feat(enrichment): record status/errors and trust high-confidence AcoustID
Docker Build & Publish / build (push) Has been cancelled
Docker Build & Publish / push (push) Has been cancelled
Docker Build & Publish / Prune old image versions (push) Has been cancelled

Two related gaps surfaced from "uploaded a track, nothing changed / no status":

- A track could stay stuck on `pending` forever (an unexpected worker error
  rolled back the run without recording anything), and `failed` carried no
  reason. Add `tracks.metadata_error` + `tracks.enriched_at` (migration), stamp
  the outcome in apply_enrichment, add TrackRepository.mark_enrichment_failed,
  wrap enrich_task to persist crashes as `failed` in a fresh session, and emit a
  human-readable no-match reason. Expose metadata_error/enriched_at in TrackOut.

- The tag-first merge let junk embedded tags (e.g. "Music Track"/"Sound_13958")
  override even a 0.99-confidence AcoustID match. Add acoustid_trust_score
  (default 0.85): above it the acoustic identity wins for title/artist/album/
  year, tags are fallback; below it, tag-first as before.

Add a license-free real-file fixture (Scarlet Fire / Otis McDonald) whose junk
tags AcoustID overrides, with an always-on tag-reader test plus fpcalc/AcoustID/
network-gated identity + full-pipeline tests (skip on host, run in the container).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Senko-san
2026-06-13 13:29:08 +03:00
parent 30cb8901f2
commit 73d7da440f
17 changed files with 468 additions and 33 deletions
+20
View File
@@ -0,0 +1,20 @@
# Test fixtures
## `scarlet_fire_otis_mcdonald.mp3`
"Scarlet Fire" by **Otis McDonald** — a royalty-free / license-free track
(YouTube Audio Library; distributed via Pro-Sound.org). Used as a real-world
audio fixture for the enrichment pipeline.
What makes it a good fixture: its **embedded ID3 tags are junk**
(`title=Sound_13958`, `artist=Music Track`, `album=Музыка`, `genre=Hip Hop & Rap`)
while AcoustID identifies it with very high confidence as *Scarlet Fire /
Otis McDonald*. So it exercises both:
- the offline tag reader (deterministic, always runs), and
- the "trust a high-confidence AcoustID match over junk tags" path
(`acoustid_trust_score`), which only runs when `fpcalc` + an AcoustID API key
+ network are available — see `tests/test_real_file_enrichment.py`.
Because it's license-free, it may also seed a built-in demo track for fresh
instances.
Binary file not shown.