Homography::apply, Affine::apply, Warp::to_corrected,
Warp::corrected_to_source and embedded_warp are now generic over
lightcraft_geom::Real (private *_real variants in pipeline, public
apply_real in geom); the f64 API is a thin wrapper, so callers are
unchanged. Real is implemented for f64 (plain IEEE operators) and for
Interval, which rounds every bound one ulp outward and turns division by
a range containing zero, the homography's 1e-300 denominator clamp and
non-finite bounds into "undecided" (NaN). Because the interval runs the
same operations as the f64 path, it bounds the rounded f64 results, not
only the exact ones.
The f64 path keeps the same operations in the same order (only commuted
operands, which IEEE + and x leave bit-identical; no fma). Proof:
warp_f64_results_are_bit_identical_to_the_pre_generic_formulas hashes the
bits of 2000 random warps x 40 points through every formula (plus the
clamp branch, including +-1e-300 exactly) and matches the digest computed
by the same test on thumb-split (419261c) before the change; reassociating
one expression, or `<=` for `<` in the clamp, changes it.
The framing chain (output px -> transformed -> lens-corrected -> green
source, then inside the edges) is written once: Warp::framing, generic.
Warp::frame (f64, returns the corrected and green source points) serves
both the CPU resample (sample_warped) and Warp::covers; block_coverage
evaluates the same chain with intervals over a whole block of output
pixels and compares the bounds against the same edges() inside uses, with
no margin: the bounds enclose every pixel's f64 result, so the decision
equals the per-pixel one; else None.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The fixed four jobs were safe on 8 GB but too many for 4 GB, and slower than
needed on large machines. `cargo xtask ci` now uses one build job per 2 GB of
installed RAM, capped at the CPU count, and 4 if the RAM can't be read
(/proc/meminfo, Win32_ComputerSystem, hw.memsize). With line-table debuginfo,
4 jobs measured about 3 GB of RAM in use and 2 jobs about 1.5 GB.
Test threads follow but stop at 4: more barely shortens CI, and at 16 threads
on a loaded machine the wall-clock frame-budget test
offline_originals_never_block_a_frame failed (one 303 ms frame). Explicit
CARGO_BUILD_JOBS / RUST_TEST_THREADS still win.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
keyword_list_filters_renames_and_suggests failed intermittently under
full-suite load (1 photo instead of 2 after clicking travel|italy, or no
travel|italy row at all). No metadata was lost: at the failure every
photo still had its keywords, but the filter was `travel|france`.
Local's rows (Downloads, Home, ...) are added only once their is_dir
checks come back from fs_cached worker threads. Those start when the left
panel is first drawn, which in this test is right before the clicks.
Headless::busy() did not count them. Under load a row could land between
the frame that read a widget's rect and the frame that delivered the
press. The keyword rows then moved down 29 px and the click hit the row
above (travel|france), or the toggle click missed travel. Diagnostics
showed the click aimed at y=1307 and travel|italy drawn at 1365 after
the two Local rows landed. The same failure reproduces on 5cee0c0 (2 of
20 runs), so the thumbnail work did not cause it.
fs_cached now keeps a per-context count of checks in flight
(fs_cached_running), and busy() includes it, so settle() waits for them.
The keyword test settles after opening the left panel. The new test
settle_waits_for_file_system_checks holds a check open and fails without
the busy() change.
Full ui-egui lib runs with 4 test threads: 2 failures in 25 before the
change, none in the runs after it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An unedited raw first shows its embedded preview (ThumbQuick), then the
grid queues the real thumbnail at BACKGROUND_THUMB_PRIORITY. The grid asks
at its own priority (10 on screen), so thumb_current found no match while
the stand-in ran (nothing pending for the real slot) and while the real job
was queued (pending at 3, asked at 10): request_thumb rebuilt the job every
frame without changing anything. thumb_inputs now records whether the
request went through a stand-in; such a request is current while its
stand-in is pending or its real job is pending/rendered at the background
priority. Direct requests keep their priority check, so a higher priority
still promotes a queued job.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Changing the disk cache size (library.preferences cacheMb) re-attached a
new PreviewCache object, and the renderer treats a new cache object as a
library reset: every texture, the loupe included, was dropped and
re-rendered. Thumbnails are keyed by content, so a new budget leaves them
valid. The budget now changes in place on the attached cache (same object,
same generation); the next write trims to it, as before. A preview clear,
a generation bump and a new library still invalidate.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Hiding a kept Local folder beneath Home while browsing it (Remove from
Local, local.hide) expanded Home down to that folder and pushed the
"Show hidden locations" row out of view. local_places picked the
innermost listed folder containing the browsed one as the tree to
reveal; once the kept row was hidden that became Home, so folder_tree
opened Home, AppData, Local, Temp, ... to show the hidden folder again.
