yt-dlp now solves YouTube JS challenges via external EJS scripts (yt-dlp-ejs package). A bare yt-dlp from PyPI does not include them, and remote fetching is disabled by default, so images built after the EJS migration download nothing and fail with HTTP 403 while metadata and playlists keep working. Pin yt-dlp[default] so yt-dlp-ejs ships in the image, and pass --remote-components ejs:github so solver scripts refresh at runtime when YouTube rotates challenges.
Fix AzuraCast integration against verified API behavior
Verified against AzuraCast source (Files\BatchAction, FlowUploadAction,
Files\ListAction, Paginator, StationPlaylistMediaRepository):
- Rotation read a nonexistent 'mtime' key; the files API returns
'timestamp' (unix epoch). Age buckets now work, so tracks actually
migrate Currents -> Recurrents -> Gold as intended.
- Rotation now filters on type == 'media' (dirs are 'directory', not
'dir') and skips rows without a path/timestamp.
- Follow ?page= pagination on /files: the default (no params) response
is an unpaginated JSON array, but any paginated response is now
traversed instead of only page 1 being visible.
- POST /files/upload returns only {success, message}; it never returns
unique_id/path, and the file lands at the uploaded filename. Removed
the dead unique_id search; playlist+queue now use the known path.
- Added success/error checks for batch ops (playlist/queue) instead of
discarding responses, and queue no longer runs when playlist assign
or Currents lookup failed.
- Replaced curl subprocess shelling with aiohttp (already a matrix-nio
dependency): single helper for API calls, real multipart upload,
timeouts everywhere, no temp JSON/pickle files in DATA_DIR.
Functional test (mock AzuraCast mirroring verified API shapes):
15/15 checks pass - upload contract, pagination, batch ops, and age
bucketing (fresh/6d/10d/30d -> Currents/Currents/Recurrents/Gold,
directories and non-media excluded).
Diane