7fc76058874d4eb03dc2e8849168e6a3f354e94a
6330
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 vicrodh
|
d5ce8bcaeb |
fix(qt): offline-cache rows, Jellyfin playback, and media art on the bar
Three defects the owner found by using it, all mine, all from the same
root: a source word that stops at Plex.
1. AN OFFLINE-CACHE ROW PLAYED NOTHING.
audible: file not available at …/audio/tracks-cmaf/426056576
`SourceId::from_word` folds `qobuz_download` into LOCAL because the row
IS a file — but for a CMAF-cached download that "file" is the offline
cache's own container and only `qbz-core`'s offline tier can read it.
IC-12's first pass declared those rows ours and handed them to the
local audible step. They now stay on the Qobuz funnel, which is where
they were before stage 3 and where their `QueueTrack.id` — the real
catalog id — is meant to be resolved. The symptom was the bar, the
queue and MPRIS all changing while the previous track kept playing.
2. A JELLYFIN ROW CHANGED THE BAR AND PLAYED NOTHING.
`local_queue_track` folded everything that was not Plex or a download
into "local", so the row entered the queue wearing a source word that
disagreed with its namespaced id and carrying no server id at all.
Both media words are now arms of their own, and the hint carries the
SERVER's item id — which `claim` prefers, because it survives a cache
rebuild that renumbers rowids.
3. MEDIA ART NEVER REACHED THE BAR, THE QUEUE OR MPRIS.
Those surfaces resolve an `artwork_url` string with no row in hand, so
`classify` has only the SHAPE to go on — and a Jellyfin image tag or a
Subsonic coverArt id has no shape worth reading. `local_queue_track`
now stamps the token SOURCE-TAGGED (`jellyfin:<albumId>/<tag>`) and
`classify` grows one arm that DELEGATES it to
`registry().artwork_token` instead of guessing. One arm, not 111 call
sites, and it shares its cache entry with the grid because it is the
same call.
Verified in the running window: a Jellyfin album streams and plays —
`[JELLYFIN/STREAMING] 331 MB @ 12 MB/s`, decoder up, 44.1 kHz to the DAC.
429/429.
cc
|
||
|
vicrodh
|
a3aff5f1b3 |
fix(qt): resolve artwork per VISIBLE row again, not per row of the document
The albums grid got slower and the owner was right about it. This is the cause and it is mine. Stage 4 moved the artwork decision from window time to MAP time so that a row's provenance would still exist when it was made. That fixed bug 3 and introduced a worse one: a remote source answers "what is this token" by reading its CREDENTIALS, and the map runs over every row of the document, not over the ~50 on screen. `local_plex::settings()` is a `load_last_user_id` file read, a mutex and a SQLite query — per Plex row. Measured on the owner's library, release, same grid each time: before stage 4 (a string clone) ~2 µs/row after stage 4 (resolve per row) 39-92 µs/row <- the regression + memoised settings ~9 µs/row + resolve at window time again ~1.8 µs/row 1703 rows: map 67-71 ms -> 3.0-3.7 ms The art index now holds `(SourceId, raw token)` instead of a resolved `ArtRef`. Provenance still survives — nobody sniffs a string, which was the whole point — but the resolution happens where it always did, in `resolve_window_blocking`, over the keys actually on screen. Same information, right place. The settings memo stays, in both `local_plex` and `media_servers_qt`: those accessors are cheap gates that several callers ask repeatedly, and every writer invalidates so a connect, a disconnect or a user switch is still seen on the next read. `search_local` keeps resolving inline on purpose — it is the cortinilla's bounded 76-row result set, and the url/path split it needs is per row by definition. 428/428. Grid verified in the running window: Plex, Jellyfin and local badges all correct, covers resolving. cc |
||
|
vicrodh
|
f2d8111721 |
perf(qt): instrument the albums load — sql, map, serialize
The albums document is republished on every mount of Local Library (five
times in a 37-second screencast) and there was NO instrument anywhere on
that path, so "the grid got slower" could only be argued about rather
than measured.
Three segments, because three different things could cost: the SQL (which
now ATTACHes two mirrors and UNIONs three sources), the row mapping
(which resolves an ArtRef per row since stage 4), and the JSON handed to
QML.
Measured on the owner's library, release build:
1703 rows (local + Plex + Jellyfin) sql 59-88ms map 67-71ms ser 2ms
1271 rows (local + Plex) sql 42-88ms map 50-117ms
29 rows (local only) sql 3-7ms map 0.06ms
The two large configurations overlap within noise: 432 more rows cost
nothing measurable in Rust, and serialisation of 593 KB is 2 ms. Whatever
the grid pays for row count is paid AFTER this line — the QML parse and
the delegate rebuild — which is exactly what Rust could not see before
and what these numbers now bound.
cc
|
||
|
vicrodh
|
3cfce8a769 |
feat(qt): the media-server settings panel, driven and verified in the window
ONE component instantiated twice, because it is one form twice. Every difference between the protocols is a property set at the call site, so they are visible in one place instead of drifting across two files: "Test" proves the address for Jellyfin and only "a Subsonic server answered" for Subsonic (which has no credential-free endpoint), the password is discarded for Jellyfin and kept for Subsonic (no session, md5(password+salt) per request), and the sync cost is stated up front — 45.8 s per 4924 tracks against 0.81 s per 6678 — so a long sweep does not read as a hang. The credential never leaves Rust; the panel gets `hasCredential`. Assets: the colour marks for cards, the monochrome ones for rows. Both mono files arrived as BLACK ink and are invisible on a dark theme untinted, so they were normalised to the #ffffff master ink and joined `icon_tint_qt::MASTERS` — the port's only shader-free way to recolour an SVG per theme. The Jellyfin colour asset had a white rounded-rect plate behind the mark; it was stripped, because every other brand mark here is transparent and a white tile beside them reads as a rendering bug. Rendered with Qt itself and looked at before any of it was wired. CAVEAT, recorded in SourceIcon: the Subsonic mark is NAVIDROME's logo, and `subsonic` covers Gonic, Airsonic, Astiga and Ampache too. `ping` reports the server's own type and it is already persisted, so branching later is cheap — a known wrongness, not an unnoticed one. THREE BUGS THAT ONLY THE RUNNING WINDOW COULD SHOW, all found by driving it under Xvfb rather than by reading the log: 1. `MediaServerSettings is not a type` — a .qml missing from build.rs is not in the QML module. It resolves fine on disk, compiles nothing, and blanks the WHOLE settings page at runtime. All three existing audits passed. `qml_module_registration_audit.py` is the missing check and is now in the gate. 2. The toggle would not move. The panel reads its state from the settings DOCUMENT, unlike Plex's toggle which reads a bridge property, so every mutation has to republish it — connect, enable, disconnect and the end of a sweep now do. 3. Every Jellyfin album in the grid drew the LOCAL HARD-DRIVE badge: `local_rows::badge_source` had never been taught the two words and folded them into `_ => local`. Not cosmetic — that string is what the source chips filter on. `play_single_from_source` also grew the arm it needed: a media-server row from the recently-played rail has no library.db row to mint a LocalTrack from, so it resolves through `registry.tracks_of` instead. END TO END, from the UI, against the real server: toggle on -> type the address -> Connect -> "jellyfin sync: 500/4924 ... 4924 saved, 0 pruned, 4924 cached" -> Local Library shows 461 albums with their covers and the Jellyfin badge. The cached quality distribution is identical to what the server reported in the original protocol sweep. 428/428. cc |
||
|
vicrodh
|
63a4f2daf1 |
feat(qt): the Local Library union, the settings store and the sweep
Three pieces that only mean anything together: rows in the shared cache,
a union that shows them, and settings that decide whether either
happens.
THE UNION. `get_albums_metadata_page` grows a second, INDEPENDENT
attachment. `plex_cache` stays; `remote_cache` is the shared mirror, and
they attach separately on purpose — a missing Plex cache must not cost
the user their Jellyfin rows. `attach_best_effort` is extracted because
the failure mode of a botched ATTACH is a silently local-only library,
and a failure stays non-fatal: refusing to list any albums because one
server is off turns "Jellyfin is unplugged" into "your library is
empty".
The union takes an ENABLED-SOURCES list, not just a path. The mirror
holds every remote source, so the master toggle has to be a filter
rather than a detach: turning Jellyfin off must hide Jellyfin's rows and
leave Subsonic's alone. That predicate is interpolated into SQL, so it
is VALIDATED rather than escaped — only [a-z] survives, an all-invalid
list yields `0` (show nothing, the safe direction), and a test feeds it
`x'; DROP TABLE local_tracks; --`.
The remote group key is `<source>:<album id>`, matching Plex's
`plex:<hash>`. That namespaces two servers using the same album id AND
lets a source claim an album card that arrives with no source word,
which is every card that round-trips through a QML string property —
`recognises` grew that arm and `claim` strips the prefix, because
carrying it inward is how Plex ended up with two album keys only one of
its queries understood (IC-2).
THE SETTINGS. One store for both servers, keyed by a `server` column,
mirroring the shared-cache decision. What each protocol needs differs
and `is_configured` is the one place that knows: Jellyfin needs its
issued token, Subsonic needs a password because it has no session and
re-derives `md5(password + salt)` per request.
Two identifiers are minted once and persisted, and both are
load-bearing. Jellyfin's DeviceId, because re-authenticating under the
same one revokes the previous token — a constant would make two installs
log each other out, a fresh one per launch drops the last run's token.
Subsonic's salt, because rolling it per request makes every cover URL
unique and the artwork cache keys on the URL.
THE SWEEP. In the frontend, where the runtime and the progress UI are,
writing through the cache handle the SOURCE owns so sweep and reads can
never point at different users. Jellyfin gets progress and a delta path
(45.8 s for 4924 rows); Subsonic needs neither (0.81 s for 6678).
`prune_stale` runs only after a COMPLETED full sweep — a delta never
prunes, because it deliberately does not see unchanged rows and would
read every one of them as deleted. `last_sync_at` is stamped only on
success for the same reason.
Verified against the running binary, twice:
source registry bound: local_tracks=617 plex=connected
ephemeral_store=yes remote=none configured
jellyfin: minted a device id for this install
subsonic: minted a salt for this install
...and on the second launch, zero re-mints and the same values on disk —
which is the property that keeps a Jellyfin session alive.
qbz-qt 428/428, qbz-app 224/224, qbz-library 100/100, qbz-source 88/88.
Release build clean, zero warnings.
cc
|
||
|
vicrodh
|
51aa8ba081 |
test(source): walk the whole chain against both real servers
The unit tests prove each method in isolation against a hand-made row.
They cannot prove the thing that actually broke this project once: that
the pieces are CONNECTED. `source_wiring` shipped a registry whose
sockets were empty and every test stayed green, because a pure-function
test has no bound user and a structural gate only proves the OLD path is
gone.
So this walks it: sweep a real page off the server, write it through the
same mapping the sync will use, bind the source exactly as the app binds
it, then claim -> tracks -> meta -> artwork -> playback, and finish by
pulling real bytes through the url the source produced. Nothing is
asserted from a type; everything is asserted from a server's answer.
