Enumerate PipeWire sinks via `pw-dump` (the daemon's native socket) as the
primary path, falling back to the existing `pactl` parser only when pw-dump is
absent or empty. pw-dump needs only `pipewire-bin` — not `pipewire-pulse` (the
Pulse server pactl talks to) nor `pipewire-alsa` (the bridge PCM CPAL relies
on), which are commonly missing on PipeWire-only Ubuntu and produced the empty
device list. As a bonus it yields the exact `alsa_output.*` node.name that the
HiFi wizard otherwise asks the user to paste by hand.
Discovery code only: `create_output_stream` and the protected sample-rate
passthrough path are untouched. max_sample_rate stays None from pw-dump because
its EnumFormat reports the currently negotiated rate, not the device maximum —
the /proc/asound capability probe remains the honest source.
Pure `parse_pw_dump_sinks` helper + 2 unit tests (fixture: USB DAC + PCI card +
filtered-out capture source; garbage/empty JSON).
cc
Album and local-album track lists render a "Disc N" header above the
first track of each disc when the album spans more than one disc
(media_number-driven; single-disc albums render flat), matching Tauri.
Each header carries a context menu (new — Tauri has none) with Play all /
Shuffle / Add to queue / Play next scoped to that disc's tracks only; the
whole-album header buttons are unchanged.
The menu lives in a DiscHeaderMenu wrapper because Slint rejects a
PopupWindow placed directly under a for/if delegate — wrapping it makes
the popup a non-root child (same pattern as TrackRow + TrackContextMenu).
Every TrackItem constructor gains disc-header-number: 0; artist
top-tracks (a flat cross-album list) default the disc to 1.
cc
Single-cover Discover playlist cards now letterbox the cover with
image-fit: contain over its dominant colour (computed host-side from the
decoded cover), matching Tauri, and show the first tag as an accent
category subtag. Add a multi-select category filter dropdown (sm toolbar
variant, no pills per ADR-008) that re-filters the cached playlists row
client-side; an empty selection shows everything. Raise the playlists
cap from 18 to 100 so the filter operates on a full set.
cc
Bring the Slint mix generation to Tauri parity: build a track_to_analysed
payload from up to 9 spread-sampled seeds (pick_spread -> get_track ->
{track_id,artist_id,genre_id,label_id}, drop artist_id==0) for the PRIMARY
/dynamic/suggest algorithm, limit = max(1, 50 - analysed), and fall back to
the empty-analysis call when the primary returns nothing. The listened-seed
build is factored into mix_listened_seed_ids (recents + favorites, ~120,
Qobuz-only) as a seam the reco-store swap (b3) replaces. Logs seed/analysed/
result counts so an empty mix is diagnosable.
cc
Slint reserves 'row' as a GridLayout attached property, so the configurator
modal's 'in property <ConfigRow> row;' failed to compile ("Cannot override
property 'row'"). Rename it to 'entry' — the declaration, the seven
root.row.* accesses, and the for-binding pass-through.
cc
Cleanroom port of the Tauri reco-store into qbz-app::settings::reco_store
(frontend-agnostic, no tauri::State, ADR-006) so the Slint frontend can seed
DailyQ/WeeklyQ, the WeeklyQ 7-day window, and top-genres. Opens the SAME
per-user reco/events.db that Tauri uses (cross-frontend history share); ports
the reco_events schema + indexes, log_play/favorite_event, get_recent and
get_favorite track ids, a windowed get_recent_track_ids_since, get_top_genres,
get_home_seeds, and the train() scorer verbatim. 6 headless tests green.
Tauri's own reco_store is left as-is (rewiring it is a deferred cleanup).
cc
Tauri's DailyQ/WeeklyQ /dynamic/suggest seeds a track_to_analysed payload
(~9 resolved {track_id,artist_id,genre_id,label_id} tuples) for the PRIMARY
mix algorithm; the empty-analysis call is only its fallback. The Slint port
only ever sent the empty-analysis path. Add a TrackToAnalyse model
(qbz-models) and a get_dynamic_suggest_full method on the qobuz client and
qbz-core that carries the payload; the existing get_dynamic_suggest now
delegates with an empty analysis (behaviour unchanged). Plumbing only — the
mix.rs build of the analysis tuples + the two-call fallback lands next (a2).
cc
DailyQ/WeeklyQ build their /dynamic/suggest seed from recently-played track
ids. Recently-played now includes local and Plex tracks whose ids are not
Qobuz catalog ids — numeric Plex rating keys even passed the u64 parse, so
they reached Qobuz and the mix request failed. Filter the seed to Qobuz (and
legacy pre-source) tracks before parsing, and skip the request when no Qobuz
seeds remain. Local tracks stay visible in the Recent rows.
cc
The section list used vertical-stretch inside a panel sized to
preferred-height; a stretch child there collapses to ~0 (a Flickable's
intrinsic preferred-height is ~0 — the documented Slint trap), so the list
was near-invisible. Size it to its content capped at the window minus the
modal chrome; the Flickable scrolls when the rows exceed it.
cc
The configurator offers Qobuzissimes on the Home tab (in DEFAULT_PREFS.home,
off by default), but Slice 2d had removed it from the Home cache, so enabling
it rendered nothing (descriptors_for found no backing data and dropped it).
