refactor: delegate scene.Renderer rasterization to gg.SoftwareRenderer

[kolkov]

Summary

scene.Renderer currently has its own simplified CPU rasterizer that was written when the scene package was first created — at that time, gg.SoftwareRenderer with analytic AA and the GPU acceleration pipeline (GPUAccelerator, SDF, stencil-then-cover, MSDF text) did not yet exist. The rasterizer was a minimal implementation to get the scene graph working.

Now that gg has a full-featured rendering infrastructure — analytic AA, proper stroke expansion with caps/joins, GPU acceleration, premultiplied alpha compositing — the scene renderer should delegate pixel rendering to gg.SoftwareRenderer instead of maintaining its own incomplete rasterizer.

Reported by contributor in #116: white background destroyed, circle stroke invisible.

Current Problems

The historical rasterizer has 6 known issues:

Bug	Location	Description
No anti-aliasing	renderer.go:527-538	Binary winding number test, jagged edges
Stroke ignores curves	renderer.go:572-594	Only handles LineTo, silently drops CubicTo/QuadTo
Background destroyed	renderer.go:384	clear(tile.Data) wipes user's pre-cleared background
No alpha compositing	renderer.go:687	copy() instead of source-over blending
Layers/clips stubbed	renderer.go:459-474	PushLayer/BeginClip are no-ops
Only solid brushes	renderer.go:485	Returns immediately for non-solid brushes

Architecture Decision

Research of 4 enterprise scene graph systems (Qt Quick, Skia/SkPicture, Vello, Flutter/Impeller) confirms a universal pattern: the scene graph layer handles orchestration (recording, tiles, dirty regions, parallelism), while pixel rendering is always delegated to an immediate-mode backend. None of them reimplement rasterization in the scene graph layer.

Approach: Replace the rasterizer, not the architecture.

• Keep: TileGrid, WorkerPool, DirtyRegion, LayerCache, Encoding/Decoder — all well-designed
• Replace: fillPathOnTile, strokePathOnTile, compositeTile — delegate to gg.SoftwareRenderer

This also opens the path to GPU acceleration in the scene renderer — since gg.Context already routes through GPUAccelerator when available.

Implementation Phases

• Phase 1: Core delegation — Replace fill/stroke/composite with SoftwareRenderer delegation (~200 lines)
• Phase 2: Pool optimization — sync.Pool for per-tile SoftwareRenderer/Pixmap reuse (~100 lines)
• Phase 3: Layer & clip support — Implement PushLayer/PopLayer, BeginClip/EndClip (~300 lines)
• Phase 4: GPU acceleration — Route shapes through GPUAccelerator when available (~150 lines)
• Phase 5: Image & gradient — TagImage rendering, gradient brush support (~200 lines)

Zero API Changes

All modifications are internal to renderer.go. Public API (NewRenderer, Render, RenderDirty, Stats) remains unchanged.

References

• Internal research: SCENE-RENDERER-DELEGATION-RESEARCH.md (covers Qt/Skia/Vello/Flutter delegation patterns)
• Contributor report: SceneBuilder.WithTransform causes invisible rendering #116 (comment from 2026-02-23)

[kolkov]

Phase 1 (core delegation) completed and released in v0.29.4 (PR #125).

All 6 bugs from the original report are fixed:

1. No AA → analytic AA via SoftwareRenderer
2. Stroke ignores curves → full CubicTo/QuadTo support
3. Background destroyed → preserved (no more clear(tile.Data))
4. No alpha compositing → premultiplied source-over blending
5. Layers/clips stubbed → deferred to Phase 3 (SCENE-003)
6. Only solid brushes → deferred to Phase 5 (SCENE-005)

Remaining phases (pool optimization, layer/clip support, GPU acceleration, image/gradient) tracked in internal kanban.