You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Revision: 18
Last modified: 2026-06-16T20:05:00Z
Authority: Derived from docs/features/.feature_inventory_raw.md (Revision 1). Per §11.4.45 (captured-evidence Status doc), §11.4.44 (revision header), §11.4.56 (two-audience summary → Status_Summary.md), §11.4.60 (always-sync), §11.4.6 (no-guessing — every status reflects what the inventory actually found in source).
Scope: Every one of the 417 inventoried features across CLI, API, Web, gRPC, Library, Submodule and Infra. helix_qa deliberately not included.
Anti-bluff note (§11.4.6 / §11.4.2 / §11.4.107):Video-confirmation is set to a real file only where a recording exists AND that recording genuinely demonstrates the feature working (each cited .mp4 was ffprobe-verified for non-zero duration+frames and content-verified against its claimed behaviour — e.g. real DeepSeek output in the target language, real gRPC translated text, live WebSocket count change). 66 feature rows are video-confirmed (Wave 9 / Wave 8 / Wave 7 of 2026-06-16 added 12 net-new (4 Wave 7 + 5 Wave 8 + 3 Wave 9), each a real input → real LLM (default in-process bridge, NOT -use-verifier) → real output clip, window-scoped (§11.4.154), helix_translate--prefixed (§11.4.155), frame-content-read so source-frame ≠ output-frame per §11.4.107. Wave 7 (4):EPUB→EPUB round-trip EN→ES ("The old man walked along the shore at dawn…" → "El anciano caminaba por la orilla al amanecer…", output EPUB document — helix_translate-epub-to-epub-en-es-wave7-20260616-181058.mp4; headline re-confirm — the existing EPUB→EPUB-conversion row was already Confirmed, this clip is an additional distinct-scope confirmation, NOT a new row), standalone -script Cyrillic→Latin (pkg/script, NO LLM: "Добар дан свима…ђ…џ" → "Dobar dan svima…đ…dž", zero Cyrillic remaining — helix_translate-script-cyrillic-to-latin-wave7-20260616-181012.mp4; headline-only re-confirm of an already-Confirmed row), preparation-translator analysis pass (genre "science fiction", chars Sarah Mills/Admiral Chen, untranslatable Captain Sarah Mills/Admiral Chen/starship Aurora/New Geneva — helix_translate-preparation-analysis-en-es-wave7-20260616-181123.mp4; headline-only re-confirm — the preparation-translator row is already Confirmed) and REST POST /api/v1/translate/string (local ./build/server :8080, JSON translated_text = "Las montañas estaban cubiertas de nieve y el río fluía tranquilamente por el valle.", provider deepseek — a real translation, NOT a session_id — helix_translate-rest-translate-string-en-es-wave7-20260616-181155.mp4; ROW-FLIPPED PENDING→Confirmed). §11.4.138 / §11.4.146 honest finding: Wave 7 ALSO recorded the markdown-translator EPUB→MD empty-payload data-loss BUG (step-3 translate sends empty strings to DeepSeek, the LLM's "you sent an empty message" boilerplate written AS IF the translation, chapter text lost) — found broken (wave7) → fixed at source in commit 4fbe581 (this session's PART-1 Translate-arg data-loss fix) → a re-record is owed before any confirmation; it was NOT a previously-confirmed row so there is NO demotion, and it is NOT counted as a confirmed net-new. Wave 8 (5):GROQ provider path EN→ES (groq/llama-3.3-70b-versatile, real Groq API: "La vieja farola se erguía sola sobre el acantilado rocoso…" — helix_translate-groq-en-es-wave8-20260616_182021.mp4; ROW-FLIPPED the Groq client library row PENDING→Confirmed), and 4 language-pair confirmations through the deepseek path (headline-only — covered by the already-Confirmed generic multi-language-pair coverage, no fabricated per-pair rows): EN→KO Korean/Hangul ("새벽에 용이 잠든 숨겨진 성을 찾기 위해 용감한 기사가 고대 숲을 가로질렀다." — helix_translate-en-ko-korean-wave8-20260616_182059.mp4), DE→EN reverse ("The scientist worked all night in his lab…" — helix_translate-de-en-reverse-wave8-20260616_182101.mp4), FR→ES non-English↔non-English ("El anciano marinero contaba historias de tormentas y tesoros ocultos…" — helix_translate-fr-es-pair-wave8-20260616_182111.mp4) and EN→SR Latin script ("U srcu planina živela je mudra stara žena…", all-Latin, no Cyrillic — helix_translate-en-sr-latin-wave8-20260616_182113.mp4). §11.4.6 nit (not a bug): -h provider help omits groq though groq IS fully wired + works end-to-end. Wave 9 (3):REST POST /api/v1/translate/batch (live server-w9 :8091 TLS, 3 EN→SR in one request, 0 errors, provider llm-novita — helix_translate-rest-batch-version-wave9-20260616_184823.mp4; ROW-FLIPPED PENDING→Confirmed), REST GET /api/v1/version (real substantive data: codebase 2.3.0, git_commit c2aa7c8…, go1.26.2, build_time + components map, NOT a bare status — same clip; headline re-confirm of the already-Confirmed combined endpoints row) and preparation→translate two-stage pipeline (stage-1 real DeepSeek analysis: "short story"/"literary fiction", chars Captain Marlowe + Eli, 3 untranslatable Stormcrag Point/Captain Marlowe/Eli, 1974 tokens; stage-2 output EPUB preserves exactly those names in Spanish "…El viejo capitán Marlowe había sido farero de Stormcrag Point durante cuarenta años…" — helix_translate-prep-translate-two-stage-wave9-20260616_184823.mp4; headline-only — covered by the already-Confirmed preparation-translator row). §11.4.6 / §11.4.124 honest non-confirmable findings (recorded as gaps, NOT confirmed, never faked):pkg/api/server.go alt Server (/api/translate, /api/batch, /api/languages) is UNWIRED — no cmd entry instantiates api.NewServer, no runnable binary (did NOT fabricate a runner); and GET /api/v1/providers returns a static hardcoded list (openai/anthropic/zhipu/deepseek), NOT substantive real upstream data, so NOT confirmed as substantive per the Wave-9 criterion. [Both addressed post-Wave-9, this session: the pkg/api/server.go alt Server is now documented as honest §11.4.124 NEVER-WIRED test-only-scaffolding N/A (see the dedicated section header note below, kept-not-deleted per operator review); and the /api/v1/providers static-list was FIXED in commit 388a2eb — it now serves the real configured/available provider set (§11.4.115 RED→GREEN + §11.4.135 guard HTQ-FIX-006), with a live re-record of /api/v1/providers against the rebuilt nezha image owed before flipping that aspect to a fresh Confirmed.]** Running tally (§11.4.6 honest arithmetic): 54 (prior authoritative total) + 4 (Wave 7) + 5 (Wave 8) + 3 (Wave 9) = 66. Of the 12 net-new, 3 are genuine row-flips (PENDING→Confirmed: POST /api/v1/translate/string, the Groq client library row, POST /api/v1/translate/batch); the remaining 9 are headline-counted additional distinct-scope confirmations on rows that were already Confirmed (EPUB→EPUB round-trip, -script Cyrillic→Latin, preparation analysis, GET /api/v1/version, preparation→translate two-stage) or language-pair/provider confirmations covered by already-Confirmed generic rows (EN→KO, DE→EN, FR→ES, EN→SR-latin) — counted in the operator-sanctioned running total but honestly distinguished here so no row is double-flipped. Wave 6 of 2026-06-16 added 5 net-new via the default in-process direct DeepSeek CLI path, each ffprobe+frame-verified per §11.4.107 (790×560 yuv420p, advancing frames, English input-frame ≠ translated output-frame): HTML→HTML round-trip EN→ES (helix_translate-html_html_roundtrip-wave6-20260616-175612.mp4, real Spanish HTML), PDF input accepted EN→ES (helix_translate-pdf_input-wave6-20260616-175612.mp4, from a real PDF 1.7), DOCX input accepted EN→ES (helix_translate-docx_input-wave6-20260616-175612.mp4, from a real Word 2007+ file), EN→ZH (Chinese) (helix_translate-en_zh_translate-wave6-20260616-175612.mp4) and EN→RU (Russian) (helix_translate-en_ru_translate-wave6-20260616-175612.mp4); Wave 6 also recorded an honest MINOR finding (HTML/EPUB chapter-title duplication on output, PENDING_FORENSICS, cosmetic — core translation correct). Wave 5 of 2026-06-16 added 4 net-new (the -multipass row was NOT counted — see §11.4.138 demotion below): EN→PT (Portuguese) (helix_translate-en-pt-translate-wave5-20260616-173814.mp4), EN→JA (Japanese, non-Latin) (helix_translate-en-ja-translate-wave5-20260616-173814.mp4), REST POST /api/v1/convert/script (helix_translate-rest-convert-script-wave5-20260616-173814.mp4, real Cyrillic↔Latin JSON both