250 Commits
Author SHA1 Message Date
zarzet 30ec85296b fix(usb): convert when the endpoint is too small, and queue 1.5 s at any rate
A DAC limited to 48 kHz has an endpoint too small for 96 kHz. That check
runs first and was classed as an unsupported DAC, so the song went to
Android's mixer instead of being converted. It now counts as a refused
format and the next rate is tried, without touching the DAC.

The output queue held 200 ms or 256 KiB, whichever is more: 1.5 s for a
CD-rate song but a third of a second at 96 kHz in 32-bit, so a short stall
of the source thread (another app opening) cut high-rate songs only. It
now holds 1.5 s at every rate. Pause and seek still flush it at once.
2026-11-18 21:59:04 +07:00
zarzet da608a6638 fix(usb): never set a rate the DAC's clock does not list, and add a rate limit
A song whose rate the DAC refuses is already converted, but the DAC was
still asked for that rate first, for every song, and a DAC that takes the
request and then plays silence was never converted at all.

- Read the UAC2 clock's RANGE before setting a rate and refuse one it
  does not list, without touching the clock.
- Remember per connected DAC the rate a song rate played at, so the
  refused rate is not requested again on each song.
- Add "Highest sample rate" to the Direct USB settings (Automatic by
  default): rates above it are never requested and such songs are
  converted down.
2026-11-18 21:48:21 +07:00
zarzet 208bd75696 feat(usb): resample a song whose rate the DAC refuses
A DAC that tops out at 48 kHz refused 96 kHz songs, so they fell back to
Android's mixer and played through a DAC still set to the low level the
direct driver leaves it at, which sounded like silence.

Direct USB now tries the other standard rates when the DAC refuses the
song's own: lower rates first, the song's own family ahead of the rest,
then higher ones. The song is converted with a Kaiser windowed sinc
polyphase filter (130 dB stopband, passband to 0.907 of the lower
Nyquist), with triangular dither for a 16-bit output. The status line
names both rates and Audio info no longer calls such a song bit-perfect.

DSD is never converted, and a DAC is asked at most four times.
2026-11-18 01:00:13 +07:00
zarzet e5bc806a91 chore(link): drop a comment main already removed from this test 2026-11-17 03:09:49 +07:00
zarzet b52a28b85f feat(link): serve songs the app lends to the Link server
A catalog path that is a URI, such as an Android storage access
document, opens through the app: the server forwards server/files/open
without a device, duplicates the descriptor the app lends and forwards
server/files/close when the response ends, the client goes away or any
step fails. Lent files are bounded, opened one at a time, and serve
streams with ranges, downloads, embedded covers and lyrics. Phones
cannot send the server's own ops, and plain paths take the old route.
2026-11-17 03:09:29 +07:00
zarzet b648cd2e07 feat(link): let the app advertise a phone host itself
linkStart takes advertise: false so a phone host announces itself
through the platform's service discovery instead of raw mDNS.
2026-11-17 03:09:29 +07:00
zarzet 4308677dbb feat(link): share the Link server glue with the phone backend
Move the desktop's link* glue into mornye_link::bridge so the phone
crate can run the same server, and expose it to Android and iOS as the
UniFFI object LinkBridge. linkStart takes an optional platform and
features list, announced in the info, the welcome and the mDNS TXT
record (pf), so clients can tell a phone host and the abilities it has.
2026-11-17 03:09:29 +07:00
zarzet 45d4c2d411 chore: remove redundant comments and unused import 2026-11-17 02:05:45 +07:00
zarzet af2198504b Merge branch 'feat/authenticity-report'
# Conflicts:
#	lib/widgets/material_player_backdrop.dart
2026-11-16 18:09:51 +07:00
zarzet ede25f8409 feat(authenticity): find MP3 sources at any bitrate with the exact filterbank
The scan for an MP3 source approximated the layer 3 filterbank with one
plain lapped transform, which shows the frame grid at 128 kbps and little
above it: on the test set it found 44 of 51 files at 128 kbps, 16 at
192 kbps and none at 320 kbps or V0.

