feat: reliable mobile recording (wake lock + Whisper anti-hallucination)

[Endika]

Makes recording reliable on phones: keeps the screen awake while recording and stops Whisper's silence-hallucinations (the repeated "Subtítulos realizados por la comunidad de Amara.org") from polluting the transcript.

Spec & plan: mintza-notes/2026-06-04-reliable-mobile-recording-{design,plan}.md (outside the repo).

What changed

1. Screen Wake Lock (auto + override)

• New ScreenWakePort + WakeLockAdapter (re-acquires on visibilitychange, no-ops when unsupported).
• Auto-acquired while recording, released on pause/stop and on leaving the page; a toggle button overrides it (persisted as keepScreenAwake, default on).

2. Anti-hallucination (3 layers)

• Skip silence at source: MediaRecorderAdapter measures per-chunk peakLevel via an AnalyserNode; TranscribeChunkUseCase skips near-silent chunks (ok(null)) so Whisper is never called on silence (no hallucination, no wasted cost).
• Confidence filter (Whisper-specific): WhisperClient parses verbose_json segments and drops non-speech ones by no_speech_prob/avg_logprob.
• Blocklist (provider-agnostic): HallucinationFilter strips known credit/sign-off/tag phrases; single-word tags (music/applause) match whole-line only so real speech mentioning them is kept.
• New skipped HUD counter distinguishes silent/non-speech from real failures (EN/ES/EU).

Quality

• TDD throughout; in-memory fakes, no mocks. Full gate green: lint + typecheck + 129 tests + build.
• Includes fixes from an adversarial code review: stop peak-sampling on pause (battery), release wake lock on dispose (leak), and guard the silence gate when AudioContext is unavailable.

Needs real-device verification

Web Audio peak metering, the Wake Lock API, and the mobile recording/pause flow can't be exercised headless — please smoke-test on a phone (record with the screen lock, confirm no Amara lines, check the toggle).