* fix(stream): defer M4A probe until moov tail or fast-start prefix
Ranged moov-at-end M4A started Symphonia format probe as soon as ~384 KiB
linear data arrived, before the parallel tail prefetch filled the moov
atom — probe hit end of stream and skipped the track. Wait for tail_ready
or detect fast-start moov in the prefix; do not arm playback from linear
bytes alone when tail prefetch is active.
* fix(isomp4): skip EOF-spanning mdat after moov-at-end is parsed
Second-pass header scan with moov already loaded still tried to read
through mdat→EOF on RangedHttpSource holes, causing format probe
end of stream. Re-check play generation after moov wait.
* fix(isomp4): fix AtomIterator overread after seek in patched demuxer
`AtomIterator::new_root(reader, len)` treats `len` as bytes available
from the current `reader.pos()`, not the absolute file length. After
`mss.seek(resume_at)` we were passing the absolute `total_len`, so the
iterator thought there were `total_len` bytes left and tried to read
past EOF on the next iteration, returning "end of stream".
Fix: pass `total_len.map(|tl| tl.saturating_sub(resume_at))` (remaining
bytes from the new position) in both branches:
- `moov.is_none()` (moov-at-end layout, seek to moov offset)
- `moov.is_some()` (fast-start layout, skip bounded mdat body)
This caused Symphonia to fail probing moov-at-end M4A files read from
local disk (hot-cache) and from in-memory buffers — every decode attempt
returned "end of stream", analysis fell back to `byte_envelope_no_ebu`
(no EBU R128 loudness), and rodio produced distorted audio.
Also in this commit:
- `resolve_playback_format_hint()` helper to resolve hint from URL,
stream suffix, Content-Disposition, or byte sniff
- ISO-BMFF diagnostic helpers (`isobmff_buffer_looks_complete`,
`log_isobmff_buffer_diagnostic`, `mp4_suspect_zero_holes`)
- Probe-fallback path for ranged-stream failures now uses these helpers
to decide whether to refetch or wait for the in-flight download
* chore(audio): remove redundant hint recomputation and add missing blank line
`bytes_hint_for_wait` in the ranged-stream fallback path was an exact
duplicate of `effective_hint` already in scope — reuse the existing binding.
Also add missing blank line after `wait_for_ranged_mp4_probe_ready` in mod.rs.
* chore(release): CHANGELOG and credits for PR #757
Add Fixed entry for M4A moov-at-end probe fix and credits line in
settingsCredits.ts under existing cucadmuh contributions.
* fix(audio): extract BuildSourceArgs to fix clippy::too_many_arguments
`build_playback_source_with_probe_fallback` had 12 parameters, exceeding
the clippy limit of 7. Group url/gen/hints/fade/hi-res/duration into
`BuildSourceArgs` so the function signature stays at 4 arguments.
* feat(playback): stream buffering UI, ranged M4A tail prefetch, demuxer fix
Defer seekbar/progress until HTTP stream is armed for both legacy and
RangedHttpSource; show buffering overlay on cover art. Add MP4 tail
prefetch and Symphonia isomp4 bounded-mdat/moov-at-EOF probing so
moov-at-end M4A can start without reading the full mdat.
* feat(hot-cache): spill large ranged streams to disk for promote
When a ranged HTTP download completes above the 64 MiB RAM promote cap,
write the existing buffer once to app-data stream-spill/ and register it
for hot-cache promote (rename) and replay via fetch_data. Analysis seeds
from the spill file up to the local-file cap (512 MiB).
* fix(ui): stream buffering — grayscale cover and static clock icon
Desaturate player and queue cover art while isPlaybackBuffering; keep a
non-animated clock overlay for visibility without the spinning animation.
* fix(playback): review follow-up — tests, i18n, spill cleanup, changelog
Clippy and test layout fixes; stream spill orphan cleanup on startup;
buffering flag guard in progress handler; bufferingStream in all player
locales; CHANGELOG and contributor credits for stream/M4A work.
* docs: attribute stream buffering and M4A streaming to PR #737
* test(audio): avoid create_engine in stream spill unit test
CI runners have no audio output device; test spill take/consume via
the Mutex slot only, matching install_stream_completed_spill tests.
Improve normalization consistency by separating cache refresh for non-playing tracks, hardening track-id cache lookups, and keeping UI gain state aligned with actual playback. Also reduce startup stalls by preferring non-seekable streaming when format hints are missing and by tuning ALSA/analysis paths to avoid underrun-prone churn.
