feat(transcript): declarative extractor tier for flat event streams

[slowdini]

Summary

Closes #139 (part of #122, data-driven harness descriptors phase 2).

Adds a declarative [transcript.extract] tier to harness descriptors: flat event streams (every tool event self-contained) now get full transcript ingest — transcript_check grading, parsed invocations in run.json, token/duration backfill into timing.json — from descriptor data alone, with zero Rust. Streams that need cross-event state (keyed tool_use/tool_result joins, content coercion) remain named code capabilities; claude-stream-json is untouched.

User-facing changes

• New descriptor surface: [transcript] now takes exactly one of parser (named capability) or extract (declarative tier). The extract block holds exactly the ticket's five primitives — equality where filter, last-match final-text pick, flat tool-item mapping, token sum, duration via field pick or timestamp spread — all subtle rules fixed and non-configurable.
• Codex flipped: harnesses/codex.toml ingests through the extract block. codex-items stays as the reference implementation and remains referenceable by BYOH descriptors; a differential test pins the extract spec byte-equivalent to the code parser on the full fixture corpus.
• Validation: new load-time invariants (exactly-one of parser/extract, non-empty extract block, exactly-one duration variant) with actionable messages, including the layering trap: overlays can't delete an inherited parser, so moving a harness to extract needs a new label.
• Docs: worked example + primitive reference in docs/byoh.md ("The declarative extractor"), two-tier story in docs/progressive-enhancements.md, updated harness init template scaffold, and the descriptor PR template now counts extract blocks as data-only.

Before/after (BYOH)

Before — a harness with a novel flat stream had no transcript ingest without a Rust parser PR:

--harness cool-custom-harness declares no transcript parser — transcript_check
assertions will grade as unverifiable ... tokens/duration go unrecorded.

After — pure descriptor data:

[transcript]
events_filename = "cool-events.jsonl"

[transcript.extract.tools]
where = { kind = "tool.call" }
name_field = "tool"
result_coalesce = ["output"]

→ no warning; run.json carries parsed tool_invocations, timing.json carries tokens/duration (proven end-to-end by the new tests/run/byoh.rs integration test).

Verification

• cargo test — 619 tests pass (453 lib / 96 cli / 70 run), including the new engine units, the codex differential-equivalence test, validation rejects_* suite, BYOH end-to-end, and lint CLI tests; all existing codex pipeline/golden tests pass unchanged through the extract engine (no golden re-bless needed)
• cargo fmt --check / cargo clippy --all-targets -- -D warnings — clean
• Manual: harness lint codex passes; harness show codex round-trips the resolved extract block as loadable TOML; harness list still declares transcript for codex

Schema / docs callouts

• schema/harness-descriptor.schema.json: transcript.required relaxed to events_filename; new fully-described extract sub-schema (exactly-one rules live in Rust invariants, matching the slug_template/slug_capability precedent)
• Follow-up candidate: OpenCode has no transcript ingest today — if its stream is flat, it's now a pure-data descriptor change

🤖 Generated with Claude Code