directions) and REST POST /api/v1/translate/fb2 (helix_translate-rest-translate-fb2-wave5-20260616-173814.mp4, FB2→real Serbian EPUB; note: handler hardcodes ru→sr — tracked fix). Wave 4 of 2026-06-16 added 3 net-new multi-language-pair translations from the same EN source proving genuine per-language LLM calls (distinct outputs): EN→FR (helix_translate-multilang-en-fr-wave4-20260616-173316.mp4, "Le soleil matinal se leva…"), EN→DE (helix_translate-multilang-en-de-wave4-20260616-173319.mp4, "Der morgendliche Sonnenaufgang…"), EN→IT (helix_translate-multilang-en-it-wave4-20260616-173419.mp4, "Il sole mattutino sorse…"); Wave 4 also re-confirmed (NOT double-counted) RU→SR-Cyrillic (helix_translate-multilang-ru-sr-cyr-wave4-20260616-173445.mp4). §11.4.138 DEMOTION (−1): the pkg/verificationMulti-pass polisher (-multipass) row, previously PASS via helixtranslate-cli-multipass-verify-20260615.mp4, is DEMOTED to PENDING_FORENSICS — the wave-3e regression (docs/qa/wave3e-multipass-regression/FINDING.md) and wave-5 evidence prove the default-model path (gpt-4) fails the ValidModels whitelist and silently no-ops, and even with an explicit model the polish path can write an LLM meta-response instead of polished prose; a tracked source fix + re-record is owed before any re-confirmation. Net video-confirmed: 43 (Rev 14) − 1 (multipass demote) + 3 (Wave 4) + 4 (Wave 5) + 5 (Wave 6) = 54. Wave 3f of 2026-06-16 added 2 net-new — EPUB→EPUB conversion (helix_translate-epub-output-wave3e-20260616.mp4: EN .epub → real deepseek run → OUTPUT EPUB carries real Spanish "La Antigua Farola la vieja farola se alzaba sola en la rocosa costa…") and TXT→TXT translation (helix_translate-txt-translation-wave3f-20260616.mp4: EN .txt → real deepseek run → ES .txt "El río fluía tranquilamente a través del valle verde. Los pájaros cantaban bajo la luz de la mañana."), each ffprobe+frame-verified per §11.4.107 (790×560 yuv420p, ≥6 frames, start-frame English-only ≠ output-frame Spanish); Wave 3f also re-confirmed (NOT double-counted) the already-confirmed FB2→FB2 (helix_translate-fb2-convert-wave3e-20260616.mp4, namespace preserved + real Spanish), format-detection (helix_translate-format-detect-wave3e-20260616.mp4, extensionless .dat detected-by-content), -script Cyrillic↔Latin (helix_translate-script-conversion-wave3e-20260616.mp4, RUN A latin "Ovo je testni…" ≠ RUN B cyrillic "Систем треба…") and markdown-translator EPUB↔MD round-trip (helix_translate-markdown-roundtrip-wave3f-20260616.mp4, "# El Faro Viejo / El viejo faro se alzaba solitario en la costa rocosa.") rows with fresh frame-verified clips; running total 41 (Rev 13) + 2 = 43. Wave 3d of 2026-06-16 added 4 net-new, prioritising feature-type diversity — the cmd/verify-models CLI (real LLMsVerifier discovery/scoring/selection: 14 env-keyed providers with real per-provider reachable/auth/HTTP-status/candidate/verified counts, summary total verified: 4, then model-bridge list selection ranking #1 novita/Sao10K/L3-8B-Stheno-v3.2 score=0.91875 … #4 siliconflow score=0.88125 — helix_translate-verify-models-wave3d-20260616.mp4), the cmd/workable-items CLI (real §11.4.93 SQLite SSoT: validate → OK: DB … matches markdown (81 items), list → real ATM-NNN rows + descriptions — helix_translate-workable-items-wave3d-20260616.mp4), and 2 more providers — nvidia ("…Un anciano pescador caminaba a lo largo de la orilla." via integrate.api.nvidia.com, helix_translate-provider-nvidia-translate-wave3d-20260616.mp4) and openrouter ("…Un anciano pescador paseaba a lo largo de la orilla." via openrouter.ai/api, helix_translate-provider-openrouter-translate-wave3d-20260616.mp4), the two provider outputs distinct (caminaba/paseaba) proving genuine per-provider models, all four Wave-3d clips ffprobe+frame-verified per §11.4.107 (790×560 yuv420p, ≥6 frames at fps=2, CLI clips show advancing distinct sections / provider clips show start-frame English-only ≠ output-frame Spanish); the Wave-3d providers reach the llm_provider/OpenAI-compat per-provider path; upstage was tested but is an honest SKIP (auth=false HTTP 403), together SKIPped (no key) — never faked. Wave 3c of 2026-06-16 added 2 net-new — the Hyperbolic library client (pkg/translator/llm) and the Fireworks provider in the llm_provider submodule (llm_provider/pkg/providers/fireworks/fireworks.go), each a real-use Wave-3c .mp4 showing a distinct EN→ES translation through the OpenAI-compat per-provider path — helix_translate-provider-hyperbolic-translate-wave3c-20260616.mp4 ("Al amanecer, el sol salía sobre el tranquilo pueblo. Un anciano pescador caminaba por la orilla…") and helix_translate-provider-fireworks-translate-wave3c-20260616.mp4 ("Al amanecer, el sol ascendió sobre el tranquilo pueblo. Un anciano pescador paseaba por la orilla…"), the two outputs distinct from each other (salía/caminaba vs ascendió/paseaba) proving genuine per-provider models, ffprobe+frame-verified per §11.4.107 (790×560 yuv420p, ≥6 frames at fps=2, start-frame input-only ≠ output-frame Spanish). Wave 3c also re-confirmed (no double-count) the already-confirmed REST POST /api/v1/translate and GET /health · /api/v1/version · /api/v1/providers rows against a freshly-started live cmd/server TLS instance (port 9443) — helix_translate-rest-translate-wave3c-20260616.mp4 (real JSON {"translated":"El sol de la mañana se alzó sobre el tranquilo pueblo.","provider":"llm-deepseek",…}) and helix_translate-rest-endpoints-wave3c-20260616.mp4 (real health/version/providers JSON). The GET /api/v1/verified-models 404 remains an honest PENDING gap (LLMsVerifier disabled in config — not confirmed). Wave 3b of 2026-06-16 added 3 net-new — the Gemini, Zhipu (GLM) and Cohere library clients, each backed by a real-use Wave-3b .mp4 showing a distinct EN→SR-latin translation through the per-provider path — helix_translate-provider-gemini-translate-wave3b-20260616.mp4 ("Staro svjetiljko stajalo je samotočno na kamenitoj obali…"), helix_translate-provider-zhipu-translate-wave3b-20260616.mp4 ("Stara svjetiljka stojeća samostalno na kamenitoj obali…"), helix_translate-provider-cohere-translate-wave3b-20260616.mp4 (provider=cohere via -base-url https://api.cohere.ai/compatibility/v1, "Stari svjetionik stojeo je sam na kamenitoj obali… svijeća njegova vodila brodove…"), all four outputs (incl. the re-confirmed deepseek clip helix_translate-provider-deepseek-translate-wave3b-20260616.mp4 "Stari svjetionik stajao je sam na kamenitoj obali…") distinct from one another proving genuine per-provider models, ffprobe-verified (non-zero dur+frames, 790×560 even-dims yuv420p) + ≥6 frames extracted with one content-verified per §11.4.107; the deepseek clip is a re-confirmation, no double-count; Wave 3a of 2026-06-16 added 2 net-new — the Cerebras and SambaNova per-provider LLM clients, each backed by a real-use Wave-3a .mp4 showing a distinct EN→ES translation through unified-translator -provider <name> — helix_translate-provider-cerebras-translate-wave3a-20260616.mp4 ("El sol de la mañana se alzó sobre el tranquilo pueblo.") and helix_translate-provider-sambanova-translate-wave3a-20260616.mp4 ("El sol matutino se alzó sobre el tranquilo pueblo."), the two outputs distinct from each other proving genuine per-provider models, ffprobe+frame-verified per §11.4.107 with start-frame ≠ output-frame liveness; Wave 2 of 2026-06-16 added 5 net-new — DOCX output, PDF output, HTML→Markdown conversion, ProviderDiverseTranslators, EnsembleFactory — each backed by a real-use Wave-2 .mp4 independently validator-verified ffprobe+frame; Wave 2 also re-confirmed the already-confirmed gRPC TranslationService row with an additional clip, no double-count; Wave 1 added 2 — bridge-ensemble best-model selection + bridge provider-diverse catalogue) across unified-translator (DeepSeek translate, EPUB→TXT / HTML→EPUB / FB2→EPUB / FB2→FB2 conversions, Serbian-Cyrillic, -script Cyrillic↔Latin, format-detection, help/version, -verify, -multipass multi-pass polishing engine), the REST API server (POST /api/v1/translate real DeepSeek + target_lang FIXED, TLS startup FIXED, /health · /api/v1/version · /api/v1/providers), the gRPC server (real EN→ES), markdown-translator (EPUB↔MD