WebGL2's mip chains keep alpha coverage by the same rule as WebGPU

[pasquelin]

Why

Split from #748 (635 lines): WebGL2's mip chain builder applies the coverage-preserving alpha rule after #748 brings it to WebGPU.

To do

• WebGL2 keeps coverage: the WebGL2 float-blend chain builder applies Runtime mip chains keep alpha coverage on WebGPU and WebGL2 #748's rule, with the glTF effective cutoff (alphaCutoff ÷ factor alpha).

Code context

The rule and the WebGPU builder #748 lands; the textures lead's local branch (WebGL2 builder, ~215 lines).

Proof

Coverage within ±2.5 % at every level on sponza's masked maps and leaves-cut-by-alpha on WebGL2; the measurer's branch proof.

Links

Blocked by #748. Split from #748.

[pasquelin]

To-do added by the lead (found in #748): WebGL2 multiplies by the surface opacity only on transparent surfaces (packages/sdk-browser/src/webgl/cluster/materialBinding.ts:73), so a MASK surface under a baseColorFactor alpha below 1 is cut raw on WebGL2, unlike glTF 2.0, #44's cutoff byte and WebGPU after #748. Cut opacity × texel alpha × vertex alpha against alphaTest on every masked surface, test failing on develop. The GLSL builder is kept on local branch 769-webgl-coverage-mips in the textures lead's checkout.

[pasquelin]

taken by lead textures (built on #748's branch)

[pasquelin]

To-do added (CTO, 2026-09-26, found in #748's proof)

• Coverage on the filtered cut, not texel counts. Multi-resolution benchmark and temporal stability verification for foliage #44's rule matches the share of texels ≥ C; the reference practice (Castaño, "Computing alpha mipmaps") matches the coverage of the bilinearly filtered cut, which is what the sampler draws. On leaves-cut-by-alpha's leaf (C = 128) the texel-count rule is exact per level but the cut area is 7.5 % under level 0's at 4² and up to +8.5 % over at 8². Move the rule to filtered-cut coverage in one place for the three builders (compiler texture_preview/coverage.rs, WebGPU texture/coverageRule.ts WGSL, WebGL2 GLSL), shared table updated, docs/FORMAT.md updated. Proof: per-level filtered-cut coverage within ±2.5 % of level 0 on the leaf and sponza's masked maps; the measurer's high-DPR level-0 reference on leaves-cut-by-alpha.
• Size note (lead): WebGL2's mip chains keep alpha coverage by the same rule as WebGPU #769 is ~440 lines already; with this item it will need the CTO's split.