A rule's − and + were laid out right to left in whatever room the row had
left. The dialogs share one window, which keeps the width the previous
dialog left it at, so after a narrower one the rule editor only grew to
the end of its longest value: a date "is between" then had no room left,
and the − and + were drawn over its end date and the end date's calendar
button. That calendar was there all along, but the + on top of it took
the click. The buttons now follow the value (pushed to the right edge
when there is room, as before), so the dialog widens to the row.
Every dropdown list drew its items with the theme's resting button
border. An unhovered item gives up that border's width of padding, so a
hovered one needed 2 points more, and the longest name of a submenu
sized to it (Date ▸ Capture Date, Keywords & People ▸ Keyword Count, and
the longest name of every other group) wrapped onto a second line. The
dropdowns' lists and their submenus now drop the resting border, as
egui's own menus do.
Fixes#2353Fixes#2352Fixes#2354
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A max-pool, a channel-copy 2x2 transposed convolution (mode 1) and the
channel-copy 1x1 layer give an exact answer, so a store fault in mode 1
can't hide in tolerance (CodeRabbit review of #2380).
On the GTX 1070 this test passes even without mode 1's stepped stores:
FXC merges them the same way, but the driver runs that store right, so
only mode 0 was shown wrong. Mode 1 keeps the stepped stores as a
precaution; the shader comment and docs/gpu-pipeline.md now say so.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On DX12, FXC merged the conv kernel's four output stores (modes 0 and 1)
into one 16-byte store_raw. The DXBC is valid and aligned, but NVIDIA's
driver ran it wrong on a GTX 1070: every second pixel's channels came
out shifted, so AI Denoise on the GPU was silently wrong on Windows.
DXC and Vulkan were right on the same card.
The stores now step by `1 + p.zero` (the always-0 uniform, renamed from
`pad`), which FXC cannot fold, so it keeps them apart. A two-layer
channel-copy network checks the stores exactly; it fails on that card
without the fix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The top bar's left margin holds the traffic lights' room on macOS, and full
screen gives that room back (#1763). The search field was centred on the rect
inside the margins, so it sat 33 px right of centre in a window and 2 px right
of it in full screen, and its width changed by 62 px as well: the field jumped
sideways whenever that margin changed.
The field is now placed against the whole bar, so neither its centre nor its
width depends on the traffic lights. The left buttons still move into their
space in full screen. `window_handle` takes that bar rect instead of building
the same rect from the content and its margins.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
GPU handles(): only output-device proofs stay on the CPU; Show All Masks now renders zoom windows (upstream); proof-boundary tolerance in the GPU equivalence test; device-profile draft test.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
covers_camera tried only the first-space split, so NIKON CORPORATION NIKON Z5_2
stayed on the relative scale in the catalog and editor. Try every split, as
spectral::find_camera does.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A badge click no longer selects the photo first: stack.toggle opts into explicit_targets, so
`ids` skip the selection guard and the command refuses ids with no stacked photo. Compare keeps
its pair. Both badges announce Expand Stack / Collapse Stack.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review follow-up:
- the badge is painted after the region window renders, which zoomed in
covered it (the very case the fix is for);
- with a filter active, the pill in the canvas' corner has that corner:
the badge sits below it, also when the photo's corner is just under it;
- a wide label near the pane's right edge stays 8 points inside it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AbegtYaPuwqZ81pGfELsFK
Found while testing crates/engine/src/cmd/library.rs:1723 (parse_gps)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Found while testing crates/engine/src/cmd/library.rs:1723 (parse_gps)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The folder's mtime may not move from the save's listing to the list that follows (same clock tick on a fast machine), so the stale cache hid the bad files and problems came back empty - deterministic on a warm build. The test drops the cache as its own wild.json case already did (and as templates_changed documents).
The folder tree writes FolderNode.path with forward slashes on every platform, but tests_folder_drop built its widget ids from to_string_lossy, which gives backslashes on Windows: the 8 tests never found their rows there. row_path() normalizes the way tests_sync already does. Same class of fix in tests_export (path comparison) and an unused-import guard in catalog's staging tests (the module's only test is unix-only).
Review on #1468: main's #2099 split out text_matches() and lets unknown field
prefixes fall back to plain text. On top of it, text_matches() now also matches
Photo::people(); the free-text token is unquoted before text_matches(); an
unknown field's quoted value is unquoted in the fallback (`10:"30 am"` is the
text `10:30 am`); and the field split keeps the !t.starts_with('"') guard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
From review on #2327: the move is in the engine, so a script, the CLI or an
MCP client that rates the active photo out of a filtered view without `ids`
finds the next photo selected afterwards. The parameter text of the eleven
rating, flag and label commands now says so, where commands.list and the MCP
tools show it: without `ids`, the selection moves on when the edit takes the
active photo out of the view.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With a filter showing one-star photos, taking the star off a photo took it
out of the view but left it the active photo: the grid highlighted nothing,
Detail kept showing it, and the next step (→, Shift+0–5 or Auto Advance)
started from the view's first photo, because `library.next` finds no place
in the view for an active photo the view doesn't show.