round-trip), preparation-translator (FIXED real analysis), cmd/cli (real DeepSeek), monitor-server (live WS hub), the DeepSeek library client, and the LLMsVerifier→agent bridge CLI (cmd/model-bridge — helixtranslate-bridge-bestmodel-translate-20260615.mp4: in-process verify→SQLite store→select strongest novita/Sao10K/L3-8B-Stheno-v3.2 score 0.919 → live invoke EN→ES "Good morning, friend." → "Buenos días, amigo."; ffprobe 8.0s/80 frames, ≥6 frames content-verified) and the LLMsVerifier→agent bridge MCP-stdio surface (cmd/model-bridge mcp — helixtranslate-bridge-mcp-stdio-20260615.mp4: a real JSON-RPC MCP session over stdio — initialize→tools/list→bridge_best_model→bridge_invoke — where tool bridge_invoke returned a live verified-model translation "Le pont relie deux rives." (isError:None); ffprobe 11.88s/10 frames, ≥6 frames content-verified). Now video-confirmed (Wave 2):DOCX output (Implemented; pure-Go OOXML writer pkg/ebook/docx_writer.go, commit 87cd2be) — Wave-2 real-use video helix_translate-docx-output-20260616-112943.mp4 (unified-translator DOCX output, openai/gpt-4, the produced .docx carries real Spanish text; validator ffprobe+frame-verified). PDF output (Implemented, commit fb265e7; pkg/ebook/pdf_writer.go Book→HTML5→weasyprint→Cyrillic-faithful PDF, wired into unified-translator .pdf output, main.go:1160) — Wave-2 real-use video helix_translate-pdf-output-20260616-113057.mp4 (valid PDF with real Spanish; validator ffprobe+frame-verified). Wave-1 bridge-translator-path blocker RESOLVED (caveat 8): the ValidModels-whitelist runtime failure that blocked the DOCX/PDF/gRPC/per-provider videos in Wave 1 is fixed — the verified-bridge translator path now produces real per-provider translations (Wave-2 provider clips: novita/mistral/groq/siliconflow). Honest gaps recorded, NOT confirmed:cmd/translator + translate-ssh + ebook-translator are Obsolete (→ Fixed.md), Rev 7 (feature-removed — the local-runtime + SSH-local paths were removed in bridge phase-2 R-2..R-4; the default path now sources the LLMsVerifier pkg/bridge; the earlier STUB/OPERATOR-BLOCKED wording is stale and corrected); GET /api/v1/verified-models returns 404 because LLMsVerifier is disabled in config. Everything else is PENDING (a recording is owed) or N/A (no user-visible video applies — e.g. CLI flags, internal library types). Absence of a runtime recording is stated honestly, never papered over.
Operator-blocked items (read first, §11.4.45)
Two binaries are OPERATOR-BLOCKED — both require a remote SSH worker host running llama.cpp, which is not available in this environment. The binaries build and run as far as the missing host allows (captured in helixtranslate-cmd-blocked-binaries_20260615-172456.mp4):
cmd/translate-ssh — SSH worker 4-step FB2→EPUB translator. Unblock condition: a reachable SSH host with llama.cpp + a model.
cmd/ebook-translator — FB2 remote-translate workflow (syncs binary to remote, runs remote translate, downloads outputs). Unblock condition: same remote SSH/llama.cpp host.
cmd/translator (remote path) — remote SSH/llama.cpp translate path is OPERATOR-BLOCKED (same host requirement); its LOCAL path is a separate STUB ("local translation not yet implemented").
Coverage summary
Metric
Value
Headline feature total (inventory's deduplicated per-category tally)
417
Enumerated feature rows in this doc
497 (Rev 16 = 496 Rev 14 + 1 new cmd/grpc-translate-probe row — standalone gRPC translate-round-trip probe client, Implemented + Wired + Source-confirmed, Video-confirmation PENDING; Rev 14 = 494 Rev 13 + 2 new unified-translator rows added Wave 3f for the net-new videos: EPUB→EPUB conversion + TXT→TXT translation; Rev 9 = 493 Rev 8 + 1 new unified-translator HTML→Markdown conversion row added that wave for the Wave-2 html-to-md video; §11.4.6 — re-derive with awk over ^| 9-column rows. The earlier "478" was NOT re-derivable from the live table — it was a hand-tracked running tally that had drifted; Rev 8 added the 7 missing pkg/bridge ensemble-seam rows: BestTranslator, BestTranslatorFunc, EnsembleFactory, ProviderDiverseTranslators, ProviderDiverseClients, ProviderDiverseModels, BestClient)
Implemented
446 (Rev 16: 445 Rev 14 + 1 new cmd/grpc-translate-probe row, Implemented + Wired; Rev 14: 443 Rev 13 + 2 new unified-translator rows EPUB→EPUB conversion + TXT→TXT translation, both Implemented + Wired + Wave-3f video-confirmed; Rev 9: 442 Rev 8 + 1 new unified-translator HTML→Markdown conversion row; Rev 8 was 435 Rev 7 + 7 new pkg/bridge ensemble-seam rows; was 471 Rev 6, −36 flipped to Obsolete Rev 7 — local-runtime + SSH-local features removed in bridge phase-2 R-2..R-4)
2 (POST /api/batch on standalone pkg/api/server.go — returns queued batch_id only, no translation; vision_engine OpenCV — stub default, real impl behind build tag) — was 3 Rev 6; cmd/translator flipped to Obsolete Rev 7
Operator-blocked
0 (was 2 Rev 6 — translate-ssh + ebook-translator flipped to Obsolete Rev 7: the remote SSH/llama.cpp worker capability was removed, not merely host-blocked)
Video-confirmed
66 (real, content-verified recordings — see Anti-bluff note. Rev-16 / Wave 9 + Wave 8 + Wave 7 of 2026-06-16 added 12 net-new (4+5+3), of which 3 are genuine row-flips PENDING→Confirmed (POST /api/v1/translate/string Wave 7, the Groq client library row Wave 8, POST /api/v1/translate/batch Wave 9) and 9 are headline-counted additional distinct-scope confirmations on already-Confirmed rows / language-pair coverage (EPUB→EPUB round-trip, -script Cyrillic→Latin, preparation analysis, GET /api/v1/version, preparation→translate two-stage, EN→KO, DE→EN, FR→ES, EN→SR-latin — see Anti-bluff note for the full per-feature mapping + the Wave-7 markdown-translator bug found→fixed-at-4fbe581→re-record-owed + the Wave-9 honest non-confirmable gaps). §11.4.6 cell reconciliation: this cell had a known sync-lag behind the authoritative headline running total (it read 43 at Rev 15 while the Anti-bluff-note prose was at 54); it is corrected here to the true running total — 54 (prior authoritative) + 4 (Wave 7) + 5 (Wave 8) + 3 (Wave 9) = 66. Rev-14 / Wave 3f of 2026-06-16 added 2 net-new video confirmations prioritising feature-type diversity — EPUB→EPUB conversion (helix_translate-epub-output-wave3e-20260616.mp4, EN .epub → real Spanish EPUB) and TXT→TXT translation (helix_translate-txt-translation-wave3f-20260616.mp4, EN .txt → real Spanish .txt), each ffprobe+frame-verified per §11.4.107; Wave 3f also re-confirmed (NOT double-counted) FB2→FB2 / format-detection / -script Cyrillic↔Latin / markdown EPUB↔MD round-trip with fresh frame-verified clips; running total 41 (Rev 13) + 2 = 43. Rev-13 / Wave 3d of 2026-06-16 added 4 net-new video confirmations prioritising feature-type diversity — the cmd/verify-models CLI (helix_translate-verify-models-wave3d-20260616.mp4, real LLMsVerifier discovery/scoring/selection + ranking), the cmd/workable-items CLI (helix_translate-workable-items-wave3d-20260616.mp4, real §11.4.93 SQLite validate+list), and 2 more providersnvidia (helix_translate-provider-nvidia-translate-wave3d-20260616.mp4, "…caminaba a lo largo de la orilla.") + openrouter (helix_translate-provider-openrouter-translate-wave3d-20260616.mp4, "…paseaba a lo largo de la orilla."), each ffprobe+frame-verified per §11.4.107; running total 37 (Rev 12) + 4 = 41. upstage HTTP-403 SKIP + together no-key SKIP, never faked. Rev-12 / Wave 3c of 2026-06-16 added 2 net-new video confirmations — the Hyperbolic library client (helix_translate-provider-hyperbolic-translate-wave3c-20260616.mp4, "…el sol salía…caminaba por la orilla…") and the Fireworksllm_provider submodule client (helix_translate-provider-fireworks-translate-wave3c-20260616.mp4, "…el sol ascendió…paseaba por la orilla…"), each a distinct real EN→ES translation via the OpenAI-compat per-provider path, ffprobe+frame-verified per §11.4.107 (790×560 yuv420p, ≥6 frames at fps=2, start-frame input-only ≠ output-frame Spanish), the two outputs distinct proving genuine per-provider models; Wave 3c also re-confirmed (NOT double-counted) the already-confirmed REST POST /api/v1/translate + /health · /api/v1/version · /api/v1/providers rows against a live cmd/server