TabularReplacement, TextReplacement};
-use nvisy_core::{Error, Result};
-
-use crate::redaction::anonymizer::encrypt::NONCE_LEN;
-
-const TARGET: &str = "nvisy_toolkit::redaction::decrypt";
-
-/// AES-256-GCM decrypt operator. Inverse of [`Encrypt`].
-///
-/// [`Encrypt`]: crate::redaction::anonymizer::Encrypt
-#[derive(Clone)]
-pub struct Decrypt {
-    key: Key<Aes256Gcm>,
-}
-
-impl Decrypt {
-    /// Construct from the same 32-byte key passed to [`Encrypt`].
-    ///
-    /// [`Encrypt`]: crate::redaction::anonymizer::Encrypt
-    pub fn from_key(key: [u8; 32]) -> Self {
-        Self { key: key.into() }
-    }
-
-    /// Parse the base64 envelope, split nonce from ciphertext, run
-    /// AES-256-GCM, and decode the plaintext bytes as UTF-8.
-    fn decrypt_blob(&self, value: &str) -> Result<TextData> {
-        let blob = STANDARD
-            .decode(value)
-            .map_err(|e| Error::validation(format!("invalid base64 ciphertext: {e}"), TARGET))?;
-
-        if blob.len() < NONCE_LEN {
-            return Err(Error::validation(
-                format!("ciphertext too short for nonce (got {} bytes)", blob.len()),
-                TARGET,
-            ));
-        }
-
-        let (nonce_bytes, ciphertext) = blob.split_at(NONCE_LEN);
-        let nonce = Nonce::from_slice(nonce_bytes);
-        let cipher = Aes256Gcm::new(&self.key);
-        let plaintext = cipher.decrypt(nonce, ciphertext).map_err(|e| {
-            Error::runtime(
-                format!("decryption failed (wrong key or tampered): {e}"),
-                TARGET,
-                false,
-            )
-        })?;
-
-        let text = String::from_utf8(plaintext).map_err(|e| {
-            Error::validation(format!("decrypted bytes are not valid UTF-8: {e}"), TARGET)
-        })?;
-
-        Ok(TextData::new(text))
-    }
-}
-
-#[async_trait::async_trait]
-impl Deanonymizer<Text> for Decrypt {
-    async fn revert(
-        &self,
-        _entity: &Entity<Text>,
-        replacement: &TextReplacement,
-    ) -> Result<Option<TextData>> {
-        let TextReplacement::Substituted { value } = replacement else {
-            return Ok(None);
-        };
-        self.decrypt_blob(value).map(Some)
-    }
-}
-
-#[async_trait::async_trait]
-impl Deanonymizer<Tabular> for Decrypt {
-    async fn revert(
-        &self,
-        _entity: &Entity<Tabular>,
-        replacement: &TabularReplacement,
-    ) -> Result<Option<TextData>> {
-        let TabularReplacement::Substituted { value } = replacement else {
-            return Ok(None);
-        };
-        self.decrypt_blob(value).map(Some)
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::entity::{TrailStep, builtins};
-    use nvisy_core::modality::{TabularLocation, TextLocation};
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-    use crate::redaction::Anonymizer;
-    use crate::redaction::anonymizer::Encrypt;
-
-    const KEY: [u8; 32] = [
-        0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f,
-        0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e,
-        0x1f, 0x20,
-    ];
-
-    fn text_entity(start: usize, end: usize) -> Entity<Text> {
-        Entity::builder()
-            .with_label(builtins::EMAIL_ADDRESS.label_ref())
-            .with_location(TextLocation::new(start, end))
-            .with_confidence(Confidence::new(1.0).unwrap())
-            .with_trail(Vec::<TrailStep>::new())
-            .build()
-            .unwrap()
-    }
-
-    fn tabular_entity(row: u32, col: u32) -> Entity<Tabular> {
-        Entity::builder()
-            .with_label(builtins::EMAIL_ADDRESS.label_ref())
-            .with_location(TabularLocation::new(row, col))
-            .with_confidence(Confidence::new(1.0).unwrap())
-            .with_trail(Vec::<TrailStep>::new())
-            .build()
-            .unwrap()
-    }
-
-    #[tokio::test]
-    async fn text_round_trip() {
-        let enc = Encrypt::from_key(KEY);
-        let dec = Decrypt::from_key(KEY);
-        let source = TextData::new("alice@example.test");
-        let entity = text_entity(0, 18);
-
-        let ciphertext = enc.apply(&entity, &source).await.unwrap();
-        let recovered = dec.revert(&entity, &ciphertext).await.unwrap();
-        assert_eq!(
-            recovered.expect("decrypts cleanly").text.as_str(),
-            "alice@example.test"
-        );
-    }
-
-    #[tokio::test]
-    async fn tabular_round_trip() {
-        let enc = Encrypt::from_key(KEY);
-        let dec = Decrypt::from_key(KEY);
-        let source = TextData::new("alice@example.test");
-        let entity = tabular_entity(1, 1);
-
-        let ciphertext: TabularReplacement = enc.apply(&entity, &source).await.unwrap();
-        let recovered = dec.revert(&entity, &ciphertext).await.unwrap();
-        assert_eq!(
-            recovered.expect("decrypts cleanly").text.as_str(),
-            "alice@example.test"
-        );
-    }
-
-    #[tokio::test]
-    async fn wrong_key_fails_decryption() {
-        let enc = Encrypt::from_key(KEY);
-        let mut bad_key = KEY;
-        bad_key[0] ^= 0xff;
-        let dec = Decrypt::from_key(bad_key);
-        let source = TextData::new("alice@example.test");
-        let entity = text_entity(0, 18);
-
-        let ciphertext = enc.apply(&entity, &source).await.unwrap();
-        let err = dec.revert(&entity, &ciphertext).await.unwrap_err();
-        assert!(err.to_string().contains("decryption failed"));
-    }
-
-    #[tokio::test]
-    async fn removed_text_replacement_yields_none() {
-        let dec = Decrypt::from_key(KEY);
-        let entity = text_entity(0, 0);
-        let out = dec
-            .revert(&entity, &TextReplacement::Removed)
-            .await
-            .unwrap();
-        assert!(out.is_none());
-    }
-
-    #[tokio::test]
-    async fn column_dropped_tabular_replacement_yields_none() {
-        let dec = Decrypt::from_key(KEY);
-        let entity = tabular_entity(1, 1);
-        let out = dec
-            .revert(&entity, &TabularReplacement::ColumnDropped)
-            .await
-            .unwrap();
-        assert!(out.is_none());
-    }
-}
diff --git a/crates/nvisy-toolkit/src/redaction/deanonymizer/mod.rs b/crates/nvisy-toolkit/src/redaction/deanonymizer/mod.rs
deleted file mode 100644
index 733c06642..000000000
--- a/crates/nvisy-toolkit/src/redaction/deanonymizer/mod.rs
+++ /dev/null
@@ -1,19 +0,0 @@
-//! Built-in [`Deanonymizer<M>`] implementations shipped with the
-//! toolkit.
-//!
-//! Each operator recovers the original payload an [`Anonymizer<M>`]
-//! wrote. Two recovery shapes ship today (see [`Deanonymizer`]):
-//! audit-keyed (no impl yet) and self-contained (e.g. `Decrypt`,
-//! gated behind the `encrypt` feature).
-//!
-//! [`Anonymizer<M>`]: crate::redaction::Anonymizer
-//! [`Deanonymizer<M>`]: crate::redaction::Deanonymizer
-//! [`Deanonymizer`]: crate::redaction::Deanonymizer
-
-#[cfg(feature = "encrypt")]
-#[cfg_attr(docsrs, doc(cfg(feature = "encrypt")))]
-mod decrypt;
-
-#[cfg(feature = "encrypt")]
-#[cfg_attr(docsrs, doc(cfg(feature = "encrypt")))]
-pub use self::decrypt::Decrypt;
diff --git a/crates/nvisy-toolkit/src/redaction/id.rs b/crates/nvisy-toolkit/src/redaction/id.rs
deleted file mode 100644
index 086bc4108..000000000
--- a/crates/nvisy-toolkit/src/redaction/id.rs
+++ /dev/null
@@ -1,125 +0,0 @@
-//! [`AnonymizerId<M>`]: the lookup handle for a custom
-//! [`Anonymizer<M>`] in the [`RedactionRegistry<M>`].
-//!
-//! Identifies a deployment-registered custom operator by name. The
-//! `M` phantom binds the id to a single modality at the type level —
-//! a `AnonymizerId<Text>` cannot be handed to the image registry.
-//!
-//! The built-in operators bundled with the toolkit
-//! ([`Replace`], [`Mask`], [`Hash`], [`Redact`],
-//! [`Keep`], [`Encrypt`]) do **not** carry an `AnonymizerId` —
-//! they are instantiated per-call from policy-supplied params and
-//! never round-trip through the registry. `AnonymizerId` exists
-//! solely for the `Custom` escape hatch used to plug stateful Rust
-//! operators (KMS clients, token vaults, …) that can't be expressed
-//! as a config blob.
-//!
-//! [`Anonymizer<M>`]: super::Anonymizer
-//! [`RedactionRegistry<M>`]: super::RedactionRegistry
-//! [`Replace`]: super::anonymizer::Replace
-//! [`Mask`]: super::anonymizer::Mask
-//! [`Hash`]: super::anonymizer::Hash
-//! [`Redact`]: super::anonymizer::Redact
-//! [`Keep`]: super::anonymizer::Keep
-//! [`Encrypt`]: super::anonymizer::Encrypt
-
-use std::borrow::Cow;
-use std::fmt::{self, Debug, Display};
-use std::hash::{Hash, Hasher};
-use std::marker::PhantomData;
-
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-
-use super::Modality;
-
-/// Lookup handle for a custom [`Anonymizer<M>`] in the
-/// [`RedactionRegistry<M>`]. Phantom-typed over `M` so the compiler
-/// rejects cross-modality use at the call site.
-///
-/// Ids are author-supplied strings (deployment chooses the
-/// vocabulary). Equality and hashing ignore the modality marker —
-/// the name is the actual key.
-///
-/// [`Anonymizer<M>`]: super::Anonymizer
-/// [`RedactionRegistry<M>`]: super::RedactionRegistry
-#[derive(Serialize, Deserialize, JsonSchema)]
-#[serde(transparent)]
-pub struct AnonymizerId<M: Modality> {
-    name: Cow<'static, str>,
-    #[serde(skip)]
-    #[schemars(skip)]
-    _modality: PhantomData<fn() -> M>,
-}
-
-impl<M: Modality> AnonymizerId<M> {
-    /// Build an id from a `&'static str` without allocating.
-    #[must_use]
-    pub const fn from_static(name: &'static str) -> Self {
-        Self {
-            name: Cow::Borrowed(name),
-            _modality: PhantomData,
-        }
-    }
-
-    /// Build an id from any string-like value.
-    pub fn new(name: impl Into<Cow<'static, str>>) -> Self {
-        Self {
-            name: name.into(),
-            _modality: PhantomData,
-        }
-    }
-
-    /// Borrow the underlying name.