The assertion with the most teeth is the identity round trip. A queue
row as `cached_to_queue_track` builds it — namespaced id plus the
server's id in the hint — must claim back to the server's id, and the
same row must ALSO resolve with no hint at all, through the cache. Those
are the two shapes a producer can hand over, and a mismatch between them
is bug 2 wearing a new source's clothes.
`assert_serves_audio` issues a Range request, which is what the
progressive feeder does: it checks seekability, refuses a chunked
(transcoding) answer, runs the Subsonic error-envelope guard over the
body, and keeps the test to 64 KB instead of a 100 MB FLAC.
Against the bench, 3/3:
jellyfin swept 500 of 4924 -> album "...And Justice For All",
9 tracks, tier "hires", cover image/jpeg (tokenless,
self-keying), stream 206 audio/flac
subsonic navidrome openSubsonic=true, swept 500 -> album
"POWERS OF VOICE (Disc 1)", 13 tracks, tier "cd",
cover image/webp keyed on subsonic:al-…_0 rather than on
the credentialed url, stream 206 audio/flac
registry each namespace routed to its owner, and a Plex id still
goes to Plex
The third needs no server and runs everywhere.
cc
|
||
|
vicrodh
|
b09f03b739 |
feat(source): JellyfinSource and SubsonicSource — §10's acceptance test, paid
Design 02 §10 predicted a fourth source would be "one impl plus one
registration line, zero edits in any view". Two of them landed, and the
prediction held: no arm in any view, none in the playback router, none
in the artwork pipeline, no entry in a url taxonomy. Every one of those
became structural during stages 3 and 4, which is what this commit is
the receipt for.
One honest correction to that claim, now written in sources/mod.rs: it
was true only AFTER stages 3 and 4. From stage 1 the seam existed but
playback and artwork never asked it anything, so a source added earlier
would have claimed correctly and then been silent and coverless.
`sources/remote.rs` holds what the pair shares — the cache handle and
the row mappers — so they cannot drift the way the three copies of
play_audible did. The two conversions that are easy to get backwards
(ms -> s, Hz -> kHz) live there once, with a test.
Identity, and it is the part with teeth. A queue row carries BOTH the
namespaced id the grid published and the server's own id in the hint,
and `claim` prefers the hint: rowids are reassigned by a cache rebuild,
the server's id is not, so a queue that outlives a re-sync still
resolves. Recognition is by source word or namespace bit ONLY — never by
the shape of an opaque id. A Jellyfin item id is a 32-char hex GUID and
a Subsonic one is base62, but neither is guaranteed, and inferring a
source from a string's alphabet is the class of guess the seam deletes.
Both sources are tested against every neighbouring floor: Plex's 1<<40,
each other's, the ephemeral store's 1<<48, and a plain local rowid.
The Subsonic brand spellings — navidrome, gonic, airsonic, astiga — fold
to ONE source. They speak one API and QBZ reads them through one impl,
so a row stamped with the server's brand must still resolve; returning
None would make it Unclaimed and it would refuse to play.
Artwork differs between the two, and only because the protocols do.
Jellyfin's cover url carries no credentials, so it is stable and its
cache_key can be the url itself. Subsonic's carries the salt and the
size, so its key is the opaque coverArt id instead — keying on the url
would re-download every cover whenever either moved, and would put a
credential in a cache key.
Both playback tickets are Stream with the bit-perfect contract spelled
out: `?static=true` for Jellyfin, `format=raw` for Subsonic, and tests
that fail if `audioCodec` or `maxBitRate` ever appear.
One test had to be rewritten rather than deleted: `from_word("jellyfin")
== None` was the example for "unknown words are not guessed into Qobuz".
The property still holds and is now spelled with "emby".
qbz-source 84/84, qbz-media-cache 11/11, qbz-qt still builds clean.
cc
|
||
|
vicrodh
|
0dfec607c2 |
feat(media-cache): one cache table for every remote server
The Local Library union works by ATTACHing a remote server's mirror and UNIONing it with local_tracks, and today that mechanism names Plex inside a shared crate: `get_albums_metadata_page(..., plex_cache_path)` with `FROM plex_cache.plex_cache_tracks` hardcoded twice. Copying it per source means three Option<&Path> parameters and six UNION arms, and every future source pays again. This is the generic form: one table with a `source` discriminator, one ATTACH, one UNION arm. The schema is `plex_cache_tracks` with its columns renamed — all 19 map 1:1 onto Jellyfin and Subsonic, which is what made "generic" a measurement rather than a guess. Plex is deliberately NOT migrated here. It is a live table with 17145 rows on the owner's install, and folding it in alongside two new sources would make one commit where a regression could come from either. This table proves itself with Jellyfin and Subsonic first; Plex moves in as its own change with its own verification. The cost until then is a second ATTACH in one SQL builder, which is cheaper than the risk. Identity: the namespaced Local Library id is `FLOOR | rowid`, with Jellyfin at 1<<41 and Subsonic at 1<<42 — disjoint from Plex's 1<<40 and the ephemeral store's 1<<48, asserted against both. The payload is this table's AUTOINCREMENT rowid, NOT a hash of the server's id. Plex had to hash because its ids predate its cache, which is why resolving one costs a table scan; a new table has no such debt, and at ~50k tracks a 40-bit hash carries ~0.1% odds of a collision that would play the WRONG TRACK. Writes are UPSERT on (source, item_id), never delete-then-insert: the row id is published identity — it sits inside queue entries and session.db — so a re-scan must not mint a new one for a track that was already there. AUTOINCREMENT for the same reason: a freed id must never be handed to a different track. Both are pinned by tests. `prune_stale` keys on updated_at rather than a set of ids, so a 50k sweep needs no giant IN(...) — with the documented precondition that it only runs after a sweep that COMPLETED, since a dropped connection halfway through would otherwise read as "the server deleted everything it did not get to". 11 tests, including the id-namespace matrix against the two existing floors and the untagged-track ordering (75 of 4924 measured Jellyfin rows carry no track number and must not sort above track 1). cc |
||
|
vicrodh
|
0f41c44dac |
feat(subsonic): the protocol crate, with the HTTP-200 trap enforced in code
Stage 5's second protocol crate. One client for Navidrome, Gonic,
Airsonic, Astiga and Ampache — which is why the owner scoped this as
"Subsonic, not just Navidrome". Only Navidrome 0.63.2 was on the bench,
so the places servers are likely to differ are named rather than
smoothed over.
The dangerous property of this protocol is that it reports FAILURE with
HTTP 200. `error_for_status()` passes on a refusal. Three defences, all
in code rather than in a reviewer's memory:
- `parse_envelope` is the only way a body becomes a payload, and it
checks `subsonic-response.status` first. It also accepts XML on the
failure path, because the measured no-credential getCoverArt reply came
back as XML DESPITE `f=json` — a JSON-only parser would report "not
understood" for a perfectly clear "missing parameter: 'u'".
- `looks_like_audio` exists because a refused stream is a ~200-byte JSON
blob served as the audio body. That is the one way this feature could
reach the protected decoder with garbage.
- code 40 and 70 become their own variants, so a caller can act on
"re-authenticate" without matching on prose.
Two mapping decisions worth stating: `duration` is SECONDS on the wire
and milliseconds in QBZ, and `bitDepth` is **0** for a lossy track where
Jellyfin reports null — both sentinels fold to None so downstream sees
one vocabulary rather than two.
The salt is FIXED per install, not rolled per request. Rolling looks
more careful and would make every cover URL unique per request, which
re-downloads every cover on every pass since the artwork cache keys on
the URL. It buys nothing: the password is not transmitted either way.
The consequence is recorded where it matters — a cover URL carries
credentials, so callers must key their cache on the opaque coverArt id.
`detect_sweep_mode` ASKS rather than checks the server's name: search3
with an empty query is 14 requests for this library and getAlbumList2 +
getAlbum is 675, but the fast path is the behaviour most likely to
differ on a server that was never tested, and probing one page costs
~100 ms.