TLS instance (helix_translate-rest-translate-wave3c-20260616.mp4 + helix_translate-rest-endpoints-wave3c-20260616.mp4); count reconciliation (§11.4.6): this cell read 32 through Rev 10/11 — a sync lag behind the Anti-bluff-note prose; it is corrected to the true running total here — 30 (Rev 9) + 2 (Wave 3a Cerebras/SambaNova) + 3 (Wave 3b Gemini/Zhipu/Cohere) + 2 (Wave 3c Hyperbolic/Fireworks) = 37. Rev-10 / Wave 3a of 2026-06-16 added 2 net-new video confirmations — the Cerebras and SambaNova per-provider LLM clients (helix_translate-provider-cerebras-translate-wave3a-20260616.mp4 + helix_translate-provider-sambanova-translate-wave3a-20260616.mp4), each a distinct real EN→ES translation via unified-translator -provider <name>, ffprobe+frame-verified per §11.4.107 with start-frame ≠ output-frame liveness, the two outputs distinct proving genuine per-provider models; Rev-11 / Wave 3b of 2026-06-16 added 3 net-new — the Gemini, Zhipu (GLM) and Cohere library clients (helix_translate-provider-{gemini,zhipu,cohere}-translate-wave3b-20260616.mp4), distinct EN→SR-latin per-provider translations, frame-verified per §11.4.107; Rev-9 / Wave 2 of 2026-06-16 added 5 net-new video confirmations, each backed by a real-use Wave-2 .mp4 independently validator-verified ffprobe+frame §11.4.107: DOCX output (helix_translate-docx-output-20260616-112943.mp4), PDF output (helix_translate-pdf-output-20260616-113057.mp4), HTML→Markdown conversion (helix_translate-html-to-md-20260616-114228.mp4), ProviderDiverseTranslators + EnsembleFactory (the 4 per-provider clips novita/mistral/groq/siliconflow — helix_translate-provider-{novita,mistral,groq,siliconflow}-translate-20260616-114*.mp4). Wave 2 also re-confirmed the already-confirmed gRPC TranslationService row with an additional clip helix_translate-grpc-translate-20260616-113415.mp4 (NOT double-counted — that row was already video-confirmed). DOCX output and PDF output are no longer "real-artifact-only" — they are now video-confirmed. §11.4.108 note: the Wave-1 bridge-translator-path runtime blocker (caveat 8) is RESOLVED, which is what made the DOCX/PDF/gRPC/per-provider Wave-2 recordings genuinely demonstrate working features (recording them in Wave 1 would have been a §11.4.107 bluff; they were honest SKIPs then). Wave 1 added 2: bridge-ensemble best-model selection + bridge provider-diverse catalogue. ≈ 8.3% of 494 enumerated rows (≈ 9.8% of the 417 headline)
Video-confirmation PENDING
runtime/user-visible features without a recording yet
Video-confirmation N/A
flags, internal library types, infra middleware (no standalone user-visible video)
§11.4.6 reconciliation: The inventory's headline is 417 — its own note states this is an approximate, deduplicated per-category tally where "some rows are sub-features... counted once, and a few rows span two categories (counted once, in the first/primary category)." The inventory's detailed tables actually enumerate 470 feature rows. This document transcribes all 470 detailed rows 1:1 so no feature is dropped (which would itself be a bluff). The 417 headline and the 470 enumerated rows are both reported honestly; the 53-row gap is dedup, not missing coverage.
"Static-return" endpoints (e.g. GET /api/stats hardcoded zeros, GET /api/v1/status/:id hardcoded completed, GET /api/v1/metrics static, monitor-server GET /api/v1/status static monitoring_active) are marked Implemented at the route level but their Real-use cell states the static/hardcoded behaviour honestly so no false "working backend" impression is created.
Per-category counts
Category
Count
CLI
130
API
70
Web
23
gRPC
7
Library
81
Submodule
95
Infra
11
TOTAL
417
Legend
Implementation:Implemented (real code path present) / Stub (returns mock, no real work) / Partial (does some but not the named work).
Wiring:Wired (reachable from a binary / route / public API) / Unwired (present but not reached).
Real-use: what the code actually does at runtime (honest, §11.4.6).
Tests:Covered (unit/integration/challenge present in repo) / None (no test found / not asserted in inventory).
Validation: strongest evidence available.
Video-confirmation:Confirmed:<file> only with a real recording; else PENDING / N/A.
Per §11.4.6, where the inventory did not read test files line-by-line, Tests is recorded Not-inventoried (honest unknown) rather than guessed as Covered or None. The repo is known to carry extensive _test.go suites and a test/ tree, but per-feature test attribution was not part of the raw inventory.
1. CLI Binaries — cmd/
cmd/api-server
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
api-server
REST translation server
API/Web
Implemented
Wired
Gin HTTP server (:8080) proxying to gRPC backend; serves dashboard + static files
Registers + runs all anti-bluff challenges, prints PASS/FAIL/SKIP, writes results dir
Not-inventoried
Source-confirmed
PENDING
challenge-runner
-parallel
CLI
Implemented
Wired
Run challenges in parallel
Not-inventoried
Source-confirmed
N/A
challenge-runner
-max-concurrency
CLI
Implemented
Wired
Max concurrent challenges
Not-inventoried
Source-confirmed
N/A
challenge-runner
-stop-on-failure
CLI
Implemented
Wired
Stop on first failure
Not-inventoried
Source-confirmed
N/A
challenge-runner
-verbose
CLI
Implemented
Wired
Verbose output
Not-inventoried
Source-confirmed
N/A
challenge-runner
-timeout
CLI
Implemented
Wired
Per-challenge timeout
Not-inventoried
Source-confirmed
N/A
challenge-runner
-filter
CLI
Implemented
Wired
Comma-separated challenge IDs to run
Not-inventoried
Source-confirmed
N/A
challenge-runner
-category
CLI
Implemented
Wired
Limit to a category
Not-inventoried
Source-confirmed
N/A
challenge-runner
-results-dir
CLI
Implemented
Wired
Results output directory
Not-inventoried
Source-confirmed
N/A
cmd/cli
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
cli
Universal ebook translator
CLI
Implemented
Wired
Parse ebook → translate → atomic-write EPUB/FB2/TXT; multi-llm/distributed support
Not-inventoried
PASS — real DeepSeek run captured (real Serbian TXT output). Title path-leak FIXED (d53e085, sanitizeBookTitle). Write-safety FIXED (87cd2be): format validated before any FS write, written to temp sibling + atomic os.Rename only on full success — never a partial/clobbered output. Guard: cmd/cli/no_partial_output_test.go (5 tests, mutation-proven)
gRPC server (:50051): Start/Status/List/Cancel/Stream/GetProviders/SubscribeEvents
Not-inventoried
PASS — real EN→ES translation over gRPC captured (StartTranslation → real Spanish DeepSeek output). Wave 2 re-confirms via the verified-bridge path (EN→ES "Hola, mundo. Este es un pequeño test.")
Pure wire-protocol client (cmd/grpc-translate-probe/main.go): dials a live TranslationService (default nezha.local:50061, insecure creds), calls StartTranslation(input_file,output_file,src,dst,provider,model,script) then polls GetTranslationStatus to a terminal state, printing START/POLL/COMPLETED/FILE lines + generated-file paths. Does NO server-side file I/O — input must pre-exist on the server fs (staged via podman cp), output read back out-of-band; flags -addr/-input/-output/-src/-dst/-provider/-model/-script/-session/-timeout; exit 0 completed / 1 failed-cancelled-error / 2 usage-dial-error
Not-inventoried
Source-confirmed (cmd/grpc-translate-probe/main.go); no real-use recording yet
PASS — EPUB↔MD round-trip captured (Wave 3f re-confirm: EPUB→source MD "# The Old Lighthouse / The old lighthouse stood alone…" → translated MD "# El Faro Viejo / El viejo faro se alzaba solitario en la costa rocosa.", real deepseek round-trip, frame-verified §11.4.107)
EPUB parse → DeepSeek translate → EPUB writer (round-trip, real translated text in produced .epub)
Not-inventoried
PASS — captured (Wave 3f net-new: EN .epub input → real deepseek run → OUTPUT EPUB carries real Spanish "La Antigua Farola la vieja farola se alzaba sola en la rocosa costa…", frame-verified §11.4.107 — start frame English-only ≠ output frame Spanish). Wave-7 additional EPUB→EPUB round-trip clip (EN→ES): "The old man walked along the shore at dawn…" → "El anciano caminaba por la orilla al amanecer…", output verified EPUB document (re-confirmation of this already-Confirmed row, no double-count).