-    #[must_use]
-    pub fn as_str(&self) -> &str {
-        &self.name
-    }
-}
-
-impl<M: Modality> Clone for AnonymizerId<M> {
-    fn clone(&self) -> Self {
-        Self {
-            name: self.name.clone(),
-            _modality: PhantomData,
-        }
-    }
-}
-
-impl<M: Modality> Debug for AnonymizerId<M> {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_tuple("AnonymizerId").field(&self.name).finish()
-    }
-}
-
-impl<M: Modality> Display for AnonymizerId<M> {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        Display::fmt(&self.name, f)
-    }
-}
-
-impl<M: Modality> PartialEq for AnonymizerId<M> {
-    fn eq(&self, other: &Self) -> bool {
-        self.name == other.name
-    }
-}
-
-impl<M: Modality> Eq for AnonymizerId<M> {}
-
-impl<M: Modality> Hash for AnonymizerId<M> {
-    fn hash<H: Hasher>(&self, state: &mut H) {
-        self.name.hash(state);
-    }
-}
-
-impl<M: Modality> From<&'static str> for AnonymizerId<M> {
-    fn from(name: &'static str) -> Self {
-        Self::from_static(name)
-    }
-}
-
-impl<M: Modality> From<String> for AnonymizerId<M> {
-    fn from(name: String) -> Self {
-        Self::new(name)
-    }
-}
diff --git a/crates/nvisy-toolkit/src/redaction/mod.rs b/crates/nvisy-toolkit/src/redaction/mod.rs
deleted file mode 100644
index d8ec4a2a1..000000000
--- a/crates/nvisy-toolkit/src/redaction/mod.rs
+++ /dev/null
@@ -1,49 +0,0 @@
-//! Redaction layer: per-modality anonymizer/deanonymizer operators
-//! plus the generic [`RedactionRegistry<M>`] that holds
-//! deployment-registered custom operators.
-//!
-//! The toolkit ships:
-//!
-//! - The [`Anonymizer<M>`] / [`Deanonymizer<M>`] traits operators
-//!   implement.
-//! - Built-in forward operators in [`anonymizer`] (`Replace`,
-//!   `Mask`, `Hash`, `Redact`, `Keep`, `Encrypt`) and reverse
-//!   operators in [`deanonymizer`] (`Decrypt`). Each is a plain
-//!   struct; consumers construct one per rule with whatever params
-//!   they need.
-//! - [`AnonymizerId<M>`] + [`RedactionRegistry<M>`] for the
-//!   `Custom` escape hatch: stateful Rust operators (KMS clients,
-//!   token vaults, …) that can't be expressed as a config blob.
-//!   Built-in operators do **not** round-trip through the registry.
-//!
-//! Per-modality replacement shapes ([`TextReplacement`],
-//! [`ImageReplacement`], [`AudioReplacement`], [`TabularReplacement`])
-//! describe what an operator emits at the entity's location.
-//!
-//! Dispatch (entity → operator) lives one layer up, in the engine
-//! crate: each rule's `Action::Redact(ModalityRedactions)` carries
-//! per-modality operator specs. Each operator spec either names a
-//! built-in (constructed inline) or a `Custom(AnonymizerId<M>)`
-//! (looked up in the registry).
-//!
-//! [`Anonymizer<M>`]: Anonymizer
-//! [`Deanonymizer<M>`]: Deanonymizer
-//! [`AnonymizerId<M>`]: AnonymizerId
-//! [`RedactionRegistry<M>`]: RedactionRegistry
-
-mod id;
-mod registry;
-mod store;
-
-pub mod anonymizer;
-pub mod deanonymizer;
-
-pub use nvisy_core::modality::Modality;
-pub use nvisy_core::redaction::{
-    Anonymizer, AudioReplacement, Deanonymizer, ImageReplacement, LeakProfile, Memoized, Store,
-    TabularReplacement, TextReplacement,
-};
-
-pub use self::id::AnonymizerId;
-pub use self::registry::RedactionRegistry;
-pub use self::store::MemoryStore;
diff --git a/crates/nvisy-toolkit/src/redaction/registry.rs b/crates/nvisy-toolkit/src/redaction/registry.rs
deleted file mode 100644
index 5c783275b..000000000
--- a/crates/nvisy-toolkit/src/redaction/registry.rs
+++ /dev/null
@@ -1,355 +0,0 @@
-//! [`RedactionRegistry<M>`]: per-modality lookup tables of
-//! [`Anonymizer<M>`] instances.
-//!
-//! The registry exposes two independent indexes plus an optional
-//! catch-all:
-//!
-//! - **`by_label`** — keyed by [`EntityLabelRef`]. The dispatch
-//!   the toolkit-only pipeline uses: "this entity has label
-//!   `email_address`; what operator do I run?". Populated by
-//!   callers with `insert_label`.
-//! - **`by_id`** — keyed by [`AnonymizerId<M>`]. The dispatch the
-//!   document-side policy layer uses when a policy rule resolves to
-//!   `Custom { name }` and the named operator must be looked up by
-//!   string id. Populated by callers with `insert_id`.
-//! - **`fallback`** — the operator to use when `by_label.get(label)`
-//!   misses. Optional; when absent, unregistered labels skip.
-//!
-//! The two indexes are independent: registering the same operator
-//! both by kind and by id is a deliberate call-site choice, not an
-//! automatic mirroring.
-//!
-//! ```ignore
-//! use nvisy_core::entity::builtins;
-//! use nvisy_core::modality::Text;
-//! use nvisy_toolkit::redaction::anonymizer::{Mask, Redact, Replace};
-//! use nvisy_toolkit::redaction::{AnonymizerId, RedactionRegistry};
-//!
-//! const KMS_ENCRYPT: AnonymizerId<Text> = AnonymizerId::from_static("kms_encrypt");
-//!
-//! let registry = RedactionRegistry::<Text>::new()
-//!     .insert_label(builtins::EMAIL_ADDRESS.label_ref(), Replace::new("[EMAIL]"))
-//!     .insert_label(builtins::PAYMENT_CARD.label_ref(), Mask::new('#').with_keep_suffix(4))
-//!     .insert_id(KMS_ENCRYPT, MyKmsOperator::new(client))
-//!     .with_fallback(Redact);
-//! ```
-//!
-//! [`Anonymizer<M>`]: super::Anonymizer
-
-use std::collections::HashMap;
-use std::fmt;
-use std::sync::Arc;
-
-use nvisy_core::Result;
-use nvisy_core::entity::{Entity, EntityLabelRef};
-use nvisy_core::extraction::DataAt;
-use nvisy_core::modality::Modality;
-use nvisy_core::redaction::Redactions;
-
-use super::{Anonymizer, AnonymizerId};
-
-/// Per-modality registry of [`Anonymizer<M>`] instances, indexed by
-/// both [`EntityLabelRef`] (toolkit-side per-label dispatch) and
-/// [`AnonymizerId<M>`] (policy-side custom-operator resolution).
-///
-/// [`Anonymizer<M>`]: super::Anonymizer
-pub struct RedactionRegistry<M: Modality> {
-    by_label: HashMap<EntityLabelRef, Arc<dyn Anonymizer<M>>>,
-    by_id: HashMap<AnonymizerId<M>, Arc<dyn Anonymizer<M>>>,
-    fallback: Option<Arc<dyn Anonymizer<M>>>,
-}
-
-impl<M: Modality> RedactionRegistry<M> {
-    /// Build an empty registry. Use [`insert_label`], [`insert_id`],
-    /// or [`with_fallback`] to populate it.
-    ///
-    /// [`insert_label`]: Self::insert_label
-    /// [`insert_id`]: Self::insert_id
-    /// [`with_fallback`]: Self::with_fallback
-    #[must_use]
-    pub fn new() -> Self {
-        Self {
-            by_label: HashMap::new(),
-            by_id: HashMap::new(),
-            fallback: None,
-        }
-    }
-
-    /// Register `op` as the operator the toolkit pipeline picks
-    /// when it encounters an entity carrying `label`.
-    /// Re-registering the same label replaces the previous
-    /// instance.
-    #[must_use]
-    pub fn insert_label(mut self, label: EntityLabelRef, op: impl Anonymizer<M> + 'static) -> Self {
-        self.by_label.insert(label, Arc::new(op));
-        self
-    }
-
-    /// Register `op` under `id` for policy-side `Custom { name }`
-    /// resolution. Re-registering the same id replaces the previous
-    /// instance.
-    #[must_use]
-    pub fn insert_id(mut self, id: AnonymizerId<M>, op: impl Anonymizer<M> + 'static) -> Self {
-        self.by_id.insert(id, Arc::new(op));
-        self
-    }
-
-    /// Install a catch-all operator used when [`resolve`] misses.
-    /// Setting it again replaces the previous fallback.
-    ///
-    /// [`resolve`]: Self::resolve
-    #[must_use]
-    pub fn with_fallback(mut self, op: impl Anonymizer<M> + 'static) -> Self {
-        self.fallback = Some(Arc::new(op));
-        self
-    }
-
-    /// Resolve an entity label to its registered operator, falling
-    /// back to the catch-all when no per-label binding exists.
-    /// Returns `None` only when neither a per-label operator nor a
-    /// fallback was registered.
-    #[must_use]
-    pub fn resolve(&self, label: &EntityLabelRef) -> Option<&Arc<dyn Anonymizer<M>>> {
-        self.by_label.get(label).or(self.fallback.as_ref())
-    }
-
-    /// Resolve an [`AnonymizerId<M>`] to its registered operator.
-    /// Used by the document-side policy layer for `Custom { name }`
-    /// dispatch; does **not** consult the fallback.
-    #[must_use]
-    pub fn resolve_id(&self, id: &AnonymizerId<M>) -> Option<&Arc<dyn Anonymizer<M>>> {
-        self.by_id.get(id)
-    }
-
-    /// Number of distinct entity labels registered.
-    #[must_use]
-    pub fn labels_len(&self) -> usize {
-        self.by_label.len()
-    }
-
-    /// Number of distinct ids registered.
-    #[must_use]
-    pub fn ids_len(&self) -> usize {
-        self.by_id.len()
-    }
-
-    /// `true` when neither index nor a fallback are registered.
-    #[must_use]
-    pub fn is_empty(&self) -> bool {
-        self.by_label.is_empty() && self.by_id.is_empty() && self.fallback.is_none()
-    }
-
-    /// Run [`resolve`] + [`Anonymizer::apply`] over every entity, and
-    /// collect the produced replacements into a [`Redactions<M>`] batch
-    /// ready to hand to a [`RedactAt<M>`] implementation.
-    ///
-    /// `source` is a [`DataAt<M>`] resolver: for each entity, the
-    /// substring/segment **at the entity's location** is pulled from
-    /// the resolver and handed to the anonymizer — the anonymizer
-    /// never sees the whole document. Entities whose location can't
-    /// be resolved by the source are skipped.
-    ///
-    /// Entities whose label has no per-label operator and where no
-    /// fallback was registered are skipped (counted as a debug-level
-    /// tracing event); the rest are applied in iteration order.
