Refresh per-source rows under the reconciliation mutex before reset decisions and chunk queue-visibility checks to respect SQLite bind limits.
Cover a blocked pre-completion emission and source deletion at the bundled SQLite variable limit.
Consume captured reset points on the first collector emission and leave pending sources uninitialized during reconciliation.
Build ordered baselines from queue-visible completed downloads and ignore pruned-row transitions while retaining all-row state and recovery.
Prepare queue and playback reconciliation inside the reset transaction without publishing listener, cache, or revision state. Publish through a non-suspending, non-cancellable post-commit step so committed resets are never retried over later edits.
Keep the pending-origin clear under the cursor lock without a redundant single-statement transaction, and cover listener and post-clear write failures.
Move pending-origin clearing into the queue mutation lock, acquire the cursor lock before Room, and clear pending rows before cache, revision, or listener side effects. Keep sync listener reconciliation transactional for retries while preserving direct reset commit visibility.
Run transition bookkeeping under the queue mutation lock, retain detached playback when its final survivor is removed, and restore offline repeat wrapping without a forward survivor. Centralize detached survivor rebuilding and add regressions for each defect.
Preparing the restored window on launch was meant to make the paused
session reachable from Bluetooth and Automate. The follow-up on the issue
shows the notification appears, but the car's play command still does not
reach the app until the phone is unlocked, and the speculative prepare
buffered up to 50 seconds of the restored track on every launch while
claiming the media controls whenever the app opened.
Removing the prepare also removes the error gate it needed, which could
hide genuine playback errors while paused.
Reverts #230.
A restore now prepares the saved window before any play gesture, so a
network failure at launch reached the error listener and showed a toast
even though the user had not asked to play. Only report playback errors
when playback was requested.
The new tests now wait for the prepare operation rather than the window
contents, which can be observed a poll before the prepare lands.
Restoring the saved queue left the player idle, so the media session had no
notification and Bluetooth or Automate saw no active player until playback
began. Prepare the restored window without autoplaying, which surfaces the
paused session when the app opens. A queue from another source stays stopped,
and an offline queue without downloads has no window to prepare.
Closesaustinried/subtracks#202
awaitActiveSourceId() watched the DAO's activeSourceId flow, which turns
non-null as soon as the source row is written. SourceRepository sets its
in-memory active source in a separate collector, so artistImage() could
still see null and the online URL assertion failed. Wait on the
repository's own activeSource readiness signal instead.
SQLite gives a repeated named parameter one index, so rawQuery can pass the
SQL through unchanged and bind the distinct names in first-appearance order
instead of rewriting them to positional placeholders.
Use named arguments for the repository dispatch, since the adjacent
descending/downloaded booleans were easy to transpose, and add a repository
test that collects the downloaded album grid. Fix the query-timing labels so
the unqualified rows measure the same default path as before.
Rebase follow-up for the new SortNameMode preference: pass it into the
shared album/artist order functions so downloaded pages honour it too,
cover the new _plain indexes in the downloaded artist plan test, and add a
downloaded-mode ordering test.
Add tests for the downloaded artist and playlist paging sources now that
their observed entities are hand-picked, and assert the downloaded
search/starred filters directly. The downloaded artist plan test uses the
real select list and documents the computed album count's in-memory sort,
which the hand-written query also had.
Replace the thirty hand-written order variants with one query builder per
entity plus a default and a downloaded @RawQuery entry point. The
repository's sort dispatch trees and the duplicated positional filters in
LibraryPaging are gone, and the SQL now lives in multi-line strings with
named parameters.
The generated statements are identical to the previous ones modulo
whitespace: the downloaded variants keep the same CASE/download conditions,
and the observed entities match what Room inferred for the old queries. A
small helper rewrites the named parameters to RoomRawQuery's positional
form, including the repeated :sourceId bindings.
Regenerate the store screenshots from the demo library. Only the artist
page changed meaningfully, so keep that capture, using Nine Inch Nails,
which has a biography and three albums.
The demo test now fetches the artist biography and passes artist art to
the home, library and search screens, so a regeneration keeps their art
instead of falling back to letter placeholders.
The captures now come from a 393x786 dp, 440 dpi phone: Pixel XL's
640x360 dp renders text far larger than a real phone shows it. That
phone renders 1080x2161, one pixel over the 2:1 limit both stores
enforce, so captures are written out at 1080x2160.
Track info now lists the Artist field and the plural Artists field as
their own rows, hiding Artists when it would repeat the Artist value.
Keep one Artist sort row, also hidden when it matches the artist name.
austinried