Plain-text parse → DeepSeek translate → TXT writer (real translated text, no format conversion)
Not-inventoried
PASS — captured (Wave 3f net-new: EN .txt "The river flowed quietly through the green valley…" → real deepseek run → ES .txt "El río fluía tranquilamente a través del valle verde. Los pájaros cantaban bajo la luz de la mañana.", frame-verified §11.4.107)
HTML parse → DeepSeek translate → HTML writer (round-trip, real translated text in produced .html)
Not-inventoried
PASS — captured (Wave 6 net-new: EN .html → real deepseek run → valid Spanish <html> "la farola guiaba a los barcos a través de la noche neblinosa…", frame-verified §11.4.107 — input-frame English ≠ output-frame Spanish). NOTE Wave-6 MINOR finding: chapter-title duplication on single-section chapters (cosmetic, PENDING_FORENSICS — core translation correct).
PDF parse → DeepSeek translate (real PDF 1.7 input)
Not-inventoried
PASS — captured (Wave 6 net-new: from "PDF document, version 1.7" → real Spanish "El bosque se hizo silencioso a la caída del sol. Los pinos altos susurraban…", frame-verified §11.4.107)
DOCX parse → DeepSeek translate (real Word 2007+ input)
Not-inventoried
PASS — captured (Wave 6 net-new: from "Microsoft Word 2007+" → real Spanish "Los barcos de pesca regresaron al puerto al atardecer. Gaviotas volaban sobre las redes…", frame-verified §11.4.107)
Source/target language selection driving real per-language LLM translation
Not-inventoried
PASS — video-confirmed (Waves 4–6 drove the SAME EN source into 7 distinct target languages producing 7 genuinely different, language-correct outputs — proving real per-language LLM calls, not echo/cache: FR "Le soleil matinal se leva…", DE "Der morgendliche Sonnenaufgang…", IT "Il sole mattutino sorse…", PT "O antigo farol erguia-se…", JA "古い灯台は…", ZH "一列列车在黎明时分…", RU "Библиотека стояла тихо под дождем…", each ffprobe+frame-verified §11.4.107)
Target script (cyrillic/latin) script normalization on real run
Not-inventoried
PASS — captured (Wave 3f re-confirm: RUN A -script latin "Ovo je testni paragraf…" ≠ RUN B -script cyrillic "Систем треба да преводи…" — same text, genuinely different script, frame-verified §11.4.107)
Built-in per-step verification: input/markdown/translation-quality/valid-EPUB verdicts in session report. NOTE: this is the unified-translator's own verify pass, NOT the pkg/verification multipass engine (that engine is not wired into this CLI — see gap row below)
LLMsVerifier→component+agent bridge (no local models): selects the best LLMsVerifier-scored model, exposes it to agents/components via a CLI and an MCP stdio server wired in .mcp.json. Source: cmd/model-bridge/{main.go,mcp.go}, pkg/bridge/, commit ab1bed3 (+ routing seam a5860b2).
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
model-bridge
LLMsVerifier→agent bridge CLI
CLI
Implemented
Wired
Selects best LLMsVerifier-scored model (top-1 + fallback chain) and translates/invokes through it; no local models
Covered (pkg/bridge unit+integration, go test -race green this session)
PASS — real run captured (best-model selection novita/Sao10K/L3-8B-Stheno-v3.2 score 0.919; Invoke EN→ES "Good morning, friend." → "Buenos días, amigo.")
CONFIRMED — real-use video /Volumes/T7/Downloads/Recordings/helixtranslate-bridge-bestmodel-translate-20260615.mp4 (asciinema→agg→ffmpeg pipeline; ffprobe 8.0s / 80 frames / 790×560; ≥6 frames readable — selected model novita/Sao10K/L3-8B-Stheno-v3.2 (score 0.919, base_url https://api.novita.ai/v3/openai) + live OUTPUT "Buenos días, amigo." visible, not frozen/spinner per §11.4.107)
model-bridge
MCP stdio server (.mcp.json)
CLI/Infra
Implemented
Wired
Exposes the selected model to MCP clients over stdio; wired as model-bridge in .mcp.json (./build/model-bridge)
Covered (cmd/model-bridge/mcp_test.go)
PASS — real MCP JSON-RPC round-trip captured (initialize→tools/list→bridge_best_model→bridge_invoke over stdio; tool bridge_invoke returned live verified-model translation "Le pont relie deux rives.", isError:None)
CONFIRMED — real-use video /Volumes/T7/Downloads/Recordings/helixtranslate-bridge-mcp-stdio-20260615.mp4 (asciinema→agg→ffmpeg; ffprobe 11.88s / 10 frames / 790×560; ≥6 frames readable showing the progressive REAL MCP-stdio session — tools/list=[bridge_invoke, bridge_best_model, bridge_list], bridge_best_model strongest novita/Sao10K/L3-8B-Stheno-v3.2 score 0.919 + fallback chain, and tool bridge_invoke real verified-model OUTPUT "Le pont relie deux rives." with isError:None — not frozen/spinner per §11.4.107)
cmd/workable-items
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
workable-items
§11.4.93 workable-items SQLite tool
CLI
Implemented
Wired
Bidirectional sync between Issues/Fixed markdown and SQLite single-source-of-truth
Not-inventoried
PASS — real DB CLI output captured (./build/workable-items validate -db docs/workable_items.db → OK: DB docs/workable_items.db matches markdown (81 items), then list printed real ATM-NNN rows with status + descriptions — ATM-001 splitText dropped paragraph separators…, ATM-005 Distributed ResultCache spurious eviction…, ATM-014 Preparation failed-chapter analysis misattribution…; Wave-3d, asciinema→agg→ffmpeg, ffprobe 5.0s/790×560 yuv420p + 10 frames at fps=2, early validate-OK frame ≠ late list frame proving live/advancing per §11.4.107)
Exercised in unified-translator real run (re-confirmed Wave-3b: -provider deepseek -model deepseek-chat EN→SR-latin "Stari svjetionik stajao je sam na kamenitoj obali…"; ffprobe 4.7s/790×560, f_03 content-verified — additional clip, no double-count)
Pure-Go OOXML writer pkg/ebook/docx_writer.go (stdlib archive/zip + encoding/xml, zero new deps); emits the 4 required OOXML parts; XML specials auto-escaped via xml.Marshal chardata. Wired into unified-translator generateOutput (case docx). Real DeepSeek EN→ES → garden_es.docx = Microsoft Word 2007+, well-formed WordprocessingML with the real translation. Guards: pkg/ebook/docx_writer_test.go + .docx subtest in cmd/unified-translator/output_format_test.go (mutation-proven)
Covered
PASS — real artifact verified + Wave-2 video (produced .docx = Microsoft Word 2007+, 4 OOXML parts, real translated content + correct XML escaping; commit 87cd2be; Wave-2 real-use video: unified-translator DOCX output, openai/gpt-4, the .docx carries real Spanish text — validator ffprobe+frame-verified)
PDF writer pkg/ebook/pdf_writer.go (commit fb265e7): renders the in-memory Book to UTF-8 HTML5 then hands it to weasyprint to emit a VALID, Cyrillic-faithful PDF (Standard-14 fonts can't render Cyrillic — weasyprint embeds glyphs, avoiding a silent-glyph-drop bluff). Wired into unified-translator generateOutput (case .pdf → ebook.NewPDFWriter().Write, main.go:1160); .pdf is a supported -output extension. Honest typed ErrWeasyPrintUnavailable when weasyprint absent (§11.4.3/§11.4.6). Guards: pkg/ebook/pdf_writer_test.go (TestPDFWriter_ProducesValidPDFCarryingTranslatedText, _UnavailableWeasyPrint)
Covered
PASS — go test ./pkg/ebook green this session (TestPDFWriter_ProducesValidPDFCarryingTranslatedText asserts a valid PDF carrying the translated text; weasyprint present at /opt/homebrew/bin/weasyprint); Wave-2 real-use video: unified-translator PDF output, openai/gpt-4, valid PDF with real Spanish (validator ffprobe+frame-verified)
Wired into unified-translator (-multipass, d53e085)
Iterative multi-LLM polishing w/ note-taking & consensus. Now reachable via the unified-translator opt-in -multipass flag (+ -multipass-passes, -multipass-db; default OFF, zero overhead/behavior change when absent). NOTE: distinct from -verify (the CLI's own per-step output check) — -multipass invokes this pkg/verification engine. Guard: cmd/unified-translator/multipass_wiring_test.go (httptest + real DeepSeek). §11.4.150 research: docs/research/20260615_video_surfaced_fixes/verification_multipass_wiring.md
Not-inventoried
PENDING_FORENSICS (§11.4.138 demoted) — the prior helixtranslate-cli-multipass-verify-20260615.mp4 PASS is superseded: the default model gpt-4 fails the ValidModels whitelist for deepseek/most providers and the polish pass silently no-ops keeping the base translation (wave-5 docs/qa/video_wave5_20260616-173814/EVIDENCE.md), and even with an explicit valid model the path can write an LLM meta-response/UNCHANGED-marker instead of polished prose (wave-3e docs/qa/wave3e-multipass-regression/FINDING.md). Tracked source fix + re-record owed before re-confirmation.