14 unit tests + 6 live. The live suite runs in PARALLEL by design —
Subsonic auth is stateless, so unlike Jellyfin nothing here can revoke
anything else's session. Against Navidrome 0.63.2, three consecutive
runs, 6/6 each:
ping ......... navidrome 0.63.2, openSubsonic=true
wrong pass ... Unauthorized, and the raw request really answered 200
sweep ........ Search3 detected; page 0 = 500 rows, 344 bit depths
per-album .... fallback returns 500 album ids, 17 tracks in the first
stream ....... 206 audio/flac, Content-Range total == the API's own
reported file size, so nothing was re-encoded
cover ........ authed image/webp 69 KB; unauthed HTTP 200
application/xml 251 B, rejected by the guard
cc
|
||
|
vicrodh
|
910f7aec4d |
fix(jellyfin): one session per test binary — a repeat auth revokes the token
Reported as a flake: four live tests passed and one failed Unauthorized, while other tests doing the same call succeeded. It is not a flake and not the server being slow. Jellyfin keys the SESSION on DeviceId, and a second AuthenticateByName under the same DeviceId REPLACES the first — the earlier token starts answering 401. Measured directly against 10.11.11 rather than inferred: same DeviceId: auth -> T1, auth -> T2; T1 = 401, T2 = 200 different DeviceId: auth -> A, auth -> B; A = 200, B = 200 Five tests authenticating in parallel under one DeviceId were therefore invalidating each other, and whichever finished first was left holding a dead token. Running with --test-threads=1 hid it, which is why the first run looked green. The tests now share one lazily-initialised session, which is also what the app must do. Two production obligations follow and are now written on `auth_header` where somebody will read them: the DeviceId must be stable PER INSTALL (persisted like local_plex::client_id — a constant would make two QBZ installs log each other out, a fresh id per launch drops the previous run's token and litters the server's devices page), and the app must authenticate once and hold the token rather than concurrently. Verified in the parallel configuration the report used: 10 unit + 5 live, three consecutive runs, 5/5 each time. cc |
||
|
vicrodh
|
e578d8a54a |
feat(jellyfin): the protocol crate, verified against a live server
Stage 5's first half. Sibling of `qbz-plex` and deliberately shaped like it, so `qbz-source`'s JellyfinSource will read like PlexSource: frontend-agnostic, HTTP + DTOs, no cache schema and no settings store. Three measured facts shape the file, and each one is a comment because each one is a decision somebody could otherwise undo: - `?static=true` is bit-perfect. Verified by md5 against the file on the server's own disk, not inferred from a content type. `is_direct_ response` encodes the tell a transcode leaves — chunked with no Content-Length — so a server-side policy silently resampling a hi-res track is caught rather than played. - Quality costs 45.8 s per 4924 tracks and there is no cheaper path. `Fields=MediaStreams` trims 29% of the bytes and saves nothing: the cost is server-side media-info hydration. The scan will own that knowingly, which is why `track_count` exists to give a progress denominator. - Cover urls need no credentials, so unlike a Plex thumb they are STABLE and their cache key can be the url itself. `stream_url` / `image_url` are free functions rather than methods. Two reasons and both are load-bearing: `Source::artwork_token` may not touch the network by contract, and a `reqwest::Client` cannot even be CONSTRUCTED without a rustls provider — which a unit test does not have. The mapper picks the audio stream BY TYPE, not by position: an embedded cover rides in the same MediaStreams array as an `mjpeg` with BitDepth 8, and taking [0] would badge a 24-bit FLAC as 8-bit. A lossy row keeps an absent bit depth rather than defaulting to 16, which would badge every MP3 as CD quality. 10 unit tests, plus 5 LIVE tests that skip unless QBZ_JELLYFIN_URL names a server. The unit tests prove the JSON shape is understood; only the live ones can prove the endpoint still exists, the token still opens it, and the stream is still direct. Against Jellyfin 10.11.11: 5/5 — 500-row page with 221 real bit depths of 4924 total, a Range request answered 206 with a `fLaC` bitstream, a tokenless cover, and a future-dated delta returning zero. cc |
||
|
vicrodh
|
604f336e24 |
fix(qt): wire the source registry to its stores — it never was
This is the regression report from the previous build, and the cause is that stage 1 built three injection sockets and nothing ever plugged into them. `set_creds`, `set_ephemeral` and `bind_user` had zero call sites in the entire frontend. It stayed invisible for two stages because nothing needed them: `claim` is pure, and `PlexSource::tracks`/`meta` read plex_cache.db through qbz-plex's own global. The first consumers that DO need them are playback and artwork — stages 3 and 4 — so routing those onto the seam turned a latent gap into a live break: - every Plex row: "plex: not configured (no Plex credentials configured)", so Plex playback died and every Plex cover in the Local Library grid resolved to ArtRef::Unavailable and vanished; - every LOCAL row: an unbound DbPool refuses all reads, so LocalSource found no track and local playback died too; - ephemeral rows had no session store to resolve against. Wired now, each to the ONE authority that already existed: PlexCreds over local_plex's settings store (read per call, never cached, so a connect/disconnect mid-session is picked up), EphemeralTracks over the session store, bind_user as the first statement of bind_per_user_stores, teardown on logout, and the unbound-window fallback so pre-login reads resolve the way local_state::with_db already resolves them. Also wired: the album identity mode. The grouping IS the query, so a metadata-mode user's album play would have resolved a different track list than the grid rendered. `report_wiring` logs one line at bind time asking each socket a question only a wired one can answer. The failure this fixes is silent by construction — an unbound pool and a disabled Plex look identical to a user with no local tracks and Plex switched off — and it stayed hidden because the tests are pure-function tests and the structural gate only proves the OLD path is gone, never that the new one is connected. Verified against the running binary, not the test suite: source registry bound: local_tracks=617 plex=connected ephemeral_store=yes qbz-qt 423/423, qbz-source 59/59. cc |
||
|
vicrodh
|
68235a4012 |
refactor(qt): the Local Library art index carries an ArtRef, not a string
Design 02 §9 stage 4, the Local Library half. `local_state.art_index` becomes `HashMap<String, ArtRef>`: the row mappers register what the row's OWN source says its artwork token means, at the last moment that provenance still exists, and the windowed pipeline reads that answer instead of re-deriving it. `artwork_qt::classify` is gone from every caller outside artwork_qt itself. It recognised a Plex thumb only because `local_plex:: is_thumb_path` was bolted onto it, so a token from any future source that starts with `/` — a Jellyfin `/Items/<id>/Images/Primary` — matched its filesystem-path arm and the pipeline opened it from disk: blank cover, no error, nothing in the log. That is bug 3. Producers wired: both row mappers, the folder tree (two sites), the ephemeral pane, the album-detail header and the Artists rail. The Artists rail needed real work rather than a cast. Its portrait comes from one of three chain tiers (custom image, Plex portrait, borrowed album cover) while `MergedArtist.source` describes the artist's LIBRARY provenance and is legitimately "mixed" — which is not a source a token can be interpreted by. `AlbumCredit` now carries its album's source word so a borrowed cover keeps that cover's provenance, and `MergedArtist` reports `image_source` beside `source`. `fetch_plex_misses` carries `(artKey, fetch url, cache key)` instead of collapsing the last two. They are not the same string — a Plex url is re-tokenized every pass and a Qobuz one carries a size query, while the key stays put — and collapsing them re-downloads covers already on disk. `artwork_qt` grows `cached_path_for` / `download_pairs`, the classification-free entries, sharing one download loop with `download_missing` rather than a second copy of the pool and the error-page guard. Still on the old path and NOT in this commit: the Home / queue / now-playing surfaces, where the string arrives as `HomeCard.art_url` or `QueueTrack.artwork_url`. Those need IC-7 (HomeCard has no source field) and the eight hardcoded `"source": "qobuz"` QML payloads first. 423/423. cc |
||
|
vicrodh
|
ffb35085c9 |
feat(source): each source interprets its own raw artwork token
Design 02 §9 stage 4 needs the art index to carry an ArtRef instead of a String, and the producers of that index are the row mappers, which hold no MediaRef — the windowed grid pipeline is keyed by art_key (§3.4). The two ways to give them one were both bad: sniff the string later (that is `artwork_qt::classify`, i.e. bug 3), or claim per visible row and let the source look the token up again, which for Plex is one plex_cache.db query per row on a cold grid — on the exact grid whose scroll cost is the open perf item. `Source::artwork_token` is the artwork analogue of `claim`: the source says what its OWN token means, at the moment provenance is still known. It is pure — no network, no DB, no filesystem — because the token is the whole input. This is what makes stage 5 free. `classify` recognises a Plex thumb only because somebody bolted `local_plex::is_thumb_path` onto it, so a Jellyfin token (`/Items/<id>/Images/Primary`) reads as a filesystem path — it starts with `/` — and the pipeline opens it from disk: blank cover, no error, nowhere to look. Dispatching by owner removes the guess. Three impls, each carrying the arm of `classify` that was really about it: Local (file:// and absolute paths, plus the CDN url an offline row can keep), Plex (the two declared thumb prefixes, Unavailable rather than None when there are no credentials), Qobuz (CDN url, cache_key on the RAW url so downscaling does not re-download per size class). An unregistered source returns None, not Unavailable: a row whose source word this build does not know has no art to wait for, and Unavailable would park it in the retry-when-connected bucket forever. 59/59. cc |
||
|
vicrodh
|
22033672a1 |
fix(qt): claim the row before the Qobuz funnel, and before casting it
Two remaining stage-3 consumers. `play_resolved_offline_aware` is the QOBUZ entry — the tier walk, the offline cache, the quality request — and every MyQBZ and album play arm sent its anchor there regardless of who owned the row. A local, ephemeral or Plex first track resolved and then failed AT THE PLAYER with "No Qobuz client available", which `myqbz_play_qt`'s own header documented as open. It was also unsafe: the offline cache is keyed on the raw u64, and a local_tracks rowid or a Plex rating key shares that small-integer space with older Qobuz ids, so a collision played the wrong track rather than failing cleanly (#638 fix 3's sharp edge). The funnel now claims the row first and hands a non-Qobuz one to audible_qt, which fixes all six call sites without touching a signature. The claim is gated on the core's cursor already sitting on that track — true for every caller, since they all set the queue and then play its anchor — so a caller that plays some other id behaves exactly as before instead of guessing. `cast_qt::resolve_local_path` becomes a PlaybackTicket consumer (IC-10). It was an ephemeral test plus a bare db.get_track on a QueueTrack.id, which is only sometimes a library.db rowid: a Plex row carries a namespaced id and an offline row carries the Qobuz catalog id, so both missed and reported "Local file not found" — an answer naming the wrong problem. Now File/DsdFile is servable and everything else is a named refusal, which covers every remote source by construction rather than the one word somebody thought to list. The Plex arm stays an early refusal so it does not pay a network round trip to build a url it would throw away. Structural gate: no plex_resolve_track_media / plex_rating_key left in qbz-qt outside comments. Four source-word matches remain on purpose — local_queue_track STAMPS the word (producer side, not consumer), play_single_from_source looks up a ROW rather than playing it, cast's early refusal above, and batch_all_qconnect_castable, which asks "does this row have a real Qobuz catalog id?" — a different question from ownership, and one SourceId cannot answer since qobuz_download maps to LOCAL while being castable. 423/423. cc |
||
|
vicrodh
|
a55e426330 |
refactor(qt): collapse the three audible steps into one PlaybackTicket match
Design 02 §9 stage 3. `local_playback::play_audible`,
`local_album_actions::play_audible` and `local_ephemeral::play_file` were
three near-copies of the same routine, each matching on `track.source`
by hand, and they had already drifted:
- `local_playback` carried the `plex_rating_key` ladder — one
engineer's reconstruction of a five-id table nobody wrote down, whose
fallback is right for the MyQBZ path and wrong for the LocalLibrary
one. `local_album_actions` held a copy of it, kept in sync by comment.
- both handed `track.id` to `play_local_file` for anything not Plex,
which is wrong for OFFLINE rows: their `QueueTrack.id` is the Qobuz
catalog id while the `library.db` row id rides in
`source_item_id_hint`, so the lookup missed.
- the CUE fast path existed in two shapes. `local_playback` compared
track ids, and since every virtual track of a CUE album is its own
row that test only passed when replaying the same virtual track — it
never fired for the case it exists to serve. `local_ephemeral`
compared file paths, which is the right question, but only for rows
in the session store.
`SourceRegistry::playback` now claims the row once and `audible_qt`
performs the ticket. The CUE test becomes exact for every source:
`audible_qt` records what its last `play_data` actually read, so "is
this container already loaded?" is answered against a real path with no
source lookup on the hot path.
`play_current_if_local` stops matching `Some("local") | Some("plex")`
(IC-12). Six source words occur on a queue row and the four it missed —
`user`, `ephemeral`, `qobuz_download`, `qobuz_purchase` — all name
files on disk, so auto-advance onto one fell through to the Qobuz
tier-walk, died with "No Qobuz client available", and
`auto_skip_unavailable` then declined to skip because that error is not
terminal-unavailable: one ephemeral track could stop playback dead.
The whole-body fallback in the streaming arm checks Content-Type and a
minimum length before the bytes reach the player. Plex cannot trip it,
but a Subsonic server answers a refused request with HTTP 200 and a
~200-byte JSON error envelope (measured against Navidrome 0.63.2), and
that must never reach the decoder.
