STT live verification & per-model floor calibration on real audio (closes the #92/#93/#94 verification gap)

[realproject7]

Goal

The three STT behavior fixes shipped in v1.1 (#92 phantom-caption gate, #93 auto-detect CJK suppression, #94 source-language forcing) were verified only via say TTS and tiny-model integration tests — never against real varied audio on the production model. Additionally the confidence floors must be calibrated per model with real evidence, and small vs medium vs large-v3-turbo accuracy compared so the settings copy and default rest on measurement.

Prereqs

• Dev-verify ergonomics: debug dev-flags file (no resign) + stable dev signing identity #108 (dev-flags loop — fast iteration without resign/TCC churn)
• STT: per-model confidence-floor table + env override (floors currently hardcoded, seeded on tiny) #109 (env-override floors — recompile-free tuning)
• Settings: whisper model selection (small / medium / large-v3-turbo) with download-on-switch #110 nice-to-have for multi-model runs (models can also be forced via PipelineConfig in a dev build)

Verification matrix (each row = observable pass/fail, evidence via ui-heartbeat.json + saved session files)

1. Silence/ambient: 10 min with no speech, per model → zero phantom captions (STT: whisper emits phantom captions on silence/ambient noise #92).
2. Forced Spoken=English on English audio (operator's sample video, YouTube 6ARteKgyEtw) → zero CJK output (Session start: choose spoken (source) language + target language #94).
3. Auto mode on a JA clip and a ZH clip → correct language detected, no cross-language phantoms (STT: wrong source-language detection (English transcribed as Chinese) #93).
4. Floor calibration on small (then medium/large-v3-turbo): log mean per-token confidence for real speech vs any phantoms; choose floors with margin; check short-phrase real speech ("Yes." / "좋아요") survives.
5. Accuracy comparison small vs medium vs large-v3-turbo on the same 5-min sample-video segment: transcript diff, error notes, latency + RSS per model.

Deliverables

• PR updating the per-model floor table (STT: per-model confidence-floor table + env override (floors currently hardcoded, seeded on tiny) #109's table) with calibrated values + a short docs/CALIBRATION.md (method, audio set, numbers, model recommendation).
• Recommendation: keep or change the default model + final settings copy for Settings: whisper model selection (small / medium / large-v3-turbo) with download-on-switch #110.
• Follow-up tickets for any gaps found (VAD padding, bleed, etc.).

Operator-assist (called out up front)

• One system-audio (SecurityAgent) Allow click after the first re-sign — cannot be automated.
• Optional 2-min real-voice session for the mic channel (TTS is eaten by bleed suppression); scheduled when convenient.

Routing

agent/claude (PO direct on this Mac — device-bound; verification work per manual §9).

Part of #107. Depends on #108, #109.