Visibility is now governed by the prefs (off by default), so keep the data in
the Home cache pool; enabling Qobuzissimes on Home now renders.
cc
The prefs-driven Home skeleton gated on home-sections/editor-sections being
empty, but the descriptor list already carries the non-album rows at seed
time, so the gate never fired and the user saw empty-state placeholders
during the initial fetch instead of the skeleton. Gate the skeleton on
HomeState.loading alone, and suppress the Recently-Played placeholders while
loading, so only the skeleton shows during the fetch (1:1 Tauri).
cc
Add the "Customize" modal that drives the Slice 5 prefs: reorder + show/hide the
active tab's Discover sections, live-persisted (no Save button).
- DiscoverConfigModal.slint: scrim + centered panel (CreatePlaylistModal idiom),
title + close-X, an amber triangle-alert perf-warning banner, an "{} of {}
enabled" count, a scrollable section list, and a Reset-to-defaults footer
pinned left. ConfigRowView = [QbzCheckbox] [label] [up] [down]: the row tap and
the ADR-008 checkbox both fire toggle-section; the 28px ghost ReorderButtons
fire move-section and dim at the list boundary (index 0 / last). Section labels
come from DiscoverState.config-rows (resolved in Rust); chrome strings are @tr
literals.
- HomeView.slint: bind the previously-dead gear to
DiscoverActions.open-configurator() (opens for the active tab).
- AppShell.slint: mount DiscoverConfigModal after the other modals so it sits on
top of them (ADR-009 z-stack by declaration order).
cc
Replace the fixed Discover layouts with the prefs-driven descriptor repeaters
that read DiscoverState's ordered ENABLED section lists, so the configurator's
visibility + reorder choices drive what renders.
- HomeView.slint: drop `for section in HomeState.sections` + the 5 hardcoded
typed blocks; render one repeater over the active tab's descriptor list
(editor-sections when active-tab == "editorPicks", else home-sections). The
single-root delegate dispatches on desc.id: generic album carousels bind the
embedded desc.section; qobuzPlaylists / continueListening / recentlyPlayedAlbums
/ mostStreamed(slim) bind HomeState fields and keep their length gates +
placeholders. `if forYou: ForYouView` stays a static sibling.
- ForYouView.slint: replace the fixed if-sequence with a descriptor repeater
over foryou-sections; each arm reads the typed ForYouState field by id and
keeps its own length gate (qobuzMixes excepted — always rendered). The
skeleton stays a static sibling. essentialsByGenre has no arm (Slice-2c-blocked;
Rust drops it from the list).
- home.rs: select_tab writes DiscoverState.active-tab (single source of truth,
alongside the kept HomeState.active-tab pill bindings) and re-renders the
active tab from the cache via the descriptor lists; apply_home pushes the
descriptor lists once the cache is populated.
- main.rs: the initial discover load fires album-section artwork via the
descriptor-targeted jobs returned by select_tab (DiscoverSectionAlbum), spawned
on the UI thread; slim-grid / recent-album / playlist covers still target the
HomeState fields the fixed arms render.
Build-time verifications to confirm (documented in the .slint comments): the
for-over-ternary-model form compiles, and an empty delegate does not double the
outer 40px spacing (Rust already drops data-empty album ids — policy b).
cc
Add the Slice 5 foundation for the Discover section configurator: a single
DiscoverState/DiscoverActions global pair owning both the prefs-driven render
descriptor lists and the modal payload, plus a thin frontend controller over
the headless qbz_app discover_prefs store.
- state.slint: SectionDescriptor + ConfigRow structs and the DiscoverState /
DiscoverActions globals (active-tab lives here; render loop + modal read it).
- app.slint: re-export the new globals/structs so the Rust layer can bind them.
- discover_prefs.rs: per-user store lifecycle (init_for_user/teardown mirror
fav_cache), in-memory authoritative prefs, descriptor + config-row push
helpers, id->kind/label maps, and toggle/move/reset/open/close handlers
(mutate -> persist -> re-push -> re-render, re-firing artwork for the active
Home/Editor tab). Labels resolve in Rust (English today) since Slint @tr needs
a literal key and the gettext pipeline is unwired.
