Backfill library.artwork_url via LML (precondition for #643)

[jakebromberg]

Context

Precondition for #643 (and for fix surface 1 of #628). The bin-pick branch of addEntry denormalizes library.artwork_url onto the new flowsheet row at INSERT time, eliminating the metadata race for bin-picks across all client surfaces — but only for library rows that have artwork_url populated.

Per the precondition comment on #628 (issuecomment-4338014422), the production query quoted there reports that the current population is 155/64163 rows (0.2%). So #643's denormalization SELECT is shipped (held in draft) but does ≈nothing in practice today. This issue tracks the backfill that gives it teeth.

Proposal

A one-shot backfill job under jobs/library-artwork-backfill/, modeled after jobs/flowsheet-dj-name-backfill/ (the canonical pattern; see Backend-Service CLAUDE.md for the rule that bulk DML is always a separate job, never inside a migration).

Approach:

1. For each library row with artwork_url IS NULL and a non-null artist_id + album_title, call LML's /lookup endpoint (same call path as metadata.service.ts:fetchMetadata).
2. On a single high-confidence match, set library.artwork_url = artwork.artwork_url.
3. Batched UPDATEs with BACKFILL_BATCH_SIZE (default 5000), synchronous_commit=off per the bulk-update playbook, idempotent (WHERE artwork_url IS NULL filter resumes naturally).
4. Phase A observability tags (repo, tool=library-artwork-backfill, step, run_id).
5. Cooperate with LML rate limits — the job dwarfs LML's request volume from the live insert path, so will need throttling.

Out of scope

• Streaming-link backfill (spotify_url, apple_music_url, etc. on library). The current schema (shared/database/src/schema.ts:280-290) only has artwork_url on library; the other LML metadata fields land on flowsheet per the inline-metadata model. If we want to denormalize more fields onto library, that's a schema change worth a separate proposal.
• Backfilling flowsheet.artwork_url from history (covered by [Epic] Historical metadata backfill for ~1.86M flowsheet rows #631).

Acceptance criteria

• jobs/library-artwork-backfill/ exists, builds, runs in dry-run mode locally, and emits Phase A logs.
• One-shot run on a representative slice (e.g. top-1k most-played albums) reaches a sane hit rate (≥50% on the heavy-rotation slice; the long tail will be lower).
• After a full pass: library.artwork_url non-null fraction is materially higher than the 0.2% baseline reported on Re-emit liveFs SSE event after metadata enrichment UPDATE lands #628. Quote the post-run number in the closing comment.
• Denormalize library.artwork_url onto flowsheet on bin-pick #643 (the bin-pick denormalization SELECT) is then unblocked and can be merged.

Cross-repo links

• Re-emit liveFs SSE event after metadata enrichment UPDATE lands #628, Denormalize library.artwork_url onto flowsheet on bin-pick #643 — the consumers of this backfill's output.
• LML /lookup — the same endpoint the live insert path already uses.
• jobs/flowsheet-dj-name-backfill/ — reference implementation for the one-shot job pattern.

[jakebromberg]

Triage finding: this duplicates #637, which was filed earlier today and is better-scoped. #637 narrows the backfill set to library rows whose artist has a Discogs identity (artists.discogs_artist_id IS NOT NULL), expected ~18,500 rows post-reconcile, where LML lookups can resolve from cached Discogs metadata without live API hops. This issue's body proposed the broader "all NULL rows" set without that gate.

The "precondition for #643 / fix surface 1 of #628" framing is also captured by #637 implicitly (any non-zero population lift unblocks the bin-pick denormalization), so the cross-link can move to a #637 comment.

Recommend closing this and folding the #643/#628 reference into a comment on #637.