Date: 2026-04-05 Branch: wgpu-device-renderer Last commits: P11 commits + uncommitted P12 changes

Current (P12): 323/413 pass (78.2%) — 90 failures Previous (P11): 309/413 pass (74.8%) — 104 failures Net change: +14 tests passing (+3.4 percentage points)

backface (10/10), crash (2/2), scrolling (26/26), snap (3/3), split (16/16), tiles (5/5), text (48/48), transforms (23/23)

Two new categories reached 100% this session.

Files: renderer/mod.rs

Root cause: When compositing picture cache tiles onto the final surface, the wgpu backend passes CompositeInstance structs to the composite shader. Each tile may have a clip_index referencing a CompositorClip (a rounded rect that clips the tile during compositing — used for CSS border-radius on the root element or picture cache surfaces).

The wgpu compositor path at render_composite_instances_to_view() was passing None as the clip parameter to CompositeInstance::new() for ALL tiles, completely ignoring tile.clip_index. This meant the rounded_clip_rect and rounded_clip_radii fields in the instance data were always zeroed, and the composite shader never applied any rounded clip.

The GL path correctly resolves tile.clip_index via composite_state.get_compositor_clip(index) and passes the clip to the instance constructor.

Fix (Part 1 — pass clip data):

let compositor_clip = tile.clip_index.map(|idx| {
    composite_state.get_compositor_clip(idx)
});
// ... in CompositeInstance::new():
compositor_clip,  // was: None

Fix (Part 2 — use correct shader variant): The composite shader has two variants:

• CompositeFastPath (FAST_PATH,TEXTURE_2D): Skips UV clamping and rounded clip evaluation. Used for unclipped tiles.
• Composite (TEXTURE_2D): Full path with vColor modulation, UV bounds clamping, and SDF-based rounded clip evaluation.

The wgpu backend used CompositeFastPath for ALL texture-backed tiles. The #ifndef WR_FEATURE_FAST_PATH block in composite.glsl (lines 227-242) contains the rounded clip logic — never executed in the fast path.

Fixed by splitting textured tile batches into two groups:

• Unclipped tiles: continue using CompositeFastPath
• Clipped tiles (with compositor_clip.is_some()): use Composite

Impact: Fixed rounded-rect clipping at the compositor level. This is the clip path used when CSS border-radius applies to picture cache tile compositing, not to individual primitive rendering.

Direct fixes:

• conic-color-wheel: max diff 255 -> 4 (circular clip now applied)
• radial-border-radius-large: max diff 255 -> 4 (rounded clip now applied)
• clip-empty-inner-rect: now passes (was failing)
• compositor/rounded-corners-3: max diff improved to 1, 2 pixels

Files: renderer/mod.rs

Added clip_mask texture binding from batch.key.textures.clip_mask to the TextureBindings struct in all three draw locations:

1. record_alpha_batch macro (all_targets / color target path)
2. record_batch macro (picture_cache path)
3. MixBlend pass 2 standalone draw

This is technically correct for brush-path batches that use per-batch clip mask textures, but currently all observed batches have clip_mask=Invalid because quad-path primitives handle clips through QuadRenderStrategy (Tiled/Indirect/NinePatch) rather than per-batch clip textures.

Removed all eprintln! debug statements:

• [PC-O] and [PC-A] prints from picture cache batch loops in mod.rs
• [PQ] print from prepare_quad() in quad.rs

Impact: The [PQ] print was firing for every prepare_quad() call, producing enormous stderr output. This was causing 13 spurious test failures (11 text, 2 transforms) due to I/O contention or timing effects. The clean run without debug prints showed these tests passing.

Clips for Quad(*) primitives (ConicGradient, RadialGradient, etc.) are NOT handled via per-batch clip_mask textures. Instead:

1. Frame building (prepare_quad in quad.rs): get_prim_render_strategy() examines the clip chain and returns one of:

   • Direct: No mask needed (clip_chain.needs_mask == false)
   • Indirect: Render to intermediate texture, apply mask via MaskSubPass
   • Tiled: Decompose into grid tiles; masked tiles rendered indirectly
   • NinePatch: Nine-patch decomposition with rounded corner segments

2. Indirect/Tiled tiles: Rendered as render tasks via add_render_task_with_mask(), which creates a MaskSubPass. The mask is applied by ps_quad_mask / ps_quad_mask_fast_path shaders during build_sub_pass().

3. Picture cache compositing: Clips at this level use CompositorClip (rounded rect with radii), applied by the composite shader's SDF-based clip evaluation. This is what was broken and fixed in this session.

The test has a clip element with radius: 300 on a 300x300 rect. The NinePatch strategy check requires max_corner_width <= 0.5 * rect.width (300 <= 150 = false), so NinePatch is rejected. The primitive size check for tiling requires size > MIN_QUAD_SPLIT_SIZE, but in this case clip_chain.needs_mask is false because the clip is applied at the picture cache surface level, not the primitive level. The clip appears as a CompositorClip on the tile, not in the primitive's clip chain.

All are dithering/precision issues with max diff 3-4. Two exceptions:

• gradient_cache_clamp (max 255): state leakage bug (passes in isolation)
• conic-large-hard-stop (max 255): texture cache tile leakage

• Backdrop filter: 7 tests (max 255), need two-pass approach like MixBlend
• SVG filters: 6 tests (max 255), may need filter graph processing
• Other: blur chain, scaled blur, component transfer (max 1-147)

External surface compositing issues.

• clip-mode, clip-ellipse, clip-corner-overlap: max diff 1 (AA precision)
• stacking-context-clip: ColorTargets check failure (not image-based)
• raster-roots-tiled-mask: max diff 128 (likely rendering bug)

Snapshot tests (3), tile-with-spacing, segments, rgb_composite.

• rounded-corners-1: ColorTargets check failure
• rounded-corners-3: max diff 1, 2px (nearly passing)
• rounded-rgb-surface, tile-occlusion: max diff 255

boxshadow (3), mask (4), blend (3), border (2), aa (1), performance (1)

1. Backdrop filter — 7 tests. Similar to MixBlend two-pass rendering. Currently the highest-impact category to fix.

2. SVG filters — 6 tests. May need additional shader variants or filter graph processing.

3. State leakage — investigate gradient_cache_clamp failure in full suite.

4. Text shadow rendering — all text tests pass now, but some were marginal. Monitor for regressions.

5. C4 render pass sharing — optimization plan at logical-whistling-mochi.md

• render_composite_instances_to_view():
  • Resolve tile.clip_index via composite_state.get_compositor_clip()
  • Pass compositor clip to CompositeInstance::new() (was None)
  • Split textured tiles into unclipped (CompositeFastPath) and clipped (Composite) batches
  • Updated diagnostic logging to include clipped tile count
• record_alpha_batch macro: Added clip_mask texture binding
• record_batch macro: Added clip_mask texture binding
• MixBlend pass 2: Added clip_mask texture binding
• Removed [PC-O] and [PC-A] debug prints

• Removed [PQ] debug print from prepare_quad()