Deleted: `play_local_file`, `play_plex_track`, `plex_rating_key` and its
three tests (the same cases are already pinned on `PlexSource`),
`local_ephemeral::play_file`, and the duplicate `play_audible`.
cc
|
||
|
vicrodh
|
7113ef5794 |
feat(source): add PlaybackTicket::Stream and emit it from PlexSource
Design 02 §3.3 predicted a `Stream` variant "when the documented progressive-streaming follow-up lands". It landed first: `local_playback::play_plex_track` already resolves the part URL and Range-streams it through `qbz_player::remote_stream`, keeping the whole-file download only as the fallback for a server that refuses Range. Stage 3 cannot route playback through the ticket while the ticket has no way to say "stream this". `PlexSource::playback` now calls `plex_resolve_part_url` instead of `plex_resolve_track_media`: the ticket resolves a LOCATION, it does not fetch a body. Dropping the old `resolve_track_media` fallback is not a behaviour loss — both functions run the same metadata request, the same parse and the same part-key extraction, and only then does the second also GET the body, so every way the first can fail the second fails identically. Carrying the url on the ticket also removes the duplicate metadata round trip the fallback performed. `duration_secs` comes from the queue row rather than the core's current track: a caller may be priming a row that is not current yet, where the core would answer about a different one. The url embeds credentials (Plex's token, and Jellyfin's / Subsonic's when they arrive), so `Debug` redacts the query string while keeping the origin — telling a LAN address from a plex.tv relay is exactly what the perf log lines exist to answer. Three tests cover the redaction and the `play_id()` arm. cc |
||
|
vicrodh
|
321d8fe753 |
fix(qt): align the playlist manager's card grid with its own headings
The grid delegate carried x: 32 on the delegate ROOT. A ListView owns its delegates' position, so an x set there is not reliably the one that survives layout, and it did not: the cards sat about 24 px left of "Folders (12)" and "Playlists (9)", which PmPageHead draws at 32 through its own Column. Headings at one inset and cards at another, on the same page. The list and tree delegates right below it already use the shape that works — a full-width delegate with the inset on the child — so the grid now does too. The inset lives where the view has no opinion about it. cc |
||
|
vicrodh
|
7e233104f5 |
feat(import): playlist files, JSON, ListenBrainz and Last.fm as import sources
Implements qbz-nix-docs/deferred-2.0.3/playlist-importer-expansion-design.md, whose §11 the owner ratified 2026-07-31. PlaylistSource is an enum, not a trait (ADR-007): the set is closed, there is no plugin story, and every arm resolves to the same ImportPlaylist the URL scrapers already produce — so the matcher, the 2000-track split, the sink and the summary never learn a second kind of source exists. import_prepared_playlist takes the playlist BY VALUE and import_public_playlist becomes a shim. That kills the double scrape the Tauri re-fetch-on-execute left behind, and it is not optional for the new sources: the rfd bytes of a picked file are gone by execute time and the dialog cannot silently reopen. It also closes a real race — open() resets the session, so a modal reopened between the Import click and the task's read handed a detached task None. FILES read the track list ONLY. The paths inside are never opened, copied or added to the Local Library, and the modal says so in a persistent row rather than a tooltip. Content sniffing beats the extension, with an HLS guard because .m3u8 is also the stream-manifest extension, and a utf8-then-Windows-1252 fallback because .m3u and .pls predate UTF-8 and "Björk" has to stay matchable. JSON locates the track list BY SHAPE. Depth is a DoS bound, not a locator: the array sits at 0, 1, 2 or 3 levels depending on the export, and a hard two-level rule drops an ordinary GraphQL envelope. Container demotion keeps a discography from importing album names as track titles. Walls at 16 MiB, depth 8, 2M nodes and 10k tracks. SERVICES are public reads with no authentication; a connected account only prefills the field. No qbz-integrations edge — its ListenBrainz client returns tracks with no duration and drops the playlist title, both verified in its source, and a direct GET /1/playlist/<mbid> reads all three in one request. A Last.fm STATION IS A RADIO, and ?page=N is not pagination. Measured against a live profile: pages 1 and 2 overlapped and batch sizes varied (20, 20, 36). The first cut fetched 20 pages, never hit its empty-page exit, and returned 469 rows holding 198 distinct tracks. It now dedupes as it draws, stops after two dry draws, and stops at MAX_STATION_TRACKS = 200 — the number shown in "Found N tracks" has to be one we can deliver, and a station is dynamic enough that a bigger single import buys little (owner ruling: no false expectations of "traeré casi 500" when we bring half). A failed draw no longer fails an import that already has rows; Last.fm rate-limits a burst of these. The summary reports THREE DISJOINT NUMBERS that sum to the total. skipped is now counted from the matches that found nothing, not inferred as total - matched, which folded duplicates into failures: that same import told the user 271 tracks were skipped when 255 of them had matched perfectly well and been deduplicated before the add. 23 new strings across the seven translated locales; en is the sparse fallback catalogue and needs none. cc |
||
|
vicrodh
|
c11bdf0654 |
feat(qt): bring back the New playlist modal, and delete the shortcut
The sidebar "+" called create_playlist(), which made a playlist named literally "New Playlist" and navigated to it. Four of the modal's five fields had no other door, and one had NO door at all: name hardcoded reachable in Edit playlist description always empty reachable in Edit playlist folder always root reachable in the sidebar row menu public always false NOWHERE offline-only from connectivity NOWHERE for an online user That last row is why this counts as a regression rather than a rough edge. offline_only decides whether the playlist is a local_playlists row (local:<uuid>, library.db, no Qobuz account, never uploaded) or a Qobuz entity, and the shortcut derived it from connectivity alone — so someone with a Qobuz session could not create a local playlist without pulling the network cable. Ported against CreatePlaylistModal.slint plus on_create_playlist (main.rs:21107) and on_submit (:21496). The submit branches where the reference branches, and a creation forced local by being offline KEEPS the flag, as the reference's own note says, so Edit can unmark it later to enable "Upload to Qobuz". Two calls made here: - Create rides the QbzPlaylistEdit singleton rather than getting its own, where the reference splits them. Same domain — a playlist's own metadata — with two documents that never interact, and a second QObject is a second boot() to forget; that bridge's header already documents a missing boot() swallowing publishes silently and forever. - The folder picker hides on the offline-only arm. folders_qt::move_playlist takes a u64 and a local:<uuid> has none, so the reference shows the control and discards the value. The shortcut is deleted rather than left beside the modal: a second creation path that skips it is the fork the port rule forbids. All thirteen strings were already in the catalogues. cc |
||
|
vicrodh
|
e1876c15f0 |
feat(qt): the artist page's radio disc goes live, with a start-up indicator
The disc was btnEnabled: false with a comment saying neither engine had a seam on a Qt bridge. Both do now. ArtistPageView.slint:426-461 opens a two-row dropdown — QBZ Radio (the local qbz-radio pool builder) and Qobuz Radio (/radio/artist) — and that is what this restores. The smart arm already shipped behind Discover's spotlight RADIO card and simply was not wired here; the Qobuz arm is new (get_radio_artist + the same enrich_radio_tracks the album radio uses, because its track objects are just as bare). Both labels already exist in all eight catalogues, so no new msgid. AND THE WAIT IS VISIBLE NOW. A radio is a fetch, an enrich and a queue write; until the first track resolves nothing on screen moves, so the click read as dead. QbzCircleAction gains the `loading` arm the reference already has (CircleAction.slint:18/:77) and QbzHome.radioPending carries the KEY of the radio being built — "album:<id>" or "artist:<id>" — so the Discover rail spins only the station that was clicked instead of all six. The spinner rides QbzShell.pulseMs, NOT a RotationAnimation. Qt Quick has no partial redraws, so a continuous animator presents the whole window at display rate (~1.2% GPU per present on the owner's hybrid laptop); a 60 Hz spin for the seconds a radio takes is exactly the bill the pulse law refuses. 12 degrees per ~30 Hz tick, and the Connections is disabled unless the button is spinning and on screen, so at rest it writes nothing. theme/QbzSpinner.qml still uses a RotationAnimation and is deliberately not reused. On the cards, `loading` also pins the hover disc up: it only exists on hover, so it vanished the moment the pointer left after a click, which is what made the click look swallowed. A RadioGuard disarms the key in Drop, so a bad id, a network error and an empty result all clear it. A disc spinning forever is worse than no indicator. cc |
||
|
vicrodh
|
e816436fd0 |
feat(qt): remember where each page was scrolled, across back and forward
The port of the block Slint repeats in 27 views (NavState.restore-scope /
scroll-restore / report-scroll). A .slint file cannot hand a component a
reference to a Flickable, which is why the reference inlines it; QML can, so
controls/ScrollMemory.qml exists once and 31 containers mount it.
Two divergences from the reference, both forced by the same fact — tabs are not
history entries in this port:
1. THE SCOPE IS NAMED BY THE VIEW, not derived from the entry by a scope_for
table. Slint can derive it because its entries carry the tab
(Favorites { tab }); here an entry is a bare route id, so a derived scope
would read "library" for all five Library tabs and pour the Albums grid's
offset into whatever tab mounted. Letting the reporter name its own scope
buys tab granularity for free, and the mismatch case degrades the right way:
an unmatched scope restores NOTHING instead of restoring the wrong thing.
2. THE RESTORE RE-ASSERTS instead of firing once. A Qt item view resets
contentY to 0 whenever its model is replaced, and several pages assign their
model from a binding whose inputs arrive over the network after the mount —
the album page rebuilds listCells from three deferred rails, so it hands its
ListView a new array up to four times in its first seconds and snaps back to
the top over a restore that already ran correctly. The offset is held and
re-applied until a 2.5 s deadline, or until the user takes over (a wheel, a
drag or a flick raises `moving`, and from then the position is theirs).
Diagnostics are off by default: QT_LOGGING_RULES="qbz.nav.scroll.info=true" for
the QML half, RUST_LOG=qbz=debug for the stamp/arm half.
