Hardens the renderer auto-detect so users never need the manual
options (the anti-videogame-settings goal).
Ladder (phase 1): a failed start now walks one rung down instead of
jumping to GL — wgpu (default adapter) -> wgpu (opposite
PowerPreference) -> GL -> software. The alt rung exists because the
#542 family is an ADAPTER mixup, not a wgpu failure; its survival is
version-stamped at disarm time (renderer_wgpu_alt) so later starts arm
it directly. GL-rung failures continue to software through the
non-auto branch (the persisted 'gl' carries the 'auto-gl' sentinel
value; a user-picked gl arms the literal 'gl' — unambiguous).
Auto-degradations are version-keyed (renderer_auto_degraded): a new
build re-probes auto once, the ladder re-degrades within a start if
the stack is still broken; manual picks clear both markers.
Sentinel disarm reworked: was a fixed 5s timer that lost the race
against late startup crashes (#558's swapchain error loops forever);
now first real input / close request / clean exit, with a 30s
no-touch fallback. The 5s timer keeps only its DPR-persist duty.
User-picked 'software' no longer arms (it cannot fail to start).
Watchdog (phase 2, #555 diagnosis aid): background thread probes the
event loop's cross-thread dispatch latency every 2s (30s startup
grace, clamped against suspend, 10 consecutive >250ms probes flag
once). Detection only — never switches tiers; surfaces as the 'UI
Loop Latency' Diagnostics row + bundle line so field reports carry a
number. The #555 lag is renderer-independent, so this measures the
layer where it actually lives.
One msgid x7 (the degradation toast is now rung-generic).
cc
A slider drag delivers dozens of SetVolume commands per second and
each logged an info line from the audio thread — 758 of 1025 lines in
a #555 field log. Log-level change only; the audio path is untouched.
cc
zbus 5 logs every dispatched/read D-Bus message as a multi-KB Debug
dump at INFO (through the tracing-log bridge, which also emits
tracing::span events). With any MPRIS applet polling GetAll this
flooded the file sink within ~1s of startup and drowned everything
else — both #555 field logs confirmed it (one diagnostics upload
covered exactly one second of runtime). Filtered records now cost
nothing (log! checks the level before formatting). RUST_LOG still
replaces the whole default, so full zbus tracing stays available.
cc
One sentence under the subscription notice so the always-visible
'Start offline' link is self-explanatory: QBZ works as an offline
player without a Qobuz account, with no access to the Qobuz library.
One msgid x7 locales.
cc
The 'Start offline' link was gated on has-previous-session and the
shell backstop refused to open without a recorded user, so a user who
never logged in (e.g. a purchases-only Qobuz account, #553) could not
use the app at all. The link is now always visible and the offline
shell falls back to the GUEST profile — user 0, the Tauri-era fallback
this port used to refuse — building a real per-user store set (local
library, mixtapes, prefs) under users/0. AppRuntime::activate_offline
already resolved user 0 and never persists it as the last-user marker,
so the guest never masquerades as a real session.
First real login takes possession: AppRuntime::activate renames the
users/0 data+cache dirs to the account's id when that account has no
profile on this machine yet (existing profiles always win — the guest
dirs stay parked, no merge). Best-effort, login never breaks on a
failed rename.
cc
serde_json::from_value::<Vec<T>>(...).unwrap_or_default() nuked the
ENTIRE tab when one favorite among thousands failed to fit the model
(a delisted or odd-shaped item): the count badge kept the separately
parsed total while the list rendered empty — 'Tracks 6125' over 'No
favorite tracks yet' (#556, 6125-track library on macOS; a 101-track
library never hit the poison item). Items now deserialize one by one;
misfits are skipped with a warn carrying the item id and error, plus a
summary line, so the reporter's log will name the exact offender.
Applies to all four get_favorites tabs (tracks/albums/artists/labels).
cc
The log uploader was reachable only through Settings > Developer >
Open log viewer, so users filing issues never found it and reports
arrived without logs. A pinned entry at the bottom of the settings
sub-nav (visible from every section) now opens the same log viewer
(copy + upload). One new msgid x6 locales.
cc
Tauri parity regression: the old app shipped
tauri-plugin-single-instance; the Slint rebuild lost it, so every click
on a pinned shortcut or launcher entry spawned another full player
(reported on Hyprland in #544 and KDE in #559). The first instance now
owns the com.blitzfc.qbz session-bus name (auto-grantable inside
Flatpak — the name is the app id, no finish-args change); a second
launch asks the owner to raise its window through the already-exported
MPRIS Raise method (full name org.mpris.MediaPlayer2.com.blitzfc.qbz)
and exits. Blocking zbus 5 + async-io on purpose (the tokio feature is
forbidden graph-wide); any D-Bus failure falls through as primary so
the guard never blocks startup.
