The ThemeTest failure was cross-test contamination, not a theme bug. A Room invalidation-tracker query keeps running on a cancelled flow collector after the in-memory DB is closed or finalised, throws "no such table: sources", and kotlinx-coroutines-test records it and fails the next Compose test with UncaughtExceptionsBeforeTest. Which test it lands on is just ordering/load luck.
Close repositories the way DownloadRepository already did: cancel and join the scope, so Room work cannot outlive close. Stop closing the in-memory test DBs; the DB is dropped and collected once the test releases it, instead of being torn down under a pending query. Wait for the stored ref in offlineModeStillServesDownloadedArtwork instead of asserting once, since it was waiting on a different flow than the state coverArt reads. Print full stack traces for failed tests in CI.
Trim the explanatory comments to short quirk notes, drop the module headers, and fail cleanly: navidrome/gonic now kill their process if they never become ready and the entrypoint always tears down. The prune test waits for the removal to appear in the album list rather than for the scan flag. Correct the AGENTS token-auth note (lms answers 42, Nextcloud 41) and give CI more headroom.
tools/integration-test.nu is now just the entrypoint. The server lifecycle and fixture preparation live in tools/integration/servers.nu, the Subsonic request helpers in subsonic.nu, and shared waiting/port helpers in util.nu. No behaviour change; the assertions still live in the Kotlin tests.
Run the integration harness from its own devshell so the server packages stay out of the everyday shell. Add lms and a Nextcloud Music instance to the harness and extend the suite with auth-fallback, media, error, write round-trip, multi-source and prune tests.
The auth test caught that lms answers OpenSubsonic error 42 (not 41) when it refuses token auth, so the client now falls back on both codes.
The app itself and the MPL-2.0 public-suffix data OkHttp embeds fall
outside the automatically collected Gradle dependency list, so add both
via the manual config. The Subtracks entry links to its source.
desugar_jdk_libs is bundled by core library desugaring rather than the
runtime classpath, so it never reached the generated licence list. Add it
plus its GPL-2.0-with-Classpath-Exception text via the plugin's manual
config directory.
Add AboutLibraries, whose Android Gradle plugin collects the licence and
attribution of every dependency into a generated resource at build time,
and surface it under a new About section in Settings.
1.15.0 is the highest minor Airsonic and Airsonic-Advanced accept; they
reject a newer client with error 30. The OpenSubsonic response fields do
not depend on the declared version.
The active source becomes one immutable id+source pair, so cover art and
stream/download URLs can no longer read a torn state mid-switch. The root
navigation watches the active source id and, on a real switch, closes the
overlays and pops an open album/artist/playlist detail back to the library.
The pager's settled-page collector reports its initial value, so a cold start
wrote Albums back before the stored tab had been read - racing the very setting
this branch added, able to pin the app to Albums and overwrite the choice. The
screen waits for the stored tab before composing now, so the pager latches the
right page and the initial report matches it, which also drops the spurious
scroll and the search clear that came with it. The report is guarded anyway.
The data layer no longer names a UI tab from a coincidentally same-named enum:
it returns the raw stored value and the view model owns the Albums fallback.
The selected tab lived in rememberSaveable, which survives a rotation or a
process-death restore but not a cold start, so the app always came back on
Albums. It goes through DataStore now, like the per-tab sort and filters, which
were already durable - landing on a different tab than the one you had filtered
is what made the filters look reset.
Verify the Subsonic token hash and salt freshness, and fall back to per-album
songs when the empty-search probe fails with an I/O error instead of
aborting the whole sync.
Add a real-server sync test that mirrors navidrome and gonic into Room and
reads the rows back, plus SyncManager coverage for the success, coalescing
and mid-sync-failure paths.
Cover the sort mapping and delete/ping ViewModels, the download engine's
request flags, the queue helpers, download prefs and migration data
preservation.
Drop the leaked collectors in the ViewModel/queue tests, document the
screenshot tests as render smoke tests, bind a real search term in the
trigram plan test, and assert a positive in-flight invariant alongside the
concurrency timing checks.
The failure was only recorded as SyncStatus.Failed, which nothing shows, so a
bad username or an unreachable server folded silently. It goes through the same
toast the rest of the data layer uses, with the server's own message.
Only treat a rewind as a new listen when it lands in the first few
seconds, so a small seek-back no longer re-arms and duplicates a
scrobble. Fall back to the player's resolved duration when the queue
item has none, and stop the position flow publishing a value for a
track the state has not caught up to. Submit off the collector so an
offline write cannot stall counting.
Scrobbling can be turned off in Settings (on by default). Now-playing,
scrobbles and stars all go through a ServerActionSink so a persistent
offline outbox can replace the network implementation later without
touching the callers.
Send a now-playing notification when a track starts playing and one
submission once it passes half its duration or four minutes, using the
Subsonic scrobble endpoint's submission and time parameters.
A transcoded stream is chunked with no length, so the app declares an upper-bound
length, and on the first request the server is still producing the stream: a seek
near that bound lands past the data and comes back as POSITION_OUT_OF_RANGE
(Navidrome answers 416 until it has cached the transcode). That first request is
what warms the cache, so give the load one retry after a second: the retry sees a
real Content-Length and the stream plays and seeks. Further out-of-range errors
are given up on rather than looped.
A decoder often gives no bitrate, so the row vanished; for a downloaded file the
size over the track length fills it in. The AAC mime Android and media3 use is
audio/mp4a-latm, which read as "MP4A-LATM" rather than AAC.
austinried