cc
|
||
|
vicrodh
|
47cab502f4 |
feat(qt): remember the immersive shader scene, not just the panel triple
"Remember the last immersive view" covered view_mode / mode / split_panel and NOT the scene, so Plasma, the Spectral Ribbon, the Line Bed and Tunnel Flow were exactly the views it could not remember. Worse, apply_open force-reset the scene to Off on EVERY open edge, which is Slint's behaviour (main.rs:10300-10301) and is only correct because Slint never stored one. The scene now persists through a custom WRITE on the property, so every writer funnels through it: the menu rows, the `g` ring and ShaderSceneLayer's shader-error latch. That last one persisting Off is intended — a scene whose effect failed to compile on this box must not be what the next open restores. The open edge restores it behind three gates: the default-view pref is "remember", the box is on the shader tier (off-tier the scenes render BLACK, so restoring one would open immersive onto a black rectangle), and the stored number is a scene this build ships. Any of them failing falls back to the Off reset, which stays the fallback rather than disappearing. The pref key is new to the shared ui_prefs.json and unknown to the Slint typed struct. Worst case there is that a Slint write drops it and this falls back to Off; the document is not poisoned. cc |
||
|
vicrodh
|
221572b36e |
fix(qt): hydrate the view a session restore lands on
"Startup page = where you left off" opened Library on an EMPTY page: the feed was still [] until the user clicked Library in the sidebar. Same hole for Mixtapes and Collections. navigate_to is not the only door. A restore mounts its view by calling nav_qt::record directly from on_session_entered — it must not clear the landing tab or the detail latch of a session that has not started — and record is history bookkeeping, nothing more. The per-view lazy load lived inside navigate_to, so the restore path never ran it. Split that load into hydrate_view and call it from both doors. Home never showed the bug because load_home_once is its own unconditional phase of shell entry; `local` is absent from the arm on purpose and is not a gap, since LocalLibraryView loads its tab from Component.onCompleted. cc |
||
|
vicrodh
|
d1d96e6918 |
fix(qt): the sidebar playlists menu stops going translucent under ambient
Sidebar.slint:920-922 gives that panel the sidebar's OWN ambient treatment (surface-card at 50%), and the port copied it. That made it the one menu surface in the app that reads through the artwork while the dynamic background is on. Every other menu and flyout paints flat and opaque: QbzContextMenu and therefore CardMenu, SidebarRowMenu, ViewModeMenu and the audio flyout, plus NavFlyout, SidebarFolderFlyout, QconnectFlyout, GenreFilterPopup and LocalFilterPopup. Swept all of them; this was the only outlier. A menu is text being read against art that moves underneath it, not chrome, so the outlier loses. cc |
||
|
vicrodh
|
da8fc598f8 |
fix(qt): the immersive trigger only moves right on macOS
|
||
|
vicrodh
|
3366cc3997 |
fix(qt): Metal sides with OpenGL on the RHI mirror, not with Vulkan
The rule was derived from two data points and generalised into a negation:
owner's Vulkan + mirror=false -> upside-down (Vulkan needs the mirror)
Xvfb Mesa GL + mirror=true -> upside-down (GL does not)
"mirror on everything EXCEPT OpenGL" fits both, and it was wrong about the one
API nobody had run. On the Mac mini M2 the Line Bed comes out upside-down, so
Metal wants mirror=false like OpenGL rather than =true like Vulkan — the
generalisation reached a third platform on the strength of a shape it had never
been tested against.
Kept as a NEGATION with Metal excluded rather than rewritten as "Vulkan ||
D3D". Naming the APIs that DO mirror would silently stop mirroring on any enum
this file failed to list: an unknown member is `undefined` in QML and compares
false, so a typo or a future backend would degrade to "no mirror" with no error.
This way only Metal's behaviour changes and every other backend keeps exactly
what it was verified with.
All four RHI scenes carry the line, because the convention belongs to the
graphics API and not to the scene — a per-scene split would be the surprising
result, not the expected one. Only LineBed made it observable: it is the one
directional scene, so a flip reads instantly, where Plasma and TunnelFlow are
near-radial and hide it. RIBBON IS THE ONE TO CHECK NEXT — it is a spectrogram
with a time axis, so if the convention really is per-API it should have been
flipped too, and if it now looks flipped the rule is per-scene after all.
The three sibling headers said "mirror on Vulkan/Metal/D3D" and were corrected
in place; they had become comments that contradict their own code.
Built and smoked on Linux (gate green). The observation this fixes is the
owner's, on the Mac.
cc
|
||
|
vicrodh
|
9c47d5b2e7 |
fix(qt): the immersive view trigger moves beside the window controls
It sat at the band's LEFT edge, which on macOS is the traffic lights' corner — x ~ 7-85, the pill's exact spot. The band was therefore being inset sideways to make room for one 44px control, and that inset is what read as wrong: the shell HeaderBar's rule is right for the shell, where the lights head a full toolbar row, and wrong here, where the row is three floating glass pills over artwork. Moving the pill instead of the row is the smaller change and the better one. The left half of the band goes empty and the lights get that corner to themselves; the search field is CENTRED on the band rather than laid out after the trigger, so nothing else shifts. Every other platform is unaffected — the trigger simply changed sides. Its right margin RESERVES the capsule's expanded width (112) plus a 12px gap rather than anchoring to the capsule's live left edge. The capsule grows leftward on hover, so anchoring to it would slide this pill sideways every time the pointer crossed the other one — motion with no meaning. Reserving the maximum keeps both still. The radius becomes height/2, the same silhouette as the capsule it now sits next to. An earlier attempt dropped the WHOLE band below the lights; it is reverted here. It moved the search field and the capsule down to solve a problem only the trigger had. The capsule stays. Worth recording why, since it looks redundant on macOS and is only half so: its fullscreen toggle writes `Window.visibility`, which the green light already does — but its X calls `QbzImmersive.open = false`, which EXITS IMMERSIVE, and no traffic light can do that. It is one of three exits (Escape and the player bar are the others) and the only visible one in the chrome. Built and smoked on Linux, built on the Mac mini M2. The macOS geometry needs eyes on a Mac. cc |
||
|
vicrodh
|
d1fd99f947 |
refactor(i18n): qbz-i18n owns its catalogues; macOS immersive band drops below the lights
QBZ-I18N WAS NOT FRONTEND-AGNOSTIC, whatever its own header said. No slint or tauri entry in its Cargo.toml — and eight `include_str!` reaching into `crates/qbz-ui/translations/`. Deleting the Slint crate, which is planned before the Qt release, would have failed to compile `qbz-i18n` and with it `qbz-qt`, i.e. the entire app, over a data directory nobody read as a dependency. The extraction that created this crate took the code and left the data behind. The eight .po files move to `crates/qbz-i18n/translations/` (git tracks them as renames, so history follows) and the includes become `../translations/`. `qbz-ui/build.rs` borrows them back through `../qbz-i18n/translations` for as long as that tree exists — the direction of the dependency is what matters, and it now points the way the deletion needs. MACOS: THE IMMERSIVE BAND GOES BELOW THE TRAFFIC LIGHTS, not beside them. It was inset to their right with the shell HeaderBar's rule, which is right for the shell — the lights sit at the head of a full toolbar row there — and wrong here, where the row is three floating glass pills over artwork and a 44px pill parked against the lights reads as debris. Vertical clearance comes from `SafeArea.margins.top`, the same question the left inset asks in the other axis: with ExpandedClientAreaHint the client area covers the titlebar region and SafeArea reports how much of it is unsafe, with a 28px floor for a 0 reading. The band keeps `leftMargin: 24` on every platform, which puts the trigger's leading edge in the same 20-24 band the lights start in, and the 16px gap under them is the band's own margin reused rather than a new number. The view trigger matches the cluster's width on macOS (SafeArea left margin, 78 floor): stacked, a 44px pill under a ~78px cluster reads as misaligned even with the left edges lined up. Its radius becomes height/2 rather than a constant 18, or the widened pill would have been a rounded rectangle beside three circles. Every other platform keeps the natural 44px, and the glyph was already centred so it needed nothing. The menu opens downward from a band that now has more room under it, so the flyout has more clearance than before rather than less. Built and smoked on Linux (gate green) and built on the Mac mini M2 (3m32s, no errors). The macOS geometry itself needs eyes on a Mac. cc |
||
|
vicrodh
|
1352e499f9 |
fix(qt): the hand-written RHI items now compile on macOS
`build_rhi_items()` compiles the four QQuickRhiItems in `cxx/` with its own
cc::Build, and its macOS arm had never run: those files landed in
|
||
|
vicrodh
|
931cb06d38 |
feat(qt): ease the visualizers on the pulse, and make Plasma react
TWO FINDINGS, both of which falsified something that was being assumed.
FLUIDITY IS NOT FRAME RATE. Driven at QBZ_PULSE_MS=10 the shell presented ~100
frames/s — qt-batches confirmed it, and the script now prints the pulse it ran
with so two logs cannot be confused — and the bars looked exactly as they had at
30. They would: every published frame was applied as a STEP, so the extra
presents repainted the same discrete values three times over. What the Slint
build does differently is not rate, it is `animate bar-h { 90ms }` — it moves
BETWEEN values.
So the interpolation is back, ON THE PULSE, and it costs nothing. The old
interpolator was expensive because of its private 16 ms Timer, not because it
interpolated: it presented at its own rate, unsynchronised with everything else.
Easing on the pulse presents exactly as often as stepping on the pulse did — one
repaint per period either way, only the value differs. Self-parking survives: a
converged ease writes nothing and schedules no repaint. The price is lag, and it
is the honest one — k = 0.45 is ~83% there in three pulses (~100 ms), matching
the reference's 90 ms, against 46 ms of FFT window and 77 ms of Rust smoothing
already in the chain.
The from/to/progress triple is gone with it, and its consumers moved to `cur`.
Note for the next reader: `SpectrumBand` drove all three of its modes from
`Connections` on those properties, and a Connections handler naming a property
that no longer exists is NOT a compile error — QML warns at runtime and simply
never fires. That is how the whole band went dead at once; grep for
`on<Name>Changed` when removing a property, not only for direct reads.
PLASMA WAS DRIVEN BY THE SLOWEST SIGNALS IN THE PIPELINE while the fast ones
travelled beside it unconnected. `transientAmp`, `bandsLo` and `bandsHi` were
published, parsed and stored, and never handed to the item — the shader declared
them and received zeros — so everything ran off the energy bands, which decay
x0.85 per frame (~473 ms) against the bars' x0.65 (~178 ms).
Worse, `beat` stops being a beat on dense material: it decays x0.88 per tick, so
above ~3 hits/s it never falls far enough to come back as a spike and settles
into a shallow ripple around 0.7. A splat multiplied by it becomes a steady
glow, and the busier the music the flatter the picture — the most frenetic input
produced the least motion. `beatAc` subtracts a slow density floor, so a hit
reads as a hit at any tempo.
The scene itself was rebuilt around the reference the owner filmed (Butterchurn
on a percussion-heavy track). Its recipe is three parts: one thin annulus
injected per frame that the feedback warp marches inward, so the concentric
rings ARE the last N frames of audio laid out in space; retention high enough
for that to stack (this held 0.85, where a trail died in ~15 frames and nothing
ever accumulated); and angular structure from reading the spectrum BY ANGLE.
But a CLOSED ring was wrong, and the owner named why: a perfect inner
circumference reads as a drawn shape, hard-edges the field, and at this
retention successive rings pack into one solid disc — which is what killed the
ambient feel. The annulus is gated by angle now, so a hit paints arcs the
conveyor carries inward, the circle never closes, and the gaps stay open for the
background. Noise bends, frays and thins those arcs so they are made of plasma
rather than of compass.
The worms are back at full strength with tighter cores and a contrast curve — a
raw gaussian is all shoulder and reads as haze — and their orbits now reach past
the arc band instead of being trapped inside it. Colour comes from ONE
album-derived triad, so everything used to be a shade of the same hue; a
Rodrigues hue turn splits it into worms, arcs and an ambient wash that sit apart
whatever album is playing. That wash is what stops the background being void for
the foreground to be a cut-out against. A soft limiter replaces the hard clamp:
at 0.945 retention a clamp pins the middle to white and the structure vanishes.
