The 2^28 pixel limit applied per image; one composite or layer decode kept
the ZIP output, a Vec<f32> per channel and the result alive together, and the
bake size came from the file's quad. A decode now charges every buffer
against a byte budget (4 GiB, 1 GiB on 32-bit), checks its whole expected peak
before allocating, writes channels straight into the result and returns an
error when over budget. Bakes are at most 4x the document across and 2^24
pixels.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Raw channels are borrowed, ZIP planes are un-predicted in place in the inflated buffer,
the 32-bit un-predictor uses one scratch row, the layer mask tests the rect per row
instead of per pixel (and skips a constant factor of 1.0), and a placed-layer bake
solves the quad once and hoists per-triangle invariants. Output is bit-identical;
tests compare against the replaced implementations and count the avoided work.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
One decode_plane for raw/RLE/ZIP (decode_channel and composite), placed::eight and
quad_from for the SoLd/PlLd transforms, the writer's big-endian and UTF-16 helpers
from descriptor.rs, one delimiter const in engine_data.rs, and the dead binding in
write.rs. No behaviour change; adds ZIP and RLE composite tests.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
nonAffineTransform and the warp descriptor (named styles with bend / distortion, custom
envelope quilt meshes via ObAr meshPoints) are baked as placed over the placed bounds
(SourceLayer::placed), at up to 4x resolution.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brandon Thomas