Description

The audit skill already generates youtube title/description/tags/chapters into _data/interviews.yml and _data/video_assets.yml. This task pushes that existing metadata to the live YouTube videos via the Data API — generation is out of scope here, this is sync only. Outward-facing: nothing goes live without an approved diff.

Acceptance Criteria

• #1 Reads title/description/tags/chapters from _data
• #2 Diffs against live YouTube state and produces a dry-run report of changes
• #3 Applies updates only after explicit human approval of the diff
• #4 Idempotent — re-running with no changes is a no-op

Implementation Plan

Sync already-generated YouTube metadata to the channel (no generation)

Source-of-truth discrepancy (READ THIS)

The task text says metadata lives in _data/interviews.yml + _data/video_assets.yml. Investigation shows it actually lives in _data/transcripts/*.yml under the youtube: block (title, description, tags, chapters). chapters exist ONLY there. Read from the transcripts, not from interviews/video_assets. (video_assets.yml per-platform description is stale platform copy, not the generated metadata.)

Data flow / join

1. For each _data/transcripts/<stem>.yml with a youtube: block → transcript_id = <stem>.
2. Resolve _data/video_assets.yml item where transcript_id == <stem> → platforms[] where platform == youtube → asset_id (the live YouTube video id). Skip assets with no youtube platform.
3. Desired video snippet: title → snippet.title; description + chapters → snippet.description; tags → snippet.tags.
   • Chapters are NOT a field. YouTube renders chapters from timestamped lines in the description BODY. Append/merge the chapters list as lines (00:00 Title) into the description text; first line must be 00:00, need ≥3 chapters, each ≥10s apart — validate and warn if the generated set violates this.

Endpoints + quota

• Read: videos.list(part=snippet, id=<asset_id>) = 1 unit each.
• Write: videos.update(part=snippet) = 50 units each.
• Corpus aggregate: ~178 primary-youtube assets × (1 + 50) ≈ 9.1k units for a full pass — nearly the entire 10k/day default; the full 367 youtube occurrences would exceed it. So: dry-run is list-only (cheap, ~few hundred units) and can run whole-corpus; apply must batch across days or request a quota increase.

Dry-run + approval gate (per TASK-245 two-phase pattern)

• youtube_metadata dry-run job: list live snippet, diff desired-vs-live (normalize whitespace/tag order), emit a per-video diff artifact + live-state fingerprint (hash of live title+description+tags). No writes.
• Human approves. Apply job re-fetches, confirms fingerprint unchanged, then videos.update. Abort on drift.

Idempotency (AC#4)

• Desired snippet is deterministic from _data. Skip any video whose live snippet already equals desired (no-op). Fingerprint-idempotent job key = asset_id + sha(desired_snippet).

Verification

• Re-run dry-run after apply → zero diffs (proves no-op idempotency + successful push).

Files

Reads: _data/transcripts/*.yml, _data/video_assets.yml. Uses bin/lib/youtube_client.rb (245). No repo writes.

Definition of Done

• #1 AC criteria is completed and the change has been verified