Release build clean, offscreen smoke green, 421/421.
cc
|
||
|
vicrodh
|
730ae6998e |
feat(qt): windowed View-all artwork, a real per-rail default, Play later, MyQBZ tabs
VIEW-ALL PAGES GOT THE LIBRARY > ALL TREATMENT, in both halves. PlaylistBrowse windowed nothing: every card mounted, and a View-all of playlists runs to hundreds, each a PlaylistCard carrying a four-tile collage. AlbumCollection had solved that for the album pages and this one never got it. The band is sampled on a timer rather than bound to contentY — a direct binding re-evaluates every delegate on every scroll frame — and out-of-band cards keep their footprint as bare Items, so the scroll geometry and the infinite-scroll trigger never move. Covers arrived in ONE batch: Rust downloaded every missing cover of the page and then republished the whole document to attach the paths, so a page had no artwork until it had all of it, and the republish rebuilt every delegate on arrival. Both browse pages now ask for what is near the viewport and take the answers one key at a time through the existing sidebarArtworkWindow -> libraryArtworkReady seam. `attach_card_art` stays — it is a pure disk-cache lookup, so anything already downloaded still renders on the first frame with no request. The two now-dead republish helpers are gone with it. The AlbumCollection change is ADDITIVE for its nine hosts: those that pass no `flick` never report, the map stays empty and `artOf` falls through to the `artPath` they already publish. ITEMS PER CAROUSEL now has a real default instead of a lie. The first entry read "All", which was wrong twice over: an "all" in Qobuz terms is what the View-all page holds, and /discover/index hands each rail a different count anyway — measured on a live response, new_releases 26, ideal_discography 26, playlists 20, most_streamed 15, and qobuzissims / album_of_the_week / press_awards 10. No single label could be honest about all of them, so the default became a real uniform cap of ten and the entry says so. That is a behaviour change and not only a label one: the rails that used to arrive at 26 now arrive at 10, and the select is how a rail gets more. Ten therefore stops being a separate option because it IS the first one, and the number comes from Rust so the label and the cap cannot drift. The select also went 82 -> 124 wide: "Default (10)" elided to "Defaul…" in both the closed control and the popup. PLAY LATER on the album page's track rows. It was off because `enqueue_album_track` had only two arms — "next" and "everything else appends" — so "later" and "queue" both plain-appended and offering both would have been two menu entries doing one thing. The third arm lands on the manual block's tail via `add_track_later`, which is the whole distinction, and the QConnect route stops collapsing "later" into "queue" (route_track_to_peer has had a real later arm all along). Only then does the entry turn on. MIXTAPES <-> COLLECTIONS as a tab bar, the Discover/Library pattern where the tabs stand in for the page heading. They are separate ROUTES sharing one component, so a tab click navigates rather than flipping a local property: the sidebar highlight, the back stack and the `kind` Binding then agree on where the user is, and the page transition is the one every route change gets. Eight locales. Release build clean, offscreen smoke green, 421/421 and 216/216. cc |
||
|
vicrodh
|
e9f09a59a4 |
feat(qt): opt-in — a click on a section row opens its first tab
Settings > Appearance, OFF by default. With it on, clicking Discover lands on Home and clicking Library lands on All, instead of only opening the section's flyout and waiting for a pick. The rule lives in NavFlyout because the catalog does: "the first tab" is `entries[0]`, and routing it back through `activate` means the My QBZ section — whose entries are plain views carrying no tab — needs no special case. Both hosts call it from their row's click next to the open they already do, so the sidebar and the title bar cannot drift. The pref rides `settingsJson` like the two Purchases toggles beside it, so nothing is wired beyond the write and the republish every settings toggle already does. The read is a guarded parse: a bare JSON.parse in that binding throws on the pre-publish frame and would take the whole nav catalog — and therefore both hosts — down with it. Eight locales. Release build clean, offscreen smoke green. cc |
||
|
vicrodh
|
81e172457f |
fix(qt): migrate playlist_meta, break a router binding loop, zebra the playlist
RECENTLY PLAYED PLAYLISTS STAYED EMPTY THROUGH EIGHT TRACKS, a restart and a manual refresh — while the plays were sitting in the database the whole time. Read off the live file rather than reasoned about: `playlist_play_events` held 3, 9 and 10 rows for three playlists, so the context edge was working perfectly; `playlist_meta` was empty, and its `.schema` said why — no `covers`, no `own_image`. `covers` and `own_image` were added by inlining them into the CREATE. `CREATE TABLE IF NOT EXISTS` does NOTHING when the table already exists, so anyone who had run the previous build kept the six-column table, every upsert failed on "no such column: covers", and the rail JOINs meta against events — so it rendered its empty state with the evidence one table away. The failure was invisible by construction: swallowed into a warn, and the UI's empty state is also its correct state for a user who has played nothing. A brand-new table is only new ONCE. The moment a build runs anywhere its schema is legacy and every later column is a migration, so the pair moves to the same idempotent ADD COLUMN the album history already uses, and the upsert's failure becomes an error rather than a warn. Verified against a copy of the real database: the columns land and all 22 events survive. A regression test builds the exact pre-covers schema, records a play into it, migrates, and asserts the upsert works and the old event still counts — it reproduces the bug without the fix, and it also asserts a second `init_schema` is a no-op, since every launch calls it. The meta is written by `playlist_qt::load`, so the three playlists already in the events table surface as soon as each is opened once. BINDING LOOP IN THE ROUTER, and this one was self-inflicted: "QML Loader: Binding loop detected for property source". `_stampRoute` runs inside the `source` binding, and it read the last path from a `property string` and then wrote it — a binding depending on a property it mutates. QML detects the cycle, warns, and stops re-evaluating, which is a broken router and not merely a noisy log. The holder becomes an object whose MEMBER is mutated: no change signal, so the dependency is satisfied once and never fires again. (The second warning on the same file, the `when` loop on the nav-flyout `activeTab` Binding, is untouched by any of this work and predates it. Left alone deliberately — it is a warning on a working feature, and rewriting that Binding is its own change.) ZEBRA on the playlist detail's track rows, like the local album page. TrackRow already owns the stripe, so it is one property. Release build clean, offscreen smoke green, 421/421 and 216/216. cc |
||
|
vicrodh
|
907d366d71 |
fix(qt): Library artwork loaded one row; add Refresh content to the configurator
LIBRARY ALL / ALBUMS LOADED EXACTLY ONE ROW OF COVERS and then nothing until the user nudged the scrollbar. The grid and the list reported their artwork window on contentY, model, width, visible and completion — but NOT on height, and their height is gated on `visible` (`height: visible ? parent.height - 16 : 0`, the 2026-08-17 change that stops a hidden body from refilling). QML gives no ordering between "the `visible` change handler runs" and "the `height` binding that depends on `visible` re-evaluates", so when the handler wins the reporter reads height 0 and computes `lastRow = ceil(0 / cellHeight) + 1 = 1` — a window of exactly `cols` items. One row. Nothing fires again until a scroll, which is why moving the scrollbar "fixed" it. Playlists were unaffected and that asymmetry is the tell: PlaylistCollage dispatches its own urls per delegate, independent of the window, while album cells live off `artMap`, which only the window report fills. The control case is in the same tree — LibraryAlbumsList, the playlists list and the Artists rail all report on height and none of them has the defect. The two bodies that lack the handler are exactly the two whose height is gated on visibility; `width` had one because width always changed, `height` never did before the gate. REFRESH CONTENT in the Discover configurator. The index is fetched once per shell entry (`HOME_LOADED` latches and `reload_home` is never called again), so a session shows the page it opened with — no new catalogue rows, and none of the plays made since, because the recently-played rails are built from the local history in that same pass. Everything else in that modal re-renders from the cached candidates, which is precisely what cannot produce fresh ones. The invokable already existed for the error-state retry; this is the door. The Reset button's markup became a shared FooterButton so the pair cannot drift. cc |
||
|
vicrodh
|
f0d753a751 |
fix(qt): fade on tab switches, playlist recents artwork, per-rail sizes, wheel
Four defects from the smoke of
|
||
|
vicrodh
|
26ceeab828 |
feat(qt): mixed playlists, page fade, category filter, per-context recents
Six parity fronts against the Slint/Tauri references, plus the applied-filters
tooltip the app never had.
MIXED PLAYLISTS — the bug. A mixed playlist is a QOBUZ playlist carrying local
and Plex rows through library.db's playlist_local_tracks / playlist_plex_tracks
sidecar tables. This port could WRITE those rows (the picker) and COUNT them
(the Playlist Manager badge), but playlist_qt::load built its rows exclusively
from the Qobuz API response, so a track added to a playlist never appeared on
the page it was added to. Two comments elsewhere claimed the merge already
happened; they were describing the reference. Ported literally:
read_sidecar_rows_blocking (healing + bulk Plex resolve) and interleave_rows
(absolute-slot claim, same-slot collisions all emit), merged in load, with the
merged queue adopted as the open-detail snapshot.
Routing had to split with it. local_playlist_qt::open_id() was the one
predicate for every action, and a mixed detail adopts that same snapshot — so
it now means "a local:<uuid> detail OR a mixed one", which is right for
play/shuffle/row-play and wrong for the writes. Remove and reorder route on the
new local_detail_open() instead; without that, removing a local row from a
Qobuz playlist wrote into the local-playlist repo, found nothing, and reported
success. Removal by source is a third arm: the Qobuz arm parses the display id
as a u64 playlist_track_id, and a local row's display id is a library.db rowid
in the same numeric space. Reorder is refused on a mixed detail for the same
collision, in Rust and in the view.
PAGE FADE. 300 ms OpacityAnimator on the mounted page, OutCubic, honouring
reduceMotion. Zero added latency: the router's Loader is synchronous, the UI
thread is frozen through the mount and no frame exists in which a fade-OUT
could be shown, so the page snaps transparent in the same turn as the rebuild
and only the reveal is animated. An Animator and not a NumberAnimation because
it runs on the render thread — measured 0.53 vs 0.11 of progress with the GUI
thread blocked — and not on the shared pulse because the pulse arrives on the
very thread the mount blocks.
The first cut faded a veil coloured from the pane, and that was wrong in a way
that only shows in the configuration the app actually runs in: AppShell paints
the pane surface-main@0.22 while the ambient background is on, so a veil at
full opacity covered 22% of the page and the fade read as a flicker. Fading the
page has no such failure mode and keeps the ambient art visible.
SPLASH COPY. The login screen was already 1:1 (18 msgids, verified by diff).
The splash was missing the legal disclaimer the reference pins to the bottom of
the screen, drew "QBZ" / "QOBUZ PLAYER" as bare literals (English in all eight
locales), and hardcoded the dark palette — black on the other 34 themes.
CATEGORY FILTER on the Home playlists rail. Slugs now travel per card and the
offered set comes from the same /discover/index response, so the filter is
client-side over the cached 40 with no extra request. The selection lives in
Rust because HomeView is destroyed on every navigation; the reference keeps its
own for the same reason. New trigger + popup, and a `leading` slot on
QbzSectionHeader to hang it off the title line.
