Make Liquify redraw the brush, not the layer

[IAmJSD]

Liquify felt laggy because every pointer move re-warped the whole layer: flatten the entire snapshot to f32, resample all of it, clone the tile map, write every tile back, and damage the layer rect so the compositor recomposited the canvas. On a 12-megapixel layer that was ~1 s per mouse move, and Enter cost another 1.3 s.

What changed

• Render the dab's footprint, not the layer. A dab only moves the mesh vertices under the brush, and a destination pixel interpolates the four vertices around it, so nothing outside radius + one cell can change. Mesh::dab_rect returns that rect; warp_into resamples into the layer's own tiles instead of rebuilding them from a clone; damage shrinks to the same rect, so the compositor invalidates a handful of tiles.
• Crop what the kernel is handed. src_token != 0 still means "the whole snapshot, keep it resident" — Puppet Warp's case, and the one an accelerated backend can use. No token now also means the plane is cropped to what the displacement over the region can reach. Mesh::subgrid does the same for the offsets, which were a megabyte and a half copied per pointer move.
• Commit stops re-sweeping. The layer already holds the finished warp, so Enter moves the tiles out and puts the snapshot back. Tiles no dab reached are still Arc-shared with the snapshot and replace_layer_tiles diffs by pointer, so undo history carries only what moved.
• The conversions either side of the kernel, which dominate a big dab once the sweep is gone: flatten goes tile-wise and parallel instead of a map lookup per pixel, warp_cpu puts its rows across the cores (both also help Puppet Warp and any GPU-declined fallback), and picking the tool up takes tile-granular bounds rather than scanning twelve megapixels of alpha.

Numbers

cargo run --release -p schist-tools-warp --example liquifybench (new):

layer	brush	pick up	pointer move	commit
1024×768	100	3.8 → 0.0 ms	50.7 → 0.9 ms	52.4 → 0.2 ms
4000×3000	100	51 → 0.1 ms	1045 → 0.8 ms	1356 → 1.0 ms
4000×3000	1000	97 → 0.3 ms	1318 → 21.5 ms	1320 → 0.3 ms

Testing

tests/incremental_warp.rs holds the invariant the speedup rests on: a stroke rendered dab by dab is pixel-identical to one sweep of the final mesh. Plus a dab touching nothing outside its footprint, a cropped plane reaching as far as the displacement does, and tool-level tests for damage size, commit/undo and Escape. Workspace tests and clippy are clean, and the GPU warp-parity tests still pass.

Also driven end to end through the headless MCP server on a real image — smooth warp, no seams or stale patches, undo restored the file byte-identically.

Notes

• Liquify no longer goes near the GPU warp; a dab is far below what a round trip pays for. The comments naming it as the resident-source client now name Puppet Warp, which still is.
• Pre-existing edge case, left alone: if another edit changes the layer while a Liquify session is open but idle, the snapshot goes stale. Before, the next dab reverted the whole layer; now it reverts only the dab's footprint.

🤖 Generated with Claude Code