Fix: Separate mic/system transcription — segmentation rewrite + WebRTC AEC3 (Phases 1+2)

[ghnmqdtg]

Phase 1 of mic/system transcription separation

Fixes cross-attribution between microphone and system-audio transcriptions at the logic level. The acoustic echo cancellation (Phase 2) lands separately.

Root cause

Not a labeling bug — sourceType was always carried correctly. The cross-talk is signal-level: the old per-worklet RMS VAD had no cross-stream arbitration, and (Phase 2) the mic physically picks up computer audio on speakers.

What this PR does

• New pure, deterministic SegmentationController (frame-count based, fully unit-tested) replacing the naive in-worklet RMS VAD + ad-hoc chunk buffering.
• Worklets simplified to pure frame emitters; segmentation now runs on the main thread, one controller per stream (attribution correct by construction).
• transcription.ts: chunk-buffer/VAD-event path replaced by sendSegment(pcm); dead code removed.
• Offline harness proving per-stream attribution incl. double-talk.

Tests

• 59/59 pass (pnpm test:run), incl. 13 new segmentation tests.

Scope note

This is Phase 1 (no new dependency). Phase 2 adds WebRTC AEC3 (WASM) for the speaker/acoustic case. Design + plan: plans/2026-06-19-separate-mic-system-transcription-design.md.

Do not merge without review.

[ghnmqdtg]

Phase 2 added to this PR (same branch)

WebRTC AEC3 (vendored @ennuicastr/webrtcaec3.js v0.3.0, BSD-3, under src/renderer/public/vendor/webrtcaec3/ — no npm/lockfile change) now provides the actual echo cancellation for the speaker case.

Architecture (post-spike):

• AEC runs on the main thread (the renderer loads the wasm normally; an AudioWorklet can't). Verified by Node spike: loads + cancels a delayed echo (residual energy ratio 0.0000) and resamples 48k→16k internally — so no hand-written decimator.
• apmWasmAdapter.ts: loadAec3Module() + createAec3Adapter(module, {inRate, outRate}) → AEC-only process(near, far): Float32Array.
• RealTimeAnalysis.tsx: when the wasm loads and a system stream is present → AudioContext @48 kHz + two cross-referenced AEC3 instances (mic far-end = system; system far-end = mic) → a pump() that pairs mic/system 48 kHz blocks so the far-end reference is real (naive 'either-side' draining would feed silence = zero cancellation) → cleaned 16 kHz → existing SegmentationController + RMS VAD. Falls back cleanly to the Phase-1 16 kHz path if the wasm fails to load or there's no system stream. Muted mic feeds silence to keep the pump paired.

Tests: 62/62 pass, incl. a real-wasm adapter test (cancellation + 48k→16k resample) and ideal-mock AEC harness tests (bleed removed → no false segment).

Still pending: manual speaker smoke test on real hardware (console logs AEC3 enabled (48kHz); confirm computer audio stays on the computer side and the user's voice on the mic side during double-talk).