-    ///
-    /// [`resolve`]: Self::resolve
-    /// [`Anonymizer::apply`]: super::Anonymizer::apply
-    /// [`DataAt<M>`]: nvisy_core::extraction::DataAt
-    /// [`RedactAt<M>`]: nvisy_core::redaction::RedactAt
-    pub async fn apply_all<'a, I>(
-        &self,
-        entities: I,
-        source: &(impl DataAt<M> + ?Sized),
-    ) -> Result<Redactions<M>>
-    where
-        I: IntoIterator<Item = &'a Entity<M>>,
-        M: 'a,
-    {
-        let mut out = Redactions::<M>::new();
-        let mut skipped = 0usize;
-        let mut unresolved = 0usize;
-        for entity in entities {
-            let Some(op) = self.resolve(&entity.label) else {
-                skipped += 1;
-                continue;
-            };
-            let Some(data) = source.data_at(&entity.location).await else {
-                unresolved += 1;
-                continue;
-            };
-            let replacement = op.apply(entity, &data).await?;
-            out.push(entity.location.clone(), replacement);
-        }
-        if skipped > 0 {
-            tracing::debug!(
-                target: "nvisy_toolkit::redaction::registry",
-                skipped,
-                "RedactionRegistry::apply_all skipped entities with no per-label operator and no fallback",
-            );
-        }
-        if unresolved > 0 {
-            tracing::debug!(
-                target: "nvisy_toolkit::redaction::registry",
-                unresolved,
-                "RedactionRegistry::apply_all skipped entities whose location resolved no source payload",
-            );
-        }
-        Ok(out)
-    }
-}
-
-impl<M: Modality> Default for RedactionRegistry<M> {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
-impl<M: Modality> Clone for RedactionRegistry<M> {
-    fn clone(&self) -> Self {
-        Self {
-            by_label: self.by_label.clone(),
-            by_id: self.by_id.clone(),
-            fallback: self.fallback.clone(),
-        }
-    }
-}
-
-impl<M: Modality> fmt::Debug for RedactionRegistry<M> {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        f.debug_struct("RedactionRegistry")
-            .field("labels", &self.by_label.len())
-            .field("ids", &self.by_id.len())
-            .field("fallback", &self.fallback.is_some())
-            .finish()
-    }
-}
-
-#[cfg(test)]
-mod tests {
-
-    use nvisy_core::entity::{Entity, builtins};
-    use nvisy_core::modality::{Text, TextData, TextLocation};
-    use nvisy_core::primitive::Confidence;
-
-    use super::*;
-    use crate::redaction::{LeakProfile, TextReplacement};
-
-    struct StubAnonymizer(&'static str);
-
-    #[async_trait::async_trait]
-    impl Anonymizer<Text> for StubAnonymizer {
-        fn leak_profile(&self) -> LeakProfile {
-            LeakProfile::Irrecoverable
-        }
-
-        async fn apply(
-            &self,
-            _entity: &Entity<Text>,
-            _source: &TextData,
-        ) -> Result<TextReplacement> {
-            Ok(TextReplacement::substituted(self.0))
-        }
-    }
-
-    /// Minimal in-memory `DataAt<Text>` for tests: slices the source
-    /// string at the entity's byte range, like the codec does.
-    struct StubSource(String);
-
-    #[async_trait::async_trait]
-    impl DataAt<Text> for StubSource {
-        async fn data_at(&self, location: &TextLocation) -> Option<TextData> {
-            self.0
-                .get(location.start..location.end)
-                .map(|s| TextData::new(s.to_owned()))
-        }
-    }
-
-    fn entity(label: EntityLabelRef, start: usize, end: usize) -> Entity<Text> {
-        Entity::<Text>::builder()
-            .with_label(label)
-            .with_location(TextLocation::new(start, end))
-            .with_confidence(Confidence::new(0.9).unwrap())
-            .build()
-            .expect("test fixture entity must build")
-    }
-
-    #[test]
-    fn empty_registry_resolves_nothing() {
-        let r = RedactionRegistry::<Text>::new();
-        assert!(r.is_empty());
-        assert!(r.resolve(&builtins::EMAIL_ADDRESS.label_ref()).is_none());
-        assert!(r.resolve_id(&AnonymizerId::from_static("kms")).is_none());
-    }
-
-    #[test]
-    fn insert_label_then_resolve_returns_operator() {
-        let r = RedactionRegistry::<Text>::new().insert_label(
-            builtins::EMAIL_ADDRESS.label_ref(),
-            StubAnonymizer("[EMAIL]"),
-        );
-        assert_eq!(r.labels_len(), 1);
-        assert!(r.resolve(&builtins::EMAIL_ADDRESS.label_ref()).is_some());
-    }
-
-    #[test]
-    fn insert_id_then_resolve_id_returns_operator() {
-        let id = AnonymizerId::<Text>::from_static("kms");
-        let r = RedactionRegistry::<Text>::new().insert_id(id.clone(), StubAnonymizer("[KMS]"));
-        assert_eq!(r.ids_len(), 1);
-        assert!(r.resolve_id(&id).is_some());
-    }
-
-    #[test]
-    fn fallback_covers_unregistered_labels() {
-        let r = RedactionRegistry::<Text>::new().with_fallback(StubAnonymizer("[REDACTED]"));
-        assert!(r.resolve(&builtins::PAYMENT_CARD.label_ref()).is_some());
-    }
-
-    #[test]
-    fn per_label_wins_over_fallback() {
-        let r = RedactionRegistry::<Text>::new()
-            .insert_label(
-                builtins::EMAIL_ADDRESS.label_ref(),
-                StubAnonymizer("[EMAIL]"),
-            )
-            .with_fallback(StubAnonymizer("[OTHER]"));
-        assert!(r.resolve(&builtins::EMAIL_ADDRESS.label_ref()).is_some());
-        assert!(r.resolve(&builtins::PAYMENT_CARD.label_ref()).is_some());
-    }
-
-    #[tokio::test]
-    async fn apply_all_uses_per_label_with_fallback() {
-        let r = RedactionRegistry::<Text>::new()
-            .insert_label(
-                builtins::EMAIL_ADDRESS.label_ref(),
-                StubAnonymizer("[EMAIL]"),
-            )
-            .with_fallback(StubAnonymizer("[OTHER]"));
-        let entities = [
-            entity(builtins::EMAIL_ADDRESS.label_ref(), 0, 5),
-            entity(builtins::PAYMENT_CARD.label_ref(), 6, 10),
-        ];
-        let source = StubSource("abcdefghij".to_owned());
-        let rs = r.apply_all(entities.iter(), &source).await.unwrap();
-        let items = rs.into_items();
-        assert_eq!(items.len(), 2);
-        assert_eq!(items[0].1, TextReplacement::substituted("[EMAIL]"));
-        assert_eq!(items[1].1, TextReplacement::substituted("[OTHER]"));
-    }
-
-    #[tokio::test]
-    async fn apply_all_skips_unmatched_entities_without_fallback() {
-        let r = RedactionRegistry::<Text>::new().insert_label(
-            builtins::EMAIL_ADDRESS.label_ref(),
-            StubAnonymizer("[EMAIL]"),
-        );
-        let entities = [
-            entity(builtins::EMAIL_ADDRESS.label_ref(), 0, 5),
-            entity(builtins::PAYMENT_CARD.label_ref(), 6, 10),
-        ];
-        let source = StubSource("abcdefghij".to_owned());
-        let rs = r.apply_all(entities.iter(), &source).await.unwrap();
-        assert_eq!(rs.len(), 1);
-    }
-}
diff --git a/crates/nvisy-toolkit/src/redaction/store.rs b/crates/nvisy-toolkit/src/redaction/store.rs
deleted file mode 100644
index 8582fd65d..000000000
--- a/crates/nvisy-toolkit/src/redaction/store.rs
+++ /dev/null
@@ -1,114 +0,0 @@
-//! In-memory [`Store<V>`] implementation suitable for batch jobs
-//! and tests.
-//!
-//! Backing is an `Arc<RwLock<HashMap>>` so the same store can be
-//! cloned cheaply and shared across multiple anonymizers without
-//! external synchronisation.
-
-use std::collections::HashMap;
-use std::sync::Arc;
-
-use nvisy_core::Result;
-use nvisy_core::redaction::Store;
-use tokio::sync::RwLock;
-
-/// In-memory implementation of [`Store<V>`].
-///
-/// Cloning is cheap — every clone shares the same backing map. Use
-/// [`Self::len`] / [`Self::is_empty`] for inspection; iterate keys
-/// via [`Self::tokens`] when you need to drain.
-#[derive(Debug)]
-pub struct MemoryStore<V> {
-    inner: Arc<RwLock<HashMap<String, V>>>,
-}
-
-impl<V> MemoryStore<V> {
-    /// Build an empty store.
-    pub fn new() -> Self {
-        Self::default()
-    }
-
-    /// Number of token → value mappings currently held.
-    pub async fn len(&self) -> usize {
-        self.inner.read().await.len()
-    }
-
-    /// `true` if no mappings have been recorded yet.
-    pub async fn is_empty(&self) -> bool {
-        self.inner.read().await.is_empty()
-    }
-
-    /// Snapshot the token list. Useful for tests; not a stable order.
-    pub async fn tokens(&self) -> Vec<String> {
-        self.inner.read().await.keys().cloned().collect()
-    }
-}
-
-impl<V> Default for MemoryStore<V> {
-    fn default() -> Self {
-        Self {
-            inner: Arc::new(RwLock::new(HashMap::new())),
-        }
-    }
-}
-
-impl<V> Clone for MemoryStore<V> {
-    fn clone(&self) -> Self {
-        Self {
-            inner: Arc::clone(&self.inner),
-        }
-    }
-}
-
-#[async_trait::async_trait]
-impl<V> Store<V> for MemoryStore<V>
-where
-    V: Clone + Send + Sync,
-{
-    async fn put(&self, token: &str, value: V) -> Result<()> {
-        self.inner.write().await.insert(token.to_owned(), value);
-        Ok(())
-    }
-
-    async fn get(&self, token: &str) -> Result<Option<V>> {
-        Ok(self.inner.read().await.get(token).cloned())
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use nvisy_core::modality::TextData;
-
-    use super::*;
-
-    fn text_of(d: Option<TextData>) -> Option<String> {
-        d.map(|d| d.text.as_str().to_owned())
-    }
-
-    #[tokio::test]
-    async fn put_then_get_round_trips() {
-        let store: MemoryStore<TextData> = MemoryStore::new();
-        store.put("tok", TextData::new("alice")).await.unwrap();
-        assert_eq!(
-            text_of(store.get("tok").await.unwrap()),
-            Some("alice".to_owned())
-        );
-    }
-
-    #[tokio::test]
-    async fn unknown_token_returns_none() {
-        let store: MemoryStore<TextData> = MemoryStore::new();
-        assert!(store.get("missing").await.unwrap().is_none());
-    }
-
-    #[tokio::test]
-    async fn put_overwrites_existing() {
-        let store: MemoryStore<TextData> = MemoryStore::new();
-        store.put("tok", TextData::new("alice")).await.unwrap();
-        store.put("tok", TextData::new("bob")).await.unwrap();
-        assert_eq!(
-            text_of(store.get("tok").await.unwrap()),
-            Some("bob".to_owned())
-        );
-    }
-}
diff --git a/crates/nvisy-toolkit/src/validation/check.rs b/crates/nvisy-toolkit/src/validation/check.rs
deleted file mode 100644
index 10b42567d..000000000
--- a/crates/nvisy-toolkit/src/validation/check.rs
+++ /dev/null
@@ -1,152 +0,0 @@
-//! Abstract [`Check`] trait + [`CheckContext`] + [`Finding`] /
-//! [`Severity`] types every concrete check produces.
-
-use std::marker::PhantomData;
-
-use nvisy_core::entity::Entity;
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::Modality;
-use schemars::JsonSchema;
-use serde::{Deserialize, Serialize};
-use uuid::Uuid;
-
-/// Read-only context every [`Check::check`] call receives.
-///
-/// Built once per call by the caller and passed to each check in the
-/// pipeline. `P` is the resolver type, mirroring the dedup
-/// [`LayerContext`]. Generic so the resolver call
-/// (`ctx.resolver.text_at(...)`) is monomorphised. Object safety on
-/// [`Check<M, P>`] still holds — `P` is a type parameter, not a
-/// generic method.
-///
-/// [`LayerContext`]: crate::deduplication::LayerContext
-pub struct CheckContext<'a, M, P>
-where
-    M: Modality,
-    P: TextAt<M> + ?Sized,
-{
-    /// Resolver for "what value sits at this location?"
-    pub resolver: &'a P,
-    /// Concatenated post-redaction output that checks like
-    /// [`LeakCheck`] substring-search against. The caller streams the
-    /// (already-redacted) handle chunks once and hands the assembled
-    /// text in — checks never touch the codec directly.
