PR #23 merged: Restore transcript punctuation (initial_prompt) + ASR_THREADS=16

Punctuation fix (the regression)

crunch's large-v3-turbo was emitting all-lowercase, zero-punctuation text. Confirmed live vs a punctuated reference. This predates the recent deploys (the no-punctuation transcript was saved at 10:50, before the first deploy at 12:29) — it's a pre-existing faster-whisper/turbo behaviour, not collateral from this session's work.

Fix: seed the decoder with a short punctuated initial_prompt ("Okay, so let's take a look at this.") — it's prior context, not transcribed, and reliably restores casing + punctuation. Crucially this keeps the turbo model, so word accuracy is preserved: turbo correctly hears "capture" where the older punctuated reference mis-heard "CAPTCHA". Configurable via ASR_INITIAL_PROMPT (clear it if it ever leaks on very short clips).

ASR_THREADS=16

Set on the asr service (16-core box; the PHP queue runs one transcribe at a time, so a job can use all cores briefly).

Verify post-deploy: same audio → punctuation + capitals back, and "capture" still correct (not "CAPTCHA").

🤖 Generated with Claude Code