The audio is now analysed with the filterbank the format defines (the
512-tap window into 32 subbands, the 36-point transform per subband and the
alias butterflies) at all 576 granule alignments. At the encoder's own
alignment the coefficients are the stored spectrum again, which shows in
two ways: their sizes sit on the lattice the format quantizes to, and lines
that are loud everywhere else are zero. Both must peak at the same
alignment for a source to count as found. Excerpts with clipped samples
are avoided, since one clipped sample spoils the granules around it.

On 255 decodes of MP3 at 128, 192 and 320 kbps, V2 and V0 all are found,
and none of 317 originals and AAC or Opus decodes is taken for MP3. Other
decoders, mono, a gain, dither, a trimmed start and 24-bit output keep it
found; audio resampled after decoding does not, and the report says so.

The report now carries the score a pattern of the measured kind counts as
found from, so the page shows the right limit for each codec.
2026-11-16 18:01:39 +07:00
zarzet c163b8858e fix(cast): serve the TV from an address on its own network
The media server asked the OS which local address routes to the TV.
Android routes each app through an active VPN, and through mobile data
when Wi-Fi is not the default network, so the TV was handed an address
it could not reach: it connected, showed the Cast screen and never
played, ending in Google Cast timeout 15 or the app's own 25 second
limit.

Pick the interface address whose network holds the TV and fall back to
the route only when there is none. Rebind when this device's address
changed, say what to check when the TV cannot reach the device, and let
the platform's answer arrive before the app's own limit on a load.
2026-11-16 17:41:55 +07:00
zarzet 4c1f2495e4 feat(authenticity): share the report and check several songs at once
Share report sends the outcome and every section as plain text through the
system share sheet. Check authenticity, in the library track menu and among
the selection actions, reads the selected FLAC and WAV files one after
another and lists each song with its outcome, a summary of how many need a
look and a note when they differ in bit depth or sample rate; a finished
row opens its full report in place.

The words of a report now come from one place (HiResReportContent), so the
card, the shared text and the list cannot disagree. The new strings are
translated to Indonesian, five strings the first report left unused are
removed, and a C ABI test covers the desktop route of the whole-file
check.
2026-11-16 16:39:06 +07:00
zarzet 5fbda37180 fix(network): use only a valid cached DoH answer inside the resolver window
Inside the 15 s window after the system resolver failed for a host, the
lookup called the DoH fallback, which queries the DoH servers again when
the cached answer has expired. On a network where the system resolver
works but the DoH servers are blocked, that connection waited for the DoH
timeouts before the system resolver was asked. The window now uses only
a still-valid cached DoH answer; otherwise the system resolver is asked
first, exactly as before the window existed.
2026-11-16 16:24:05 +07:00
zarzet d06ae920eb perf(lyrics): append new cache results to a journal instead of a full rewrite
Every newly cached song re-serialized up to 500 entries and rewrote the
whole snapshot with a file sync, about once per new song played. A new
result (and any eviction it causes) is now appended as one line to a
journal beside the snapshot, and the snapshot is rewritten only after 64
journal changes, on clear, after a memory drop, or when the journal
cannot be used. The snapshot names its journal by ID, so a journal left
by a crash or by an older version that rewrote the snapshot is ignored,
and older versions read the snapshot as before. A torn last line keeps
the readable changes and starts a new snapshot. The provider wait and
the result selection are unchanged.
2026-11-16 16:24:05 +07:00
zarzet b0f3319842 perf(reenrich): ask every provider at once in the single-song picker
A picker, and any search that may not stop at the first confident
provider, asked metadata providers one after another, so it waited for
the sum of their latencies. Each extension has its own engine, so the
providers are now asked on scoped threads sharing the request lease and
the results are joined in provider order. The candidate list, the
per-provider 30 s timeout, cancellation and the verification_required
report (first provider in order, only when nothing was found) are the
serial search's. Searches that stop early keep the serial loop.
2026-11-16 16:24:05 +07:00
zarzet 7352d369bf perf(network): skip the system resolver briefly for hosts only DoH answers
When the system resolver failed for a host and the DNS-over-HTTPS
fallback answered, every new connection to that host still asked the
system resolver first and waited for it to fail again. Such a host now
goes straight to the cached DoH answer for 15 seconds. A failure DoH
cannot answer is not remembered, so the system resolver keeps being
asked as before, and if the DoH path fails during the window the system
resolver is tried again.
2026-11-16 16:24:05 +07:00
zarzet cb830ea147 perf(desktop): wake the FFmpeg wait when the process exits
The wait polled try_wait every 200 ms, so each command finished on
average about 100 ms after FFmpeg exited, several times per download
with conversion and embedding. The output drain threads now signal when
their pipes reach end of file, which happens as FFmpeg exits, and the
wait checks the exit status right away. Cancellation is still checked at
least every 200 ms and kills the process as before.
2026-11-16 16:24:05 +07:00
zarzet 054201a7e1 build(rust): drop the unused rusqlite hooks feature
No crate or test calls an update, commit, rollback, WAL, progress or
authorizer hook, so the feature only added compile time and code size.
2026-11-16 16:24:05 +07:00
zarzet 8708180376 fix(library): bound folder artwork and retry failed folder image reads
The folder artwork memo lived in the folder record that every song of the
folder holds, and the scan keeps its whole file list until it returns, so
every folder's artwork stayed in memory for the whole scan. A folder's
artwork is now released when the last song that uses it finishes (songs
are announced before the workers start), artwork over 8 MiB, the thumbnail
cache budget, is not kept at all, and CUE songs, which run before any song
is announced, never keep it. Songs of a folder are scanned next to each
other, so at most a few folders hold artwork at once.