A tree no longer opens on the way down through a hidden folder: the
owner is chosen only among places whose path to the browsed folder
passes no hidden one. Hiding Home itself still lets Pictures reveal a
folder browsed inside it.
local_location_can_be_hidden_and_restored again hides a browsed
temporary folder (beneath Home on Windows), kept as Browse Folder...
does, while it is still browsed, and checks the restore row is inside
the viewport before clicking it. It failed before this fix. New unit
test: hidden_folder_is_not_revealed_in_an_outer_tree.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
thumb_request_bench measures UI-thread time and allocations for grid
thumbnail requests (with and without pending work), using the same counting
allocator as the other bench examples. REAL_PHOTOS=1 samples up to four
photos from PHOTO_ROOT, which must now be set explicitly; the original
defaulted to a local photo library path.
Split from a327b98 (T0 thumbnail work).
Co-authored-by: Codex <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Skip unchanged thumbnail jobs and reuse bounded settings hashes (both maps
capped at 1024 entries). Reject stale render results and cache writes after
edits, library changes and preview clears: render requests carry a request
id and cache generation, PreviewCache writes and reads hold the generation
lock that clear() takes, and library.clearPreviews also drops cached
sources. The web worker index follows the same generation rule.
The decoder guards, Windows path fixes and test fixture changes that were
bundled in a327b98 are separate commits; the benchmark follows this one.
Co-authored-by: Codex <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Folder listings arrive from background threads that Headless::busy() does
not track, so a fixed number of frames was not always enough under
full-suite load. Two tests failed intermittently and passed alone:
kept_local_root_stays_while_browsing_below_and_elsewhere and
local_folder_tree_expands_and_browses. They now wait for the expected
widget (wait_for_widget, test-only) and use their temporary folders as kept
roots. The kept-root test hides Home (in the test's in-memory session) so
the tree does not depend on the user's real folders.
local_location_can_be_hidden_and_restored now hides a kept root, reached
via "All Photographs". It no longer covers hiding a browsed temporary
folder: hiding a folder beneath Home expands the whole parent tree and
pushes the restore button out of view. That UI bug is not fixed here.
Fixture changes split from a327b98 (T0 thumbnail work), plus follow-ups.
Co-authored-by: Codex <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Catalog paths may mix `/` and `\` on Windows, so raw string comparison saw
`N:/photos/IMG.jpg` and `N:/photos\IMG.jpg` as different files:
- Rename: planning, the move list, move_file_with and file_moves now compare
as paths. A separator-only "rename" is a no-op: no file move, no catalog
change and no undo entry, alone or batched with a genuine rename.
Previously it attempted a plain rename onto the existing file, failed, and
rolled back the rest of the batch. Case-only renames still work (#95).
- Import: the already-in-library index is keyed by PathBuf.
- Tests build expected paths with Path::join, and the CLI test escapes the
device path with serde_json instead of string formatting.
Split from a327b98 (T0 thumbnail work), with the follow-up rename fix that
the committed version needed.
Co-authored-by: Codex <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Non-JPEG thumbnail fallbacks decode at source resolution before resizing, so
the output size does not bound memory. ThumbnailOptions gains max_pixels
(default 64 MP, previously about 1 GP via DecodeOptions::fit); larger sources
return Error::TooLarge. Nothing in the workspace calls decode_thumbnail yet.
WebP RIFF chunk lengths are untrusted: set_memory_limit(bytes.len()) stops a
tiny truncated file requesting gigabytes for metadata. In image-webp 0.2.4
the limit applies only to ICC, EXIF and XMP chunks. Fuzz tests cover both;
the roundtrip test only switches to struct update syntax.
Split from a327b98 (T0 thumbnail work), where it was bundled.
Co-authored-by: Codex <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The lens + perspective warp classified source-edge pixels in f32 on the GPU.
A position 0.000035 px outside the edge rounded inside, so the GPU sampled
photo pixels where the CPU reference draws blank canvas (geometry_variants,
"embedded lens + perspective", max 91 LSB). The CPU now evaluates the f64
framing decision into a one-bit-per-pixel mask; sampling and colour stay on
the GPU. The embedded-lens cases move to their own test,
embedded_lens_perspective_edges_match, with the same tolerances.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A full-debuginfo workspace test build with one rustc/link.exe per CPU peaked
at 14-20 GB commit on a 32-thread Windows machine before any test ran; the
test binaries themselves peak at about 3.3 GB each. `cargo xtask ci` now sets
CARGO_PROFILE_DEV_DEBUG=line-tables-only, CARGO_BUILD_JOBS=4 and
RUST_TEST_THREADS=4 unless they are already set. Ordinary cargo commands are
unchanged; GPU coverage is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Looked at in the running app: without a configured mirror the dialog
offered Download / Try Again that could only fail; it now shows the
manual-install hint and Close (an agent confirming still gets the
reason). "A one-time download of about 3.4 GB, saved in <folder>"; a
cancelled download is reported as stopped (resumable), not failed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Engine (crates/engine/src/segment/):
- One worker thread owns the model and the encoded photo; the session
sends jobs (with the render job for the model input) and gets results
over channels, never locking what the model holds. In the app
(Segmenter::background) clicks, descriptions, prepares and detail
passes return at once and Session::segment_poll applies results (one
undo step per settled selection, pending clicks shown on the photo);
the CLI/MCP wait for them. Newer requests for the same selection
supersede queued ones.
