feat(transcription): add on-device local transcription via FluidAudio

[tenequm]

Summary

Add on-device speech recognition and speaker diarization using FluidAudio (Parakeet TDT v3 ASR + offline diarization). Users get speaker-attributed transcripts without any API key, account, or internet connection. Soniox cloud transcription remains available as an alternative.

What changed

New file

• LocalTranscriptionService.swift - FluidAudio integration: dual-track extraction via AVAssetReader, ASR on both tracks, diarization on both tracks, temporal overlap matching for speaker assignment, merge with sequential integer speaker IDs

Modified files

• Package.swift - Added FluidAudio v0.12.4 dependency
• TranscriptionService.swift - TranscriptionProvider enum (.local, .soniox), per-provider sidecarURL/load/save, speakers field on TranscriptDocument, legacy transcript.json migration with fallback
• MainWindowView.swift - Segmented tab switcher between Local/Soniox, per-provider transcript and status state, split transcription methods, speaker name resolution (transcript.speakers -> metadata.speakers -> default)
• SettingsView.swift - Default provider picker, model download button with progress
• AudioMonitor.swift - Auto-transcription hook after recording stops (fires when local provider is default and models are downloaded)

Architecture

Recording finishes
  -> auto-transcribe (if enabled + models ready)
  -> LocalTranscriptionService:
      1. Extract Track 0 (system) + Track 1 (mic) from M4A
         (prefers audio-processed.m4a when AEC ran)
      2. AVAssetReader per track -> 16kHz mono Float32
      3. ASR each track with AsrManager (Parakeet TDT v3)
      4. Diarize each track with OfflineDiarizerManager
      5. Assign speakers to ASR segments via temporal overlap
      6. Merge: mic speakers first ("You"), then remote ("Speaker N")
      7. Save as transcript-local.json
  -> Single-track fallback if only 1 track exists

What does NOT change

• Soniox API client (zero modifications)
• Recording pipeline (AudioRecorder, SCStream, AVAudioEngine)
• AEC post-processing
• Playback, waveform, track switcher
• Transcript renderer (provider-agnostic, takes any TranscriptDocument)
• Transcript export
• Onboarding, HUD

Per-provider transcript storage

Each provider writes its own sidecar file:

• transcript-local.json - local on-device transcription
• transcript-soniox.json - Soniox cloud transcription
• Legacy transcript.json is migrated to transcript-soniox.json on first load

Speaker names are stored per-provider in TranscriptDocument.speakers, independent of metadata. Renaming a speaker on one tab does not affect the other.

Smoke test checklist

• Open app with existing recordings that have transcript.json - verify they appear under Soniox tab
• Local tab: click Transcribe on a recording - verify model download + diarized transcript
• Soniox tab: enter API key, Transcribe - verify works exactly as before
• Switch between Local/Soniox tabs - verify each shows its own transcript
• Rename a speaker on Local tab - verify Soniox tab speakers are unaffected
• Settings: change default provider, download models
• Record a call with local as default + models ready - verify auto-transcription
• Single-track recording - verify transcription still works
• Cancel mid-transcription - verify clean state