PENDING_FORENSICS (was Confirmed:/Volumes/T7/Downloads/Recordings/helixtranslate-cli-multipass-verify-20260615.mp4 — superseded by §11.4.138)
Correctness fix (commit 97a8afd): a model with affirmative_response=0 is now a HARD disqualifier — rejected by the verification gate regardless of other scores
Covered + §11.4.115 polarity guardpipeline_affirmative_gate_test.go (GREEN-default: the standing suite asserts the defect ABSENT; RED reproduction is OPT-IN only under RED_MODE=1)
PASS — observed this session: default suite GREEN (ok internal/verifier); the defect reproduction FAILs only under explicit RED_MODE=1 (by-design §11.4.115)
N/A (§11.4.6 — internal verification gate, no user-visible surface)
In-process bridge (no local models) that selects the best LLMsVerifier-scored model and routes component/agent calls to it; backs cmd/model-bridge CLI + MCP. Source: pkg/bridge/bridge.go, commit ab1bed3 (+ routing mutation seam a5860b2).
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
bridge
Best-model selection
Library
Implemented
Wired (via cmd/model-bridge CLI+MCP)
Picks top-1 LLMsVerifier-scored model with a provider-diverse fallback chain; no local models
bridge.go:242 — returns the single strongest verified translator.Translator plus its fallback-chain model IDs for the task; the non-ensemble single-translator source
Covered at unit level (pkg/bridge/bridge_test.go, pkg/bridge/ensemble_test.go, go test ./pkg/bridge green) — but unit tests use whitelisted models, masking the runtime bug below
PARTIAL — UNIT PASS, RUNTIME BLOCKED (§11.4.108):go test ./pkg/bridge → ok, BUT a real run via unified-translator/gRPC FAILS at translate time — the verified strongest model (novita/Sao10K/L3-8B-Stheno-v3.2) is rejected by the stale ValidModels whitelist in VerifiedFactory.CreateTranslatorWithFallback→NewLLMTranslatorWithConfig (pkg/translator/llm/llm.go:237), unlike the BestModel/invoke path which correctly bypasses it. See Wave-1 discovered blocker below
PENDING — NOT recordable until the whitelist blocker is fixed (recording a "working" video here would be a §11.4.107 bluff)
ensemble.go:37 — context-only closure yielding the single strongest verified translator (fallback chain discarded); plain function value passed to single-translator sites WITHOUT importing the bridge (§11.4.28 decoupling); no API key captured/logged (§11.4.10)
Unit-covered (pkg/bridge/ensemble_test.go, -race green) — masks the runtime bug
PARTIAL — UNIT PASS, RUNTIME BLOCKED (§11.4.108): delegates to BestTranslator; same ValidModels-whitelist runtime failure (see Wave-1 discovered blocker)
PENDING — NOT recordable until the blocker is fixed
bridge
EnsembleFactory (provider-diverse ensemble seam)
Library
Implemented
Wired (R-1 ensemble redirect)
ensemble.go:25 — context-only closure yielding the provider-diverse verified translator set (one per distinct provider, strongest-first); structural shape of the per-component EnsembleTranslatorFactory seams, passed to NewMultiLLMCoordinatorWithFactory / NewPreparationCoordinatorWithFactory / NewBookPolisherWithFactory WITHOUT those packages importing the bridge (§11.4.28); no API key captured/logged (§11.4.10)
PASS — Wave-1 blocker FIXED, runtime confirmed (§11.4.108): backed by ProviderDiverseTranslators; the ValidModels-whitelist blocker (caveat 8) is resolved — verified non-whitelisted models now translate. Wave-2 provider clips show real per-provider translations through this ensemble seam
PASS — Wave-1 blocker FIXED, runtime confirmed (§11.4.108): the ValidModels-whitelist blocker (caveat 8) is resolved — ProviderDiverseTranslators now yields working verified translators per provider. Wave-2 provider clips show real EN→ES translations through each provider's verified translator (novita "Hola, mundo. Esta es una pequeña prueba." · mistral "Hola, mundo. Este es un pequeño test." · groq "Hola, mundo. Este es un pequeño prueba." · siliconflow "Hola, mundo. Esta es una pequeña prueba.") — validator ffprobe+frame-verified
bridge.go:492 — returns provider-diverse verified low-level LLM clients (one per distinct provider, strongest-first) for callers needing the client rather than a translator
PASS (client path) — builds via the OpenAI-compatible clientBuild (NOT the ValidModels whitelist), so it is NOT affected by the translator-path blocker; go test ./pkg/bridge → ok
PENDING (recording owed — distinct from the catalogue/best-model videos already CONFIRMED)
bridge
ProviderDiverseModels
Library
Implemented
Wired (backs the ProviderDiverse* set)
bridge.go:455 — returns the provider-diverse verified ModelInfo list (one per distinct provider, strongest-first) underlying the ProviderDiverse clients/translators
Covered (pkg/bridge/bridge_test.go, -race green)
PASS (go test ./pkg/bridge → ok)
CONFIRMED — real-use video /Volumes/T7/Downloads/Recordings/helix_translate-bridge-verified-catalogue-translate-20260616-105327.mp4 (Wave 1; model-bridge list renders the provider-diverse verified ModelInfo catalogue — novita/mistral/groq/siliconflow each with real base_url + score — then a real invoke returns "Le pont relie deux rives."; ffprobe 13.71s / 40 frames / 6 distinct frame-hashes; ≥6 frames readable per §11.4.107)
bridge
BestClient (single strongest low-level client)
Library
Implemented
Wired
bridge.go:413 — returns the single strongest verified low-level LLM client for the task (client-level analogue of BestTranslator)
Covered (pkg/bridge/bridge_test.go, -race green)
PASS (go test ./pkg/bridge → ok)
N/A (internal client helper — surfaced via cmd/model-bridge / ProviderDiverse*)
pkg/version
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
version
App version
Library
Implemented
Wired
Singleton version constant (2.3.0)
Not-inventoried
Source-confirmed
N/A
version
Codebase hasher
Library
Implemented
Wired
SHA256 codebase versioning over dirs/extensions w/ exclude patterns
Not-inventoried
Source-confirmed
N/A
pkg/challenge_runner — anti-bluff test harness (uses digital.vasic.challenges)
POST /api/v1/translations with provider config — triggers backend translation
Not-inventoried
Source-confirmed
PENDING
dashboard.html
Cancel session
Web
Implemented
Wired
DELETE /api/v1/translations/:id
Not-inventoried
Source-confirmed
PENDING
Web Dashboard backend handler — pkg/api/dashboard.go
Backend that serves the dashboard page (/, /dashboard, /monitor) and the translations endpoints. Was previously UNROUTED (returned 404); now wired so the dashboard genuinely translates for end users. Source: pkg/api/dashboard.go, commit f6ba5cc.
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
api/dashboard.go
Dashboard page + translations endpoints
Web
Implemented
Wired (was Unwired/404 pre-fix)
Serves the dashboard HTML at /, /dashboard, /monitor and the GET/POST /api/v1/translations path; Start-Translation now runs a real translation through the wired UI path (no longer a 404)
Covered + §11.4.115 RED→GREEN guarddashboard_test.go (default GREEN asserts pages 200 + real translation start; RED_MODE=1 reproduces the pre-fix 404) — go test -race ./pkg/api GREEN this session
PASS (unit/integration) — dashboard_test.go observed green this session; live run produced real translation "Dobro jutro, prijatelju." through the dashboard handler
SKIP (§11.4.3/§11.4.52) — autonomous browser+video capture needs the HelixQA web/video backend, off-limits to this session; tracked migration item owed (a helixtranslate-web-dashboard-* recording requires that backend). NOT a faked confirmation
§11.4.124 honest N/A — NEVER-WIRED test-only scaffolding (do NOT delete; operator-review item).
The pkg/api.Server type (pkg/api/server.go) is a parallel, simpler gin server that is
NEVER instantiated by any cmd/ entry point — git-history (git log --follow, pickaxe
across all history) confirms api.NewServer( was never present in any cmd/ file; it was born
as test-targeted scaffolding alongside *_comprehensive_test.go and is exercised today ONLY by
its own test suites (server_test.go, server_lifecycle_test.go,
server_realhttp_integration_test.go, test/security/input_validation_new_test.go). The two
production HTTP servers are wired through different types: cmd/server → pkg/api.Handler
(the rich /api/v1/* surface below) and cmd/api-server → its own APIServer. Per §11.4.124
it is NOT blind-deleted (a §11.4.115 bug-fix commit 85d9849 + live security/integration tests
depend on it); the correct closure if it is to be removed is the §11.4.90 tracked-Obsolete path
(superseded-by-design-change) WITH prior migration of the dependent security tests onto
Handler — an operator-review decision, not an autonomous removal (§11.4.122). Until then it is
documented here as honest N/A (no user-visible binary ships it, so no real-use video applies)
rather than silently presented as a shipping API surface. Evidence:
docs/qa/deadcode_providers_investigation_20260616_160012/FINDING.md TASK 1.
