Transcript Fetch Authority

Carve-Out Claimed

• jbaruch/coding-policy: testing-standards Platform-Bound Untestable Carve-Out.
• This rule is the authority of record satisfying precondition 3 — it names the carve-out, the exempt artifact, and where the manual validation procedure lives.
• Qualifying condition: mlx-whisper requires Apple Silicon and cannot install on the project's Linux CI runners.

Covered Artifact

• The mlx_whisper imports, actual provider.transcribe() call, native model access, spectrogram construction and detect_language() call inside _transcribe_with_mlx() in skills/vault-ingress/scripts/local_media_transcription.py are exempt.
• Both YouTube --method whisper and non-YouTube --audio delegate to that worker. Their orchestration, media admission, resource supervision, download, result validation, cleanup, and receipt writes are not exempt.
• Sampled-word acquisition delegates to the same native boundary. Read its receipt contract and validation procedure in skills/vault-ingress/references/sampled-word-evidence.md.
• mlx-whisper is declared as the optional whisper extra in pyproject.toml, never a base dependency. The caption path and every validator work without it.
• Route non-YouTube talks through fetch-transcript.py --audio. Never call mlx_whisper.transcribe() from skill prose.

Precondition 1 — CI-Runnable Pieces Are Extracted and Tested

• validate_transcript(), build_quality_policy(), count_words(), resolve_video_id(), segments_to_text(), receipt validation, and atomic writers are deterministic and carry tests in tests/test_fetch_transcript.py and tests/test_transcript_timing.py.
• local_media_contract.py, local_media_evidence.py, local_media_process.py, local_media_download.py, and local_media_transcription.py each have a corresponding tests/test_local_media_*.py suite. Native synthetic workers exercise authenticated framing, process limits, capped pipes, container inspection, source-generation changes, and cleanup without provider access.
• YouTube duration comes from identity-bound yt-dlp metadata. Local duration comes from supervised ffprobe and binds the exact source SHA-256.
• Tests cover empty/error/VTT/non-speech artifacts, the fixed and WPM floors, low---min-words bypass attempts, source-duration matching, Unicode word counting, both segment shapes, invalid UTF-8, exact provenance, no-segment quality receipts, CRLF/LF byte drift, transactional rollback, enumerated optional-library failures, malformed-timing downgrade, media snapshot mutation, destination symlinks, VTT path safety, and caption fall-through.
• Caption and Whisper imports, constructors, API calls, and provider result objects are optional-lane boundaries. ImportError, OSError, RuntimeError, AttributeError, TypeError, ValueError, and KeyError make that lane unavailable; auto mode continues to its next source. Unknown exception classes reach the tokenized outer process boundary. Provider stdout routes to stderr so stdout remains one JSON object. Process-control signals are not swallowed.

Transcript Quality Authority

• transcripts/<id>.segments.json owns timing and acquisition provenance only. It never owns quality policy.
• Its current closed schema is version 2 with exactly schema_version, transcript_sha256, source, provenance, and segments. YouTube caption and Whisper timing bind an 11-character video ID plus trusted provider duration; local Whisper timing binds exact media SHA-256 plus trusted probe duration; VTT timing binds a safe transcript-relative regular-file path, exact artifact SHA-256, and final cue extent.
• Every schema-v2 segment must be canonical, its joined text must equal the transcript modulo Unicode whitespace layout, and its time range must fit the source-owned duration (or exact VTT cue extent). Schema v1 is archival only: it cannot supply timing or relabel provenance and must be regenerated from a source whose owner and bounds can be proved.
• transcripts/<id>.quality.json owns the exact policy that validated the transcript. Its closed schema contains schema_version, the SHA-256 of exact .txt bytes, exact policy, and exact provenance.
• Policy is exactly {schema_version: 1, min_words: int, duration_seconds: number|null}. A low --min-words never lowers the fixed 400-word floor or a trusted duration-derived floor; any value above the derived floor tightens it.
• Provenance is exactly one of {kind: "fixed_default"}, {kind: "youtube_duration", video_id, duration_seconds}, or {kind: "local_media_duration", media_sha256, duration_seconds}. A duration-bearing provenance value must equal the policy duration exactly.
• --duration-seconds is caller expectation, not authority. It can be used only when it matches provider/yt-dlp duration for the exact YouTube ID or ffprobe duration for the exact local-media digest. Worker-returned or talk analysis metadata can never lower a threshold.
• Missing timing does not invalidate a current quality receipt. Missing or stale quality authority does make a transcript ineligible for the current pattern-scoring generation until the fetcher validates it and writes a receipt.
• Both receipt readers hash raw transcript bytes. Any byte replacement, including CRLF→LF with identical decoded words, invalidates both receipts.
• The talk's recorded transcript_source is canonical. A sidecar may confirm matching owner-bound timing, but captions cannot promote manual, Whisper, or unknown text, and a timing source never rewrites the talk's provenance.
• Existing transcript bytes are never replaced without explicit --force. Invalid existing text fails closed and asks for that authorization. A caught bundle-write failure restores the prior transcript and both receipts exactly. Missing, malformed, mismatched, or over-bound optional timing removes stale timing while valid transcript text and its quality receipt still commit.
• Local --audio admission rejects unavailable or unsupported source metadata before byte I/O. A bounded worker copies and hashes one admitted generation into an owner-cleaned private read-only snapshot for ffprobe.
• Whisper reuses established media or video probe facts without another copy, hash, or probe. Its bounded worker checks the original source's pathname and open-descriptor generation before and after transcription.
• The fetcher checks source generation after staging and before replacing a transcript bundle or refreshing existing quality. Source drift or worker failure leaves prior artifacts untouched. Failed local probing never replaces an existing quality receipt with fixed-default provenance.
• Transcript, timing, and quality final-component symlinks are forbidden, including dangling links. VTT imports require lexical and resolved containment below the transcript directory, no symlink components below that root, and a regular artifact before read/open.