- Every job runs under catch_unwind: a candle panic becomes an error and
the model is reloaded next time; busy counters are reset by drop
guards; a dead worker is restarted. The model unloads after 10 min idle.
- segment.model.status / .download {acknowledged: true} / .cancel: the
download runs on its own thread; status via atomics (try_lock only).
Without the model every AI request fails at once with "The SAM 3 model
is not installed…"; renders/exports never need it; `seg` params make
AI masks from stored segmentations without it.
- Hostile documents: at most 4 detail patches per component (capped on
deserialize and in the renderer), at most 64 Object clicks.
UI: Object / Describe without the model open "Download the SAM 3
Model?" (size, folder, SAM License link; Not Now / Download, progress,
cancel, Try Again with the reason, Continue once installed); progress in
the Masking panel when the dialog is closed. Reverted the PR's unrelated
default change (mask overlay off + hover-red) to main's behaviour.
tools/install-sam3.*: --fail so error pages never land in files, 416
restarts instead of looping, the weights are verified before they are
moved into place (a bad file is deleted, never resumed from).
Docs: docs/ai-masks.md (in-app download, mirrors, SAM License summary,
no weights in the repo), NOTICE, parity LR-MASK-OBJECTS, What's New,
ROADMAP.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The crate ports Apache-2.0-only Hugging Face transformers code: it is
now `license = "Apache-2.0"`, and every ported file carries the
modification notice Apache-2.0 §4(b) asks for.
- Weights::open checks every tensor's byte range (and the header)
against the file's length: a truncated or damaged download is an
error at load, never a short read later; regression test with
hostile offsets.
- fetch/: the model download, only when asked for. An ordered mirror
list (env LIGHTCRAFT_SAM3_MIRRORS, a mirrors file, DEFAULT_MIRRORS —
empty until LightCraft's CDN exists, TODO for the maintainer),
fallback to the next mirror, resume via Range, redirects, connect and
stall timeouts, cancel, size caps, exact size + SHA-256 (pinned from
tools/install-sam3.sh) before an atomic rename; damaged files are
deleted, HTTP error bodies never written.
- HTTP/1.1 client over std sockets + rustls with the RustCrypto provider
(rustls-rustcrypto) and webpki-roots: no ring / aws-lc-rs, nothing
compiled from C (no new `cc` users in the app's build graph).
- Tests against a local TcpListener server (no internet): fallback,
redirects, chunked bodies, resume, 416, hash mismatch, too much data,
stalls, cancel.
- *.safetensors are gitignored.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A built-in Sony ILCE-7M4 colour profile (assets/camera-profiles/ILCE-7M4.json, compact JSON, 39 KB), fitted by lightcraft-cli calibrate to the contributor's own 597 raw photos against their in-camera JPEGs; aggregate colour statistics only, MIT OR Apache-2.0. Compiled in via camera_profiles::BUNDLED; a local profile replaces it.
Co-authored-by: Harald Hoyer <harald@hoyer.xyz>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Conflicts with issue #150 (NEF colour from the camera JPEG): the per-file
look, its hue/saturation/value table and chroma curve, and camera
profiles now apply to every `file_local_look` format (ARW, NEF, NRW);
`calibrate` also collects NEF/NRW. Render cache version 11 (main 9,
branch 10).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
One photo can show too little of a colour for its own fit: a lime shirt
covering a few dozen proxy pixels next to foliage at the same hue stayed
yellow (hue 68 deg vs the JPEG's 82 deg).
- `lightcraft-cli calibrate [--max N] [--out DIR] FOLDERS...` pools the
colour pairs of many ARWs per model (each against its embedded JPEG,
192 px proxies, <= 4000 pairs per photo) and fits one matrix from
white-balanced camera RGB plus a hue/saturation/value table.
- Profiles live in $LIGHTCRAFT_CAMERA_PROFILES or
<config>/camera-profiles/<model>.json; validated on load, damaged files
ignored with a warning, part of the render cache keys.
- Photos of a profiled model take its colour and fit only their own tone
and chroma curves; the acceptance gates still apply.
