An accelerator hook in the face interpreter takes Conv nodes; lightcraft-gpu implements it with a WGSL
implicit-GEMM convolution (any stride, padding, grouping; split-K for small maps; a direct kernel for depthwise
layers), each model's weights resident on a device of its own. After loading, detection and recognition check
the GPU in the background (agreement within 1e-3 and faster than the CPU) and otherwise stay on the CPU, which
remains the reference; the first GPU error drops it. Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Docs recounted with docs/raw/coverage.py: 241 verified, 35 unsupported. The MCP
headless extension list now comes from engine::import::EXTENSIONS upstream, which
already carries mef.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Keep main's spots.rs implementation and preserve the catalog reader and
updated capability documentation. Readers are byte-identical to the PR.
Validation: 48 catalog and 9 spot tests pass on candidate and main.
Formatting, Clippy, parity, layers, assets and WASM checks pass.
Full CI remains blocked by the same mask-overlay test failure reproduced
on main; no assertions or tests were changed to hide it.
Refs #798.
Fitted with `lightcraft-cli calibrate` to 144 usable ILCA-99M2 ARWs (of
150, seven shooting days, Standard creative style, DRO Auto) and their
in-camera JPEGs; aggregate colour statistics only.
On 30 withheld photos (profile fitted to the other 120) mean dE76 against
the camera JPEG goes 5.11 -> 4.87; 25 of 30 improve, the worst
regression is 0.24. All 150 photos stay accepted. The public raw.pixls.us
ILCA-99M2 file, not used for fitting, stays accepted (5.57 -> 5.73).
A corpus test checks that this public sample's fit keeps exactly the
built-in profile's matrix and table; the ILCE-7CR test now shares its
body. LOOK_VERSION 5 -> 6, as the profile changes the fit of this model.
Setting up Synchronize Folder, an import review or an import made a table
of every photo in the library on the session's thread before the first
file was looked at, and naming one photo's sidecar worked out who owns
every shared sidecar name in the library. In a library of 3.5 million
photos (`scan_setup_bench`, release, a folder of 500): 5.8 s before a
Synchronize scan, 2.6 s before an import review, 1.7 s before an import,
2.4 s for each sidecar written.
- A scan keeps the library as it is (photos are shared, not copied) and
looks up the files it finds (`library_files`) and the contents it probes
(`library_hashes`) on its own thread.
- An import looks the library up with its first files
(`ImportJob::look_up_library`), from the library and the options as they
were when it was set up.
- The owners of shared sidecar names are worked out for the photos asked
about (`StemOwners::among`): all five callers name only those photos.
- Synchronize Folder names its photos' sidecars in the scan, after it has
seen that the folder is there.
Now 0.9 s, 0.05 s, 0.05 s and under 1 s. What is found, and how sidecars
are named, is unchanged: the tests hold the owners against the table of
the whole library as it was made before, and the key made without
allocating against the key, for every string of up to six of the
characters that decide it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brandon Lim