A rating, flag or label applied to the selection now does what a delete
already does when it takes the active photo out of the view: the photo after
it becomes active (the one before it at the end). That photo is the next one,
so `advance`, Shift+key and Auto Advance (keys and the Photo menu) don't step
past it. Commands that name their photos with `ids` leave the selection
alone, so Compare and Survey still show the photo they rate; a held view
keeps the photo, and the selection with it.
photo.pick, photo.reject, photo.unflag and photo.label now take `advance`
like photo.rate and photo.flag.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The test checked that Restore All was listed but clicked it only when the
machine already had a hidden location, so CI never exercised that path.
It now hides two places and restores them with Restore All, then hides
one and checks the row restores it directly, without a list.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AbegtYaPuwqZ81pGfELsFK
write_file_durable creates parent folders, which broke the ui-egui test
exporting_keywords_says_why_it_failed. Write atomically with
safe_file::write_atomic so a missing folder is still an error.
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Zoomed in, the photo's corner is off the canvas, and the "Before" label
(\ key or the Before view) left the screen with it. The label's pill now
sits at the canvas' corner when the photo reaches past it, on whichever
axes it does; at Fit nothing changes. Same rule for the hovered preset's
name. Stacked on #2274 (the dark pill behind the label, same lines).
Tests: the pill geometry with the photo inside and past the pane, and a
headless check that at 100 % the registered label rect stays inside the
canvas.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AbegtYaPuwqZ81pGfELsFK
The menu bar turned a command's error into a toast only for exports, so
File ▸ Back Up Library… did nothing visible when the host refused (changes
still unsaved, or nothing stored in this browser); the refusal reached only
the control channel (found in a real-browser check of PR #1987). Backup and
restore refusals are toasted too. Menu items are registered for automation
as `menuitem:<command id>`, and a headless test clicks File ▸ Back Up
Library… against a refusing host and reads the toast.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AbegtYaPuwqZ81pGfELsFK
Found while testing crates/meta/src/tags.rs:148 (rational)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
"Show N hidden locations" put every location removed from Local back at
once. With several hidden, the row now lists them (name, path on hover):
each comes back on its own, or all with Restore All. With one hidden, the
row restores it directly, as before. The command side already had
local.restoreHidden {path}; only the panel called it without a path.
The headless test hides a second place, opens the list from the row,
restores one, checks the other stays hidden, then restores the rest.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AbegtYaPuwqZ81pGfELsFK
`cargo test -p <changed>` resolves features for that selection only, so its artifacts differ from
the full `ci`'s workspace build (and from a worktree warmed from one): it recompiled ~130 crates,
dependencies included. Now `--quick` runs `cargo test --workspace --no-run` (fresh when the
workspace was built before) and executes the test binaries of the changed packages, with the
package directory as cwd like cargo does. Doc tests stay with the full `ci`. A one-file change
in lightcraft-raw in a warmed worktree: 170 s for the whole `ci --quick`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Queued embedded grid-preview requests now change priority without restarting their work. Undo to an embedded preview retires obsolete thumbnail requests so a late edited render cannot replace the restored preview.
Partially addresses #348.
Co-authored-by: Codex <codex@openai.com>
`face_download::Downloads::drop` gave a running download a second to stop,
with `Instant::now()` and `thread::sleep`, which panic on wasm32 ("time not
implemented on this platform"). The web host replaces its Session when the
library can't be opened, so with this branch an unreadable `catalog.log`
killed the page with the generic panic screen instead of showing the
library-problem window (found in a real-browser check of PR #1988). The wait
is skipped on wasm32, where there are no download threads; the cancel flags
are still set. Native test: an idle Downloads drops at once.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AbegtYaPuwqZ81pGfELsFK
- .github/workflows/ci.yml: the full `cargo xtask ci` on every pull request and push to main
(Linux x64, craft-fonts, rust-cache for dependencies), so it becomes the merge gate.
- `cargo xtask ci --quick [--base REF]`: fmt, clippy, the changed packages' tests, layers and
assets, plus heif / parity / wasm when the changes touch them, measured against the merge base
with origin/main (xtask/src/ci_quick.rs). Workspace-wide files test everything.
- A machine-wide lock (xtask/src/ci_lock.rs): a second `ci`, from another worktree, waits and
names the run it waits for. LIGHTCRAFT_CI_SLOTS=N, LIGHTCRAFT_CI_LOCK=off|<dir>.
- README, CLAUDE.md quality gate, gaps.md gap 9, target-app-parity.md stability note.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every level of Albums, Local, By Date, Folders and Keywords still steps one
level right of its parent, but the step is chosen each frame from the panel's
width and the deepest open row: 16 pt while that row keeps about 80 pt for its
name and room for its count, else 14, 12, 10 or 8 pt. Past 8 pt the sidebar
scrolls sideways as before. Callers keep passing indents in 16 pt steps;
`indent_x` maps them to the drawn step for the icon, name, chevron, album drop
line and the width a row asks for.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A bad `orientation` was refused only on the refit path: "current" and an
unlocked "toggle" returned before the check, so {"aspect": "current",
"orientation": "bogus"} succeeded silently, and a value that isn't a string
fell back to the default. It is parsed first now, so every path refuses a
wrong word, a number, a boolean or null, and nothing changes (CodeRabbit on
PR #2280).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AbegtYaPuwqZ81pGfELsFK
