Extraction always used Matroska (.mka) because it takes essentially any
audio codec. That is right for most codecs but wrong for AAC: MP4 records
the AAC encoder priming (the ~1000 warm-up samples every AAC encoder emits)
in an edit list, and the edit list does not survive into Matroska. The
remuxed track then decodes ~23 ms longer than the source, with the priming
samples played at the head instead of discarded.
Measured on a 5s 640x480 fixture: source audio decodes to 962980 bytes of
PCM, the .mka to 967440 — 4460 bytes / ~23 ms extra, peaking at -66 dBFS.
Inaudible in practice, but it also puts the track fractionally out of step
with its own reported duration, for no reason.
Pick the container by codec instead: aac -> .m4a (keeps the edit list),
everything else -> .mka as before. If the preferred container refuses the
codec, retry into .mka before falling back to keeping the whole video. AAC
is worth the special case because mp4 / mov / m4v — what people actually
upload — almost always carry it.
Adds the strongest available test of the "lossless" claim: decode the audio
straight out of the source mp4, decode the stored extract, assert the PCM is
byte-for-byte equal. Forcing .mka fails it with exactly the 4460-byte delta.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
After extracting the audio track, uploadAudio assigned the record's `size`
from the new .mka BEFORE deleting the source video:
size = statSync(extracted).size;
rmSync(filePath, { force: true }); // can throw EBUSY/EPERM on Windows
filePath = extracted;
rmSync with force:true only swallows ENOENT — a briefly locked file (exactly
what the existing scheduleRetry machinery in this file exists to handle)
throws. The catch then discards the extract and keeps playing the original
container, which is correct, but `size` had already been overwritten with the
much smaller extracted size while the whole video stayed on disk. That makes
totalBytes() under-count and lets the upload directory grow past its quota.
Commit filePath and size together, only once the source is actually gone.
Adds a regression test that partially mocks node:fs to make rmSync throw for
the source .mp4 and asserts the persisted record (index.json — `size` is not
exposed through search()/toSong) still describes the retained file. With the
old ordering it records 27894 bytes for a 104544-byte file.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
