[EPIC] LiveCap v1.2 — STT quality, coaching polish, release readiness

[realproject7]

Vision

v1 (MVP, #1) and v1.1 (meeting intelligence, #76) are shipped and merged. This EPIC covers the gap between "feature-complete" and "a product the operator trusts in real meetings": STT accuracy/calibration on the production model, coaching latency + persistence, local-engine verification, design-system debt, and everything that can be prepped before the Apple-signing operator gate (#14).

Source: handoff audit 2026-07-05 — every item below was verified against current main (fa26d60) with file:line evidence before ticketing.

Problem inventory (why these tickets)

1. STT floors are tuned for the wrong model. CONFIDENCE_FLOOR=0.5 / AUTO_DETECT_CONFIDENCE_FLOOR=0.6 (crates/livecap-core/src/whisper/engine.rs:35,48) are hardcoded consts seeded against tiny, but production default is small (crates/livecap-core/src/model.rs:22). No per-model table, no runtime override — live tuning requires recompiles.
2. STT live behavior (STT: whisper emits phantom captions on silence/ambient noise #92/STT: wrong source-language detection (English transcribed as Chinese) #93/Session start: choose spoken (source) language + target language #94) was never verified end-to-end with real varied audio; only say TTS and the tiny-model integration tests.
3. No user-facing model choice. AppSettings has no STT model field; better models (medium / large-v3-turbo) exist in MODEL_NAMES but are unreachable.
4. Coaching is slow by construction: coachUtterances loops serially, one CLI turn per utterance through a shared turn mutex (packages/engine/src/extras-pipeline.ts:219-228) — ~30 s for 10 utterances.
5. Coaching rewrites are not persisted: archive CaptionEntry has no coaching fields (packages/archive/src/types.ts:8-21); the Dashboard coaching tab can only show "before" text (src/dashboard.ts:310-334).
6. Local Qwen tier never ran for real: code-complete and e2e-tested only against a fake llama-server (packages/engine/test/local-llm-engine.e2e.test.ts). Real download → real server → real quality unverified.
7. Dev-verify loop is painfully slow: capture-visible is env-only (needs LSEnvironment inject + re-sign), and every ad-hoc re-sign resets TCC → an operator SecurityAgent click per rebuild.
8. Design-token debt: 3× #0d0f12, #1a1205, #7ad17a + ~45 raw rgba(255,255,255,…) literals in src/styles.css with no CI guard.
9. Release pipeline doesn't exist yet — nothing in CI for signing/notarization/DMG; all of it except credentials can be prepped ahead of gate [GATE] Domain, signing, notarization, first release #14.

Tickets & order

#	Ticket	Route	Status	Depends on
#108	Dev-verify ergonomics: dev-flags file + stable dev signing identity	claude	✅ merged (PR #119; TCC-persist AC pending operator session)	— (do first)
#109	STT: per-model confidence-floor table + env override	quadwork	✅ merged (PR #120)	—
#110	Settings: whisper model selection + download-on-switch	claude	✅ merged (PR #127)	#109
#111	STT live verification & floor calibration (real audio)	claude (PO)	open	#108 #109
#112	Engine: batch coaching rewrites (latency)	quadwork	✅ merged (PR #121)	—
#113	Archive: persist coaching rewrites (+ round-trip, amend)	quadwork	✅ merged (PR #122)	—
#114	UI: save coaching on generate + Dashboard renders rewrites	claude	✅ merged (PR #124)	#113
#115	Verify: local Qwen engine end-to-end on device	claude (PO)	open	—
#116	Design: tokens + CI raw-color guard	quadwork	✅ merged (PR #123)	—
#117	Release prep: signing/notarization/DMG pipeline	claude (PO)	✅ merged + dry-run verified (PR #125, #130)	—
#118	Engine: BYO API key tier	backlog	unscheduled	—
#129	release.yml: fix DMG collect path	claude (PO)	✅ merged (PR #130)	#117
#131	Dashboard: search sessions (title/summary/board/transcript)	quadwork	✅ merged (PR #133)	—
#132	[R&D] Alternative STT engines feasibility	backlog	post-MVP (after #111)	—

Dispatch order: #108 first (unblocks all device verification) · QuadWork VPS stream: #109 → #112 → #113 → #116 (one ticket per batch) · local/PO stream: #110 → #111, #114 (after #113), #115, #117 · gate #14 stays open, operator-owned.

Engineering policy

Same binding rules as #1: no mock/stub/TODO in app code (CI gate), tests first, branch task/<issue>-<slug>, PR Fixes #<issue>, CI green, never log caption content.