A failed read was kept for the whole folder, so one passing I/O error on a
network or removable drive dropped the folder image for every song of the
folder. Only successful reads are kept now; a failure is read again by the
next song, as before the memo.
2026-11-16 16:20:25 +07:00
zarzet 04db843b64 perf(library): skip the snapshot sync and reuse deletion statements
The incremental snapshot TSV is read by the scan in the same session and
deleted right after it, so syncing it to disk on every incremental scan
bought nothing. Removed paths are now deleted with cached statements
instead of preparing the same two or three statements again per path.
2026-11-16 16:20:25 +07:00
zarzet bb6b435ae4 perf(library): read folder images and CUE sheets once
A scan read, hashed and compared the folder image (cover.jpg and the like)
once per song that used it, and on a large image that did not fit the
thumbnail cache it decoded and resized the image again for every song
under the global thumbnail mutex. Each folder image of a scan now keeps
the artwork it resolved to (thumbnail, type and stored-file record), so
the songs of a folder read and resize it once. The memo lives in the
scan's folder image, so the next scan reads a changed image again; a
cancelled read is not kept. The reader opens the image with
open_native_read instead of validating the path a second time.

Every incremental scan also parsed every CUE sheet to learn which audio
it owns. A sheet the scan skips as unchanged now keeps the audio
reference read at the same size and modification time in the scan state;
the reference is still resolved against the files on disk each scan, and
files the walk just found need no probe. Changed sheets and full scans
parse as before.

CUE thumbnails now use the scan's sync batch like audio thumbnails, so
on Apple platforms they are flushed with the scan's single folder sync
instead of a full sync per file.
2026-11-16 16:20:25 +07:00
zarzet d21086c3df fix(extensions): keep the 5 ms stop check for lyrics providers
A lyrics search returns only once every provider has noticed its stop, and
that stop ends nearly every search. The slower heartbeat held the chosen
lyrics back by up to 40 ms, so this path keeps its fixed 5 ms check. The
download and metadata callers keep the slower one.
2026-11-16 16:05:10 +07:00
zarzet 1327ecd696 perf(extensions): wait for busy engines and planners without 5 ms polling
A call that found the extension engine busy retried try_lock every 5 ms
until the holder finished, which for a post-process can last minutes.
The engine lock now keeps a list of waiting threads and wakes them when
any holder releases it; the waiter also subscribes to its lease, so a
cancel wakes it at once. Only manager shutdown is still checked on a
50 ms heartbeat.

