- Create lightsync/audio/features.py with extract_features and extract_features_async
- Create lightsync/audio/segments.py with detect_segments via agglomerative clustering
- Add /api/audio/segments and /api/audio/features endpoints to audio.py
- Initialize app.state.segments cache in main.py lifespan
- POST /api/audio/fetch accepts {url: string}
- Spawns yt-dlp subprocess via sys.executable -m yt_dlp
- Streams SSE progress events with PROGRESS_RE line-by-line parsing
- Uses --newline flag for reliable progress output
- Uses yt_%(id)s.%(ext)s template for predictable filenames
- Runs beat analysis on completion and caches in app.state.beats
- Returns done event with path and filename on success
- Remove engine dependency entirely (no MPVEngine, no app.state.engine)
- POST /api/audio/upload: saves file, runs librosa beat analysis, caches in app.state.beats
- GET /api/audio/beats?path=...: returns cached or on-demand beat analysis
- GET /api/audio/waveform?path=...: now takes path query param (no engine state lookup)
- GET /api/audio/files: unchanged
- Removed POST /api/audio/load and GET /api/audio/state (engine-only endpoints)
- Create lightsync/audio/waveform.py with extract_peaks() (AUD-05)
- MP3 via ffmpeg pipe, WAV/FLAC/OGG via soundfile+numpy
- Add GET /api/audio/waveform endpoint with asyncio.to_thread
- peaks param clamped 100-5000, returns {peaks, count, file}
- Create lightsync/api/audio.py with POST /load and GET /state endpoints
- Validate file existence and extension (.mp3/.wav/.flac/.ogg)
- Use asyncio.to_thread for engine.load to avoid blocking event loop
- Return 404/400/503 on error conditions
- Register audio router at /api/audio prefix in main.py
- AUD-02 (YouTube) noted as Phase 6 deferral