The issues' other symptoms (generic Wayland icon, forgotten window
position) trace to the missing app_id in 2.0.0 — already fixed by
dae6e7f6 (#544): compositors key both the icon and geometry-restore
rules on app_id.
cc
The Tracks tab is server-paginated (200 rows/page), so the new 8-option
sort dropdown (Default / Artist / Title / Year / Date added, mirroring
the Albums toolbar control) rides the SQL query through a validated
ORDER BY allowlist in search_with_filter_page (same pattern as the
albums page query), re-queries page 1 on change (gen-bumped, stale
in-flight pages drop), and persists in locallibrary_prefs. Every sort
key ends in id so LIMIT/OFFSET pagination is deterministic across
ties. The once-only Plex page-1 merge re-sorts client-side with a
comparator matching the SQL semantics. Group modes keep their
client-side visual reorder on top, unchanged. Zero new msgids — all
labels already exist.
cc
A Manifest published without its amd64 DIST entry bricks emerge for
every amd64 user, which is worse than failing the job; a 3x30s retry
absorbs release-asset propagation lag right after the tag. Also
documents the #557 root cause at the top: the overlay must never be
edited by hand — an ebuild change without Manifest regeneration is a
guaranteed digest mismatch; template changes are republished by
re-dispatching this workflow.
cc
Tauri regression: the Svelte playbackService caught TrackUnavailable play
errors and called autoSkipToNext (bounded at 5 consecutive skips, issue
Tauri#467); the Slint port only logged and cleared the spinner, so the
queue cursor parked on the dead track and playback stopped.
play_audible's error arm now detects TERMINAL unavailability (substring
contract over the flattened ApiError Display: TrackUnavailable /
NonStreamable / NoQualityAvailable — transient errors keep NOT skipping,
the client already retries those), toasts, honors stop-after-this-song
on the failed track, and advances through the real machinery
(advance_to_playable + after_track_change, async recursion boxed).
Superseded fetches (user already picked another track) never skip.
Counter resets when audio actually starts. One new msgid x6 locales.
cc
Ports the band + hysteresis scheme into SearchResultsView's bespoke
AlbumGrid (no window-changed — search artwork is dispatched per loaded
page onto the directly-rendered model) and wires the Albums tab
instance to the view's Flickable.
cc
Grouped section grids now receive the windowing inputs when the surface
is windowed; each section's content-offset is the absolute-position
delta to the collection root — scroll-invariant (both live in the same
Flickable content) and exact in plain Flickable/for chains. Section
grids do not forward window-changed: grouped artwork stays full
dispatch, so windowing there is mount-only.
The band seed moves from init to a one-shot Timer a tick after mount:
evaluating the target bindings from init would pull absolute-position
while layout is still evaluating (the documented recursion panic).
Until it fires, the default [0,0] band keeps at most one row mounted
for a frame.
cc
Flips windowed on the four AlbumCollectionView consumers with their
scroll wiring and documented content-offset approximations. Mount-only:
artwork stays per-page full dispatch (index-keyed, append-only models),
so covers are already on the model when a card mounts. Load-more
triggers untouched (grid height formula is windowing-invariant).
cc
Column band from viewport-x (visible plus one viewport of margin each
side) with the same contract hysteresis as AlbumGrid; off-band cards
are same-footprint placeholder tiles. Element-count only, always on —
section loaders dispatch artwork for the whole strip up front, so a
card mounting later picks its cover from the model instantly. Covers
every Home/ForYou/Discover/Search strip through one component.
cc
Mirrors the LL albums pattern 1:1: artwork migrates from index-keyed
FavoriteAlbum{index} to id-keyed FavoriteAlbumById{id,gen} (re-sort
immune), a FAV_ALBUMS_GEN token guards stale applies, the band from the
windowed grid routes through a thread-local BandDispatcher (180ms
leading+trailing), covers far outside the window evict back to the
placeholder, and list/grouped modes keep full dispatch. The load path
no longer dispatches every favorite cover up front — the first window
report pulls only what is near the viewport.
cc
Split the row band into pure target props and a stored mounted band.