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
api/server.go
GET /health
API
Implemented
UNWIRED (test-only)
Status + translator name (no cmd ships it)
Test-only callers
Source-confirmed; never-wired (§11.4.124)
N/A
api/server.go
POST /api/translate
API
Implemented
UNWIRED (test-only)
translator.Translate(...) (no cmd ships it)
Test-only callers
Source-confirmed; never-wired (§11.4.124)
N/A
api/server.go
GET /api/languages
API
Implemented
UNWIRED (test-only)
Static language list (no cmd ships it)
Test-only callers
Source-confirmed; never-wired (§11.4.124)
N/A
api/server.go
GET /api/stats
API
Implemented
UNWIRED (test-only)
Returns hardcoded zeros (no cmd ships it)
Test-only callers
Source-confirmed; never-wired (§11.4.124)
N/A
api/server.go
POST /api/upload
API
Implemented
UNWIRED (test-only)
Save uploaded file, detect format (no cmd ships it)
Test-only callers
Source-confirmed; never-wired (§11.4.124)
N/A
api/server.go
POST /api/batch
API
Partial
UNWIRED (test-only)
Returns batch_id queued only; performs NO translation; no cmd ships it
Test-only callers
Source-confirmed; never-wired (§11.4.124)
N/A
REST API — pkg/api/handler.go (primary Handler, gin)
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
api/handler.go
GET /health, /, /api/v1/version, /api/v1/providers, /api/v1/stats
API
Implemented
Wired
Static/info/version/provider list/cache stats
Not-inventoried
PASS — captured (/health, /api/v1/version, /api/v1/providers real provider list returned). Re-confirmed Wave-3c against a live cmd/server TLS instance (port 9443): /health→{"status":"healthy",…}, /api/v1/version→codebase 2.3.0 + git_commit 29aa0e7, /api/v1/providers→real list (openai/anthropic/zhipu/deepseek/ollama/llamacpp); ffprobe 10.63s/790×560 yuv420p, ≥6 frames content-verified, health-frame ≠ providers-frame per §11.4.107 (re-confirmation, no double-count). Wave-9 re-confirmed GET /api/v1/version against live server-w9 :8091 with real substantive data — codebase_version 2.3.0, git_commit c2aa7c841c90df2ac1519c147f61bb82a9614963, go_version go1.26.2, build_time + components map (NOT a bare status); same clip as the Wave-9 batch row (re-confirmation, no double-count). §11.4.6 gap FIXED (commit 388a2eb): GET /api/v1/providers previously returned a static hardcoded list (openai/anthropic/zhipu/deepseek), ignoring config.Translation.Providers — mis-representing capability + availability (FINDING docs/qa/deadcode_providers_investigation_20260616_160012 TASK 2). It now serves the REAL configured provider set (per-provider configured/available/requires_api_key/model from config + <PROVIDER>_API_KEY env), with a complete supported-provider catalogue fallback (configured:false) when nothing is configured; §11.4.115 RED→GREEN polarity test + §11.4.135 guard HTQ-FIX-006. A re-record of the live /api/v1/providers against the rebuilt nezha image is owed before flipping the provider-discovery aspect of this row to a fresh Confirmed.
Distributed or TranslateWithProgress; honors target_lang (FIXED a5e8866 — was hardcoded)
Not-inventoried
PASS — real DeepSeek translation captured (real Serbian output via API; target_lang=spanish FIXED run returns real Spanish). Re-confirmed Wave-3c against a live cmd/server TLS instance (port 9443): curl -k -X POST .../api/v1/translate{"text":"The morning sun rose over the quiet village.","source_lang":"en","target_lang":"es","provider":"deepseek","model":"deepseek-chat"} → real JSON {"provider":"llm-deepseek","translated":"El sol de la mañana se alzó sobre el tranquilo pueblo.","stats":{"Translated":1,…},"session_id":…}; ffprobe 4.24s/790×560 yuv420p, ≥6 frames at fps=2 content-verified, request-frame ≠ response-frame per §11.4.107 (re-confirmation, no double-count)
PASS — video-confirmed (Wave 5: FB2 English input → POST → valid EPUB 2016 bytes, mimetype application/epub+zip, extracted chapter text real Serbian "Стари свјетионич стоје сам на каменитој обали…"; ffprobe 16 frames/8.0s/790×560 yuv420p, input-frame English ≠ output-frame Serbian per §11.4.107). NOTE: handler hardcoded ru→sr (pkg/api/handler.go:84-85), ignores requested langs — tracked fix.
PASS — video-confirmed (Wave 9: live server-w9 :8091 TLS, in-process bridge — 3 EN→SR translated in one request: "The river flows to the sea."→"Река тече до мора.", "Books open windows to the world."→"Knjige otvaraju prozore u svet.", "Time heals all wounds."→"Vrijeme leči sve rane.", 0 errors, provider resolved llm-novita; ffprobe 790 frames/7.52s, early-frame ≠ late-frame per §11.4.107)
PASS — video-confirmed (Wave 5: real JSON both directions — Cyrillic→Latin {"converted":"Sistem treba prevesti","original":"Систем треба превести","target":"latin"}, Latin→Cyrillic {"converted":"Ово је тестни текст","original":"Ovo je testni tekst","target":"cyrillic"}; ffprobe 6 frames/3.0s, original ≠ converted per §11.4.107)
GET /api/v1/monitoring/version/{metrics,alerts,health,dashboard}
API
Implemented
Wired
Version-sync monitoring reads
Not-inventoried
Source-confirmed
PENDING
api/handler.go
POST /api/v1/monitoring/version/drift-check
API
Implemented
Wired
Trigger drift check
Not-inventoried
Source-confirmed
PENDING
api/handler.go
GET /api/v1/monitoring/version/alerts/history
API
Implemented
Wired
Alert history
Not-inventoried
Source-confirmed
PENDING
api/handler.go
POST /api/v1/monitoring/version/alerts/:alert_id/acknowledge
API
Implemented
Wired
Ack alert
Not-inventoried
Source-confirmed
PENDING
api/handler.go
POST /api/v1/monitoring/version/alerts/channels/{email,webhook,slack}
API
Implemented
Wired
Add alert channels
Not-inventoried
Source-confirmed
PENDING
api/handler.go
GET /api/v1/monitoring/version/dashboard.html
Web
Implemented
Wired
Serve embedded dashboard HTML
Not-inventoried
Source-confirmed
PENDING
api/handler.go
POST /api/v1/auth/login, /auth/token
API
Implemented
Wired
JWT auth (if Security.EnableAuth)
Not-inventoried
Source-confirmed
PENDING
api/handler.go
GET /api/v1/profile
API
Implemented
Wired
Return claims (behind auth middleware)
Not-inventoried
Source-confirmed
PENDING
REST API — pkg/api/batch_handlers.go (under /api/v1)
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
api/batch_handlers.go
POST /api/v1/translate/string
API
Implemented
Wired
NewLLMTranslator + Translate
Not-inventoried
PASS — video-confirmed (Wave 7: local ./build/server :8080, real DeepSeek path — JSON translated_text = "Las montañas estaban cubiertas de nieve y el río fluía tranquilamente por el valle.", provider deepseek (a real translation, NOT a bare session_id); frame-verified §11.4.107 — advancing frames, request-frame ≠ response-frame)
REST API — pkg/api/verifier_handlers.go (LLMsVerifier, under /api/v1)
Component
Feature
Category
Implementation
Wiring
Real-use
Tests
Validation
Video-confirmation
api/verifier_handlers.go
GET /api/v1/verified-models
API
Implemented
Wired (only when LLMsVerifier enabled)
List verified models (LLMsVerifier API call). GAP (§11.4.6): LLMsVerifier is DISABLED in the current config, so the route is not mounted and returns HTTP 404. Not a defect — feature-disabled-by-config; needs a verifier-enabled config to confirm
Not-inventoried
GAP — 404 captured (verifier disabled in config; route not mounted)
PASS — fireworks provider real EN→ES translation captured (the llm_provider/pkg/providers/fireworks/fireworks.go client, reached via unified-translator -provider openai -base-url https://api.fireworks.ai/inference/v1 -model accounts/fireworks/models/llama-v3p1-70b-instruct: "The morning sun rose over the quiet village. An old fisherman walked along the shore…" → "Al amanecer, el sol ascendió sobre el tranquilo pueblo. Un anciano pescador paseaba por la orilla…" — distinct verbs (ascendió/paseaba) from Hyperbolic's output (salía/caminaba), proving a genuine per-provider model; Wave-3c, asciinema→agg→ffmpeg, ffprobe 4.54s/790×560 yuv420p + 9 extracted frames at fps=2, ≥6 frames content-verified, start-frame (input only) ≠ output-frame (Spanish) per §11.4.107). Wave-3d added 2 net-new: nvidia + openrouter — nvidia (reached via unified-translator -provider openai -base-url https://integrate.api.nvidia.com/v1 -model meta/llama-3.1-8b-instruct): "At dawn, the sun rose over the quiet village. An old fisherman walked along the shore." → "Al amanecer, el sol se alzó sobre el tranquilo pueblo. Un anciano pescador caminaba a lo largo de la orilla." (helix_translate-provider-nvidia-translate-wave3d-20260616.mp4, ffprobe 6.5s/790×560 yuv420p + 13 frames at fps=2); openrouter (-base-url https://openrouter.ai/api/v1 -model meta-llama/llama-3.1-8b-instruct): same EN source → "Al amanecer, el sol se alzó sobre el tranquilo pueblo. Un anciano pescador paseaba a lo largo de la orilla." (helix_translate-provider-openrouter-translate-wave3d-20260616.mp4, ffprobe 5.5s/790×560 yuv420p + 11 frames at fps=2) — distinct verbs (caminaba/paseaba) prove genuine per-provider models, each start-frame (English only) ≠ output-frame (Spanish) per §11.4.107, ≥6 frames content-verified. The remaining 40 provider dirs remain Source-confirmed, PENDING a per-provider recording.