The download planner, installed lyrics and metadata provider callers
polled their caller check every 5 ms for the whole call. The check is
an opaque closure with no wake signal, so the heartbeat now starts at
5 ms and doubles to 40 ms: completion still wakes the caller at once,
and a cancel is noticed within 40 ms instead of 5 ms.
2026-11-16 16:05:10 +07:00
zarzet 6aac23000e perf(extensions): skip the store write when a set changes nothing
storage.set with the value already stored rewrote and fsynced the whole
store file. Such a set now returns without touching the file. The
credentials remove write for a missing key is kept for Go parity.
2026-11-16 16:05:10 +07:00
zarzet 3b6acc694b perf(extensions): reuse the compiled script when a VM reloads
A cancelled or timed-out call retires the extension VM, and the next
call loaded a new VM that compiled index.js from source again. Startup
validation also threw its compile away, so the first real call compiled
the same script a second time.

The installed entry now keeps the compiled source and hands it to
primary loads. The loader still reuses it only when the source text is
identical, each install or upgrade has its own entry, and an
interrupted VM is still never reused. Disabled packages are validated
without keeping their compile, and disable or uninstall drops it.
2026-11-16 16:05:10 +07:00
zarzet 0d9d6fa6de perf(link): free the replaced catalog outside the write lock
set_catalog dropped the previous catalog while holding the catalog write
lock, so requests waited for a large catalog to be freed. The old Arc is
now taken out and dropped after the guard.
2026-11-16 16:02:02 +07:00
zarzet 01f7376d88 perf(link): write lastSeenMs outside the authentication lock
touch wrote and fsynced devices.json while holding the mutex every
request takes to authenticate, on the two-worker Link runtime. The
in-memory record is now updated under that mutex and the file is written
after it is released, under a separate writer lock that keeps writes in
order; touches that overlap a write share it. Pairing and revocation
still write while holding both locks, so a revoked key stops working as
soon as before and a pending touch never writes a removed device back. A
failed write is retried on a later touch, as before.
2026-11-16 16:02:02 +07:00
zarzet 560eccb13f perf(link): share and bound cover resizes, list the cover cache less
Every new cover cache file listed and stat'ed the whole cache (up to
4,000 files) to enforce its limit. The count is now kept per cache folder
after one listing and the folder is listed and trimmed again only once it
passes the limit by 64; the eviction order and extension filter are
unchanged.

Uncached resizes ran with no bound and two requests for the same target
both decoded the full image. At most two resizes now decode at once and
identical in-flight requests share one run. The output bytes, filter,
cache names and headers are unchanged.

The MIME sniff of the served file also moves into the blocking tasks
that already touch the file instead of opening it on a runtime worker.
2026-11-16 16:02:02 +07:00
zarzet 9a6cdf8476 feat(hires): inspect the whole file for damage, real bit depth and lossy sources
The Hi-Res Authenticity card now reads the whole file and reports, section
by section: the audio against its stored checksum and declared length, the
real resolution shown by the spacing of quiet samples (padding, a scaled
lower-resolution source, a 16-bit grid under low-level noise), where the
spectrum is cut, and the frame grid a lapped-transform codec (AAC, Vorbis,
MP3) leaves in audio that was lossy once. It shows the average spectrum and
says what each test cannot see. The sampled check used by downloads is
unchanged.
2026-11-16 12:34:28 +07:00
re-code d847bcfcfc fix(tags): stop stale TXXX lyrics from shadowing saved lyrics 2026-11-15 23:33:45 +07:00
zarzet fa667d6a0d chore(deps): update packages, Rust 1.100.0 and Gradle 9.9.1
Dart packages move to their newest compatible releases, among them
dart_smb2 0.4.1 (libsmb2-r13: uploads that reconnect can rename their
temporary file, no wait for ever on a reply that stops, OpenSSL
signing), record 7.2.0, archive 4.7.1, file_picker 13.3.3 and
material_ui 1.11.0. Rust moves to 1.100.0 with rusqlite 0.43.0,
zip 9.6.0, notify 8.2.2 and the transitive crates. Gradle moves to
9.9.1.