Precondition 2 — Manual Validation Procedure

Run against a talk whose caption track is disabled, on Apple Silicon with the whisper extra installed. First require the configured runtime:

"{python_path}" skills/vault-ingress/scripts/check-runtime.py --lanes core,whisper,youtube-download --require-lanes core,whisper,youtube-download

1. "{python_path}" skills/vault-ingress/scripts/fetch-transcript.py {youtube_id} --out /tmp/{youtube_id}.txt --method whisper --existing-source unknown
2. Observe: exit 0, one JSON object on stdout with "method": "whisper", "timed_path": "/tmp/{youtube_id}.segments.json", "quality_path": "/tmp/{youtube_id}.quality.json", and a words count plausible for the runtime (conference delivery is 110–160 wpm).
3. jq -e --arg id "{youtube_id}" '.schema_version == 2 and .source == "whisper" and .provenance.kind == "youtube_whisper" and .provenance.video_id == $id and (.provenance.duration_seconds > 0) and (.segments | length > 0) and all(.segments[]; .start_seconds >= 0 and .end_seconds > .start_seconds and (.text | length > 0))' /tmp/{youtube_id}.segments.json exits 0.
4. test "$(jq -r .transcript_sha256 /tmp/{youtube_id}.segments.json)" = "$(shasum -a 256 /tmp/{youtube_id}.txt | awk '{print $1}')" exits 0.
5. jq -e '.schema_version == 1 and (.transcript_sha256 | length == 64) and .policy.schema_version == 1 and (.policy.min_words >= 1) and ((.policy.duration_seconds == null and .provenance == {"kind":"fixed_default"}) or (.policy.duration_seconds == .provenance.duration_seconds))' /tmp/{youtube_id}.quality.json exits 0, and its transcript_sha256 equals shasum -a 256 /tmp/{youtube_id}.txt.
6. Confirm /tmp/{youtube_id}.txt holds prose, not [Music] markers or a traceback.
7. Re-run with YT_DLP set to a nonexistent absolute path and a fresh output path. Expect exit 1, a path-neutral stderr failure with the ytdlp_dependency_unavailable reason, one JSON object with "ok": false, and no transcript or receipt at that output path. Removing only yt-dlp from PATH is insufficient: the fetcher deliberately resolves the configured interpreter's console script before the compatibility PATH fallback.
8. "{python_path}" skills/vault-ingress/scripts/fetch-transcript.py local-talk-label --audio <local-file> --out /tmp/a.txt --existing-source unknown on a non-YouTube recording: exit 0, "method": "whisper", "timed_path": "/tmp/a.segments.json", "quality_path": "/tmp/a.quality.json", prose at the output path, timing schema v2 with local_media_whisper provenance, and timing plus quality provenance whose media_sha256 equals the exact input-media digest.

A pass requires all eight checks.

Scope Limits

• The carve-out covers only the named Apple-Silicon imports and native calls. It does not extend to another script, dependency, or the caption path.
• Adding a second exempt artifact requires naming it here AND documenting its own validation procedure. Adding one without both invalidates the precondition.
• "Hard to install in CI" does not qualify — see jbaruch/coding-policy: ci-safety Install, Don't Skip. mlx-whisper cannot run on the runner's architecture and qualifies for the carve-out.