- home.rs: add the SectionData.id key; build the Home/Editor ordered descriptor
lists from the cached section data (album ids embed their DiscoverSection,
fixed ids carry an empty section); rerender_active_tab + descriptor artwork
jobs. Empty album sections are dropped (policy b); placeholder-bearing Home
sections are kept and gated in the view.
- artwork.rs: DiscoverSectionAlbum target so covers paint on the descriptor
list (DiscoverState.home-sections / editor-sections) rather than HomeState.
- main.rs/auth.rs: mod + seed at both shell entries + per-user init/teardown at
every session-activation point + the 5 DiscoverActions callback registrations.
Descriptor lists are pushed but not yet rendered (the views still draw the old
way); no behavior change until the render loop lands.
cc
The new ("playlist","play") media-action arm gave these cards a body-click
action for the first time. Without body-opens they defaulted to play, so a
playlist card body-click played instead of opening (and before the arm
existed, did nothing at all). Set body-opens:true on the Search carousel,
Search grid, and Label carousel so body-click opens the playlist while the
overlay button plays — matching the album-card convention and Tauri.
cc
The Spotlight hero portrait was 96px; enlarge to 140px (border-radius
70px to stay circular) for parity with Tauri. Layout-safe: the inner
Image is 100% image-fit:cover, the fallback glyph re-centers, the hero
HorizontalLayout has no fixed height so it grows, and the page Flickable
absorbs the vertical growth.
cc
The More From Your Library / similarAlbums row was hardcoded "More From
Your Library"; Tauri titles it "Similar to {seed}" where seed is the album
the suggestions are seeded from (discovery.similarTo).
Thread a seed_title through apply_more_from_library: the common case passes
the most-recent album title (recents_seed_title), the no-recent-history
fallback passes the first favorite album title. Empty seed falls back to
"More From Your Library" so the row is never titleless. Title stays a Rust
data string (rendered raw at Carousel.slint), matching the other For You
section titles.
cc
Qobuzissimes was pushed into the Home sections vec unconditionally, so
it always rendered on the Home tab — but Tauri's home default has it OFF.
Editor's Picks keeps it (push_section_ref, correct). Remove the Home push;
borrow-safe (containers.qobuzissims is now consumed only by the editor
push_section_ref).
cc
Add the missing For You "Favorite Albums" carousel (Tauri's
favoriteAlbums), the user's favorited albums capped at 18 in favorite
order, self-hiding when empty. Reuses the album fav-fetch already done in
albums_branch — no new endpoint.
- state.slint: ForYouState.favorite-albums: DiscoverSection.
- foryou.rs: build_favorite_albums (cap 18, unfiltered) + apply_favorite_albums
(mirrors apply_rediscover); wired into albums_branch right after fav_albums
resolves, painting in the same layer as rediscover/radio.
- artwork.rs: ForYouFavoriteAlbum { index } variant + apply arm.
- ForYouView.slint: mount the Carousel after Your Top Artists, before More
From Your Library, gated on favorite-albums.albums.length > 0.
Titled "Favorite Albums" (not Tauri's literal "More From Your Library")
to avoid the label collision with similarAlbums.
cc
Add the missing Discover "Qobuz Playlists" row, present on the Home and
Editor's Picks tabs (never For You), 1:1 with Tauri's PlaylistCardLite:
single cover (rectangle || covers[0]), name only, body=open / overlay=play.
- home.rs: PlaylistCardData + map_playlist (drops owner/duration/count);
HomeData + TabSections playlist fields; capture containers.playlists
(capped 18) in load_home; playlist_to_item (single-cover SearchPlaylistItem,
mirrors label.rs); playlist_artwork_jobs; select_tab swaps the active tab's
row + concatenates its artwork jobs; apply_home caches both sets + pushes Home.
- state.slint: HomeState.playlists: [SearchPlaylistItem].
- PlaylistCarousel.slint: forward a new body-opens prop to PlaylistCard.
- HomeView.slint: mount PlaylistCarousel after the section loop, gated on
active-tab != forYou && playlists.length > 0, body-opens: true.
- artwork.rs: HomePlaylistCover { idx } variant + apply arm (cover1).
- main.rs: reload_home extends jobs with the active tab's playlist covers.
cc
The on_media_action chain had no ("playlist","play") arm, so the
PlaylistCard play-overlay button and context-menu Play (which emit
"play") fell through to the dead _ => {} no-op everywhere they were
used (Search, Label, and now Discover).
