Whisper pool, DB contention, differential keep override (2.96.5)

[ttlequals0]

2.96.5: database contention and a keep-map hole

Feed discovery wrote a whole feed's episodes inside one transaction holding SQLite's single write lock. On a large archive feed that ran to 80 seconds. Every other writer waits on a 30 second timeout, and past that it fails rather than queues. An RSS sweep, pattern-match recording and podping stamps all reported "database is locked", and one episode lost 15 minutes of processing. Discovery now writes in chunks and takes its existing-episode snapshot outside the write path.

Two smaller contention fixes: an LLM call bumped three global counters as three statements inside its transaction, now one; and a refresh committed once per feed even when the guarded update matched no row.

A marker whose category resolves to keep is now cut when its span overlaps a measured cross-fetch differential region (#728). cross_promo covers both a guest plugging their own show and a paid injected ad for another podcast. The category cannot separate them. The differential can, since injected audio differs between two fetches and host content does not. Both passes apply it, and a toggle under Ad Detection turns it off.

Log records no longer split across lines. A logged prompt body carries newlines, and each became its own log line with no level prefix, so a scraper re-sniffed the level and filed a line reading "CRITICAL:" as a critical entry.

Five releases batched into one PR. Detail in CHANGELOG.md.

2.96.0 to 2.96.1: transcript coverage

Transcripts were losing 7-17% of each episode's speech. Whisper was seeded with a 312-character sponsor-vocabulary initial_prompt, and faster-whisper's batched pipeline prepends it to every VAD clip with no decode fallback. On roughly one clip in ten, large-v3 answered the prompt instead of the audio, and the rest of that clip was dropped. The prompt and its echo scrubber are gone on both backends. Coverage rose from 83-93% to 95.6-96.0%.

Also: the batch-size ceiling no longer sticks after a transient CUDA error; the recents feed (#721), an opt-in combined feed at /recents; publish dates stored as true UTC; chapters listed in episode descriptions (#720); an episode Download menu.

2.96.2: Whisper pool and rate-limit hold

The hold parked each 429'd episode as deferred and leaned on a maintenance pass to return it. The gate was also skipped by user requests and by API-started runs. A 429 now returns the episode to pending in place, keeping its priority, and the gate covers every entry point.

The Whisper pool is opt-in and off by default. Against a remote backend that accepts parallel requests, it runs several episodes at once under a single cap on in-flight requests. Both dials are settings.

2.96.3: stats counted reprocesses twice

The dashboard and per-podcast stats summed a join, so a reprocessed episode's time saved counted once per run (#727). Length, time saved, and episode count now come from the episode's current row. Per-run costs still count every run. The lifetime counter credits only the change since it was last credited, with a backfill.

2.96.4: probes

A rate-limit hold re-probes instead of waiting out a reset that may no longer be true, through llmUsageUrl or a single minimal completion.

A Whisper backend's optional health endpoint is sampled for replica count, model, and total concurrency, with a suggested cap. It is bounded: only while the pool is active, cached 120 seconds, a 15 second sampling budget, one probe per backend. When sampling cannot prove it wrapped the replica set, the UI says "at least N".

Version

2.96.5

Test plan

• pytest tests/ (6827 passed), ruff clean
• tsc --noEmit, npm run lint, production build, vitest (832 passed)
• CodeQL clean on actions, javascript-typescript, and python
• GPU image scanned (0 HIGH, 0 CRITICAL) and pushed; CPU workflow green on both arches
• Deployed; /api/v1/health reports 2.96.4; the health probe resolves a three-replica backend with no mismatch