-    ///
-    /// `None` when the modality doesn't produce searchable text
-    /// (image / audio at present).
-    ///
-    /// [`LeakCheck`]: super::LeakCheck
-    pub redacted_output: Option<&'a str>,
-    /// Optional correlation id used to stitch tracing spans across
-    /// the call.
-    pub correlation_id: Option<Uuid>,
-    _marker: PhantomData<&'a M>,
-}
-
-impl<'a, M, P> CheckContext<'a, M, P>
-where
-    M: Modality,
-    P: TextAt<M> + ?Sized,
-{
-    /// Build a context from the resolver alone. Checks that need the
-    /// post-redaction text attach it via [`with_redacted_output`].
-    ///
-    /// [`with_redacted_output`]: Self::with_redacted_output
-    pub fn new(resolver: &'a P) -> Self {
-        Self {
-            resolver,
-            redacted_output: None,
-            correlation_id: None,
-            _marker: PhantomData,
-        }
-    }
-
-    /// Attach the streamed post-redaction text.
-    pub fn with_redacted_output(mut self, redacted_output: &'a str) -> Self {
-        self.redacted_output = Some(redacted_output);
-        self
-    }
-
-    /// Attach a correlation id (typically a run / detection id).
-    pub fn with_correlation_id(mut self, correlation_id: Uuid) -> Self {
-        self.correlation_id = Some(correlation_id);
-        self
-    }
-}
-
-/// Severity of a single [`Finding`].
-///
-/// `Warn` causes the caller to log the finding and continue. `Fail`
-/// causes the caller to treat the result as a failure.
-#[derive(
-    Debug,
-    Clone,
-    Copy,
-    Default,
-    PartialEq,
-    Eq,
-    Serialize,
-    Deserialize,
-    JsonSchema
-)]
-#[serde(rename_all = "snake_case")]
-pub enum Severity {
-    /// Log + continue.
-    #[default]
-    Warn,
-    /// Log + fail.
-    Fail,
-}
-
-/// Discriminator on what *kind* of issue a [`Finding`] represents.
-///
-/// Each check kind extends this enum with its own variant. The enum
-/// is `#[non_exhaustive]` so adding a new check kind doesn't break
-/// existing match-on-FindingKind callers.
-#[derive(Debug, Clone)]
-#[non_exhaustive]
-pub enum FindingKind {
-    /// A redacted value remained visible in the post-redaction
-    /// output. Produced by [`LeakCheck`].
-    ///
-    /// [`LeakCheck`]: super::LeakCheck
-    Leak {
-        /// The entity whose redacted value was still found in the
-        /// output.
-        entity_id: Uuid,
-        /// The original sensitive value that should have been
-        /// redacted.
-        value: String,
-    },
-    /// Catch-all for future check kinds. Carries an opaque
-    /// human-readable message.
-    Other,
-}
-
-/// One observation emitted by a [`Check`].
-#[derive(Debug, Clone)]
-pub struct Finding {
-    /// Whether this finding should fail the call.
-    pub severity: Severity,
-    /// What kind of issue this finding represents.
-    pub kind: FindingKind,
-    /// Human-readable description of the issue, suitable for the
-    /// tracing message and (on `Severity::Fail`) the error payload.
-    pub message: String,
-}
-
-/// One stage of a validation pipeline.
-///
-/// Each check inspects `entities` and produces a list of findings.
-/// Checks that find nothing return an empty vec; checks that don't
-/// support the modality should not be registered in the first place
-/// (the pipeline simply has no check for that modality).
-#[async_trait::async_trait]
-pub trait Check<M, P>: Send + Sync
-where
-    M: Modality,
-    P: TextAt<M> + ?Sized,
-{
-    /// Inspect `entities` and emit a list of findings.
-    async fn check(&self, entities: &[Entity<M>], ctx: &CheckContext<'_, M, P>) -> Vec<Finding>;
-}
diff --git a/crates/nvisy-toolkit/src/validation/leak/mod.rs b/crates/nvisy-toolkit/src/validation/leak/mod.rs
deleted file mode 100644
index 1645f67db..000000000
--- a/crates/nvisy-toolkit/src/validation/leak/mod.rs
+++ /dev/null
@@ -1,206 +0,0 @@
-//! Substring-based leak detection: post-redaction re-scan.
-//!
-//! Public surface:
-//!
-//! - [`CheckLeaks`] — domain-specific trait. Callers who want just
-//!   leak detection (skipping the abstract [`Check`] machinery) hold
-//!   `&dyn CheckLeaks<M, P>` and call `check_leaks` directly,
-//!   getting back typed [`LeakFinding`]s.
-//! - [`LeakCheck`] — the canonical implementation. Implements both
-//!   [`CheckLeaks<M, P>`] (for direct domain callers) and
-//!   [`Check<M, P>`] (so it slots into a [`CheckPipeline`]). The
-//!   `Check` impl stamps `self.severity` on every emitted leak.
-//! - [`LeakFinding`] — typed output of [`CheckLeaks::check_leaks`].
-//!
-//! Modality coverage today: [`Text`] and [`Tabular`]. Image and
-//! audio leak detection requires visual / audio inspection that the
-//! runtime can't do — see
-//! <https://github.com/nvisycom/runtime/issues/209> (image) and
-//! <https://github.com/nvisycom/runtime/issues/210> (audio).
-//!
-//! [`Check`]: super::Check
-//! [`CheckPipeline`]: super::CheckPipeline
-
-use nvisy_core::entity::Entity;
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::{Modality, Tabular, Text};
-use unicode_normalization::UnicodeNormalization;
-use uuid::Uuid;
-
-use super::check::{Check, CheckContext, Finding, FindingKind, Severity};
-
-/// A single leak detected by [`LeakCheck`].
-///
-/// Domain-typed output of [`CheckLeaks::check_leaks`]. The bridge
-/// `impl Check for LeakCheck` wraps each [`LeakFinding`] into a
-/// generic [`Finding`] carrying the configured severity.
-#[derive(Debug, Clone)]
-pub struct LeakFinding {
-    /// The entity whose redacted value remained visible.
-    pub entity_id: Uuid,
-    /// The original sensitive value found in the redacted output.
-    pub value: String,
-}
-
-/// Domain-specific trait for leak detection.
-///
-/// Implementors return a typed [`LeakFinding`] list (one per
-/// surviving sensitive value). The abstract [`Check`] impl on the
-/// same type translates these into [`Finding`]s with the
-/// implementor's configured [`Severity`].
-///
-/// Today only [`LeakCheck`] implements this; the trait exists so
-/// custom leak strategies (LLM-based scan, n-gram check, …) can plug
-/// in without touching the abstract `Check` glue.
-///
-/// [`Check`]: super::Check
-#[async_trait::async_trait]
-pub trait CheckLeaks<M, P>: Send + Sync
-where
-    M: Modality,
-    P: TextAt<M> + ?Sized,
-{
-    /// Inspect `entities` and emit a typed list of leak findings.
-    async fn check_leaks(
-        &self,
-        entities: &[Entity<M>],
-        ctx: &CheckContext<'_, M, P>,
-    ) -> Vec<LeakFinding>;
-}
-
-/// Canonical [`CheckLeaks`] implementation.
-///
-/// Substring-scans the post-redaction text for each entity's original
-/// value. Folds via NFC + lowercase before matching to be resilient
-/// to Unicode normalization and ASCII case differences.
-pub struct LeakCheck {
-    severity: Severity,
-}
-
-impl LeakCheck {
-    /// Build a leak check whose findings are stamped with `severity`.
-    pub fn new(severity: Severity) -> Self {
-        Self { severity }
-    }
-}
-
-#[async_trait::async_trait]
-impl<P> CheckLeaks<Text, P> for LeakCheck
-where
-    P: TextAt<Text> + ?Sized,
-{
-    async fn check_leaks(
-        &self,
-        entities: &[Entity<Text>],
-        ctx: &CheckContext<'_, Text, P>,
-    ) -> Vec<LeakFinding> {
-        check_text_like::<Text, P>(entities, ctx.resolver, ctx.redacted_output).await
-    }
-}
-
-#[async_trait::async_trait]
-impl<P> CheckLeaks<Tabular, P> for LeakCheck
-where
-    P: TextAt<Tabular> + ?Sized,
-{
-    async fn check_leaks(
-        &self,
-        entities: &[Entity<Tabular>],
-        ctx: &CheckContext<'_, Tabular, P>,
-    ) -> Vec<LeakFinding> {
-        check_text_like::<Tabular, P>(entities, ctx.resolver, ctx.redacted_output).await
-    }
-}
-
-#[async_trait::async_trait]
-impl<M, P> Check<M, P> for LeakCheck
-where
-    M: Modality,
-    P: TextAt<M> + ?Sized,
-    LeakCheck: CheckLeaks<M, P>,
-{
-    async fn check(&self, entities: &[Entity<M>], ctx: &CheckContext<'_, M, P>) -> Vec<Finding> {
-        let leaks = <Self as CheckLeaks<M, P>>::check_leaks(self, entities, ctx).await;
-        leaks
-            .into_iter()
-            .map(|l| {
-                let message = format!(
-                    "redacted value {:?} for entity {} still present in output",
-                    l.value, l.entity_id
-                );
-                Finding {
-                    severity: self.severity,
-                    kind: FindingKind::Leak {
-                        entity_id: l.entity_id,
-                        value: l.value,
-                    },
-                    message,
-                }
-            })
-            .collect()
-    }
-}
-
-async fn check_text_like<M, P>(
-    entities: &[Entity<M>],
-    resolver: &P,
-    redacted_text: Option<&str>,
-) -> Vec<LeakFinding>
-where
-    M: Modality,
-    P: TextAt<M> + ?Sized,
-{
-    let Some(text) = redacted_text else {
-        return Vec::new();
-    };
-
-    let mut leaks = Vec::new();
-    let folded_text = fold_for_match(text);
-    for entity in entities {
-        if let Some(value) = resolver.text_at(&entity.location).await {
-            let folded_value = fold_for_match(&value);
-            if !value.is_empty() && folded_text.contains(&folded_value) {
-                leaks.push(LeakFinding {
-                    entity_id: entity.id,
-                    value,
-                });
-            }
-        }
-    }
-
-    leaks
-}
-
-/// Fold a string for case-insensitive substring matching across
-/// Unicode normalization forms.
-///
-/// Without normalization, the same character sequence written in NFC
-/// vs NFD won't substring-match: e.g. `"café"` as
-/// `[c, a, f, U+00E9]` vs `[c, a, f, e, U+0301]` are distinct byte
-/// sequences. We normalize to NFC, then lowercase. Turkish dotless-i
-/// and other locale-sensitive cases still misfold (`I` → `i` not
-/// `ı`), but that's a deeper Unicode rabbit hole and the redaction
-/// pipeline doesn't currently carry per-document locale info to do
-/// it properly.
-fn fold_for_match(s: &str) -> String {
-    s.nfc().collect::<String>().to_lowercase()
-}
-
-#[cfg(test)]
-mod tests {
-    use super::fold_for_match;
-
-    #[test]
-    fn nfc_normalization_collapses_combining_accents() {
-        let nfc = "caf\u{00e9}";
-        let nfd = "cafe\u{0301}";
-        assert_eq!(fold_for_match(nfc), fold_for_match(nfd));
-    }
-
-    #[test]
-    fn case_fold_substring_match_works_across_normalization() {
-        let haystack = fold_for_match("HELLO CAFE\u{0301}!");
-        let needle = fold_for_match("caf\u{00e9}");
-        assert!(haystack.contains(&needle));
-    }
-}
diff --git a/crates/nvisy-toolkit/src/validation/mod.rs b/crates/nvisy-toolkit/src/validation/mod.rs
deleted file mode 100644
index e49639fc3..000000000
--- a/crates/nvisy-toolkit/src/validation/mod.rs
+++ /dev/null
@@ -1,30 +0,0 @@
-//! Composable post-redaction check pipeline.