Add a playback::play_playlist(id) that fetches the playlist fresh,
interleaves the mixed-playlist sidecar rows (same as enqueue_playlist),
replaces the queue and starts at the first track (set_queue + context),
then wire the new ("playlist","play") arm to it. Unlike the stateful
play-all arm (which reads the open PlaylistView state), this serves a
cold card play from any carousel.
cc
Frontend-agnostic port of Tauri discovery-v2/sectionPrefs.ts (ADR-006):
the ordered per-tab {id,enabled} model for the Discover configurator
(Home / Editor's Picks / For You). Pure model logic (default_prefs,
migrate with the legacy-array / object / garbage branches, reconcile_list
with drop-unknown + first-wins dedup + append-missing, toggle, move_section
with boundary clamp, reset_tab, is_enabled, enabled_count, enabled_ordered)
plus a per-user SQLite store (dedicated discover_prefs.db, single JSON-blob
column in the Tauri-compatible shape). 10 headless tests green.
Lives in qbz-app::settings::discover_prefs so any frontend (Slint, Tauri)
drives the configurator from one source of truth.
cc
Replace the sequential ~9-round-trip load_for_you + monolithic
apply_for_you with a progressive, parallel loader. The recently-played
sections paint instantly (local, no network); the independent API calls
fan out into concurrent branches (release-watch, favorite-artists,
favorite-albums, album-suggest) and each section applies the moment its
data resolves via upgrade_in_event_loop. The loaded flag is latched only
after every branch resolves, so the ensure_for_you_loaded re-entry guard
can never strand a partially-loaded tab.
Artists to Follow keeps byte-identical dedup/cap semantics (concurrent
per-seed fetch via join_all in seed order, then sequential dedup + cap).
album-suggest runs concurrently in the common case (recent-album seed),
falling back to the first favorite album only when there is no history.
The For You skeleton now hides as soon as the first section paints.
cc
Builds in release then execs crates/target/release/qbz-slint directly instead of
via 'cargo run', so process monitors show it as 'qbz-slint' (a cargo-run child
inherits CARGO_* env and shows as 'cargo'), cleanly separate from the build
processes. Same release flags as slint-dev.sh.
cc
(cherry picked from commit 41269aabbe)
spectrum_colors derived the secondary gradient colour as dom_hue + 55deg, which
fabricated a hue the cover does not contain (a mono pink/magenta cover became
pink->orange). Use real album colours: a second genuine hue cluster when the
cover has one (>=45deg from the primary, >=35% of the peak mass), otherwise the
same dominant hue deeper + more saturated. Mono-colour covers now read as a
one-hue light->deep gradient.
cc
(cherry picked from commit 6aa8d3b3d1)
The now-playing title elided early (the wrapping Rectangle content-sized to the
meta line) and, when forced to width:100%, overflowed into the audio stamp.
Replace both with an explicit `title-avail` width = card width - cover - art-gap
- (in-card stamp + 4px, when present), mirroring the inner row's width formula.
The title now elides at precisely the free space in New / Classic / Small at any
window size (the formula is parametric on root.width + has-stamp + show-art, so
Big inherits it), leaving the intended ~4px gap before the stamp and never
overlapping it.
cc
(cherry picked from commit 2acf53ed20)
Builds in release then execs crates/target/release/qbz-slint directly instead of
via 'cargo run', so process monitors show it as 'qbz-slint' (a cargo-run child
inherits CARGO_* env and shows as 'cargo'), cleanly separate from the build
processes. Same release flags as slint-dev.sh.
cc
spectrum_colors derived the secondary gradient colour as dom_hue + 55deg, which
fabricated a hue the cover does not contain (a mono pink/magenta cover became
pink->orange). Use real album colours: a second genuine hue cluster when the
cover has one (>=45deg from the primary, >=35% of the peak mass), otherwise the
same dominant hue deeper + more saturated. Mono-colour covers now read as a
one-hue light->deep gradient.
cc
The now-playing title elided early (the wrapping Rectangle content-sized to the
meta line) and, when forced to width:100%, overflowed into the audio stamp.
Replace both with an explicit `title-avail` width = card width - cover - art-gap
- (in-card stamp + 4px, when present), mirroring the inner row's width formula.
The title now elides at precisely the free space in New / Classic / Small at any
window size (the formula is parametric on root.width + has-stamp + show-art, so
Big inherits it), leaving the intended ~4px gap before the stamp and never
overlapping it.
cc
SongCard lives in the always-mounted now-playing bar. Its idle (no-track) state
drew a sleeping cat with two animation-tick()-driven drifting "z"s, which forced
the whole window to repaint at the display rate forever while idle — ~4% CPU on
Linux and ~55% on macOS/Metal on a static Discover screen (Metal's higher
per-frame cost made it egregious). Replace with a static music glyph placeholder
(matches TrackPlayCell). Idle is now lazy (~0 fps).