SSH auto-install, model pull, lifecycle; GPU/CPU/RAM detection; distributed llama.cpp RPC
Not-inventoried
Source-confirmed
PENDING
vision_engine
OpenCV (stub + build-tag real)
Submodule
Partial
Wired
OpenCV-based CV — stub by default, real impl behind build tag
Not-inventoried
Source-confirmed; real impl gated behind build tag
PENDING
vision_engine
Config / i18n
Submodule
Implemented
Wired
Configuration (HELIX_VISION_PROVIDER etc.), i18n
Not-inventoried
Source-confirmed
N/A
Honest gaps & caveats (§11.4.6)
Video coverage is 66 feature rows. 66 features are confirmed by real, ffprobe- and content-verified recordings. §11.4.6 cell reconciliation: this caveat previously read 30 — a sync lag behind the authoritative running total carried in the Anti-bluff note (top of file) and the Coverage-summary Video-confirmed row (both 66); it is corrected here to the true running total 66 (54 prior-authoritative + 4 Wave 7 + 5 Wave 8 + 3 Wave 9 = 66). For the full per-wave / per-feature mapping (Waves 1–9 of 2026-06-16, the genuine row-flips vs headline-only re-confirmations, the §11.4.138 multipass demotion, and the honest non-confirmable gaps) see the authoritative Anti-bluff note + the Coverage-summary Video-confirmed row — this caveat is reconciled with both per §11.4.60. DOCX output and PDF output, formerly real-artifact-only, are now video-confirmed. Every other runtime/user-visible feature is PENDING a recording or N/A. This is stated, not hidden.
1a. DOCX output is Implemented (Rev 3, commit 87cd2be) — pure-Go OOXML writer; a real run produced a Microsoft Word 2007+.docx; now video-confirmed (Wave 2) by helix_translate-docx-output-20260616-112943.mp4 (openai/gpt-4, real Spanish in the .docx). PDF output is Implemented (Rev 4, commit fb265e7) — pkg/ebook/pdf_writer.go (Book→HTML5→weasyprint→Cyrillic-faithful PDF), wired into unified-translator (.pdf output, main.go:1160); pkg/ebook PDF tests green; now video-confirmed (Wave 2) by helix_translate-pdf-output-20260616-113057.mp4 (openai/gpt-4, valid PDF with real Spanish). Output formats are EPUB/FB2/TXT/HTML/MD/DOCX/PDF.
1b. -multipass ≠ -verify. The unified-translator -verify flag runs the CLI's own per-step output check; the separate -multipass flag (d53e085) invokes the pkg/verification multi-pass polisher engine — now wired and video-confirmed.
1c. cmd/translator, cmd/translate-ssh, cmd/ebook-translator are now Obsolete (→ Fixed.md), Rev 7 — RECONCILED with the per-binary table rows (lines ~136/218/233) which mark all three Obsolete (→ Fixed.md) with Reason=feature-removed. The local-runtime + SSH-local translation paths were REMOVED in bridge phase-2 R-2..R-4 (operator-approved 2026-06-15 D1/D2 + R-4 "keep distributed API, remove SSH-local"); the default path now sources the LLMsVerifier bridge (pkg/bridge). These are therefore NO LONGER open STUB / OPERATOR-BLOCKED items — the earlier "STUB / OPERATOR-BLOCKED" wording was stale and is corrected here. The KEPT distributed-API translation path is NOT Obsolete.
1e. GET /api/v1/verified-models returns 404 — LLMsVerifier is disabled in the current config so the route is not mounted (feature-disabled-by-config, not a defect).
1f. cmd/cli title path-leak FIXED (d53e085) + write-safety FIXED (87cd2be, atomic write/no-clobber, guard cmd/cli/no_partial_output_test.go).
Tests = Not-inventoried for all rows: the raw inventory did not perform per-feature test attribution. The repo carries extensive _test.go + test/ suites, but mapping each to a feature is future work — recorded as honest unknown, never guessed Covered/None.
4 STUB endpoints return mock data and do NOT do the named work: POST /api/v1/translate/ebook, POST /api/v1/preparation/analyze, GET /api/v1/preparation/result/:id, POST /api/v1/translate-with-verification.
1 Partial endpoint: standalone pkg/api/server.go POST /api/batch returns a queued batch_id only and performs no translation.
Static/hardcoded returns (not stubs, but not live): GET /api/stats (zeros), GET /api/v1/status/:id (hardcoded completed), GET /api/v1/metrics (static), monitor-server GET /api/v1/status (static monitoring_active), several static language/provider lists. Marked Implemented with the static behaviour stated in Real-use.
vision_engine OpenCV is Partial — stub by default, real CV only behind a build tag.
UNCONFIRMED LLM determinations carried verbatim from the inventory for thin OpenAI-compatible providers, several llms_verifier packages, challenges/userflow evaluators, and doc_processor/archdoc — noted inline rather than asserted.
✅ Wave-1 DISCOVERED BLOCKER — RESOLVED (Wave 2, §11.4.4 / §11.4.107 / §11.4.108 / §11.4.138). The default bridge TRANSLATOR-path runtime failure described below is FIXED: the Wave-2 recordings (DOCX output, PDF output, gRPC translate, and the 4 per-provider unified clips novita/mistral/groq/siliconflow) are real-use videos validator-verified (ffprobe+frame) showing the verified-bridge translator path producing real target-language output — the exact path that failed in Wave 1 now works, which is why those rows flip to CONFIRMED this wave. The historical forensic record of the blocker is preserved below for §11.4.6 / §11.4.138 traceability. [HISTORICAL — Wave-1 finding, now fixed] While recording Wave 1 (DOCX/PDF/gRPC), a real run of unified-translator (and, by source, the gRPC backend pkg/grpc/core_translator.go which uses BestTranslatorFunc) FAILS at translate time with: bridge: no verified translator available: model 'Sao10K/L3-8B-Stheno-v3.2' is not valid for provider 'novita'. Valid models: [meta-llama/llama-3.1-70b-instruct mistralai/mixtral-8x7b-instruct]. Root cause (FACT, §11.4.6):VerifiedFactory.CreateTranslatorWithFallback / CreateTranslator build the verified translator via NewLLMTranslatorWithConfig, which re-validates the model against the STALE hardcoded ValidModels whitelist (pkg/translator/llm/llm.go:227-241); a genuinely LLMsVerifier-verified model whose ID is not in that whitelist is rejected. The BestModel/invoke path (used by model-bridge invoke + Wave-1 videos) correctly BYPASSES this whitelist (its own code comment, bridge.go:250-258, documents exactly why a verified model must not be gated on the whitelist) — and so does NOT exhibit the bug. Impact: unified-translator (all outputs: TXT/HTML/EPUB/FB2/MD/DOCX/PDF) + gRPC translate are non-functional via the default bridge path until fixed. This BLOCKS Wave-1 videos #2 (DOCX), #3 (PDF), #4 (gRPC), and the per-provider unified-path; recording any of them as "working" would be a §11.4.107 bluff, so they are honest SKIPs this wave. Fix direction (not yet applied — needs its own §11.4.115 RED→GREEN + §11.4.142 independent review + §11.4.145 impact research + §11.4.40 full-suite retest): route VerifiedFactory's verified-translator construction through the ResolvedProvider resolver (FactoryProvider="openai" + BaseURL) the way bridge.realClientBuild already does, bypassing the ValidModels whitelist for already-verified models. Tracked as a Wave-2 / dedicated fix item (severity HIGH — release-critical default path).