Scanner and library: - Keep the last library when a configured folder cannot be walked; the previous code deleted every song, cascading away mood seeds and progress. - Re-adding a file now clears its tombstone, and a re-deletion refreshes the timestamp instead of keeping the old one. - Tombstone timestamps are unique per row, and delta pages continue from the older of the song/tombstone page ends. - Snapshot pagination accepts a stable revision for later pages so songs inserted mid-sync are not skipped forever. - Reject "%" and "_" searches that normalize to an empty LIKE needle. Downloads and discovery: - Download promotion uses the deterministic path id, one folder per track so covers do not overwrite each other, removes failed promotions, and restarts the scan after a running one finishes. - Stale downloading jobs are recovered; reject cancels queued/in-flight jobs; markDone only owns downloading rows; rejecting no longer seeds the track. - Reuse an already completed download of the same video for another mood. - yt-dlp stdout and stderr are separated, empty discovery results no longer prune undecided recommendations, and filenames are capped by UTF-8 bytes. API and data: - Malformed or missing JSON returns 400 instead of 500; optional session bodies use receiveNullable. - Playback events must belong to the session, known song and open session. - Radio refills the candidate pool after recent-song pruning. - Seed percent is required in MoodRepository.update. - Normalize legacy V4 liked/downloaded recommendation states before V5, with an upgrade regression test. CI: check out without persisting credentials and require tests before chart publishing.