cc
SongCard lives in the always-mounted now-playing bar. Its idle (no-track) state
drew a sleeping cat with two animation-tick()-driven drifting "z"s, which forced
the whole window to repaint at the display rate forever while idle — ~4% CPU on
Linux and ~55% on macOS/Metal on a static Discover screen (Metal's higher
per-frame cost made it egregious). Replace with a static music glyph placeholder
(matches TrackPlayCell). Idle is now lazy (~0 fps).
cc
The wgpu underlay spike validated the GPU-shader path, but femtovg-wgpu has a
high per-frame CPU cost on macOS/Metal: continuously-animating views
(ImmersiveView atmosphere, Discover) hit 100-120% CPU on a Mac mini M2 vs
0.5-4% for the same view on Linux. Skia is Slint's mature Metal renderer and
does not have this cost (immersive ~55%, never 100%); text renders correctly.
Static views stay lazy/idle on both renderers.
Choose the renderer per-platform: Linux/Windows keep renderer-femtovg-wgpu
(cheap per-frame there + the femtovg text fixes #5177/#11335); macOS (Intel and
Apple Silicon, CICD included) uses renderer-skia. Both keep unstable-wgpu-28 and
resolve to a *-wgpu renderer, so the shader underlay (require_wgpu_28 +
Image::try_from) works unchanged on both.
Cargo.lock updated with the Skia dependency graph.
cc
Swap the winit renderer femtovg(GL) -> femtovg-wgpu (+ unstable-wgpu-28,
slint pinned =1.16.1) and render a WGSL plasma fragment shader into a wgpu
texture composited under ImmersiveView. The shader is fed by the existing
30fps visualizer drain (sub-bass energy + decayed transient) and toggled
in-view with the 'g' key via ImmersiveState.shader-mode; when active it
replaces the atmosphere + mode panels as the bottom layer, chrome stays on
top. Device/queue captured from Slint's rendering notifier so Image::try_from
shares Slint's wgpu device.
Phase 0 spike: validates the app-wide renderer swap with no font/UI/tray
regression. Pending macOS Metal smoke before the swap lands on slint-mvp.
cc
The original RGB "most saturated sample" extraction produced a near-grey color
with no perceptible gradient on dark/metallic covers. Rework in HSL: bin
chromatic pixels by hue, pick the most ABUNDANT hue (coverage), force it vivid
(sat 0.85, mid-bright) so it reads on the black bg, and rotate +55° for the
secondary stop -> a clear gradient. True B&W covers fall back to a default
duotone. (Hue selection can still surprise on some covers; further tuning of
which tone wins is deferred.)
cc
On large windows the panels left big empty space; on small they were fine.
Make the art/covers grow to fill while keeping small-window behavior.
- AlbumReactivePanel: art cap 460->660, width frac 0.55->0.62 (grows more,
center-anchored pulse untouched).
- StaticPanel: cap 460->640, width 0.55->0.62 (fills large windows).
- CoverflowPanel: center cover cap 440->560, width 0.42->0.52; wrap the fan +
track-info in ONE center-aligned VerticalLayout so the composition is
vertically balanced (kills the "covers clustered at top, empty bottom" look
on large windows). Persistent-element model / dual repeaters / animated
scroll / z-order all untouched — only sizing + the wrapping layout changed.
- New shared ImmersiveTrackInfo.slint (Now Playing indicator + EqualizerBars +
title/artist/album/quality badge), standardized on Static's look; all three
panels use it, removing three duplicated inline copies (they can't drift).
cc
refresh_now_playing_meta cleared ImmersiveState.bg-image (set default) on every
track change; the async 300px decode+blur then re-set it a beat later, so the
immersive atmosphere blanked then reappeared. Natural advances hid it (warm
art), but a click (cover or next button) showed the flicker.
Stop clearing bg-image on track change: the previous blurred ambient bg now
persists until load_now_playing_artwork_large swaps in the new one — a brief
stale blur is imperceptible; a blank is not.
cc
Slint 1.16 has no per-element z-index (paint order = model/child order), so a
single Repeater painted the right (upcoming) side near->far = far covers on top
(disordered), and it also drew a ghost center card under the dedicated one.
- Split the fan into a LEFT Repeater (forward coverflow-tracks: paints
far-left->near-left, near on top) and a RIGHT Repeater over a new reversed
model coverflow-tracks-rev (walks far-upcoming->near-upcoming, near on top).
Both compute offset from the original flat index so geometry matches and the
±0.5 hand-off during a slide is seamless. Center band owned solely by the
dedicated top card -> no ghost.
- coverflow-tracks-rev is built only under the same id-sequence-hash rebuild
gate (no churn on advance) by cloning+reversing the forward QueueItems, so a
track's element holds the SAME decoded image on both lists: no source
reassignment on advance (no spike), and the animated `scroll` slide is intact.