- patches/cpal-0.15.3 wasapi device.rs: annotate MutexGuard return with
'_ so mismatched_lifetime_syntaxes stops firing.
- lib.rs: gate the MouseButtonState import to non-Windows targets (the
Windows tray uses DoubleClick and never pattern-matches it).
Cosmetic only, no behavioural change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
tauri-action always re-packs Psysonic.app into .app.tar.gz after tauri
CLI finishes (regardless of includeUpdaterJson), so the .sig that tauri
CLI produces never matches the final tarball — and the sig tauri-action
was supposed to produce silently goes missing ("Signature not found for
the updater JSON. Skipping upload...").
Rather than fight the repack, sign the repacked tarball ourselves in a
follow-up step using `tauri signer sign` against the same private key
(from TAURI_SIGNING_PRIVATE_KEY), then upload the fresh .sig. tauri-
action still handles the DMG + .app.tar.gz upload; we only add the .sig.
Also removes an unused `std::rc::Rc` import in the vendored cpal patch.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Two independent user-facing fixes rolled together.
── Lyrics: YouLyPlus provider (issue #172) ──
Adds word-by-word synced (karaoke) lyrics as an alternative lyrics mode.
Backed by the public lyricsplus aggregator (Apple Music / Spotify /
Musixmatch / QQ Music) — no API keys on our side, subscription costs are
borne by the backend operator. Five mirrors are tried on network failure.
Settings → Lyrics now exposes a mode radio (Standard vs YouLyPlus) plus
a static-only toggle. YouLyPlus misses silently fall back to the existing
server + LRCLIB + Netease pipeline so obscure tracks still resolve. The
drag-to-reorder source list is hidden in YouLyPlus mode since the
external aggregator manages its own source priority.
Rendering: per-word spans on each line, active word highlighted with
accent-tinted glow. Subscribed imperatively via usePlayerStore.subscribe
so 500 ms progress ticks do not re-render the whole lyrics block. The
Fullscreen Player reuses the same pattern; the 5-line rail layout is
unchanged. white-space: pre on .fs-lyric-word preserves the trailing
spaces the API embeds in each syllabus token (flex context would
otherwise collapse them).
i18n: new keys in all 8 locales (de, en, fr, nl, nb, ru, es, zh).
── macOS mic-prompt suppression ──
Ships a vendored cpal 0.15.3 at patches/cpal-0.15.3/ wired via
[patch.crates-io]. The only change is in src/host/coreaudio/macos/mod.rs:
audio_unit_from_device now unconditionally uses IOType::DefaultOutput for
output streams instead of conditionally choosing HalOutput.
AUHAL (HalOutput) is classified as input-capable by macOS TCC and
triggers the microphone-permission dialog at first launch / after each
update even for playback-only apps. Previous attempts (removing
NSMicrophoneUsageDescription, setting com.apple.security.device.audio-input
false in Entitlements.plist) did not suppress the prompt because TCC
fires at AudioUnit instantiation, not at plist level. The upstream fix
in cpal PR #1070 / cpal 0.17 only helps when the pinned device equals
the system default; always-DefaultOutput covers all cases.
Tradeoff: per-device output selection is a no-op on macOS — the stream
always follows the system default. Settings.tsx surfaces this via
audioOutputDeviceMacNotice (hides the CustomSelect on macOS) and skips
audio_list_devices + device-watcher effects there. Matches the behaviour
of Apple Music / Spotify on macOS.
Also removes the now-obsolete com.apple.security.device.audio-input
entitlement (sandbox is disabled anyway, so it was ignored).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
M4A files store the iTunSMPB value behind a 16-byte binary 'data' atom
header. The previous parser took bytes directly after the ASCII key,
landing in the binary header and producing a corrupted parts array
(total_valid=348 instead of ~13M samples), which caused take_duration
to exhaust the source in ~8ms — silent no-playback.
Fix: search for the " 00000000 " sentinel within 128 bytes of the key
to locate the actual value string, falling back to the old offset for
ID3v2/MP3 files where no binary header exists.
Also patches symphonia-format-isomp4 (local crate under patches/) to:
- Tolerate SL descriptor predefined=0x01 (older iTunes M4A)
- Convert descriptor.unwrap() panic to a proper decode error
- Gracefully skip malformed trak atoms (e.g. MJPEG cover-art streams)
instead of failing the entire probe
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
cucadmuh
Maxim Isaev
Psychotoxical