-//!
-//! Public surface:
-//!
-//! - [`Check`] / [`CheckContext`] / [`Finding`] / [`FindingKind`] /
-//!   [`Severity`] — the abstract check abstraction shared by every
-//!   validation pass.
-//! - [`CheckPipeline`] — ordered stack of checks, built with
-//!   `new().with_check(...).run(...)`.
-//! - [`CheckLeaks`] / [`LeakCheck`] / [`LeakFinding`] — the canonical
-//!   leak-detection implementation, scoped under the [`leak`]
-//!   submodule. [`LeakCheck`] implements both [`CheckLeaks`] (for
-//!   direct domain callers) and [`Check`] (so it slots into a
-//!   pipeline).
-//!
-//! Each check receives a `&[Entity<M>]` slice (the entities to
-//! verify) plus a [`CheckContext`] carrying the resolver, the
-//! optional post-redaction text, and an optional correlation id.
-//! Callers that hold a richer carrier — `Document<M>` + audit —
-//! adapt by passing the relevant entity slice in.
-//!
-//! [`leak`]: self::leak
-
-mod check;
-pub mod leak;
-mod pipeline;
-
-pub use self::check::{Check, CheckContext, Finding, FindingKind, Severity};
-pub use self::leak::{CheckLeaks, LeakCheck, LeakFinding};
-pub use self::pipeline::CheckPipeline;
diff --git a/crates/nvisy-toolkit/src/validation/pipeline.rs b/crates/nvisy-toolkit/src/validation/pipeline.rs
deleted file mode 100644
index f6cbcca05..000000000
--- a/crates/nvisy-toolkit/src/validation/pipeline.rs
+++ /dev/null
@@ -1,66 +0,0 @@
-//! Ordered stack of [`Check`]s run head-to-tail.
-
-use std::marker::PhantomData;
-
-use nvisy_core::entity::Entity;
-use nvisy_core::extraction::TextAt;
-use nvisy_core::modality::Modality;
-
-use super::check::{Check, CheckContext, Finding};
-
-/// Ordered stack of checks, run head-to-tail against an entity
-/// slice.
-///
-/// Construction is open: callers compose any sequence of
-/// [`Check<M, P>`] impls (built-in or custom).
-///
-/// `P` is the resolver type. Let the compiler infer it via `_` at
-/// the call site.
-pub struct CheckPipeline<M, P>
-where
-    M: Modality,
-    P: TextAt<M> + ?Sized,
-{
-    checks: Vec<Box<dyn Check<M, P>>>,
-    _marker: PhantomData<fn(&M, &P)>,
-}
-
-impl<M, P> CheckPipeline<M, P>
-where
-    M: Modality,
-    P: TextAt<M> + ?Sized,
-{
-    /// Empty pipeline. Use [`Self::with_check`] to append checks.
-    pub fn new() -> Self {
-        Self {
-            checks: Vec::new(),
-            _marker: PhantomData,
-        }
-    }
-
-    /// Append a check.
-    pub fn with_check<C: Check<M, P> + 'static>(mut self, check: C) -> Self {
-        self.checks.push(Box::new(check));
-        self
-    }
-
-    /// Run every check in registration order against `entities` and
-    /// collect their findings.
-    pub async fn run(&self, entities: &[Entity<M>], ctx: &CheckContext<'_, M, P>) -> Vec<Finding> {
-        let mut findings = Vec::new();
-        for check in &self.checks {
-            findings.extend(check.check(entities, ctx).await);
-        }
-        findings
-    }
-}
-
-impl<M, P> Default for CheckPipeline<M, P>
-where
-    M: Modality,
-    P: TextAt<M> + ?Sized,
-{
-    fn default() -> Self {
-        Self::new()
-    }
-}
diff --git a/crates/nvisy-toolkit/testdata/contact.csv b/crates/nvisy-toolkit/testdata/contact.csv
deleted file mode 100644
index 3f7e5da68..000000000
--- a/crates/nvisy-toolkit/testdata/contact.csv
+++ /dev/null
@@ -1,3 +0,0 @@
-name,email,phone,card,iban,ssn,host
-Alice Johnson,alice.johnson@example.com,+1 (415) 555-0142,4111 1111 1111 1111,GB29 NWBK 6016 1331 9268 19,123-45-6789,192.168.1.42
-Bob Smith,bob.smith@example.com,+1 (510) 555-0199,5555 5555 5555 4444,DE89 3704 0044 0532 0130 00,234-56-7890,10.0.0.7
diff --git a/crates/nvisy-toolkit/testdata/contact.html b/crates/nvisy-toolkit/testdata/contact.html
deleted file mode 100644
index 369be280b..000000000
--- a/crates/nvisy-toolkit/testdata/contact.html
+++ /dev/null
@@ -1,26 +0,0 @@
-<!DOCTYPE html>
-<html lang="en">
-<head>
-    <meta charset="UTF-8">
-    <title>Customer onboarding</title>
-</head>
-<body>
-    <!-- Reviewer: Alice Johnson handles bob.smith@example.com inbox -->
-    <h1>Customer onboarding</h1>
-    <p>Reach out to <strong>alice.johnson@example.com</strong> with any questions
-       about the proposal. For urgent matters, call <em>+1 (415) 555-0142</em>
-       and confirm the payment card <code>4111 1111 1111 1111</code> is on file.</p>
-    <p>Wire transfers use IBAN <code>GB29 NWBK 6016 1331 9268 19</code>.
-       The customer's US tax id is <code>123-45-6789</code>. Our staging API
-       runs at <code>192.168.1.42</code>.</p>
-    <p>
-        <img src="/avatars/alice.png"
-             alt="Avatar for alice.johnson@example.com">
-        <a href="mailto:alice.johnson@example.com?subject=Welcome">Email Alice</a> ·
-        <a href="tel:+14155550142" aria-label="Call Alice at +1 (415) 555-0142">Call</a>
-        <input type="text" placeholder="Backup contact e.g. carol.lee@example.com"
-               data-staging-host="192.168.1.42">
-    </p>
-    <p>Best,<br>Bob</p>
-</body>
-</html>
diff --git a/crates/nvisy-toolkit/testdata/contact.json b/crates/nvisy-toolkit/testdata/contact.json
deleted file mode 100644
index c13b8a2ea..000000000
--- a/crates/nvisy-toolkit/testdata/contact.json
+++ /dev/null
@@ -1,23 +0,0 @@
-{
-  "subject": "Customer onboarding",
-  "contacts": [
-    {
-      "name": "Alice Johnson",
-      "email": "alice.johnson@example.com",
-      "phone": "+1 (415) 555-0142",
-      "card": "4111 1111 1111 1111",
-      "iban": "GB29 NWBK 6016 1331 9268 19",
-      "ssn": "123-45-6789",
-      "host": "192.168.1.42"
-    },
-    {
-      "name": "Bob Smith",
-      "email": "bob.smith@example.com",
-      "phone": "+1 (510) 555-0199",
-      "card": "5555 5555 5555 4444",
-      "iban": "DE89 3704 0044 0532 0130 00",
-      "ssn": "234-56-7890",
-      "host": "10.0.0.7"
-    }
-  ]
-}
diff --git a/crates/nvisy-toolkit/testdata/contact.txt b/crates/nvisy-toolkit/testdata/contact.txt
deleted file mode 100644
index de97c5394..000000000
--- a/crates/nvisy-toolkit/testdata/contact.txt
+++ /dev/null
@@ -1,14 +0,0 @@
-Subject: Customer onboarding
-
-Hi team,
-
-Please reach out to alice.johnson@example.com if you have questions
-about the proposal. For urgent matters, call me at +1 (415) 555-0142
-and confirm the payment card 4111 1111 1111 1111 is on file.
-
-For wire transfers, use IBAN GB29 NWBK 6016 1331 9268 19. The customer's
-US tax id (SSN) is 123-45-6789. Our staging API runs at 192.168.1.42 if
-you need to verify the flow end-to-end.
-
-Best,
-Bob
diff --git a/crates/nvisy-toolkit/tests/codec_e2e_csv.rs b/crates/nvisy-toolkit/tests/codec_e2e_csv.rs
deleted file mode 100644
index a39429aca..000000000
--- a/crates/nvisy-toolkit/tests/codec_e2e_csv.rs
+++ /dev/null
@@ -1,138 +0,0 @@
-//! End-to-end CSV codec. On top of the shared tabular pipeline
-//! shape (`decode → detect-per-cell → dedup → redact + encode`),
-//! this test exercises:
-//!
-//! 1. Cell-coordinate lifting: each entity's location must point at
-//!    the right `(row, col)` plus an intra-cell byte range
-//!    (`Handler::lift_chunk` on the CSV handler).
-//! 2. Partial-cell redaction: `Mask::stars()` over a payment card
-//!    cell rewrites only the card digits, not surrounding bytes.
-//! 3. `Anonymizer<Tabular>` impls on built-in [`Replace`] and
-//!    [`Mask`] operators.
-//!
-//! [`Replace`]: nvisy_toolkit::redaction::anonymizer::Replace
-//! [`Mask`]: nvisy_toolkit::redaction::anonymizer::Mask
-
-mod fixtures;
-
-use nvisy_core::entity::builtins;
-
-use crate::fixtures::{Fixture, assert_pii_removed, assert_tabular_entity, assert_tokens_present};
-
-const FIXTURE: Fixture = Fixture {
-    path: concat!(env!("CARGO_MANIFEST_DIR"), "/testdata/contact.csv"),
-    source: include_str!("../testdata/contact.csv"),
-    extension: "csv",
-};
-
-#[tokio::test]
-async fn csv_codec_detects_and_redacts() {
-    let outcome = FIXTURE.run_tabular_pipeline().await;
-
-    // Header row 0 is name,email,phone,card,iban,ssn,host;
-    // data rows are 1 (Alice) and 2 (Bob).
-    for (kind, row, col, cell) in [
-        (
-            builtins::EMAIL_ADDRESS.label_ref(),
-            1,
-            1,
-            "alice.johnson@example.com",
-        ),
-        (
-            builtins::EMAIL_ADDRESS.label_ref(),
-            2,
-            1,
-            "bob.smith@example.com",
-        ),
-        (
-            builtins::PHONE_NUMBER.label_ref(),
-            1,
-            2,
-            "+1 (415) 555-0142",
-        ),
-        (
-            builtins::PHONE_NUMBER.label_ref(),
-            2,
-            2,
-            "+1 (510) 555-0199",
-        ),
-        (
-            builtins::PAYMENT_CARD.label_ref(),
-            1,
-            3,
-            "4111 1111 1111 1111",
-        ),
-        (
-            builtins::PAYMENT_CARD.label_ref(),
-            2,
-            3,
-            "5555 5555 5555 4444",
-        ),
-        (
-            builtins::IBAN.label_ref(),
-            1,
-            4,
-            "GB29 NWBK 6016 1331 9268 19",
-        ),
-        (
-            builtins::IBAN.label_ref(),
-            2,
-            4,
-            "DE89 3704 0044 0532 0130 00",
-        ),
-        (builtins::GOVERNMENT_ID.label_ref(), 1, 5, "123-45-6789"),
-        (builtins::GOVERNMENT_ID.label_ref(), 2, 5, "234-56-7890"),
-        (builtins::IP_ADDRESS.label_ref(), 1, 6, "192.168.1.42"),
-        (builtins::IP_ADDRESS.label_ref(), 2, 6, "10.0.0.7"),
-    ] {
-        assert_tabular_entity(cell, &outcome.entities, kind, row, col, cell);
-    }
-
-    assert_pii_removed(
-        &outcome.redacted,
-        &[
-            "alice.johnson@example.com",
-            "bob.smith@example.com",
-            "+1 (415) 555-0142",
-            "+1 (510) 555-0199",
-            "4111 1111 1111 1111",
-            "5555 5555 5555 4444",
-            "GB29 NWBK 6016 1331 9268 19",
-            "DE89 3704 0044 0532 0130 00",
-            "123-45-6789",
-            "234-56-7890",
-            "192.168.1.42",
-            "10.0.0.7",
-        ],
-    );
-    assert_tokens_present(
-        &outcome.redacted,
-        &[
-            "[email_address]",
-            "[phone_number]",
-            "[iban]",
-            "[government_id]",
-            "[ip_address]",
-        ],
-    );
-
-    // CSV structure preserved: header row and non-sensitive cells
-    // pass through verbatim.