- 3 past + center + 3 upcoming, mirror POSITIONS (scale/dim by |offset| only;
no reflection). Nav mapping (play-history / play-coverflow-upcoming) preserved.
cc
The Coverflow rebuild added an ATMOSPHERE_LAST_URL dedupe to skip re-blurring
the same cover. But the track-change path clears bg-image (set_bg_image default),
and the dedupe then SKIPPED the regen (URL recorded as unchanged) -> bg-image
stayed empty -> ImmersiveAtmosphere fell back to the raw (sharp) cover instead
of Codex's blurred ambient background. Affected ALL immersive modes.
Removed the dedupe entirely (static + atmosphere_url_changed fn + the track-clear
reset). load_now_playing_artwork_large now ALWAYS regenerates bg + glow + spectrum
colors from the decoded cover, as the original code did. The regen is the same
modest 300px-decode+blur it always was; it was never the CPU spike (that was the
coverflow re-decode, fixed separately).
cc
The album art appeared to drift sideways while pulsing; Tauri's art only scales
in place (transform: scale, intrinsically center-anchored) like a vibrating
loudspeaker cone, with no translate.
Cause: the Slint port animated `x, y` alongside `width, height` on the art +
glow. x/y are the centering offset (parent.width - self.width)/2; putting them
on their own 100ms easing timeline desynced the centering from the size, so the
box eased off-center mid-pulse - the reported lateral movement.
Fix: animate only `width, height`; x/y stay plain bindings that recompute in
lockstep with the animated size, so the art grows/shrinks from a rock-solid
center. Faithful to Tauri's center-anchored scale (scale source kept on
global-energy, not bass - bass only drives the glow in Tauri). Atmosphere,
other panels, player, and audio backend untouched.
cc
Rebuild that finally kills the per-click CPU spike, using the classic
coverflow architecture (the owner's reference: Addy Osmani / iPhone-2007).
Root cause of the spike (now removed): every navigation re-ran refresh_async
which REPLACED the coverflow VecModels, so Slint dropped + rebuilt all 7 bound
Image elements and re-uploaded/re-decoded every cover on each click.
New architecture:
- A single STABLE flat cover model, QueueState.coverflow-tracks =
[history oldest..newest, NOW, upcoming...], + an int coverflow-index. An
order-sensitive id-sequence hash gates the model setter: a pure track advance
(same sequence, pointer moves) does NOT rebuild the model - only the index
updates. So the Repeater is never rebuilt and visible covers never re-decode.
Rebuild happens only on a real content change (new queue / shuffle / add /
remove).
- CoverflowPanel: a Repeater over coverflow-tracks; each card's transform is a
pure function of offset = index - scroll. `scroll` is a float bound to
coverflow-index with `animate scroll { 320ms }` - GPU-composited slide that
repaints only during the transition then idles (no animation-tick / Timer, so
idle = no continuous render). Flat scaled cards (no rotateY/skew/reflection),
per-depth black-overlay dim, center painted last + non-interactive, ~7 visible.
- Windowed lazy decode (+-4) reusing the DECODED_PIXEL_CACHE; at most the one
newly-entering cover decodes per advance.
- Side cards navigate via queue-wide play_history / play_coverflow_upcoming.
- Bonus: atmosphere bg de-dupes by URL (ATMOSPHERE_LAST_URL) so a track change
no longer re-blurs the same cover. [coverflow-perf] debug timers added.
cc
Owner rebuild: cheap "rocola" of flat scaled cards, and eliminate the CPU spike
on every navigation click.
CPU spike (root cause): each queue navigation re-ran refresh_async, whose
to_item reset EVERY rebuilt row's artwork to Image::default(), discarding the
already-decoded handles for the 6-of-7 on-screen covers; each was then
re-decoded (the disk cache stored only compressed bytes, so even a hit ran a
full image::load + thumbnail). Dozens of full JPEG/PNG decodes per click.
Fix: (A) artwork.rs decoded-pixel cache (DECODED_PIXEL_CACHE keyed by url+size;
fetch_and_decode_ref consults it before any disk read/decode); (B) queue.rs
to_item_reuse + per-list prior-art id-diff carries unchanged decoded handles
across the model rebuild (no blank, no re-decode); load_artwork short-circuits
on a decoded-cache hit. Net: at most the one newly-revealed cover decodes.
Pyramid (CoverflowPanel.slint rebuilt): 7 flat cards (3 history left + center +
3 upcoming right), center declared last (paints on top), step-down scale
(1.00/0.80/0.64/0.50) + per-depth black-overlay attenuation on an inner Rectangle
(0/0.22/0.40/0.55). NO rotateY/skew/perspective, NO reflection, NO far-edge
gradient, NO root transparency. Removed the 500ms x/width/height reflow (it
re-laid-out clipped images every frame); only a 120ms GPU opacity cross-fade on
the dim overlay remains. Center drop-shadow static. 132px player clearance kept.