Expansion still follows the target immediately (cards mount before
reaching the visible area); contraction lags by one row, so a few-px
oscillation across a row boundary no longer mounts/unmounts a whole
row of AlbumCards per direction change. window-changed now fires only
when the mounted band actually moves. Non-windowed consumers
(Discover/Favorites/Labels/grouped) are behavior-identical.
cc
Band reports from the windowed LL Albums grid now route through a
leading+trailing throttle (new viewport::BandDispatcher, 180ms) instead
of dispatching artwork on every row-boundary crossing during a fling.
The scheduled pass captures ALBUMS_GEN and drops itself if a reload
superseded it — the reload's own derive dispatch is authoritative.
cc
Owner feedback on the first smoke: the editor stacked 11 full-width
token rows. The tokens now render as compact inline cells (swatch with
its label underneath) grouped side by side — SURFACES/TEXT/ACCENT on
one line, STATUS/OTHER on the next — and ONE shared ColorPicker
expands below the grid for whichever token is open, so editing doesn't
reflow the swatches. The picker's external-change reseed (init +
changed-value guard) already handles switching tokens while open.
cc
Per-element custom theme editor (port of Tauri's swatch editor,
extended): first-class Custom theme seedable from any active theme,
12 base tokens deriving the full coherent palette, in-app ColorPicker
primitive, JSON persistence, live apply, 7 locales.
cc
Selecting Custom for the first time seeded the editor from the OLED
default; snapshot the live palette instead (the theme the user was
just looking at) — customizing what you see is the point of the
editor. Subsequent selections keep loading the persisted base.
cc
Wire the custom-theme editor end to end. In AppearanceSettings, when the
"Custom" theme is selected, show a "Start from current theme" button, a "Dark
theme" polarity toggle, and a grouped token grid (SURFACES / TEXT / ACCENT /
STATUS / OTHER) of 11 base tokens. Each token is a swatch that toggles an
inline ColorPicker; edits route through AppearanceState.custom-set-token
(drag) and custom-set-token-hex (hex commit).
The crate::custom_theme controller reads the base straight from the editor
swatch properties (the in-memory source of truth, no per-drag disk reads),
sets the touched field, re-derives via qbz_theme::theme_from_base, pushes the
palette live and persists. seed_from_current snapshots the live Theme global
into the base; toggle_dark flips polarity; set_token_hex parses in Rust (Slint
cannot). main.rs registers the four callbacks.
Also harden the ColorPicker reseed guard: skip decompose when the fed-back
value equals our own emitted color, so a drag ending at v=0/s=0 no longer
resets the hue. New strings added to all 7 gettext catalogs (es Mexican/
neutral); msgfmt --check clean; full .slint tree validated via the slintcheck
stub.
cc
Add crates/qbz-ui/ui/primitives/ColorPicker.slint: a self-contained inline
HSV picker (no PopupWindow, so no z-order/positioning fight). A saturation/
value square (pure-hue base + white->transparent + transparent->black
overlays, draggable crosshair), a 7-stop hue strip with a draggable marker,
a live preview swatch, and an editable #rrggbb HEX field.
The picker owns its hue/sat/val so dragging s or v to an edge never loses the
hue; it seeds from `value` on mount and on external changes, guarded by a
`dragging` flag so the round-tripped color the parent feeds back mid-drag does
not fight the crosshair. HSV->RGB uses Colors.hsv; RGB->HSV and int->hex are
self-contained pure functions (Slint has no local bindings, so intermediates
are derived helper properties). Hex parsing is delegated to the parent via
`hex-committed` (Slint cannot parse a hex string). Emits `changed(color)` live
on every drag frame. Validated with the slintcheck stub.
cc
Append a "Custom" entry to the theme dropdown, right after "Auto (dynamic)"
(theme::CUSTOM_SLUG/custom_index, mirroring the AUTO_SLUG pattern). New
crate::custom_theme persists the CustomThemeBase to
<data_dir>/qbz/custom_theme.json (load/save mirror ui_prefs) and derives it
via qbz_theme::theme_from_base.