Held back on purpose: flutter_secure_storage 11 (the extension storage
key lives there), flutter_rust_bridge (smtc_windows needs 2.11.1), the
Unicode 15 crates, and desktop_drop, dbus and connectivity_plus, which
wait for cross_file 0.4 in file_picker and share_plus.
2026-11-15 21:51:31 +07:00
zarzet b572ef0166 fix(tags): write track and disc totals in their own FLAC and Ogg fields
FLAC and Ogg files got TRACKNUMBER=6/12 and DISCNUMBER=1/2, the combined
form ID3 uses. Vorbis comments keep the total in a field of its own, so
other players and tag editors showed no total.

Downloads, re-enrich and tag edits now write TRACKNUMBER=6 with
TRACKTOTAL=12 and TOTALTRACKS=12, and the same for discs. A total the new
metadata leaves out keeps the file's own. Reading also takes these total
fields, so files tagged elsewhere show their totals. ID3, APE and RIFF
keep the combined form.

Fixes #878.
2026-11-15 17:33:24 +07:00
zarzet 86f313f432 feat(lyrics): pick the best-timed provider result
The provider cascade took the first provider in order that returned any
text, so plain lyrics from the first provider beat word-synced lyrics
from a later one.

With the new prefer_best_timing fetch option (on by default) the search
still asks the enabled providers in order, at most three at a time, but
ranks each result: word-synced, then line-synced, then plain, then an
instrumental flag, then text without a line of lyrics. A word-synced
result ends the search once no earlier provider is still running;
anything less keeps waiting for the remaining providers, bounded by the
existing grace after each better result. Provider order decides between
equal results. With the option off the search keeps the old first-in-
order behavior.

The option is part of the lyrics cache key, so results cached under
provider order are not served as the best timing.
2026-11-15 17:30:49 +07:00
zarzet 6c718fd7bc fix(files): edited songs no longer turn hidden on local and mounted drives
Tag, lyrics and cover writes, in-place conversions, FFmpeg stages and
M3U writes staged next to the song under a name with a leading dot and
renamed it over the song. Samba and NTFS drivers with hide_dot_files
mark a dot name hidden when it is created and keep the attribute
through the rename, so songs edited on a pendrive, an NTFS disk or a
mounted NAS share showed as hidden in Windows Explorer.

The stages now start without a dot and keep a `.partial` part that
library scans already skip; the desktop watcher ignores `.partial.<ext>`
stages too. FFmpeg sweeps still remove dotted stages left by earlier
versions. Songs already hidden keep the attribute until it is cleared
or the song is saved again.
2026-11-11 21:21:19 +07:00
zarzet b98f81046a fix(usb): set the rate on the clock source behind a clock selector
XMOS-based DACs such as the FiiO BTR15 point their input terminal at a
clock selector, not a clock source. Direct USB sent the 4-byte sample
rate request to the selector, which answers with its 1-byte input pin,
so playback fell back with "Short USB control response".

Parse UAC2 clock source, selector and multiplier descriptors and follow
the terminal's clock to its source, reading the selector's current
input when it has more than one. Devices without clock descriptors keep
the old behavior.
2026-11-11 17:59:54 +07:00
zarzet 4b7edf89ce fix(usb): explain Direct USB fallbacks and decode compressed formats
Direct USB reduced every failure to "unsupported", so the app showed one
generic label and a tester with a working DAC could not tell why. It
also played only FLAC, WAV and DSD.

Failure reasons: the native side now reports format_unsupported,
decoder_unavailable, channels_unsupported, dac_format_rejected,
dac_unsupported, dac_open_failed and timeout, with the source codec,
rate, depth and channels. Rust errors are classified by message, and
"No compatible USB format" now names the request and the PCM layouts
the DAC offers, for the log. A file Direct USB cannot play is probed
from its headers before the USB permission prompt, so it reports its
reason without a prompt. The Android 14 mixer and AAudio paths reuse
the source reasons, and the mixer reports a refused rate or depth and
a timeout separately. A source that stops decoding during playback
pauses with decode_failed instead of route_changed.

Formats: with Direct USB, MediaCodec decodes ALAC and FLAC in MP4/Ogg
losslessly, and AAC, MP3, Opus and Vorbis as lossy sources.
Lossless sources up to 16 bits use integer output; 17 to 24-bit
sources use float output, which BitPerfectPcm checks sample by sample.
32-bit lossless sources are refused. Lossy output is rounded to the
DAC's precision without dither, at 24 bits or 16 bits when the DAC
offers nothing deeper. The decoder is primed before the device opens,
so HE-AAC reports its real rate; a later format change stops playback.
Encoder delay and padding are trimmed by frame count, and a seek flushes
the decoder and skips to the exact frame. Multichannel files are
refused, never downmixed. libFLAC and WAV sources are unchanged.