Correctness fix (review blocker): the coverflow upcoming cards index the
UNFILTERED next-3 (coverflow-upcoming), so routing their click through the
paged/filtered play-upcoming played the wrong track on a paged/searched sidebar.
Added QueueState.play-coverflow-upcoming -> queue.play_coverflow_upcoming, which
plays the queue-wide upcoming index directly via play_upcoming_at. History
already indexes queue-wide via play-history.
cc
The bars used a fixed cyan->purple default. Now they take contrast colors from
the album artwork, with a negative-fallback for dark covers (a near-black
Metallica album would vanish against the black background -> invert so the bars
pop).
- immersive.rs: spectrum_colors() extracts the most saturated sample as primary
+ a hue-distinct secondary (channel-rotated if none distinct); if the cover is
too dark/desaturated it uses the NEGATIVE of the average. ensure_visible_on_black()
inverts a too-dark color and lifts a dim one so both stops read on black.
- ImmersiveState gains spectrum-primary / spectrum-secondary (defaults cyan->purple
until art resolves); set in playback.rs from the same decoded artwork as
glow-color / bg-image.
- SpectrumPanel binds the bar gradient to those (base = primary, tip = secondary).
cc
The cubelet port instantiated ~80 Rectangles per column x2 mirror x14 bins
(~2240 retained scene-graph nodes) whose `visible` flipped every frame off an
animated `shown-cubes`. Audio updates at 30fps kept that animation from ever
settling, so ~2240 visible-bindings re-evaluated per frame and femtovg
re-rasterized hundreds of tiny rects - ~5fps and a large CPU spike. Tauri is
cheap because it is immediate-mode Canvas2D (fillRect), not a retained tree.
Fix: one solid gradient bar per column (2 nodes/pair, ~28 total) with an
animated height. Same geometry (28 mirrored columns, baseY 0.85H, maxBarHeight
0.7H, widths/offsets), same data binding (VisualizerState.bars, bounds-safe),
same cyan->purple default gradient (now a vertical fill instead of per-cube).
Dropped the discrete-cubelet texture for the perf win; it can return cheaply as
a static stripe overlay if wanted. Removed the unused cube-indices array.
cc
Replaces the provisional S2 ModeBar with the polished segmented mode pill,
copying the compact player's visual language (owner instruction).
- Pill tokens copied verbatim from ImmersivePlayerBar: background #00000080,
1px #ffffff2e border, drop-shadow blur 32 / #00000066 / offset-y 8 (the
shadow is the main upgrade - the provisional bar had none), radius 12.
- 4 tabs (icon + label): Album Reactive / Static / Coverflow / Spectrum, with
existing icon assets. States per chrome.md: idle #ffffffb3, hover #ffffff1f
bg / #fffffff2, active #ffffff33 bg / #ffffff; 13px, weight 500->600;
150ms background animate. Leading panel-left glyph + 1px divider (decorative).
- Mount site unchanged: same chrome-visible auto-hide opacity (on the instance,
not the root), same select(idx) => mode + wake() wiring, no second timer.
- The 4 mode foregrounds, player, volume fix, atmosphere, responsive sizing,
and ESC are untouched.
Logged nits: border alpha matches the player (0.18) not the Tauri header tabs
(0.15) - intentional for one design language; leading panel icon is inert
(Tauri's is a functional split/focus toggle, out of scope).
cc
New SpectrumPanel.slint: faithful port of VisualizerPanel.svelte (the 16-bar
spectrum, not SpectralRibbon/Oscilloscope/LineBed).
- 16 source bars, bins {1,15} excluded -> 14 active -> 28 mirrored columns.
Geometry exact to source: bar-gap 4, total-bar-width W/28, baseY 0.85H,
maxBarHeight 0.7H, xLeft/xRight per spec.
- Discrete 6px+2px cubelets with per-cubelet gradient (gradientRatio
min(c/20,1), primary->secondary). GPU path (ADR-011); CPU flat fallback not
ported (Slint is GPU by default).
- Heights driven by VisualizerState.bars (bounds-safe clamp, 30fps len-16
VecModel) + a short animate settle - NO Timer. Bars already smoothed+clamped
in qbz-audio (no double-EMA). Now-playing card reuses the sibling pattern.
- Mounted on mode==3; modes 0/1/2 untouched.