Startup: when the persisted theme slug is "custom", seed the editor swatches
from custom_theme.json and apply the derived palette (load_or_seed seeds the
OLED default the first time). The dropdown "theme" arm gains a Custom branch
with the same live-apply path. AppearanceState grows theme-is-custom, the 11
custom-* swatch colors, custom-is-dark, custom-open-token and the
custom-set-token / custom-seed-from-current / custom-toggle-dark callbacks
(handlers wired in a later commit).
cc
Add crates/qbz-theme/src/custom.rs: the extended port of the Tauri "edit
theme colors" feature. A CustomThemeBase of 12 hand-picked semantic tokens
(surfaces, text, accent, status hues, border, favorite + is_dark polarity),
stored as opaque #rrggbb hex strings, derives the full ThemeColors contract.
Unlike Tauri (8 raw vars, no derivation, over the auto-theme only), editing a
base token now cascades: theme_from_base derives the accent triplet
(+0.10/-0.10 lightness, contrast-picked accent_text), the danger/warning/
success families (0.1/0.3 tints, polarity hover), muted/disabled text tiers,
border_strong, the translucent edges and the polarity alpha ramp.
Reuses the existing math instead of duplicating it: generator::{tint,
pick_btn_text_for_accent_set, ensure_text_contrast_target} are now pub(crate);
PaletteColor::shift_lightness and alpha_ramp are shared as-is. base_from_theme
is the inverse seed ("Start from current theme") and round-trips exactly for
any theme this module authored. Adds Rgba::to_hex. 9 unit tests (roundtrip,
polarity, determinism, accent-text legibility, malformed-hex fallback, JSON).
cc
The random-album / random-artist / random-playlist / random-label
buttons all exist (1:1 with 1.x) but rendered as bare shuffle icons —
1.x had title tooltips, so users couldn't tell them apart from a
tracks-shuffle and reported the feature as removed (#554). CircleAction
already supports the shared tooltip overlay; set it on all seven
Library header actions. 7 msgids translated in all 7 locales.
Fixes#554
cc
Review follow-ups: extract_palette decoded the full image before the
100x100 downsample with no limits — a decompression-bomb PNG expands
to tens of GB, and this box hard-freezes on OOM; decode is now bounded
at 12k x 12k / 512 MB (any real wallpaper fits). The generating
indicator is now set through upgrade_in_event_loop so the select_image
path (which runs on a tokio thread, where a plain upgrade() returns
None) shows it too. Startup generation stays synchronous by design:
with the decode bounded it costs milliseconds and the first paint
comes up in the generated palette instead of flashing the default.
cc
Vendor i-slint-renderer-skia 1.16.1 with a second-chance typeface
workaround: skia bug 310510989 makes new_from_data fail for TTC faces
at index > 0 on macOS, and for AppleSDGothicNeo.ttc — which packs
every weight at index > 0 — the upstream read/write-fonts rebuild is
ALSO rejected by CoreText, so draw_glyph_run silently dropped every
Korean glyph run at non-regular weights (Japanese survived because
Hiragino ships one file per weight, always index 0). Resolve the same
face through CoreText by family name (typographic name preferred,
localized names resolve too) + OS/2 weight; the match points at the
same installed file so parley's glyph ids stay valid, enforced by cmap
spot checks that reject a mismatching face rather than draw wrong
glyphs. Probe-verified on the Mac Mini: matched face returns identical
glyph ids for U+C0AC across w500/600/700.
cc
Repoint the existing auto-theme Settings scaffold from the "System" theme to
the new dynamic "Auto (dynamic)" option:
- state.slint: add theme-is-auto (distinct from theme-is-system, which stays
the native OS-palette follow).
- AppearanceSettings.slint: gate the Source dropdown, image picker, detected-DE
hint and Regenerate button on theme-is-auto, and drop the placeholder
`unsupported` flag now that the rows are functional.
- translations: add "Auto theme generation failed" to all 7 catalogs and give
the es "Wallpaper Sync" string a proper Mexican-neutral translation.
cc
Integrate the dynamic "Auto (dynamic)" theme into the bin without disturbing
the static-theme path:
- ui_prefs: add auto_theme_source ("system"|"wallpaper"|"image") and
auto_theme_image_path, with serde defaults and index<->key helpers.
