Processors now emit a ProcessedAudio plan with a streamed writer instead of a whole-file AudioBinary; vault writes stream via RegisterResourceStreamingAsync. Header parsing is bounded. Wave 2 (waveform/Opus) still re-reads the full file by design.
On partial failure the old path deleted the original audio before
confirming the new write succeeded. Now: load old extension, register
new audio first (original untouched on failure), then clean up stale
backing file only on success and only when extension changed.
Swap a track's audio by EntryKey (metadata/release/position preserved, waveform regenerated); hide per-track remove on a release's sole persisted track so it can only be replaced or release-deleted.
daniel-c-harvey