Logged minors: MAX_CUBES=80 clips full-amplitude bars on very tall / 4K
windows (raise as a cheap follow-up); two-color gradient ships the Svelte
default cyan->purple (no album-derived secondary color wired yet); smoothing
is the Rust asymmetric EMA, not the Svelte symmetric 0.6 (deliberate, shared
pipeline).
cc
New CoverflowPanel.slint: honest pure-Slint 2D rendition of CoverflowPanel.svelte.
The true rotateY/perspective 3D tilt is impossible in declarative Slint and is
deferred to a future wgpu/Skia session - NO fake tilt (owner decision).
- 5-slot fan: now-playing centered + painted on top; QueueState history fanned
left, upcoming right. Depth via scale (center / *0.68 / *0.54) + per-depth
drop-shadow + distance dim overlay + far-edge shade gradient. Zero
rotate/perspective/skew (grep-confirmed).
- ~500ms reflow on track change via pure-Slint animate x/width/height/opacity
(no queue.rs change, no changed-geometry watchers).
- Side covers navigate via QueueState play-history / play-upcoming; center is
not clickable (matches Tauri). Empty/unresolved-art guarded with visible:.
- Responsive: base derived from viewport-height (reserve 364) + width*0.42 cap,
offsets clamped inside the window; song card under the fan reuses the
Static/AlbumReactive pattern. Mounted on mode==2; modes 0/1/3 untouched.
Deferred/cosmetic (logged): cover image pops on rebind instead of physically
traveling (needs a CoverflowState slide); no reflection sheen; no explicit
badge; no CPU-mode animation degrade.
cc
New StaticPanel.slint: 1:1 port of StaticPanel.svelte. The calm, centered
now-playing card over the moving atmosphere. NON-reactive (no VisualizerState
energy, no glow halo, no breathing) - the only motion is the fixed-period
4-bar equalizer (animation-tick, shown only while playing), faithful to source.
- Responsive art reused from Album Reactive (height-driven + width*0.55 cap +
clamp 180..460), reserve 364px = pad-top 52 + spacing 20 + info ~160 +
player clearance 132 (no *1.25 since Static doesn't breathe). Metadata never
clips under the player.
- Metadata: now-playing indicator / title 28/700 / artist 18 / italic album /
centered QualityBadgeFull - matching the sibling and the Svelte source.
- ImmersiveView mounts it on ImmersiveState.mode == 1 (replaces the placeholder).
Modes 0/2/3 untouched.
Accepted parity drops (match the shipped sibling + state model): explicit
badge omitted (no NowPlayingState.explicit field), no hover lift, label not
uppercased.
cc
The art was sized min(52vh, 460px) with only an 88px bottom pad, so on shorter
windows the art-block + metadata overflowed and the quality badge was clipped
under the floating player bar.
- base-art now derives from available height: it sizes the whole group
(art*1.25 growth + 40px spacing + ~160px info + 132px player clearance) to
fit viewport-height, with a viewport-width*0.55 cap for narrow windows,
clamped to 180..460px.
- padding-bottom 88 -> 132 (clears the 90px player + 24px margin + gap);
spacing 60 -> 40. Reserve constant kept in sync with base-art.
The Static panel (S3) will reuse this same responsive sizing.
cc
Builds the foreground layer on Codex's accepted atmosphere+player base.
- ImmersiveState.mode (int, default 0): 0 Album Reactive / 1 Static /
2 Coverflow / 3 Spectrum. Compact 0..3 enum for the scaffold (Tauri has no
single integer mode; it uses viewMode + activeFocusTab strings).
- ImmersiveView layer stack: atmosphere (bottom) -> mode foreground ->
auto-hide player -> ModeBar selector. Modes 1/2/3 mount no foreground yet
(atmosphere-only placeholder, TODO S3/S4/S5).
- Provisional ModeBar selector (top-center, auto-hides with chrome-visible),
to be replaced by the polished segmented bar in the header session.
- New AlbumReactivePanel.slint: faithful port of AlbumReactivePanel.svelte
(387 lines), NOT the rejected breathing-art proof. Exact maps:
global-energy = avg(energy[0..5]); bass = avg(energy[0..2]);
art-scale 1+global*0.25; glow-spread 15+bass*200; glow-opacity
min(0.1+global*1.5,1); 4-bar now-playing equalizer driven by
animation-tick() (not a Timer). Album-derived glow from the already-wired
ImmersiveState.glow-color; energy already published by visualizer.rs.
- main.rs immersive arm sets mode=0 on open (deterministic).
Flagged minors (deferred, see overnight log): glow tint not reset on
track-change until the new cover decodes; raw energy + 100ms animate vs the
Svelte asymmetric EMA; glow drop-shadow-blur only (no spread term); explicit
badge omitted (no NowPlayingState.explicit field); clamp art-scale energy to
avoid a clip at the block boundary on loud transients.
cc