fix: give a track an id the next ten thousand tracks cannot take

[ralyodio]

Fixes SQLITE_CONSTRAINT: UNIQUE constraint failed: tracks.id when building a reel from clips.

Cause

newId built every id from the first eight hex characters of a UUID — 32 bits, 4.3 billion values. Ample for projects, videos and reels; far too few for tracks, the one table that fills up on its own. Detection writes a row per track, and a single production video already carried 8,394 of them (the number is in analyze.ts, in the comment explaining why a score-only re-run reports zero). Vetting a shirt before a track may claim whoever is standing there (#50) trades spliced identities for fragments, so it multiplies that count rather than reducing it.

A project assembled from a dozen clips holds the sum, because detection clears the tracks of the video it is re-analysing, not of the project. That is why building from clips is where this surfaced.

The birthday bound, measured

tracks	P(a run dies)
8,394	0.8%
25,000	7%
50,000	25%
100,000	69%

Not a rare collision — the expected outcome of the flow that accumulates the most.

Fix

Ids carry 64 bits (randomBytes(8)), so the same 100k rows collide with probability 3e-10.

Widening rather than retrying on the constraint is deliberate: a retry loop would hide the rate instead of removing it, and it would have to sit on every insert to be honest about where else this was possible. Ids already written stay valid and stay readable — eight characters and sixteen cannot collide with each other — so no migration is owed and no existing project has to be re-analysed for the fix to hold. Re-analysing a broken project now succeeds because it re-inserts with wide ids.

Checked for format assumptions: parseFrameFileName splits on the last _ (length-agnostic), and the upload CLIENT_ID regex already accepted {8,32}. One test pinned the old width and is updated.

Verification

packages/core/src/ids.test.ts generates 200k ids and expects 200k distinct. Against the old scheme it returns 199,995 — confirmed by temporarily reverting the implementation and watching the test fail.

Full suite 691 passed / 8 skipped; pnpm lint and pnpm -r typecheck clean.

🤖 Generated with Claude Code