APPLIED-FILTERS TOOLTIP. A third QbzTooltip mode taking [{group, values}],
below the control with an upward caret. Overflow is staged summary and never
scroll — the overlay is enabled:false by construction, so a scrollable tooltip
could not be scrolled. Wired where a filter is INVISIBLE (collapsed to a count
or a glyph): the genre chip in four hosts, the Home tags, the Local Library and
Purchases funnels, the Artist Scene genres.
RECENTS BY CONTEXT. QueueTrack already carried context_kind and exactly one
function threw it away: record_queue_track wrote both album histories on every
track start whatever the user had put on, so one 40-track playlist filled a
24-entry album rail with that playlist's contents. album_play_events gains an
additive context_kind column (the other frontend names its columns on insert,
so it writes '' and '' counts as an album — no history reinterpreted), plus a
new playlist_play_history.db and a "Recently Played Playlists" rail, default on.
A separate file rather than a field in recently_played.json, which is a
destructive round trip between frontends: serde drops what it does not know and
the next write puts the truncated object back.
ITEMS PER CAROUSEL, recovered from Tauri and never ported to Slint. It lives in
a NEW TABLE of discover_prefs.db, not in the prefs blob: to_json rebuilds the
object from the keys the running binary knows and save overwrites the whole
column, so a new key would survive only until the other build's next toggle.
Default is uncapped, i.e. today's behaviour.
Two DiscoverySectionId arms were added to the frozen Slint tree, obliged by the
shared enum's exhaustive matches — same two-line follow MostPlayedAlbums took.
Not verifiable here: compiling that crate pulls qbz-ui.
Also fixes two recommendations_qt tests that had been red since
|
||
|
vicrodh
|
047db8ea9f |
perf(qt): recycle the album page's track list, on a uniform-cell ListView
DID NOT DELIVER ITS GOAL. Read this before extending it: the wheel scroll on
a 189-track album is essentially unchanged. The page is correct and the
structure is better, but the theory that motivated the work is not supported.
What the work assumed: the ceiling was the `Repeater`, which has no windowing
and so built all 189 track blocks — each a Column with a disc-header Loader, a
work-header Loader (both gated on DATA, never on the viewport, so every act and
scene header in a Wagner box was alive) and the band-gated row. Roughly a
thousand structural items before counting anything visible.
What it is now: one ListView over a UNIFORM 10px cell tape. Each piece of
content owns a run of cells, its visual item lives in the first cell of the run
and paints over the inert ones — 50px row = 5 cells, 40px header = 4, rail =
29. Uniform cells are the point: measured in a standalone probe, a ListView
with variable delegates ESTIMATES contentHeight and ignores an explicit
override (5703 then 5396 against an exact 5400), which makes the scrollbar
thumb resize and skip as sizes are discovered. With uniform cells,
count * 10 is exact from the first frame, so the scrollbar the owner had just
called the best gesture on the page keeps its behaviour.
Falling out of the design, replacing machinery that was reverted a commit ago:
`cacheBuffer` incubates upcoming delegates asynchronously in idle time, so a
heavy rail is built before it is visible without a hand-rolled band, a settle
Timer, or an unload grace — and without the "reach the bottom, then wait" that
made the owner reject that approach.
Two correctness items that came with it:
* `header` is a ListView property. The page's own `header` had to become
`albumHeader` or every `header.x` binding inside the view would silently
resolve against the view's header instead. One handler was missed in the
rename and QML refused to create the component at all — the whole page
rendered black, and no static gate caught it because the failure happens at
instantiation.
* `recycleActive` on the shared TrackRow, default true, gates the row's
Connections. Without it a pooled row keeps listening and consumes a cache
or favourite settle signal meant for a different track. Every other host
is unaffected by the default.
WHAT THIS RESULT MEANS, for whoever picks it up: delegate count and
construction cost are NOT what makes this page feel heavy. Removing ~950
structural items changed nothing the owner could feel, which points at
per-frame cost instead — consistent with his own observation that the page is
sluggish ALWAYS, not only while it loads, and with the documented finding that
this shell's bill is full-window presents rather than scene work. Do not
re-attempt a structural rewrite here on intuition. Take the QSG_RENDER_TIMING
split (polish/sync vs render) on this exact album first; the earlier reference
scene measured polish=0ms, render=2ms, swap=12-24ms.
Design and probes came from an independent review; the rebase onto the current
tree, the black-screen fix and the API verification are mine.
cc
|
||
|
vicrodh
|
3fdccae22e |
Revert "perf(qt): the album page's rails exist only near the viewport, and never build mid-scroll"
This reverts commit
|
||
|
vicrodh
|
313a7e7e4e |
fix(qt): the local album page gets its zebra, across the whole row
The shared TrackRow has had a `zebra` flag all along, off by default, and the Qobuz album page turns it on. The local one never did — and turning it on at the call site would not have been enough either, because LocalTrackRow wraps the shared row without forwarding the flag. Forwarding it is still not the fix. TrackRow tints its own background rectangle, which is `parent.width` wide, and this wrapper hands the shared row `root.width - 26 - 46`: a 26px gutter for the source glyph and 46px for the context menu the wrapper draws itself. A stripe painted from in there stops 72px short and leaves both controls outside it — which is what the smoke caught. The stripe is therefore painted by the wrapper, full width, declared before the shared row so it lands underneath and takes no pointer events. Same colour literal and same parity rule as rows/TrackRow.qml, so a local list and a Qobuz list stripe identically. cc |
||
|
vicrodh
|
5b99a6698e |
perf(qt): the album page's rails exist only near the viewport, and never build mid-scroll
The three SectionRails at the foot of the album page were gated on `visible:`
alone, and in QML `visible: false` stops an item PAINTING, never building. With
data present all three were fully constructed the moment the page mounted —
three horizontal ListViews of AlbumCards, every cover a RoundedImage with two
render targets — whether or not the user ever scrolled down to them. On a long
album nobody reaches the bottom of, that is pure cost, and it sits in the SAME
Flickable as the track list, so those layers are in the scene for every frame
the list scrolls.
Each rail is now a fixed-size `RailSlot` with the real rail behind a
`Loader { active: inBand }` — the shape `rowSlot` in this file already uses.
Three things this had to get right:
* The slot reserves 286px whether or not it is built (28 header + 12 spacing
+ 246 card band, the figure RailSkeleton already stands in with). A slot
sized to its Loader would change the page length as rails build and slide
the content under the user mid-scroll.
* The band reads the slot's OWN `y`, which the Column assigns from preceding
siblings only. The variant measured against a parent-derived height is a
dependency cycle; it cost a day on 2026-08-18, because QML breaks the loop
silently and every slot then reads as in-band.
* An inline `component` does not share the enclosing document's id scope, so
`page` / `pageFlick` / `root` are not resolvable from inside RailSlot. The
geometry and the cover map are handed in as properties instead. Left as it
was, the failure would have been silent: a band evaluated against zeroes.
Two follow-ups from the owner smoke, both about WHEN the Loader flips:
* UNLOAD IS DELAYED 600ms. This page grows in stages (primary document, then
deferred rows), and while the list is missing the page is short enough to
put these rails inside the band. They began building, the rows landed, the
page grew by thousands of pixels and the Loader tore them down again —
and since a ListView creates delegates through incubation, tearing down
mid-flight is exactly the "Object or context destroyed during incubation"
warning that appeared three times, once per rail, on every album open.
* BUILD WAITS FOR REST. A Loader creates its content synchronously, in one
frame, on the GUI thread. Building the instant a rail crossed into the band
put that spike inside a wheel scroll — measured by the owner as the wheel
getting worse exactly as the scrollbar got much better, which fits: a drag
ends before the spike lands, a wheel scroll does not. Creation now waits
for `flickY` to stop ticking. That signal works for the wheel and the
scrollbar alike, unlike `Flickable.moving`, which a scrollbar writing
contentY directly does not necessarily raise. Already-live rails stay live:
this gates the build, never the keep.
cc
|
||
|
vicrodh
|
77e9feccff |
feat(qt): album header — 260px cover, no divider rule, a gradient that lands mid-toolbar
Four owner-requested changes to the Qobuz album page, validated by smoke. 1. THE DIVIDER RULE goes invisible rather than away. The gradient band sizes itself off that item's `y` and the whole column below is positioned after it, so deleting the item would move the page; only the paint was unwanted. 2. COVER 224 -> 260. Now one constant: the literal was repeated at five sites — the real header, its skeleton twin, and the two text columns that measure `parent.width - 224 - 32`. Changing only the header would leave the skeleton at 224 and make the page jump 36px when the real one replaces it. The header Row has no explicit height, so it grows on its own. The URL tier is unchanged (`best()`, per the 2026-08-16 ruling that no existing surface is re-tiered DOWN); what decodes is RoundedImage's scaled derivative at the drawn size, so 260 over 224 is ~35% of a small image. 3. THE ATMOSPHERE REACHES MID-TOOLBAR — the 8px spacer below the divider plus half the 52px toolbar band — so it dies across the quality chip and the search box instead of stopping at the divider. A named constant rather than a `mapToItem` chain: that chain crosses three nested containers and goes stale when an ancestor moves without notifying. 4. THE FALLOFF IS PROGRESSIVE. The tint held FULL strength to 0.82 and died inside the last 18%, drawing a visible seam — and extending the band made it worse, not better, because the falloff is a FRACTION of the height: the solid part grew while the fade stayed as short. The page-surface overlay underneath started at 0.82 too, putting a second seam on the same line. Both now decay over the bottom ~60% on an ease-out curve, stepped through explicit alphas because a GradientStop interpolates linearly between the two colours it is given. cc |
||
|
vicrodh
|
c02e7713d9 |
perf(qt): stop republishing the Recommendations document for artwork
Covers that resolved after the document was published forced a second
`publish_snapshot`, and a republish hands QML a brand-new array:
`QQuickItemView::setModel()` then resets each rail's scroll offset and tears
down its QQmlDelegateModel. So every landing batch rebuilt all nine rails'
delegates, which is what a screen capture of a cold Recommendations load
shows — the rows visibly reassembling while covers trickle in, filling out of
order both between rows and within a row.
The fix is the pattern this port already applies twice: covers travel as a
url-keyed PATCH and the document is left alone. `foryou_qt::emit_foryou_art`
is the precedent for the For You rails, and `myqbz_bridge` cites it verbatim
for MyQBZ; the reco rails were the remaining holdout.
The download path is deliberately untouched. This changes only how a resolved
path is ANNOUNCED, never what is fetched or cached, so every other consumer of
the shared image cache — the now-playing bar, the queue, the `recent` rails —
sees exactly what it saw before.
Two traps this had to close, both already paid for once on the For You side:
* `sectionArtOf()` falls back to `item.artPath`. `recoArt` only ever holds
Recommendations urls, so for a Home / Editor's Picks / For You card the
lookup misses and the binding returns what it always returned. No surface
that reads a baked `artPath` can lose its cover to this — by construction,
not by care. The pinned rail, fed by its own store, is untouched.
* `anyItemArtPending` now consults the map. A patched cover never gets its
`artPath` back through a republish, so the bare "artUrl set and artPath
empty" predicate would be true forever and pin the 900ms skeleton pulse
Timer on.