- theme.rs: extract push_colors() (shared by static + auto paths), append the
"Auto (dynamic)" dropdown entry, and add AUTO_SLUG / auto_index /
is_auto_index / selected_index_for_slug so the "auto" slug maps to the
appended entry (it has no static ThemeId).
- auto_theme.rs (new controller): source_from_prefs, startup apply (OLED
fallback + toast on failure), off-thread regenerate, rfd image picker, and
source switching — all pushing the palette back on the event loop.
- main.rs: auto-aware startup seeding/application, an "auto"/"auto-theme-source"
arm in on_appearance_select, and a new on_appearance_action handler for the
image picker + Regenerate button.
cc
Port the legacy Tauri src-tauri/src/auto_theme/{palette,system,generator}
logic into a new frontend-agnostic qbz-theme::auto module. The palette
extraction (k-means), WCAG contrast math, HSL shifts and DE detection are
carried over 1:1; only the final assembly changes: the output is now the
registry ThemeColors contract instead of a CSS custom-property map.
The generator mirrors registry::StdSpec::build so a generated theme
composites identically to a static one — same polarity-driven alpha ramp,
success/focus_ring/favorite/border_muted derivations, and status tints.
Per-status text tokens (--btn-danger-text/--btn-warning-text) are dropped
since ThemeColors has no field for them.
Adds `image` (0.25, minimal codec features) to qbz-theme and unit tests for
the mapping (polarity, derived families, determinism, scheme path).
cc
The current-row equalizer animated via animation-tick() on the strong
(wgpu) tier — every step is a full-window repaint, so a playlist view
with the playing row on screen sat at ~4.5% CPU on a 165 Hz monitor
for a three-bar affordance. The bars now always use the shared ~8fps
coarse clock (already the accepted look on the GL/software tiers) and
the AppShell timer additionally runs while audio plays; when the
playing row is not on screen nothing reads the property, so the tick
dirties no binding and costs nothing.
cc
Vendor i-slint-core 1.16.1 with a one-rule patch in sharedparley's
line culling: with overflow: elide, every line whose bottom edge
exceeded max_height was dropped — INCLUDING the first one, eliding the
text to literally nothing. CJK fallback fonts report line metrics
taller than snug fixed-height boxes (Hiragino at 14px measures ~21px
inside TrackRow's 20px title line), so on macOS every CJK track title
rendered blank while Latin titles fit; Linux survived because Noto's
metrics fit the box. Root-caused with a headless render matrix on the
Mac Mini: the verbatim TrackRow structure reproduced the blank title,
removing elide rendered it, and the patched core renders all variants.
The first line still receives the ellipsis treatment as the last
visible line.
This is the real #543 mechanism. The earlier femtovg render_glyph
retry (3a1c367d) addressed a different, hypothesized layer and stays
as harmless defense in depth.
Fixes#543
cc
take_stream_error_message is a SharedState method, not a Player one —
the toast commit called it on the Arc<Player> and the release build
caught it (the bin crate can't be cheaply cargo-checked locally).
cc
register_audio inserted every cast track's complete bytes into the
entries map and NOTHING ever removed them — not stop, not disconnect;
only app shutdown dropped the server. A DLNA session therefore grew by
one full track per auto-advance (issue #550: 50 MB at launch to
4.5 GB, never released). Three changes:
- register_audio/register_file evict every other entry — only the
track being cast is servable anyway.
- stop_if_cast and disconnect clear the entries outright, so stopping
actually releases the buffers.
- MediaSource::Data is Arc-backed: the per-request entry clone no
longer copies the whole track (it did, on every GET/HEAD/range
request), and an in-flight response keeps its own Arc so eviction
can never truncate an ongoing range request.
Fixes#550
cc
A bare DLNA STOPPED is ambiguous — strict renderers hiccup mid-track —
so end-of-track now requires the max observed position (RelTime resets
to 0 on STOP; the instantaneous value is useless) to reach within 5 s
of the duration. Two escape valves keep the guard from ever blocking a
genuine end, which review flagged as the regression risk: renderers
whose RelTime never moves (position signal unreliable — common, and
duration nearly always resolves via the catalog fallback, so the
`duration <= 0` hatch alone never fires) skip the guard entirely, and
a STOPPED persisting 4 consecutive polls advances anyway (~4 s late
beats a queue stuck forever). Premature stops are logged with a
counter and recover on PLAYING.
Based on the end-of-track guard contributed in PR #547, reworked per
review findings.
cc