fix(recording): report failed audio saves and preserve recovery data

[karan68]

Problem

Failed audio finalization (FFmpeg / checkpoint / disk errors) was swallowed and reported to the user as a successful save, even when audio.mp4 was never produced.

Fix (end to end)

• Typed failure propagation: stop_recording returns a typed StopRecordingError; the recording-stopped event carries recovery_available / audio_file_available so the UI can react accurately (error vs success).
• Recovery-safe finalization: checkpoints are preserved until audio + transcripts + metadata all succeed. FFmpeg encode/concat run async, are cancellable (kill_on_drop), and publish audio.mp4 atomically after non-empty verification.
• No tail-audio loss: the accumulation worker is now an awaited JoinHandle, drained at stop after the pipeline closes the channel (replaces a fixed 200ms sleep).
• Timeout safety: a timed-out transcription worker is abort()-ed and joined before model unload / finalization, so it cannot race the save.
• Honest persistence: the frontend saves transcripts even when completion is not observed (transcriptionComplete || hasTranscripts) with a distinct warning toast, and retains a durable IndexedDB audio-repair record linked to the existing SQLite meeting (no duplicate on recovery).
• Path escaping: FFmpeg concat lines escape single quotes, so meeting names with apostrophes (e.g. Karan's Review) no longer fail finalize/recovery.

Validation

• Rust: crate compiles (0 errors); saver unit tests 4/4 pass, including new concat_escape_handles_single_quotes and failed_finalization_preserves_checkpoints.
• Frontend: tsc --noEmit exit 0; editor diagnostics clean.

Notes

• Manual FFmpeg/disk fault-injection in the running desktop app was not performed and is the recommended follow-up verification.