`refreshRecoArt()` on mount is the third: HomeView is destroyed on every
navigation while this document is published once per session, so a fresh mount
must be re-handed what already resolved or the covers come back blank.
Weekly's two rails still republish their own narrower property; their blast
radius is two rails, not nine, and widening this change was not worth it.
Verified by owner smoke: the three checks (nothing else lost its covers, the
rails stop reassembling, covers survive navigating away and back) all pass.
cc
|
||
|
vicrodh
|
101d0e25d4 |
perf(qt): one rebuild per query in the local grid, and reveal covers in order
Two changes to the local albums grid, both BUILT BUT UNSMOKED — Local Library is best-effort for the next release, so this is committed to preserve it rather than because it was validated. 1. ONE REBUILD PER QUERY. `rows` and `groups` are two bindings over one query: the host derives `albumsVisible` from the search box and `albumsGrouped` FROM `albumsVisible` (LocalLibraryView.qml:378-380). With a handler each, a single keystroke rebuilt twice, and the first of the two was WRONG as well as wasted — it chunked the new rows against the previous groups. Turning grouping off did not save it either, because `groupRows` returns a fresh `[]` every time and a new array is still a change. Coalesced onto a zero-interval Timer: one rebuild per event-loop turn, reading both new values. `Component.onCompleted` stays synchronous so a mount does not show a frame of empty grid. `[coltiming]` logs every rebuild with its cost, so a smoke run can falsify the coalescing: one line per keystroke means it holds, two means it does not. Timed around the work, never from a deferred callback. 2. ORDERED COVER REVEAL. The cold-thumbnail pool starts work in window order but finishes out of order, since decode cost varies per cover, so covers appeared scattered across the grid. Ordering the emission in Rust would have made it worse: one slow decode at index 3 would hold 4..N behind it, adding latency instead of hiding it. Instead arrivals are applied behind a front that walks the window at a fixed rate; a cover landing behind the front is applied at once. The front advances every tick unconditionally, which is what makes a stall impossible, and a key with no ordinal is released immediately. Note for whoever picks this up: this reveal covers the LOCAL path only. The uneven fill visible in ordinary navigation is HomeView's For You rails, which take a different route (`forYouArt` via `onForyouArtReady`). cc |
||
|
vicrodh
|
ed081cac6b |
fix(qt): put the local albums scrollbar and A-Z strip on the real edge
LocalAlbumsTab pads its content 32px on the right, and both the scrollbar and the jump strip lived inside that padding — so the bar sat 32px from the window edge (52 with the strip up) while LibraryView.qml:1281 puts the same control at 4, or 34 to clear its strip. Side by side the local grid's bar reads as floating away from the edge. The collection now takes two measurements, both from the WINDOW's right edge: `scrollBarInset` (where the bar should land) and `hostRightInset` (how far the component's own edge already is from it). The bar's margin is their difference, which goes negative and overhangs — safe, because the component root is a plain Item; the `clip: true` in that file is on the ListView, a sibling. Both default to 0, so LocalArtistsTab and LocalFoldersTab, which have not opted in, are untouched. The A-Z strip moves out to 12 from the window edge, matching LibraryView.qml:1269. cc |
||
|
vicrodh
|
c0fe8a80d3 |
fix(qt): silence the TypeError flood the constant-cell grid introduced
Regression from the previous commit's first fix. Pinning the inner Repeater
to `root.cols` is what made recycling free, but it also means the tail slots
of a partly-filled row are now real, live cells with no album behind them.
The delegate anticipated that — `modelData` resolves to null and the cell
sets `visible: false` — and that is not enough: in QML `visible: false` stops
an item PAINTING, it does not stop its bindings EVALUATING. Every
`cardCell.modelData.id`, `.title`, `.artKey` on an empty slot therefore threw
once per evaluation, which during a scroll is a continuous log flood (the
same ten line numbers, over and over, seen in a smoke run).
`modelData` keeps its meaning — "is this slot filled?" — so `visible` still
works. The bindings now read `cardCell.slot`, which is the same object when
the slot is filled and a frozen all-empty object when it is not.
`artWanted("")` is false and `artPathOf("")` is "", so an empty slot stays
silent through the skeleton's probe arm too.
Deliberately NOT a Loader gated on the data: `active` would flip every time a
full row recycled into a partial one, rebuilding the AlbumCard — precisely
the construction churn the constant cell count exists to remove.
cc
|
||
|
vicrodh
|
bd9aec1440 |
perf(qt): make the local albums grid actually recycle, and stop it asking
for the whole library
The owner's report is precise and is what this targets: the grid OPENS
fine; scroll, scrollbar drag and search are what feel sluggish, and "the
first two keystrokes hurt" — the ones where the filter still matches
hundreds. That shape is a model swap, not a slow filter.
UNSMOKED. Committed to save progress; no binary matches this tree.
1. FIXED CELL COUNT. `reuseItems` was already on the ListView and bought
nothing, because the pool recycles the row SHELL while the inner
`Repeater { model: modelData.items }` got a new array on every recycle
and rebuilt its AlbumCards anyway. The card is where the cost is: ~29
items, a RoundedImage with TWO render targets (its own layer plus a
separate mask layer), image probes, a skeleton. The shell was reused
and the expensive part was not. The Repeater now runs over `root.cols`
— a constant — so cells survive recycling and a scrolled or refiltered
row only reassigns `item`.
2. THE HEADER WINDOW BUG. `loIdx` fell back to 0 whenever the first
visible entry was a group header, because headers carried no `base`.
In grouped mode, scrolling with a header at the top of the viewport
asked `queueWindowReport` for artwork from album 0 through the current
letter — hundreds of covers dispatched and later evicted, per scroll.
Headers now carry the flat index their group starts at.
3. SEARCH DEBOUNCE on the albums box, which the Tracks box in the same
file already had. `albumsSearch` is not a query, it is the filter
input: writing it per keystroke refilters 1267 albums, rebuilds the
entry list and hands the ListView a new model, tearing down every
delegate. Note the tradeoff — this ADDS 200ms before any work, so the
box may feel more deliberate even as it stutters less.
Also here, from the album pages (lower priority per the owner — their
cost is paid once and the view is fine in use afterwards): the disc
header, which carried a CardMenu per TRACK, is Loader-gated; a genuinely
quadratic multi-disc scan in LocalAlbumView is hoisted out of the
per-row call; and the track rows are windowed. That windowing was
measured broken twice (a band computed against its own parent's derived
height — a dependency cycle) and the fix is unverified.
Credit where due: 1 and 2 came out of an independent review that also
judged the previous round's lazy menus and `reuseItems` "mostly beside
the main point". That is a fair call.
cc
|
||
|
vicrodh
|
99722aedaf |
feat: an unreachable track can no longer stop playback
Two owner requirements, one shape: something that cannot play must be skipped, and cleaned when there is nowhere left to skip to. The causes are outside our control — Qobuz withdraws a track overnight, or a local/Plex file sits on a share this network cannot see today. THE FUNNEL, which is where the real gap was. `play_current_if_local` returned a bool whose two `false`s meant opposite things: "not my row, run the Qobuz path" and "my row, and it cannot play". The caller could only act on the first, so an unreachable local file fell through to the Qobuz tier-walk, failed there with an unrelated error, and `auto_skip_unavailable` then declined it as non-terminal. One dead file stopped playback silently. It is now a three-state `LocalPlay`, and a local failure takes the SAME bounded walk a pulled Qobuz track takes — to the user they are the same event. CLEANING. Both dead ends of that walk (budget spent, queue edge) now call `purge_unavailable_upcoming`, which drops only rows `queue_track_unavailable` already refuses to play, walking backwards so a removal cannot shift an index still to be visited. A track on an unreachable share can never land there: it is Unreachable, not dead. PROBES THAT CANNOT HANG (`qbz-library/src/reachability.rs`, 4 tests). The machinery for network shares already existed — mount_info, `library_folders.is_network`, a populated `local_tracks.is_network_mount`, even an unreachable-folder guard in scan.rs. Every one of them probed with a bare blocking syscall, and a mounted-but-unreachable share does not make `exists()`/`read_dir()` fail, it makes them BLOCK for the mount's timeout. The guards could not fire; scan.rs hung on the very condition it existed to detect. Three states, not two, and that distinction is the whole safety property: `Missing` is the filesystem saying the file is gone and may be cleaned; `Unreachable` is it saying nothing, and may only ever be skipped. Collapsing them — which a bare `exists()` does — would delete a user's NAS library because they changed networks. The deadline goes on the PROBE only, never the transfer: a hi-res FLAC over a working-but-slow share legitimately takes longer than any probe budget, and timing the read out would turn "your network is slow" into "this track is gone". A blocked syscall cannot be cancelled, so a timed-out probe leaks its thread; the mount is therefore remembered with a cool-down, one probe per dead mount rather than one per track, which on a 2000-row queue is the difference between a defence and a thread bomb. PERSISTENCE. `busy_timeout=1000` on all 18 store pragmas. Deliberately NOT the 2-3s discussed: these writes run on the Qt thread, so a 3s wait would install exactly the stall this whole sweep has been removing, while today's 0 means contention loses a setting silently and instantly. One second resolves transient contention without a visible stall. `home_bridge::ui` announces both of its silent failure modes. Three builds went into proving it was NOT dropping the Home document during the launch freeze; one warn line would have answered it in one. Verified against the owner's real session state (1935-row queue): the app launches to a painted Discover with no QML errors and no dropped publishes. cc |
||
|
vicrodh
|
47a7641fc2 |
fix(qt): bound the queue panels, and make the gate catch what it missed
Follow-up to
|
||
|
vicrodh
|
0ed9db2f3f |
fix(qt): stop the Weekly rows rebuilding the whole Recommendations tab
Reported as "it loads, then it flashes and loads again", and only on that tab. Measured: the cached blob paints at t+0 and the two Weekly rows land at t+0.9s — and that is with BOTH of their per-week caches hitting, so the delay is ListenBrainz's playlist listing, not the tracks. Their arrival called publish_snapshot, which republishes the document, which hands QML a brand-new array; QQuickItemView::setModel() then tears down every rail's QQmlDelegateModel and rebuilds it. Nine rails destroyed and rebuilt to deliver two, a second after the tab had already settled. Not the artwork, which was the obvious suspect: the same run logged "0 artwork misses". Split them out the way HomeView already splits pinned / radio / spotlight, for the same reason and with the same shape: the document carries an ordering slot, the rows ride reco_weekly_json, and rebinding that re-creates only those two delegates. Every other rail's sectionData resolves to the same modelData reference, so a Weekly landing re-evaluates nine bindings and rebuilds two. The slots are emitted unconditionally (a new `slot()` helper) rather than self-hiding on emptiness like `section()` does — a slot that vanished while empty and reappeared when filled would change the document, which is the republish being avoided. QML hides an empty one instead, on the gate the other out-of-document rails already use. Weekly covers are attached in the same split, so a late cover republishes that property rather than the document: two rails, not nine. cc |