- The table fit buckets samples by hue and runs nodes in parallel; the
app's config_dir moves to the engine.
Profile from 300 Feb-Aug ILCE-7M4 photos (2 min); 29 held-out September
photos: mean dE vs camera JPEG 3.54 -> 3.10, saturated 6.41 -> 5.20;
the shirt's hue 68 -> 82 deg (JPEG 81.5).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Port of PR #198 onto the language table:
- Traditional Chinese (Taiwan) catalogs as a table entry: code zh-hant, name
繁體中文（台灣）, script Hant; zh-TW / zh-HK / zh-Hant tags reach it; command
app.language.traditionalChinese
- display-only helpers: builtin_label (stock preset/profile names), history_label
(generated history steps, user preset names verbatim), source_label (library
headings, album names verbatim), date_group_label / display_time
- menu, dialog and panel strings (profiles, presets, curves, crop, versions,
history, grid, date sidebar, About links) go through the catalogs
- a language command and ui.set apply the language at once; a native-menu test
checks that every language switch rebuilds the menu titles
Adapted while porting (Claude):
- date headings and capture times are localised for every language from its
*-formats.json date patterns, not only for zh-TW; ja and zh-hans gain the date
patterns, weekday names and "Unknown Date", and zh-hans the new menu messages
(ja's come from the PR)
- plain catalogs may cover different messages (English fallback); the test
reports the gaps instead of failing. The test requiring every engine command
label to be translated is dropped
- preset groups stay grouped by name: a user preset in a stock-named group is
not split into a separate group (the group header is translated when it holds
stock presets); the PR's split was a side effect of labelling
- no Traditional Chinese face in craft-fonts yet: Hant prefers the Simplified
Chinese face (which carries the Traditional characters) over the Japanese ones
- zh-hant translations for the messages main added since the PR branched
Co-authored-by: Codex <noreply@openai.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Regions are stored on the upright (EXIF-oriented) photo, but the loupe's normalized frame and the
close-up crop include the user's orientation. The loupe overlay now maps each box into that frame
(and a resize drag back out of it), and face_job maps the face and swaps the photo's size for a
quarter turn. Flips need nothing: they are applied after the normalized frame. Adds
Orientation::inverse and Orientation::map_norm_rect.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- the CJK fallback order follows the UI language's script, but the fonts were
only installed on the first frame: switching 日本語 ↔ 简体中文 at runtime kept
the previous language's Han forms. The app remembers the language its fonts
were built for and reinstalls them when it changes.
- the native menu bar rebuilds on any language change (keyed on the language
code, not on "uses CJK fonts", which is true for both CJK languages)
- restore the tests the language-table change dropped, for every language:
core catalog entries, user-named menu items stay untranslated, the window
paints each language and menu ids don't change with it, CJK glyphs in every
UI family, and the UI runs without craft-fonts
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Re-reading a sidecar whose region list was emptied by a region-aware app now clears the stale
regions. A sidecar without mwg-rs:Regions at all (an app that doesn't do regions) says nothing about
them and the photo's regions are kept.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- UI_COMMANDS still listed app.language.english/japanese next to the new
LANGUAGE_COMMANDS table, so Edit > Language showed English and 日本語 twice
- LANGUAGE_COMMANDS is `pub const`, the form xtask parity scans for; a test
reads the real menus.rs and checks every language id is found once
- the Simplified Chinese command is app.language.simplifiedChinese, in line
with the other camelCase command ids (the settings code stays zh-hans)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Restore main's compile-checked formatting, generalised to N catalogs: build.rs
writes one tr_format! arm per English message that matches on the active
language and calls format! with that language's translation (English is the
fallback arm). The runtime substitution dropped format specs ({:.1}, {:+},
{:.0}%), e.g. the smart-preview size and the merge progress percentage.
- formats catalogs use main's convention again ({0}/{1} for reordering, {:.0}
to consume an unused plural suffix); zh-hans converted, with the size and
percentage placeholders fixed by hand
- build.rs: auto-discovers locales/*-formats.json, validates the code
(^[a-z]+(-[a-z]+)*$), fails when the catalogs' keys differ, and reruns
when the locales directory changes
- zh-hans: translate the strings main added since the PR branched
- test: precision, sign and reordered positional values survive in every
language
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A region parsed from mwg-rs:Regions is now intersected with the 0..1 frame and dropped when it
misses the photo, so a hostile or stale sidecar (x=1e5, a 1e5-wide box) can't park a box far
outside the image. Lightroom's mwg-rs:Rotation of ±π (Exif orientation 3) is mapped to the upright
frame like the quarter turns; mirrored orientations can't be expressed in MWG and are taken as
written.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
radialmonster
Brandon Thomas
Harald Hoyer
sakamoto824