feat: Vello PTCL per-draw blend modes (CmdSetBlend)

[kolkov]

Vello Compute Pipeline: Per-Draw Blend Modes (CmdSetBlend)

Depends on: #439 (per-operation blend in SoftwareRenderer)

Currently, the Vello 9-stage compute pipeline (internal/gpu/tilecompute/) only supports blend modes at layer boundaries (BeginClip/EndClip). Individual fill/stroke commands are hardcoded to SrcOver. After #439 enables per-operation blend in the CPU rasterizer, the Vello compute path becomes the remaining gap.

Architecture

A new PTCL command CmdSetBlend (tag=14) will set the blend mode for subsequent color/gradient/image application commands:

CmdFill(...)        → coverage (unchanged)
CmdSetBlend(blend)  → set blend mode for next apply
CmdColor(rgba)      → apply with current blend

Zero overhead for SrcOver — CmdSetBlend is not emitted for default SrcOver draws.

Pipeline Changes (4 stages)

Stage	Change
Scene encoding	New DrawTagColorBlend (scene_size=2: color + blend)
Stage 4 (draw_leaf)	Handle new tag
Stage 7 (coarse)	Read blend, emit CmdSetBlend
Stage 9 (fine)	currentBlend state + blendMixCompose() dispatch

Research

Deep analysis of Rust Vello architecture complete. Key findings:

• No per-draw blend proposal exists in linebender/vello — our implementation would be first
• Vello fix: update wgpu v0.19.3, text fixes, coverage tests (v0.32.3 + v0.32.4) #151, fix: path clipping to prevent FDot6→FDot16 integer overflow (#148) #157: blend deliberately layer-only ("builds on clip infrastructure")
• Vello #1653: u8 fast path for blend (4-5x NEON speedup) — reference for our impl
• blend_mix_compose() in blend.wgsl has SrcOver early-exit → minimal perf impact
• PTCL is workgroup-uniform → zero GPU branch divergence

Conflation Artifacts (Known Trade-off)

Per-draw blend without layer isolation produces incorrect results for overlapping geometry with non-SrcOver modes (blend applied twice to overlap). Same limitation as Skia Graphite. For correct results: use PushLayer/PopLayer. Documented trade-off.

Phases

1. CPU path (fine.go + coarse.go + ptcl.go) — after fix: SetBlendMode() is a no-op — per-operation blend modes not wired #439
2. GPU shaders (coarse.wgsl + fine.wgsl)
3. Context integration (auto-dispatch)