The status label and Audio info mark a decoded lossy source, so a
direct transport is not shown as a bit-perfect lossless chain.
2026-11-11 12:17:01 +07:00
zarzet 5da029f162 feat(lyrics): read .lrc files next to songs and bilingual LRC translations
Now Playing, the desktop player and lock screen lyrics read only the
lyrics inside the song file, so a WAV or any song with its lyrics in a
.lrc file beside it showed no lyrics, while the song details card did.
The player now reads that .lrc file when the song holds no lyrics of its
own, for local files, Android folders and iOS alike. The song's own
lyrics still come first. Edits and saved translations of .lrc lyrics go
back into the .lrc file instead of the song.

The song details card also reads the .lrc file of an Android document,
which the lyrics reader cannot reach.

Bilingual LRC files time each translation like the line it translates.
When most moments of a song are such a pair, both lyrics parsers now
show the second line as the first line's translation. A few shared
timestamps, such as simultaneous singers, stay separate lines.
2026-11-11 00:56:24 +07:00
zarzet 5bcb9e6631 feat(store): read several extension registries in the native repository
The repository keeps an ordered list of registry sources instead of one
URL. Each source has its own TTL cache, ETag/Last-Modified revalidation,
single flight and generation, and its own cache file under store_sources/
named by the SHA-256 of its URL. Sources are fetched concurrently; a
failing source falls back to its cache or reports its error without
hiding the others. Replacing the list keeps the caches of kept URLs and
drops the rest, including their in-flight fetches. The legacy
store_cache.json is deleted on construction.

The catalog lists every supported entry of every source tagged with its
registry URL and display name (the registry's optional top-level name,
else owner/repo for GitHub, else host and path). Downloads take the source
to install from, or the first source listing the id when it is empty.

The mobile object and the desktop channel expose setRepoRegistryUrls,
resolveRepoRegistryUrl, getRepoCatalog and downloadRepoExtension with
registry_url, replacing the single-URL methods.
2026-11-09 03:40:59 +07:00
zarzet 9a9642a755 Merge branch 'fix/merge-pr846' into merge/reenrich-846 2026-11-09 02:19:17 +07:00
zarzet 71cc0c7e66 perf(reenrich): keep the early stop when a batch lists candidates
A batch preview lists candidates for every song so any of them can be
matched by hand. Asking every provider for each song multiplied the
searches of a large library, so quick_candidates keeps the stop at the
first provider with a confident match. Songs nothing fitted still ask
every provider.
2026-11-08 19:23:52 +07:00
zarzet 9dcba8bae2 fix(reenrich): ask every provider and apply one matching release
Re-enrich searched with a total limit of 5, so the first provider filled
every slot with other songs and later providers were never asked. A title
that merely contained the local title matched, another release won with
only a lower score, and the catalog track was applied before the search
result, leaving fields from two sources.

Each provider is now asked with its own limit, title plus primary artist
first and title plus album second. Tracks get a match level (exact,
strong, weak, other); automatic runs apply only exact or strong, or weak
when the file has no album. One winning track supplies every field it
has, and extras by ISRC only fill an empty genre, label or copyright.

Requests can ask for ranked candidates, apply a selected candidate as a
user choice, and limit the search to given providers. Responses carry
auto_match, and a candidate list reports a needed verification when no
provider answered.
2026-11-08 19:23:52 +07:00
zarzet a2c01d72bf feat(library): use folder images as covers for songs without one
Large libraries often keep album art as cover.jpg or folder.jpg beside
the songs instead of embedding it, so those songs showed no cover.

When a song has no embedded artwork, local scans now fall back to an
image in its folder: cover, folder, front, album, albumart (jpg, jpeg,
png, webp), case-insensitive, ignoring hidden and AppleDouble files. A
disc folder (CD1, Disc 2, ...) without one uses its parent's image.
Embedded artwork always wins.

