Replace realtime Scribe streaming with post-recording batch transcription

[tadaspetra]

What changed

Transcription is no longer a realtime WebSocket session during recording — it is now a batch pass over the finalized recording file, run after Stop.

• New main-process transcription-service: extracts audio from the finalized take via ffmpeg stream copy (source file untouched; remux also fixes MediaRecorder's incomplete WebM headers), uploads to ElevenLabs batch Scribe (scribe_v2, word-level timestamps). The API key stays in the main process — the renderer no longer receives a Scribe token.
• New batch-transcript renderer helper: groups word timestamps into speech segments using the same 1.5s silence threshold the old VAD used, so auto-cut boundaries stay familiar.
• Stop flow reordered around durability: finalized files + recovery checkpoint (.pending-recording.json) are on disk before any network call. Transcription failure or timeout falls back to full-duration sections with a visible warning — it can only ever cost the auto-cut, never footage.
• Timestamps anchored to per-file recorder start offsets (audioStartOffsetMs/cameraStartOffsetMs) instead of the old wall-clock scribeAudioOffset estimate.
• Record is blocked while the previous take is still post-processing (visible notice), so overlapping take flows cannot race on the shared recovery checkpoint or the stop guard.
• Removed dead realtime code: WebSocket/AudioWorklet/PCM pipeline, audio-processor.ts, scribe-status.ts, live segment-editing keyboard shortcuts, scribe token IPC (net −215 lines).

Why

The realtime Scribe session was killed server-side during long pauses (idle/session limits), silently ending transcription for the rest of the take. Batch transcription of the whole file removes that failure mode entirely, improves accuracy, and deletes the reconnect/offset-estimation complexity.

Behavior tradeoffs a reviewer should know

• No live transcript while recording; the panel shows a passive note and transcription progress appears after Stop.
• The Backspace-to-delete-segment shortcut during recording is gone (segments don't exist until after Stop).
• Stop now waits on the batch transcription (bounded: 2–10 min timeout scaled to take length) before entering the editor; the take is already durable on disk during that wait.

Tests

• tests/unit/batch-transcript.test.ts — segment grouping, offsets, annotation stripping, timeout bounds (15 tests)
• tests/unit/transcription-service.test.ts — ffmpeg args, temp-file cleanup on success and every failure path, missing key/file/empty-file errors, response normalization (10 tests)
• tests/unit/register-handlers.test.ts — new transcription:transcribe handler delegation + payload validation (3 tests)
• media-cleanup and scribe-service suites updated; obsolete scribe-status tests removed with the module

pnpm run check passes (lint, typecheck, unit/integration, Electron smoke, packaging smoke). Docs updated: feature-inventory, runbook, target-architecture, .env.example, AGENTS.md learned facts.

Residual risk: real-network transcription path is exercised only manually; suggested spot-check is a short take with a deliberate 30s pause to confirm auto-cut boundaries.

🤖 Generated with Claude Code