-    assert!(
-        outcome
-            .redacted
-            .starts_with("name,email,phone,card,iban,ssn,host"),
-        "header row lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("Alice Johnson,"),
-        "alice name lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("Bob Smith,"),
-        "bob name lost: {}",
-        outcome.redacted,
-    );
-}
diff --git a/crates/nvisy-toolkit/tests/codec_e2e_html.rs b/crates/nvisy-toolkit/tests/codec_e2e_html.rs
deleted file mode 100644
index 094d44410..000000000
--- a/crates/nvisy-toolkit/tests/codec_e2e_html.rs
+++ /dev/null
@@ -1,155 +0,0 @@
-//! End-to-end HTML codec. The HTML loader emits a redactable-item
-//! stream: text nodes, every element attribute, and HTML comments.
-//! `mailto:` / `tel:` URLs are scanned in place — the recognizer
-//! matches the PII substring inside the attribute value, redact
-//! rewrites only that range, and the scheme prefix + query suffix
-//! pass through untouched. This test exercises:
-//!
-//! 1. Detection inside per-tag text nodes.
-//! 2. Detection inside attributes (`alt`, `aria-label`,
-//!    `placeholder`, `data-*`).
-//! 3. Detection inside HTML comments.
-//! 4. PII matching inside URL attribute values without explicit
-//!    scheme parsing.
-//! 5. Tag preservation through encode.
-//!
-//! Note: entity locations on HTML are item-stream coords (relative
-//! to the concatenated item stream), not raw HTML source byte
-//! offsets, so this test asserts by entity-kind presence rather
-//! than slicing the source at each entity's range. See
-//! `.ignore/html-handler-improvements.md`.
-
-mod fixtures;
-
-use nvisy_core::entity::builtins;
-
-use crate::fixtures::{Fixture, assert_pii_removed, assert_tokens_present};
-
-const FIXTURE: Fixture = Fixture {
-    path: concat!(env!("CARGO_MANIFEST_DIR"), "/testdata/contact.html"),
-    source: include_str!("../testdata/contact.html"),
-    extension: "html",
-};
-
-#[tokio::test]
-async fn html_codec_detects_and_redacts() {
-    let outcome = FIXTURE.run_text_pipeline().await;
-
-    // Detection covered every PII kind in the fixture. Locations
-    // index the text-node stream rather than the raw source, so the
-    // per-needle slice assertion the txt/json tests use doesn't
-    // translate here; presence-by-kind is the right shape.
-    for expected in [
-        builtins::EMAIL_ADDRESS.label_ref(),
-        builtins::PHONE_NUMBER.label_ref(),
-        builtins::PAYMENT_CARD.label_ref(),
-        builtins::IBAN.label_ref(),
-        builtins::GOVERNMENT_ID.label_ref(),
-        builtins::IP_ADDRESS.label_ref(),
-    ] {
-        assert!(
-            outcome.entities.iter().any(|e| e.label == expected),
-            "expected at least one {expected:?} entity; got: {:?}",
-            outcome
-                .entities
-                .iter()
-                .map(|e| e.label.clone())
-                .collect::<Vec<_>>()
-        );
-    }
-
-    // Cover every PII string in the fixture — text-node, attribute,
-    // comment, and URL-body locations. If any of these survive,
-    // it's a regression somewhere in the redactable-item pipeline.
-    assert_pii_removed(
-        &outcome.redacted,
-        &[
-            // Text-node values.
-            "alice.johnson@example.com",
-            "+1 (415) 555-0142",
-            "4111 1111 1111 1111",
-            "GB29 NWBK 6016 1331 9268 19",
-            "123-45-6789",
-            "192.168.1.42",
-            // Attribute values (alt, aria-label, placeholder, data-*).
-            "Avatar for alice.johnson@example.com",
-            "Call Alice at +1 (415) 555-0142",
-            "Backup contact e.g. carol.lee@example.com",
-            // Comment body.
-            "bob.smith@example.com",
-            // URL bodies (mailto: + tel:).
-            "mailto:alice.johnson@example.com",
-            "tel:+14155550142",
-        ],
-    );
-    assert_tokens_present(
-        &outcome.redacted,
-        &[
-            "[email_address]",
-            "[phone_number]",
-            "[iban]",
-            "[government_id]",
-            "[ip_address]",
-        ],
-    );
-
-    // HTML structure preserved. The encoder rebuilds the document
-    // from the parsed tree, so attribute quoting and whitespace
-    // around tags may be normalised — assert presence of the
-    // structural elements rather than verbatim byte equality.
-    assert!(
-        outcome.redacted.contains("<!DOCTYPE html>"),
-        "doctype lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome
-            .redacted
-            .contains("<title>Customer onboarding</title>"),
-        "title lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("<h1>Customer onboarding</h1>"),
-        "h1 lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("<strong>"),
-        "strong tag lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("<em>"),
-        "em tag lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("<code>"),
-        "code tag lost: {}",
-        outcome.redacted,
-    );
-    // URL `scheme:` prefixes survive verbatim because attribute
-    // values pass through and the recognizer matches only the PII
-    // substring (`alice@example.com` inside `mailto:...`).
-    assert!(
-        outcome.redacted.contains("href=\"mailto:"),
-        "mailto: scheme prefix lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("?subject=Welcome"),
-        "mailto: query suffix lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("href=\"tel:"),
-        "tel: scheme prefix lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("<!--"),
-        "comment delimiter lost: {}",
-        outcome.redacted,
-    );
-}
diff --git a/crates/nvisy-toolkit/tests/codec_e2e_json.rs b/crates/nvisy-toolkit/tests/codec_e2e_json.rs
deleted file mode 100644
index 702d3f31a..000000000
--- a/crates/nvisy-toolkit/tests/codec_e2e_json.rs
+++ /dev/null
@@ -1,111 +0,0 @@
-//! End-to-end JSON codec. Exercises the JSON-specific guarantees
-//! on top of the shared text pipeline:
-//!
-//! 1. Lifted offsets index the raw JSON source — the quoted form
-//!    plus its escape table — so redaction edits the right bytes.
-//! 2. `Handler::encode` preserves the original indentation, key
-//!    order, and whitespace verbatim for every slot the redaction
-//!    didn't touch (the slot model in `json_handler`).
-
-mod fixtures;
-
-use nvisy_core::entity::builtins;
-
-use crate::fixtures::{Fixture, assert_pii_removed, assert_text_entity, assert_tokens_present};
-
-const FIXTURE: Fixture = Fixture {
-    path: concat!(env!("CARGO_MANIFEST_DIR"), "/testdata/contact.json"),
-    source: include_str!("../testdata/contact.json"),
-    extension: "json",
-};
-
-#[tokio::test]
-async fn json_codec_detects_and_redacts() {
-    let outcome = FIXTURE.run_text_pipeline().await;
-
-    // Lifted offsets index the raw JSON source — quoted form plus
-    // its escape table — so slicing the fixture by an entity's
-    // location yields the same bytes the recognizer matched.
-    for (kind, needle) in [
-        (
-            builtins::EMAIL_ADDRESS.label_ref(),
-            "alice.johnson@example.com",
-        ),
-        (builtins::EMAIL_ADDRESS.label_ref(), "bob.smith@example.com"),
-        (builtins::PHONE_NUMBER.label_ref(), "+1 (415) 555-0142"),
-        (builtins::PHONE_NUMBER.label_ref(), "+1 (510) 555-0199"),
-        (builtins::PAYMENT_CARD.label_ref(), "4111 1111 1111 1111"),
-        (builtins::PAYMENT_CARD.label_ref(), "5555 5555 5555 4444"),
-        (builtins::IBAN.label_ref(), "GB29 NWBK 6016 1331 9268 19"),
-        (builtins::IBAN.label_ref(), "DE89 3704 0044 0532 0130 00"),
-        (builtins::GOVERNMENT_ID.label_ref(), "123-45-6789"),
-        (builtins::GOVERNMENT_ID.label_ref(), "234-56-7890"),
-        (builtins::IP_ADDRESS.label_ref(), "192.168.1.42"),
-        (builtins::IP_ADDRESS.label_ref(), "10.0.0.7"),
-    ] {
-        assert_text_entity(FIXTURE.source, &outcome.entities, kind, needle);
-    }
-
-    assert_pii_removed(
-        &outcome.redacted,
-        &[
-            "alice.johnson@example.com",
-            "bob.smith@example.com",
-            "+1 (415) 555-0142",
-            "+1 (510) 555-0199",
-            "4111 1111 1111 1111",
-            "5555 5555 5555 4444",
-            "GB29 NWBK 6016 1331 9268 19",
-            "DE89 3704 0044 0532 0130 00",
-            "123-45-6789",
-            "234-56-7890",
-            "192.168.1.42",
-            "10.0.0.7",
-        ],
-    );
-    assert_tokens_present(
-        &outcome.redacted,
-        &[
-            "[email_address]",
-            "[phone_number]",
-            "[iban]",
-            "[government_id]",
-            "[ip_address]",
-        ],
-    );
-
-    // JSON structure preserved: the slot model encodes every
-    // untouched span byte-for-byte, so indentation, key order, and
-    // surrounding punctuation survive intact.
-    assert!(
-        outcome
-            .redacted
-            .contains("\"subject\": \"Customer onboarding\""),
-        "subject line lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("\"contacts\":"),
-        "contacts key lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("\"name\": \"Alice Johnson\""),
-        "alice name lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("\"name\": \"Bob Smith\""),
-        "bob name lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.ends_with("\n"),
-        "trailing newline lost: {:?}",
-        outcome.redacted,
-    );
-
-    // The output is still well-formed JSON.
-    serde_json::from_str::<serde_json::Value>(&outcome.redacted)
-        .expect("redacted output is valid JSON");
-}
diff --git a/crates/nvisy-toolkit/tests/codec_e2e_txt.rs b/crates/nvisy-toolkit/tests/codec_e2e_txt.rs
deleted file mode 100644
index e3425a0a4..000000000
--- a/crates/nvisy-toolkit/tests/codec_e2e_txt.rs
+++ /dev/null
@@ -1,69 +0,0 @@
-//! End-to-end TXT codec: decode → detect → dedup → redact + encode.
-//!
-//! `Handler::encode` is the canonical reassembly path — the codec
-//! handles line terminators so the test doesn't have to.