The Rust path scan (desktop, iOS, Android file paths) picks the image
from the directory listing it already makes and hands it to the
workers; the thumbnail key includes the image path, size and mtime, so
a replaced image gets a new thumbnail. CUE tracks get the same
fallback. The Android SAF scan picks the image from its own listing
through new FFI helpers and makes each image into a content-named
thumbnail once per scan.
2026-11-08 15:54:58 +07:00
zarzet a1e8e37266 chore(deps): update Rust dependencies
Bump the exact pins that had semver-compatible releases: aes 0.9.4,
mdns-sd 0.21.7 and zip 9.4.1. Refresh Cargo.lock within the existing
requirements (cc, flate2, log, smallvec, toml, wasm-bindgen and other
transitive patch releases).

Every other direct dependency is already at its newest release. The
vendored patches, the Unicode 15 pins, uniffi, rquickjs, scraper and
html5ever are unchanged. Rust 1.99.1 remains the newest stable toolchain.
2026-11-07 18:54:53 +07:00
Bartolo 62eec0f605 Hide registry entries for non-JavaScript runtimes
Registry entries may set an optional "runtime" field. Entries whose
runtime is missing or "javascript" behave as before; any other runtime
(e.g. "python" for Mornye-Python packages) is hidden from the store
and refused by download(), so one registry.json can serve both hosts.
2026-11-07 12:47:05 +01:00
Bartolo 4831b065f1 test(cover): backdate the legacy thumbnail before the regenerate marker
On filesystems with coarse mtimes the legacy file and the marker can
share a timestamp, so the legacy thumbnail stayed current and the test
got the old picture back. Backdate it like the other regenerate tests.
2026-11-07 12:45:33 +01:00
Bartolo af82f3acb6 feat(extensions): let string settings grant their URL's host
A setting marked "grantsNetwork": true adds the host of the URL the user
typed to that extension's network allowlist. A grant made with http://
allows plain HTTP for that host only; private and local addresses still
need Allow Local Network Access. Saving settings refreshes the grants on
a loaded runtime, so no reinstall or restart is needed.

This lets download extensions offer a self-hosted endpoint without
listing "*" in permissions.network.
2026-11-07 12:17:10 +01:00
zarzet bbd4332653 feat(lyrics): save AI translation word timing
A saved AI translation kept only its line text, so the word-synced
translation went back to line sync once the AI cache was gone. It now
travels with the lyrics as an [x-translation-words:<ms>:<base64 JSON>]
tag, the same JSON as [x-romaji-words].

- Both lyrics parsers read the new tag, keep it only when its words spell
  the line's translation, and move it with the file's offset. The native
  parser sends translationWords only for lines that have it, so the shared
  goldens stay as they were; a new golden covers the tag.
- Saving into the song file or a .mlrc file writes it, a new translation
  drops timing kept for the old one, and the lyrics editor moves it with
  its line like pronunciation timing.
- A shown translation whose timing the file lacks counts as unsaved, so
  the save action and automatic saving add it.
- Older versions drop the tag as an unknown x- ID tag.
2026-11-07 15:31:38 +07:00
zarzet d290a4cee3 fix(cast): stop treating every TV media error as fatal
Android failed the song on any media error message from the receiver,
even if the TV kept playing. The app then showed "TV could not play this
audio file", stopped following the position and stopped sending Pause.

Fail only when the receiver stops with an error, as the iOS and desktop
sender already does. Media errors now reach the app log with their reason
and code on every platform, so a failed seek can be traced.
2026-11-07 01:22:35 +07:00
zarzet 04e8f6cf10 fix(android): reject unknown Direct USB packet timing 2026-11-06 20:52:21 +07:00
zarzet 969ee435e2 fix(lyrics): keep spaces after word closing timestamps
Enhanced LRC can close a word before the space that follows it, as in
<00:01.00>First<00:01.50> <00:02.00>second. Both parsers treated that
space only as an end time and dropped it, so word-synced lines lost the
gaps between words.

Append the separator to the previous word in the Dart and Rust parsers,
and keep the first closing timestamp when more than one follows.
2026-11-05 19:17:53 +07:00