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>
Treat a zero-width feather transition as a hard edge so exact-radius pixels remain unchanged, including at zero opacity. Add actual image regressions for all spot modes, image corners and two clone output sizes.
Validation: both new regressions failed on the original implementation and pass here. Formatting, workspace Clippy, HEIF and a generated-photo app smoke test pass. The unfiltered workspace gate remains blocked by the existing all-masks overlay test tracked by #1599; later parity, layers, assets and WASM stages are pending. This is a source checkpoint, not a green release candidate.
Contributor: @eliets2
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>
Conflicts in ROADMAP.md, docs/gaps.md, docs/parity-checklist.md and docs/raw-parity.md resolved by
taking upstream's text and re-applying this PR's rows; coverage re-measured with docs/raw/coverage.py
(334 verified, 844 unverified, 50% of photographers on a verified body).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Coverage regenerated with docs/raw/coverage.py (258 verified, 920 unverified);
doc conflicts resolved on upstream's text with this PR's numbers re-applied.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Keeps upstream's ColorData 66 handling (white level from the data); recounts
coverage (259 verified, 919 unverified, 232 preview); drops the duplicate
raw-parity rows; notes high-QP R6 Mark II C-RAW and unverified R7 C-RAW.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
As over Lightroom's values, a press on the number next to a slider's name
that moves sideways now changes the value as dragging the thumb would: the
pointer's travel over the track's width, a tenth as far with Shift, on the
control's steps and range, as one undo step. A press that moves less than
3 points sideways is still a click that opens the value for typing
(KEY-SLIDER-TYPE), and the pointer over the value is the horizontal-resize
cursor. The tooltip says both ("Click to type a value, drag to change it"),
translated in every catalog.
Tests: scrub distance and direction, Shift, clamping, steps, one interaction
per scrub, no field or draft left behind, a jittery click still types, a
disabled slider doesn't scrub. The readout test that drags away from the
value now drags straight down, so it still checks that a cancelled press
keeps neither a value nor the keyboard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brandon Thomas