Playback queries LyricTimelineIndex; the linear activeLyricIndices only served
as a test oracle. Scenario tests now exercise the production index directly.
Production removes wishlist, loved, playlist and playlist-track entries through
the batch paths. Device benchmarks now use the production one-item calls as
their per-track baseline.
Backup restore goes through restoreFromBackupOperation with the native or
upsertBatch codec, so the streamed replaceAll path was only reached by tests.
The iOS restore test now drives the production operation with a seeded fake.
Collections live in SQLite and backup restore feeds the raw maps to the
database, so nothing parses the old JSON shape; the picker summary dates
were also never read.
Production reads playlist tracks through loadPlaylistTracksSnapshot and the
test fake keeps its own row helper, so the visibleForTesting wrapper had no caller.
Looks the snapshot render object up by its class instead of a widget type
name string, and narrows the LyricBlur doc to the changes that actually drop
the cached image.
With liquid glass every defocused lyric line ran its own blur pass on each
composited frame, all song long while word-synced lyrics play. A line whose
blur has settled is now rasterized once and drawn as an image until it
changes; lines still easing, or rebuilt, blur live as before.
The 70 dp artwork in the Mornye details sheet decoded the original cover,
up to about 36 MB for a 3000 px image. Decoding at the display size times
the device pixel ratio keeps it sharp without the large transient bitmap.
Each Library rebuild combined every loaded page, filtered playlist tracks
and regrouped the in-memory History snapshot up to three times, so the
section cache never hit and every row's month was formatted again. The
results are now memoized per input identity, with identical output.
The Songs and Singles pages wrote the same grid, table, sectioned and list
layouts twice, and the completion-bridge, special collection and playlist
drop-target cells were written once per grid and list. One builder now
serves each pair with the same widget trees.
QueueTab returns MornyeLibraryScreen for the Mornye root Library, and the
page switcher renders only on that root, so its Mornye chip branch and
_FilterChip could never build.
Incremental upserts and full source replacement created, filled and merged
their temp stage with the same code. One set of helpers now does it, with the
source replacement delete still first inside the merge transaction.
The single-row and batch upserts, replaceAll, the replaceSourceStream
wrapper, delete, getAllFilePaths and searchFtsTable have no callers in lib,
test or integration_test; scans use the staged ingestion paths instead.
Track and album sort modes differ only in the name and key columns, and the
History shadowing anti-join, page query and count summing were written out
several times. The pinned SQL test shows the generated statements and
arguments are unchanged apart from whitespace.
Records the SQL and bound arguments of every track, album, count and artist
query across sort modes, filters and page shapes, so refactors of the queue
store cannot silently change Library ordering or filtering.
Synced lyrics depended on the covered height that animates while the
controls fade, rebuilding the whole list every frame; only the loading
and empty states read it now. Position ticks and digit rolls re-recorded
the entire player because rebuilds under the body LayoutBuilder relayout
it and the seek slider's unused LayoutBuilder relaid out up to the
scaffold; the slider drops it, and the body content and timeline keep
their own layers.
The Mornye player layout and its queue repeated the same "reported index
while it still names the current item, else first match" rule; one
provider now derives it, with a test for stale, shuffled and empty queues.
Compact chrome took five backdrop reads on every scroll frame: the bar
batch plus two stacked reads per side capsule. While folding, each capsule
now has a batch of its own (three reads); at rest the capsules join the
player's batch (one read). A GlobalKey keeps their lenses mounted.
A lens outside any batch took two backdrop reads (a stock blur, then the
shader). Each one now takes a batch key of its own, so the package composes
blur and refraction into one read, as overlays and the bar already did. The
home filter chips never overlap and share one read for the whole row.
The toolbar button copied the active downloads and rebuilt on every queue
change; it now selects the count and progress, so it rebuilds only when the
ring or the count changes.
The table re-sorted every row on each build, including a selection click,
and the quality comparator parsed both labels in every comparison. Sort keys
are now computed once and the order is kept while the rows and sort stay
the same; hovering a row rebuilds only its fading actions.
Return-only switch statements in the conversion, MIME, error type,
logger and model helpers become switch expressions with the same
cases and defaults, and the log summary switch drops redundant breaks.
Fire-and-forget calls, cast results, decoded results and event stream
accessors now go through small async helpers, so each wrapper is one
expression with the same channel names, payloads, casts and decoders.
Keep the media service foreground while paused, release CPU wake locks on pause, and demote the service on stop. Await native state publication before requesting audio focus, honor focus denial, and apply the initial MediaPlayer wake mode.
Validated with 141 playback tests and an Android 17 screen-off pause, resume, and stop integration test.
The premultiplied buffer goes to the background job only when the host
quantizer needs it. The cache comment now covers failed readbacks, and the
theme README lists CoverSourceColorBuilder.
The player now resolves only the cover source colour, so album pages opened
from it no longer found a cached dark scheme and started gray. Album surfaces
read the shared source colour cache first and fall back as before.
The player background, its actions menu and the lyrics share card read only
the source colour, yet waited for one or two colour schemes during the
player's open slide. They now use the source colour path with the same final
colour.
Palette extraction now reads both pixel buffers from one decode and scores
the seed and the source colour in a single background job, instead of a seed
isolate plus a pixel loop on the UI isolate. resolveSourceColor gives
surfaces that only shade from the cover the same colour without quantizing
or building a colour scheme.
IconData overrides ==, so the switch tested up to 187 patterns in turn on
every icon build. A map keeps the first result per icon, which the switch
computes the same way each time.
Hidden tabs stay laid out, so every keyboard frame rebuilt and relaid out
their scaffolds although nothing of them paints. They now see no bottom
inset until shown, and keep the same state as before.
The button group and playback transport duplicated the press state and
squish spring, and two icon buttons repeated the disabled color pair. Also
drop three Material Symbols entries no icon in the app uses.
The theme draws determinate waves flat below 10% and above 95%, where the
wave phase is invisible, yet the package kept repainting every frame.
Muting the ticker there keeps the same pixels and stops idle frames.
MornyeArtistSurface derived three new ThemeData instances on every page
rebuild (selection, load more, title flips). It now keeps them until the base
theme or the artwork-derived colors change, so the subtree sees the same theme.
SliverLayoutBuilder passes the full SliverConstraints, which change with the
scroll offset, so library and Explore grids, album, artist and desktop
toolbars, and the extension overview rebuilt on every scroll frame. A small
SliverLayoutInfoBuilder rebuilds only when the width or collapsed flag it
selects changes; the header back buttons are built once, and headers can take
a title listenable so a title flip rebuilds only the title.
findExisting, updateAudioMetadata, getAllFilePaths, the notifier's sync
getBySpotifyId/getByIsrc and restoreFromBackup had no callers; production
restore goes through restoreFromBackupOperation.
Android SAF users waited for two sequential platform round trips before the
first frame. The checks are independent; after an installation reset the
settings are reloaded and the check restarts, as before.
History and Library migrations repeated the same add-column, backfill and
first-row query boilerplate. The helpers issue the same statements in the
same order, and a new v1 upgrade test pins the resulting history schema.
Each completion rebuilt all three history lookup maps over every loaded
item, so a long batch did quadratic key normalization. A prepended item only
wins its own keys, so copying the maps and writing those keys gives the same
lookups.