-
-mod fixtures;
-
-use nvisy_core::entity::builtins;
-
-use crate::fixtures::{Fixture, assert_pii_removed, assert_text_entity, assert_tokens_present};
-
-const FIXTURE: Fixture = Fixture {
-    path: concat!(env!("CARGO_MANIFEST_DIR"), "/testdata/contact.txt"),
-    source: include_str!("../testdata/contact.txt"),
-    extension: "txt",
-};
-
-#[tokio::test]
-async fn txt_codec_detects_and_redacts() {
-    let outcome = FIXTURE.run_text_pipeline().await;
-
-    for (kind, needle) in [
-        (
-            builtins::EMAIL_ADDRESS.label_ref(),
-            "alice.johnson@example.com",
-        ),
-        (builtins::PHONE_NUMBER.label_ref(), "+1 (415) 555-0142"),
-        (builtins::PAYMENT_CARD.label_ref(), "4111 1111 1111 1111"),
-        (builtins::IBAN.label_ref(), "GB29 NWBK 6016 1331 9268 19"),
-        (builtins::GOVERNMENT_ID.label_ref(), "123-45-6789"),
-        (builtins::IP_ADDRESS.label_ref(), "192.168.1.42"),
-    ] {
-        assert_text_entity(FIXTURE.source, &outcome.entities, kind, needle);
-    }
-
-    assert_pii_removed(
-        &outcome.redacted,
-        &[
-            "alice.johnson@example.com",
-            "+1 (415) 555-0142",
-            "4111 1111 1111 1111",
-            "GB29 NWBK 6016 1331 9268 19",
-            "123-45-6789",
-            "192.168.1.42",
-        ],
-    );
-    assert_tokens_present(
-        &outcome.redacted,
-        &[
-            "[email_address]",
-            "[phone_number]",
-            "[iban]",
-            "[government_id]",
-            "[ip_address]",
-        ],
-    );
-
-    // Non-sensitive structure preserved by the txt codec's encode.
-    assert!(
-        outcome.redacted.starts_with("Subject: Customer onboarding"),
-        "subject line lost: {}",
-        outcome.redacted,
-    );
-    assert!(
-        outcome.redacted.contains("Best,\nBob"),
-        "signature lost: {}",
-        outcome.redacted,
-    );
-}
diff --git a/crates/nvisy-toolkit/tests/fixtures/asserts.rs b/crates/nvisy-toolkit/tests/fixtures/asserts.rs
deleted file mode 100644
index 908ea9139..000000000
--- a/crates/nvisy-toolkit/tests/fixtures/asserts.rs
+++ /dev/null
@@ -1,87 +0,0 @@
-//! Entity-presence and redaction-output assertion helpers for the
-//! codec E2E tests.
-
-use nvisy_core::entity::{Entity, EntityLabelRef, builtins};
-use nvisy_core::modality::{Tabular, Text};
-
-/// Assert at least one `Entity<Text>` of `kind` matches `needle`
-/// when its location is sliced against `source`.
-#[track_caller]
-pub fn assert_text_entity(
-    source: &str,
-    entities: &[Entity<Text>],
-    label: EntityLabelRef,
-    needle: &str,
-) {
-    let hit = entities
-        .iter()
-        .any(|e| e.label == label && &source[e.location.start..e.location.end] == needle);
-    assert!(
-        hit,
-        "expected `{needle}` as {label:?}; got: {:?}",
-        entities
-            .iter()
-            .map(|e| (e.label.clone(), &source[e.location.start..e.location.end]))
-            .collect::<Vec<_>>()
-    );
-}
-
-/// Assert at least one `Entity<Tabular>` of `kind` lives in cell
-/// `(row, col)` and its intra-cell range slices to `needle`.
-#[track_caller]
-pub fn assert_tabular_entity(
-    cell_value: &str,
-    entities: &[Entity<Tabular>],
-    label: EntityLabelRef,
-    row: u32,
-    col: u32,
-    needle: &str,
-) {
-    let hit = entities.iter().any(|e| {
-        if e.label != label {
-            return false;
-        }
-        if e.location.row_index != row || e.location.column_index != col {
-            return false;
-        }
-        let start = e.location.start_offset.unwrap_or(0);
-        let end = e.location.end_offset.unwrap_or(cell_value.len());
-        cell_value.get(start..end) == Some(needle)
-    });
-    assert!(
-        hit,
-        "expected `{needle}` as {label:?} at ({row},{col}); got: {:?}",
-        entities
-            .iter()
-            .map(|e| (
-                e.label.clone(),
-                e.location.row_index,
-                e.location.column_index
-            ))
-            .collect::<Vec<_>>()
-    );
-}
-
-/// Assert every `needle` in `pii` is absent from `redacted` —
-/// detection-and-redact removed each sensitive substring end-to-end.
-#[track_caller]
-pub fn assert_pii_removed(redacted: &str, pii: &[&str]) {
-    for needle in pii {
-        assert!(
-            !redacted.contains(needle),
-            "PII `{needle}` survived redaction: {redacted}"
-        );
-    }
-}
-
-/// Assert every replacement `token` is present in `redacted` —
-/// the operator fired and wrote its substitution.
-#[track_caller]
-pub fn assert_tokens_present(redacted: &str, tokens: &[&str]) {
-    for token in tokens {
-        assert!(
-            redacted.contains(token),
-            "replacement token `{token}` missing from output: {redacted}"
-        );
-    }
-}
diff --git a/crates/nvisy-toolkit/tests/fixtures/mod.rs b/crates/nvisy-toolkit/tests/fixtures/mod.rs
deleted file mode 100644
index 818a74f8f..000000000
--- a/crates/nvisy-toolkit/tests/fixtures/mod.rs
+++ /dev/null
@@ -1,36 +0,0 @@
-//! Shared test fixtures for the toolkit's codec E2E tests
-//! (`codec_e2e_txt.rs`, `codec_e2e_json.rs`, `codec_e2e_csv.rs`,
-//! `codec_e2e_html.rs`).
-//!
-//! Each integration test declares `mod fixtures;` and pulls the
-//! pieces it needs from the re-exports below. The boilerplate of
-//! building the recognizer + redaction registries, driving the
-//! four-stage `decode → detect → dedup → redact + encode`
-//! pipeline, and asserting on the result lives here so the
-//! per-codec tests stay focused on entity-kind coverage and
-//! codec-specific structural checks.
-//!
-//! Submodules:
-//!
-//! - `pipeline` — the [`Fixture`] descriptor, the per-modality
-//!   pipeline drivers, the [`PipelineOutcome`] return type, and
-//!   the `*.redacted.*` artifact writer.
-//! - `registries` — the shipped recognizer + redaction registry +
-//!   dedup params every test uses.
-//! - `asserts` — entity-presence and PII-removed/token-present
-//!   helpers.
-
-// Each integration-test file pulls a different subset of the
-// re-exports below, so the un-used ones look "dead" from any single
-// test's perspective. Quiet both lints across the whole module.
-#![allow(dead_code, unused_imports)]
-
-mod asserts;
-mod pipeline;
-mod registries;
-
-pub use self::asserts::{
-    assert_pii_removed, assert_tabular_entity, assert_text_entity, assert_tokens_present,
-};
-pub use self::pipeline::{Fixture, PipelineOutcome};
-pub use self::registries::{dedup_params, redaction_registry, shipped_recognizer};
diff --git a/crates/nvisy-toolkit/tests/fixtures/pipeline.rs b/crates/nvisy-toolkit/tests/fixtures/pipeline.rs
deleted file mode 100644
index 3004737cc..000000000
--- a/crates/nvisy-toolkit/tests/fixtures/pipeline.rs
+++ /dev/null
@@ -1,160 +0,0 @@
-//! [`Fixture`] descriptor + the per-modality pipeline drivers every
-//! codec E2E test runs as methods on the fixture.
-//!
-//! Each driver wires the shared recognizer + redaction registries
-//! into a `decode → detect → dedup → redact + encode` flow and
-//! writes the redacted output next to the fixture as
-//! `{stem}.redacted.{ext}` for human inspection.
-
-use std::path::Path;
-use std::str::from_utf8;
-
-use nvisy_codec::{CodecRegistry, DocumentHandle};
-use nvisy_core::Result;
-use nvisy_core::entity::Entity;
-use nvisy_core::modality::{Modality, Tabular, Text};
-use nvisy_core::redaction::{RedactAt, Redactions};
-use nvisy_toolkit::deduplication::{LayerContext, LayerPipeline};
-use nvisy_toolkit::detection::{RecognizerRegistry, RecognizerRegistryExt};
-
-use super::registries::{dedup_params, redaction_registry, shipped_recognizer};
-
-/// Fixture file the codec E2E tests load. `path` is the absolute
-/// disk path the artifact-writer drops `{stem}.redacted.{ext}`
-/// next to; `source` is the compile-time-inlined fixture body;
-/// `extension` is the codec hint the registry resolves on.
-pub struct Fixture {
-    /// Absolute path to the fixture on disk; used by the artifact
-    /// writer to derive the `{stem}.redacted.{ext}` output path.
-    pub path: &'static str,
-    /// Fixture body (the bytes the codec decodes).
-    pub source: &'static str,
-    /// Extension hint the codec registry resolves on.
-    pub extension: &'static str,
-}
-
-impl Fixture {
-    /// Decode this fixture into a typed text [`DocumentHandle`].
-    pub async fn decode_text(&self) -> Result<DocumentHandle<Text>> {
-        let registry = CodecRegistry::with_builtin();
-        let untyped = registry.decode(self.source, self.extension).await?;
-        Ok(untyped
-            .into_text()
-            .expect("text-modality extension resolves to text handle"))
-    }
-
-    /// Decode this fixture into a typed tabular [`DocumentHandle`].
-    pub async fn decode_tabular(&self) -> Result<DocumentHandle<Tabular>> {
-        let registry = CodecRegistry::with_builtin();
-        let untyped = registry.decode(self.source, self.extension).await?;
-        Ok(untyped
-            .into_tabular()
-            .expect("tabular-modality extension resolves to tabular handle"))
-    }
-
-    /// Run the full text pipeline: decode → detect-per-chunk →
-    /// dedup → redact + encode → write `*.redacted.*` artifact.
-    pub async fn run_text_pipeline(&self) -> PipelineOutcome<Text> {
-        let mut buffer = self.decode_text().await.expect("text fixture decodes");
-
-        let recognizer_registry = RecognizerRegistry::new().with_recognizer(shipped_recognizer());
-        let detected = recognizer_registry
-            .detect(buffer.handler_mut())
-            .await
-            .expect("text detect succeeds");
-
-        let ctx = LayerContext::<Text, _>::new(&buffer);
-        let entities = LayerPipeline::<Text, _>::from_params(&dedup_params())
-            .expect("pipeline builds")
-            .run(detected, &ctx)
-            .await;
-
-        let redactions: Redactions<Text> = redaction_registry::<Text>()
-            .apply_all(entities.iter(), &buffer)
-            .await
-            .expect("text redactions apply");
-        buffer
-            .redact_at(redactions)
-            .await
-            .expect("text redact succeeds");
-
-        let encoded = buffer.handler().encode().expect("text encode succeeds");
-        let redacted = from_utf8(encoded.as_bytes())
-            .expect("text codec encode produces UTF-8")
-            .to_owned();
-
-        self.write_redacted_artifact(&redacted);
-        PipelineOutcome { entities, redacted }
-    }
-
-    /// Run the full tabular pipeline. Same shape as
-    /// [`run_text_pipeline`] but for `M = Tabular`.
