Preserve filter lookahead alongside main image-cache invalidation and input
bounds. Move HeldFields into its own module and finish scoped cleanup.
Charge repeated USD path names before allocating retained strings.
Reuse the blur transpose-back buffer and release scattered texels before
allocating filter temporaries.
Validated compose/engine release tests, Clippy, layering, wasm and formatting.
Preserve main's decoder, atomic-save, material-payload, and USDC limits.
Bound material reads by the remaining payload budget and trim retained
buffers. Replace quadratic image-name searches with map lookups, and
cache validated linked-material maps before synchronization so they
are not decoded twice.
Add import/undo/redo and payload-boundary regressions, extract
ImageIdentity, remove the unchecked scratch example, and complete
scoped cleanup. Correct the integrated style guide's product name.
Validated 546 release tests, Clippy, layering, wasm, formatting, and
focused allocation and lifetime probes.
The PATHS table charged 64 bytes per entry but allocates Option<PathEl> (about 80), letting a small declared count exceed the decoded-byte budget, and had no entry ceiling. Charge the real element size and cap the table at four million paths.
Imported images still used the 16384-pixel edge limit with no texel ceiling, so a small 16k square PNG expanded to multi-gigabyte RGBA f32 buffers. Default resource decoding now refuses more than 8192 square texels from the header, and the engine import path uses the same budget.
Fields kept for the coming composites stayed allocated for as long as a
stroke lasted while tiles still followed, because the end-of-pass cleanup only
runs when none do. A stroke already kept them from being used; now it also
drops them. The test keeps fields first, then starts the stroke.
A stack with a neighbourhood filter (blur, high pass, bevel, glow...) made a
recomposite spread over frames many times slower than the same recomposite in
one call: every chunk filtered its own surroundings again, and on a machine
with many threads the filter's per-texel parallel steps cost more to hand out
than to run on a region of a few tiles.
- Neighbourhood filters split their per-texel work into tasks of at least
16384 texels. Each texel is computed on its own, so results do not change.
- Session::update_for tells the compositor which dirty tiles follow the chunk
in its block (up to 64). The compositor filters the chunk and those tiles at
once and keeps the fields for the composites that follow, compared by
everything the stack reads, and drops them at the end of the pass, after an
edit, and under a live stroke.
Results are identical to the one-call composite: tests cover cold, warm and
partly evicted hot-layer caches, an edit mid-pass and a live stroke.
perf::bench_sliced_filters_4k times both paths.
* fix(formats): stop reading Alembic indexed metadata after 255 entries (#447)
Objects and properties refer to indexed metadata with one byte, and 0xff already means the metadata is stored inline, so only entries 0 to 254 can ever be used. The reader kept every entry, so a block of zero bytes cost a whole String in memory for each byte of file. It now stops once 255 entries are held.
* test(formats): verify indexed metadata during archive import (#447)
* Import glTF scenes that have no nodes array (#438)
The gltf crate requires scene.nodes, but the glTF 2.0 schema lists no
required properties for a scene. Blender writes such scenes for scenes with
no exported objects, so a file whose default scene was fine failed with
'missing field nodes'. Treat a missing array as empty before the document is
built.
* Test GLB import with an empty scene
---------
Co-authored-by: FredApps <FredApps@users.noreply.github.com>
* fix(engine): reset paint target to content when undoing mask (#307)
* fix(engine): resolve mask_owner in after_history and clear below cache in test (#307)
* fix(engine): resolve resource stem and list available names in layer.add.image (#337)
Accept an unambiguous file stem in layer.add.image alongside the full resource name with extension. When the requested resource is not found, list the available resources or guide the user to resource.import.
* fix(engine): filter image lookup stems and hints to decoded images (#337)
* fix(engine): reject unparseable colour parameters with an error (#369)
When a colour parameter cannot be parsed, fail with an error describing the expected color format instead of silently falling back to defaults.
Update color_param to return an explicit error and propagate across commands that accept color values.
* fix(engine): validate f32 overflow, pre-parse state mutations, and keep UV pairs strict (#369)
* fix(flatpak): allow release checks and self-update helpers
* test: keep Flatpak permission parsing within strict Clippy rules
* chore: bump version to 0.6.1 for the Flatpak update fix
* fix(flatpak): honor the launcher version choices
release.yml also runs nightly on main with craft-common-actions' dry-run: every release build
except macOS, with nothing signed, pushed or drafted. Breakage that only the release build hits
(riscv64, Flatpak, FreeBSD, packaging) then shows up the next morning instead of on release
day, as it did on 11 Nov 2026.
Previously, if any object in an OBJ mesh lacked texture coordinates, all UVs across all objects were discarded and replaced with a full automatic unwrap.
Skip empty models without indices, preserve existing UVs on textured objects while assigning default UVs to untextured helper/sub-objects, and report a warning without forcing a full mesh unwrap.