Runtime mip chains keep alpha coverage on WebGPU and WebGL2

[pasquelin]

Why

Split from #44 (about 600–650 lines over three chain builders). Alpha-tested mips lose coverage at coarse levels (−26 % to −57 % at L7–L8 on sponza's masked maps, #55's measurement). #44 delivers the rule and the compiler's chain; this issue brings the two runtime builders to the same rule.

To do

• WebGPU keeps coverage: the WebGPU mip chain builder applies Multi-resolution benchmark and temporal stability verification for foliage #44's coverage-preserving alpha rule.
• WebGL2 keeps coverage: the WebGL2 mip chain builder applies the same rule; one rule for the three builders, written once.

Code context

texture/mips.ts, the WebGPU and WebGL2 chain builders, the rule #44 writes in reduce.rs halve.

Proof

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

Links

Blocked by #44. Split from #44.

[pasquelin]

To do added by the CTO (textures lead, found on #44): the backends disagree on masked materials whose baseColorFactor alpha < 1: WebGL2 multiplies by it (webgl/cluster/shaders.ts:94), WebGPU does not (webgpu/row/pageRow.ts:94, maskKeep in visibility/shader/pageWgsl.ts). The glTF 2.0 rule holds on both: alpha = factor × texel alpha, compared to alphaCutoff; the coverage rule of #44 uses the same effective cutoff (cutoff ÷ factor).

[pasquelin]

taken by lead textures (built on #44's rule, branch 44-coverage-alpha-compiler, awaiting its proof)

[pasquelin]

Split by the CTO: #748 keeps the rule, the WebGPU builder and the WebGPU mask test with the factor; the WebGL2 builder moves to 769.

[pasquelin]

To-do added by the lead (found in the #44 re-review): the compiler's quality gate (asset-compiler-rust/.../blocks/quality.rs:71) checks texels that change side at the raw alphaCutoff; once WebGPU cuts factor × texel, the gate follows the same product (one cutoff_byte(cutoff, factor), #44).