-    ///
-    /// [`run_text_pipeline`]: Self::run_text_pipeline
-    pub async fn run_tabular_pipeline(&self) -> PipelineOutcome<Tabular> {
-        let mut buffer = self
-            .decode_tabular()
-            .await
-            .expect("tabular fixture decodes");
-
-        let recognizer_registry = RecognizerRegistry::new().with_recognizer(shipped_recognizer());
-        let detected = recognizer_registry
-            .detect(buffer.handler_mut())
-            .await
-            .expect("tabular detect succeeds");
-
-        let ctx = LayerContext::<Tabular, _>::new(&buffer);
-        let entities = LayerPipeline::<Tabular, _>::from_params(&dedup_params())
-            .expect("pipeline builds")
-            .run(detected, &ctx)
-            .await;
-
-        let redactions: Redactions<Tabular> = redaction_registry::<Tabular>()
-            .apply_all(entities.iter(), &buffer)
-            .await
-            .expect("tabular redactions apply");
-        buffer
-            .redact_at(redactions)
-            .await
-            .expect("tabular redact succeeds");
-
-        let encoded = buffer.handler().encode().expect("tabular encode succeeds");
-        let redacted = from_utf8(encoded.as_bytes())
-            .expect("tabular codec encode produces UTF-8")
-            .to_owned();
-
-        self.write_redacted_artifact(&redacted);
-        PipelineOutcome { entities, redacted }
-    }
-
-    /// Write `redacted` next to the fixture as `{stem}.redacted.{ext}`
-    /// for human inspection. Gitignored under
-    /// `**/testdata/**/*.redacted.*`.
-    fn write_redacted_artifact(&self, redacted: &str) {
-        let path = Path::new(self.path);
-        let stem = path
-            .file_stem()
-            .and_then(|s| s.to_str())
-            .expect("fixture has a UTF-8 stem");
-        let ext = path
-            .extension()
-            .and_then(|s| s.to_str())
-            .expect("fixture has a UTF-8 extension");
-        let parent = path.parent().expect("fixture has a parent");
-        let out = parent.join(format!("{stem}.redacted.{ext}"));
-        std::fs::write(&out, redacted).unwrap_or_else(|e| {
-            panic!("write redacted artifact {}: {e}", out.display());
-        });
-    }
-}
-
-/// Outcome of one end-to-end pipeline run. Tests pull both the
-/// entity list (to assert detection coverage) and the encoded
-/// redacted body (to assert structural preservation + token
-/// presence).
-pub struct PipelineOutcome<M: Modality> {
-    /// Entities surviving dedup (the input to redact).
-    pub entities: Vec<Entity<M>>,
-    /// UTF-8 encoded redacted output.
-    pub redacted: String,
-}
diff --git a/crates/nvisy-toolkit/tests/fixtures/registries.rs b/crates/nvisy-toolkit/tests/fixtures/registries.rs
deleted file mode 100644
index 2056a1d02..000000000
--- a/crates/nvisy-toolkit/tests/fixtures/registries.rs
+++ /dev/null
@@ -1,63 +0,0 @@
-//! Shared recognizer + redaction registry constructors and dedup
-//! params used by every codec E2E test.
-
-use nvisy_context::ContextEnhanced;
-use nvisy_core::entity::builtins;
-use nvisy_core::modality::Modality;
-use nvisy_core::primitive::ConfidenceThreshold;
-use nvisy_pattern::PatternRecognizer;
-use nvisy_toolkit::deduplication::LayerParams;
-use nvisy_toolkit::redaction::anonymizer::{Mask, Replace};
-use nvisy_toolkit::redaction::{Anonymizer, RedactionRegistry};
-
-/// Build the shipped pattern recognizer from every built-in
-/// pattern + dictionary, wrapped in its [`ContextEnhanced`] layer.
-pub fn shipped_recognizer() -> ContextEnhanced<PatternRecognizer> {
-    PatternRecognizer::builder()
-        .with_builtin_patterns()
-        .with_builtin_dictionaries()
-        .build_context_enhanced()
-        .expect("shipped recognizer builds")
-}
-
-/// Per-test redaction registry: every entity kind the shipped
-/// patterns can emit is mapped to a deterministic operator so test
-/// assertions can spot-check the replacement tokens.
-///
-/// - emails, phones, IBANs, government ids, IPs → `[{label}]`
-/// - payment cards → `Mask::stars()` (digits masked, no token)
-pub fn redaction_registry<M>() -> RedactionRegistry<M>
-where
-    M: Modality,
-    Replace: Anonymizer<M>,
-    Mask: Anonymizer<M>,
-{
-    RedactionRegistry::<M>::new()
-        .insert_label(
-            builtins::EMAIL_ADDRESS.label_ref(),
-            Replace::new("[{label}]"),
-        )
-        .insert_label(
-            builtins::PHONE_NUMBER.label_ref(),
-            Replace::new("[{label}]"),
-        )
-        .insert_label(builtins::IBAN.label_ref(), Replace::new("[{label}]"))
-        .insert_label(
-            builtins::GOVERNMENT_ID.label_ref(),
-            Replace::new("[{label}]"),
-        )
-        .insert_label(builtins::IP_ADDRESS.label_ref(), Replace::new("[{label}]"))
-        .insert_label(builtins::PAYMENT_CARD.label_ref(), Mask::stars())
-}
-
-/// Standard dedup params: a `0.35` confidence threshold sized
-/// for our shipped patterns' baseline (most regex-only matches
-/// land in 0.1–0.5 before context boost); a tighter threshold
-/// would drop legitimate weak-pattern matches the context layer
-/// is expected to lift.
-pub fn dedup_params() -> LayerParams {
-    LayerParams {
-        confidence_threshold: Some(ConfidenceThreshold::new(0.35).unwrap()),
-        ..LayerParams::default()
-    }
-}
diff --git a/crates/nvisy-toolkit/tests/recognition_registry.rs b/crates/nvisy-toolkit/tests/recognition_registry.rs
deleted file mode 100644
index 7f2c1332a..000000000
--- a/crates/nvisy-toolkit/tests/recognition_registry.rs
+++ /dev/null
@@ -1,116 +0,0 @@
-//! Integration test: `RecognizerRegistry` dispatching pattern + NER
-//! + LLM recognizers in one pass.
-//!
-//! Wires every text-modality recognizer the toolkit ships with against
-//! one input string and asserts each one fires. The NER and LLM legs
-//! call live services (Bento ner-bento + Ollama) so the test is
-//! `#[ignore]`d by default; run locally with:
-//!
-//! ```text
-//! cargo test -p nvisy-toolkit --features bento -- --ignored
-//! ```
-//!
-//! Required services:
-//!
-//! - `ner-bento` listening at `$NVISY_BENTO_URL` (default
-//!   `http://localhost:3000`). See nvisycom/inference.
-//! - Ollama listening at `$NVISY_OLLAMA_URL` (default
-//!   `http://localhost:11434`) with `$NVISY_OLLAMA_MODEL` (default
-//!   `llama3.1:8b`) pulled.
-
-#![cfg(feature = "bento")]
-
-use std::env;
-
-use nvisy_core::entity::{Entity, EntityLabelCatalog};
-use nvisy_core::modality::{Text, TextData};
-use nvisy_core::recognition::RecognizerInput;
-use nvisy_llm::backend::rig::RigBackend;
-use nvisy_llm::provider::LlmProvider;
-use nvisy_llm::{DefaultPrompt, LlmRecognizer};
-use nvisy_ner::NerRecognizer;
-use nvisy_ner::backend::{BentoBackend, BentoParams};
-use nvisy_pattern::PatternRecognizer;
-use nvisy_toolkit::detection::RecognizerRegistry;
-
-/// Sample text that triggers all three recognizers:
-/// - `pattern` catches the email (built-in email regex).
-/// - `ner` (zero-shot Bento) catches the person name + organisation.
-/// - `llm` (Ollama) is asked the same question and picks up at least
-///   one of the above.
-const SAMPLE: &str = "Alice Carter (alice.carter@acme.test) joined Acme Corp last week.";
-
-fn env_or(key: &str, default: &str) -> String {
-    env::var(key).unwrap_or_else(|_| default.to_owned())
-}
-
-fn build_registry() -> RecognizerRegistry {
-    let pattern = PatternRecognizer::builder()
-        .with_builtin_patterns()
-        .with_builtin_dictionaries()
-        .build_context_enhanced()
-        .expect("pattern recognizer builds from builtin set");
-
-    let bento_url = env_or("NVISY_BENTO_URL", "http://localhost:3000");
-    let bento_backend = BentoBackend::new(BentoParams::new(bento_url)).expect("bento backend init");
-    let ner = NerRecognizer::builder()
-        .with_name("ner")
-        .with_engine(bento_backend)
-        .with_supported_labels(
-            EntityLabelCatalog::with_builtins()
-                .iter()
-                .map(|l| l.label_ref())
-                .collect::<Vec<_>>(),
-        )
-        .build()
-        .expect("ner recognizer builds");
-
-    let ollama_model = env_or("NVISY_OLLAMA_MODEL", "llama3.1:8b");
-    let ollama_url = env_or("NVISY_OLLAMA_URL", "http://localhost:11434");
-    let rig = RigBackend::builder()
-        .with_provider(LlmProvider::ollama_with_url(&ollama_model, &ollama_url))
-        .build()
-        .expect("rig backend builds for Ollama");
-    let llm = LlmRecognizer::builder()
-        .with_name("llm")
-        .with_backend(rig)
-        .with_prompt(DefaultPrompt)
-        .build()
-        .expect("llm recognizer builds");
-
-    RecognizerRegistry::new()
-        .with_recognizer::<Text>(pattern)
-        .with_recognizer::<Text>(ner)
-        .with_recognizer::<Text>(llm)
-}
-
-fn fired(entities: &[Entity<Text>], source: &str) -> bool {
-    entities
-        .iter()
-        .any(|e| e.trail.iter().any(|s| s.source == source))
-}
-
-#[tokio::test]
-#[ignore = "requires live ner-bento + ollama; run with `--ignored`"]
-async fn registry_dispatches_pattern_ner_and_llm_against_live_services() {
-    let registry = build_registry();
-    let input = RecognizerInput::new(TextData::new(SAMPLE.to_owned()));
-
-    let entities = registry
-        .run::<Text>(input)
-        .await
-        .expect("dispatch over pattern + ner + llm succeeds");
-
-    assert!(
-        fired(&entities, "pattern"),
-        "pattern recognizer did not fire on `{SAMPLE}`; got {entities:?}"
-    );
-    assert!(
-        fired(&entities, "ner"),
-        "ner (bento) recognizer did not fire on `{SAMPLE}`; got {entities:?}"
-    );
-    assert!(
-        fired(&entities, "llm-ner"),
-        "llm (ollama) recognizer did not fire on `{SAMPLE}`; got {entities:?}"
-    );
-}
diff --git a/deny.toml b/deny.toml
index d7691e420..e726b7fbf 100644
--- a/deny.toml
+++ b/deny.toml
@@ -23,6 +23,7 @@ private = { ignore = false, registries = [] }
 unused-allowed-license = "warn"
 
 allow = [
+    "0BSD",
     "MIT",
     "MIT-0",
     "Apache-2.0",
@@ -36,9 +37,6 @@ allow = [
     "BSL-1.0",
     "CC0-1.0",
     "Zlib",
-    "OpenSSL",
-    "bzip2-1.0.6",
-    "CDLA-Permissive-2.0",
 ]
 
 exceptions = [
@@ -74,4 +72,6 @@ allow-wildcard-paths = true
 unknown-registry = "deny"
 unknown-git = "deny"
 allow-registry = ["https://github.com/rust-lang/crates.io-index"]
-allow-git = []
+# elide is consumed straight from the in-house GitHub mirror until its
+# crates publish to crates.io. Scoped to this single repo on purpose.
+allow-git = ["https://github.com/nvisycom/elide"]