Claude: Optimize SVG path rasterization (~10% faster)

[treeform]

Summary

Speeds up the tests/bench_svg.nim Ghostscript Tiger render from 6.89 ms to 6.16 ms min (~10.5%, measured with interleaved A/B runs against a master-built binary on an M-series Mac, -d:danger).

All changes are in src/pixie/paths.nim and keep output effectively identical — every test_svg/test_paths xray diff score matches master to 4 decimal places (remaining differences come from replacing divides with multiplies by the reciprocal, well below the masters' own tolerance).

Changes

• Delta + prefix-sum coverage accumulation. Instead of adding sample coverage to every covered pixel for each of the 5 AA subsample lines, each fill span now writes 2 entries into an int16 delta buffer. One prefix-sum pass per scanline (NEON-vectorized on arm64) converts deltas to the exact same coverage bytes. Long interior spans go from 5 passes over every pixel to O(1) writes + 1 pass.
• Single-pass hit gathering. computeCoverage previously rescanned all active entries once per subsample line (5x). Entries are now loaded once, their fields kept in registers, and hits bucketed per subsample line — same hit order, bit-identical sort/walk behavior. On the tiger this cuts 467k entry visits per render to 93k.
• Per-scanline coverage regions. The covered spans are tracked while walking hits, so blending only touches real coverage (small gaps fused) instead of the full path-bounds width, and the per-scanline zeroMem of the coverage buffer is gone (deltas are provably re-zeroed by the prefix pass).
• PartitionEntry stores 1/m, turning the per-hit divide in solveX/hit gathering into a multiply.
• Partition clipping via the precomputed slope instead of generic segment x line intersects().
• Thread-local scratch reuse. fillShapes and partitionSegments reuse grow-only thread-local buffers (coverages, deltas, hits, buckets, partitions + entries, segments), removing a few thousand allocations per tiger render. Partitions gained a numEntries count so entry storage can be reused without reallocating.
• Blend micro-opts. The blender() proc-pointer lookup is hoisted out of the trapezoid pair loop and blendNormal is called directly for the common NormalBlend case in the edge-coverage loops.

Testing

• test_svg, test_paths, test_images_draw, test_paints, test_contexts, test_fonts all pass; xray scores unchanged vs master.
• Benchmarked with interleaved master/branch runs to control for machine load.

🤖 Generated with Claude Code