On macOS the system-audio capture already contains the mic, so the recording embeds a full system+mic mix. The fallback also added the separate mic sidecar, doubling the mic into an echo. Use the embedded mix alone. Fixes#628.
On macOS, the screen highlight overlay's right and bottom edges were
running off-screen. macOS clamps window positions below the menu bar,
so the outward padding only appeared on the left/top while the
right/bottom extended beyond the visible area.
Use display workArea instead of full bounds for screen highlights on
macOS, and keep the border/glow within the overlay window.
Major rewrite of the audio export pipeline for stability and memory efficiency:
## Streaming Decode
- Primary decode path uses WebDemuxer + AudioDecoder (WebCodecs streaming)
- Falls back to bulk decodeAudioData if streaming fails
- Avoids holding full compressed file + decoded PCM simultaneously
## Chunked Offline Rendering
- Processes timeline in 30-second OfflineAudioContext chunks
- Memory bounded to ~30s of PCM per chunk regardless of recording length
## Chunked WAV Writing
- Writes PCM in ~256KB chunks instead of single massive ArrayBuffer
- Eliminates OOM for long recordings on the native/FFmpeg export path
## Chunked Encoding
- Single AudioEncoder kept alive across all chunks (clean AAC stream)
- Proper backpressure and error propagation per chunk
## scheduleBufferThroughTimeline chunk windowing
- Clips source nodes to chunk boundaries for correct cross-chunk audio
- Backwards compatible defaults when no chunk params provided
## Other export fixes
- Progress cap: 99% -> 100% in both exporters and UI
- Muxing errors: propagated instead of silently swallowed
- Mac recording: enhanced warning when no audio files for muxing
- Export timeout: eliminated by removing real-time rendering
- Always route audio through aresample=async=1:first_pts=0 filter during
muxing on both macOS and Windows, even when duration delta is small.
Previously skipped when delta ≤50ms, which left progressive clock drift
(from CPU load) completely uncorrected.
- Lower sync detection threshold from 50ms to 20ms so tempo correction
kicks in earlier.
- Use explicit 48kHz AudioContext sample rate in both browser recording
and export rendering to prevent sample rate mismatch drift.
- Reduce MediaRecorder timeslice from 1000ms to 250ms to reduce chunk
loss under CPU pressure.
- Tighten export audio sync: seek threshold 300ms→150ms, tolerance
15ms→8ms, correction window 2s→0.5s, max adjustment ±8%→±12%.
- Lower companion audio start delay threshold from 50ms to 25ms.
- Remove dead helperExists local in recording/windows.ts
- Hoist fs/promises import out of close handler in recording/ffmpeg.ts
- Guard fs.readdir with mkdir in recording/prune.ts (ENOENT resilience)
- Derive companion audio suffixes from COMPANION_AUDIO_LAYOUTS in prune.ts
- Guard mousemove hook registration to Linux only in cursor/interaction.ts
- Replace dynamic require('electron') with static import in cursor/monitor.ts
- Wrap nodeRequire in try/catch in ffmpeg/binary.ts for fallback safety
- Fix hardcoded timeOffsetMs: 0 in project/session.ts (use normalizer)
- Fix isPathInsideDirectory to normalize candidatePath in project/manager.ts
- Fix isAllowedLocalReadPath security: require path to be in allowlist (AND not OR)
- Derive extension regex from constants in project/manager.ts
- Consolidate duplicate Duration parsers in recording/diagnostics.ts
- Refactor ensureReadableFile to use options object instead of description string
- Make swiftc compilation async (execFile) in paths/binaries.ts
- Add socket timeout to httpsGet in captions/whisper.ts
young