v2.13.0 — visible failures, real retries, and the wrong-version fix

[rancur]

Three of the reported feature requests turned out to be different problems than described. Measuring them first is what kept this release small.

Reported	Actually
"Missing Tracks page is empty — why does it exist?"	A bug. It asked for per_page=500 from an endpoint capped at 200, got a 422, and rendered an empty table.
"Post-processing isn't running"	It is. 399/400 recent tracks have cue points; 4 of 5,612 lack them. This was a visibility gap.
"Improve fingerprint mismatch"	Not a fingerprint check. It's duration math, and the real fault was upstream in the matcher.

Highlights

api.ts was hiding server errors. It threw away the response body and raised Error("API error: 422"), so every caller's catch { setRows([]) } turned a specific, loud failure into "you have nothing". That is the only reason the Missing Tracks bug survived. The client now preserves status + FastAPI's detail and retries only transport faults and 5xx — never a 4xx, which is deterministic.

The wrong-version fix. The file-index title+artist matcher accepted any file whose title and artist agreed, so extended mixes, radio edits and live takes all matched; verify then rejected them on duration and parked them as errors. Measured across 265 such matches on the live library: 155 within 5s, 110 beyond it, 95 of those by more than 15s. The populations separate cleanly, so matches now require durations to agree within 5s (configurable, fails open when either is unknown).

Bulk retry that isn't inert. already_attempted() treats any prior fallback_attempts row as "we tried this", so resetting the pipeline columns alone produced a track that walked straight back to the same error without re-contacting Soulseek. Bulk retry clears them; the pre-existing single-track retry had the same defect and is fixed too.

One real bug in post-processing: "Auto-Analyze After Sync" silently disabled cue generation, tag lookup and cloud upload as well, despite each having its own checkbox and the toggle's description mentioning only BPM/key.

Also

• Soulseek in service health, probed by the worker and persisted to app_config (sync-api cannot import worker code)
• Dashboard month drill-down; the month filter is a half-open range so it can use an index — and tracks had no index at all beyond its implicit PK one
• The health probe called GET /v1/tracks (Lexicon's entire library) every 10 seconds; now /v1/playlists
• Bulk operations write one activity row instead of one per track

CI

The repo had ~20 test files and no workflow that ran them. There is one now — pytest (not unittest, which silently collects zero tests from the bare-function suites and reports success), plus a web typecheck and a real next build, which is what catches a useSearchParams() outside a Suspense boundary.

It immediately exposed three latent problems, all fixed here: two test modules fighting over db.py's module-level DB_PATH (decided by whichever imports first), and a Tidal test that had been failing for anyone outside the container because tiddl is installed by the Dockerfile but absent from requirements.txt.

Verification

• sync-api: 37 passed
• sync-worker: 279 passed, 0 failed
• sync-web: typecheck clean, all 11 pages build
• Duration threshold, error totals and coverage percentages all measured against the live instance

🤖 Generated with Claude Code

https://claude.ai/code/session_01HmUiLHPmKoz215WAWV5eHe