reze-engine 0.12.2 → 0.12.3

package/dist/shaders/passes/composite.d.ts

@@ -1,2 +1,2 @@
-export declare const COMPOSITE_SHADER_WGSL = "\n// Pipeline-override constant: the engine creates two composite pipelines, one\n// with APPLY_GAMMA=false (gamma=1 fast path) and one with APPLY_GAMMA=true.\n// The 'if (APPLY_GAMMA)' below is resolved at pipeline-compile time \u2014 the\n// dead branch is dropped by the shader compiler (no runtime branch, no pow\n// invocation on Safari's Metal backend in the common case).\noverride APPLY_GAMMA: bool = true;\n\n@group(0) @binding(0) var hdrTex: texture_2d<f32>;\n@group(0) @binding(1) var bloomTex: texture_2d<f32>;   // bloomUpTexture mip 0 (full pyramid top)\n@group(0) @binding(2) var bloomSamp: sampler;\n@group(0) @binding(3) var<uniform> viewU: array<vec4<f32>, 2>;\n// viewU[0] = (exposure, invGamma, _, _);  viewU[1] = (tint.rgb, intensity)\n// invGamma = 1/gamma precomputed on CPU \u2014 avoids a per-pixel divide.\n\nfn filmic(x: f32) -> f32 {\n  // Re-fit against Blender 3.6 Filmic MHC anchors (sobotka/filmic-blender\n  // look_medium-high-contrast.spi1d). Previous curve was compressed:\n  // midtones too bright, highlights too dim \u2014 flattened contrast, read\n  // as \"washed-out\" on saturated surfaces (hair especially).\n  // Reference checkpoints: linear 0.18 \u2192 ~0.395, linear 1.0 \u2192 ~0.83.\n  var lut = array<f32, 14>(\n    0.0028, 0.0068, 0.0151, 0.0313, 0.0610, 0.1120, 0.1920,\n    0.3060, 0.4590, 0.6310, 0.8200, 0.9070, 0.9620, 0.9890\n  );\n  let t = clamp(log2(max(x, 1e-10)) + 10.0, 0.0, 13.0);\n  let i = u32(t);\n  let j = min(i + 1u, 13u);\n  return mix(lut[i], lut[j], t - f32(i));\n}\n\n@vertex fn vs(@builtin(vertex_index) vi: u32) -> @builtin(position) vec4f {\n  let x = f32((vi & 1u) << 2u) - 1.0;\n  let y = f32((vi & 2u) << 1u) - 1.0;\n  return vec4f(x, y, 0.0, 1.0);\n}\n\n@fragment fn fs(@builtin(position) fragCoord: vec4f) -> @location(0) vec4f {\n  let hdr = textureLoad(hdrTex, vec2<i32>(fragCoord.xy), 0);\n  let a = max(hdr.a, 1e-6);\n  let straight = hdr.rgb / a;\n  let fullSz = vec2f(textureDimensions(hdrTex));\n  let bloomSz = vec2f(textureDimensions(bloomTex));\n  // Bloom is at half-res (pyramid mip 0). Sampler interpolates back to full-res UVs.\n  // fragCoord.xy is already at pixel center (e.g. 0.5, 0.5 for first pixel).\n  let bloomUv = fragCoord.xy / max(fullSz, vec2f(1.0));\n  let tint = viewU[1].xyz;\n  let intensity = viewU[1].w;\n  let bloom = textureSampleLevel(bloomTex, bloomSamp, bloomUv, 0.0).rgb * tint * intensity;\n  let combined = straight + bloom;\n  let exposed = combined * exp2(viewU[0].x);\n  let tm = vec3f(filmic(exposed.r), filmic(exposed.g), filmic(exposed.b));\n  var disp = max(tm, vec3f(0.0));\n  if (APPLY_GAMMA) {\n    disp = pow(disp, vec3f(viewU[0].y));\n  }\n  return vec4f(disp * hdr.a, hdr.a);\n}\n";
+export declare const COMPOSITE_SHADER_WGSL = "\n// Pipeline-override constant: the engine creates two composite pipelines, one\n// with APPLY_GAMMA=false (gamma=1 fast path) and one with APPLY_GAMMA=true.\n// The 'if (APPLY_GAMMA)' below is resolved at pipeline-compile time \u2014 the\n// dead branch is dropped by the shader compiler (no runtime branch, no pow\n// invocation on Safari's Metal backend in the common case).\noverride APPLY_GAMMA: bool = true;\n\n@group(0) @binding(0) var hdrTex: texture_2d<f32>;\n@group(0) @binding(1) var bloomTex: texture_2d<f32>;   // bloomUpTexture mip 0 (full pyramid top)\n@group(0) @binding(2) var bloomSamp: sampler;\n@group(0) @binding(3) var<uniform> viewU: array<vec4<f32>, 2>;\n// Aux mask/alpha texture. .r = bloom mask (unused here; bloom blit uses it).\n// .g = accumulated canvas alpha (what hdr.a carried before the HDR format\n// became rg11b10ufloat). We unpremultiply HDR by this alpha for tonemap, then\n// re-premultiply the tonemapped color for output so the premultiplied canvas\n// alphaMode composites the WebGPU surface over the page background correctly.\n@group(0) @binding(4) var maskTex: texture_2d<f32>;\n// viewU[0] = (exposure, invGamma, _, _);  viewU[1] = (tint.rgb, intensity)\n// invGamma = 1/gamma precomputed on CPU \u2014 avoids a per-pixel divide.\n\nfn filmic(x: f32) -> f32 {\n  // Re-fit against Blender 3.6 Filmic MHC anchors (sobotka/filmic-blender\n  // look_medium-high-contrast.spi1d). Previous curve was compressed:\n  // midtones too bright, highlights too dim \u2014 flattened contrast, read\n  // as \"washed-out\" on saturated surfaces (hair especially).\n  // Reference checkpoints: linear 0.18 \u2192 ~0.395, linear 1.0 \u2192 ~0.83.\n  var lut = array<f32, 14>(\n    0.0028, 0.0068, 0.0151, 0.0313, 0.0610, 0.1120, 0.1920,\n    0.3060, 0.4590, 0.6310, 0.8200, 0.9070, 0.9620, 0.9890\n  );\n  let t = clamp(log2(max(x, 1e-10)) + 10.0, 0.0, 13.0);\n  let i = u32(t);\n  let j = min(i + 1u, 13u);\n  return mix(lut[i], lut[j], t - f32(i));\n}\n\n@vertex fn vs(@builtin(vertex_index) vi: u32) -> @builtin(position) vec4f {\n  let x = f32((vi & 1u) << 2u) - 1.0;\n  let y = f32((vi & 2u) << 1u) - 1.0;\n  return vec4f(x, y, 0.0, 1.0);\n}\n\n@fragment fn fs(@builtin(position) fragCoord: vec4f) -> @location(0) vec4f {\n  let coord = vec2<i32>(fragCoord.xy);\n  let hdr = textureLoad(hdrTex, coord, 0);\n  let alpha = textureLoad(maskTex, coord, 0).g;\n  let a = max(alpha, 1e-6);\n  let straight = hdr.rgb / a;\n  let fullSz = vec2f(textureDimensions(hdrTex));\n  // Bloom is at half-res (pyramid mip 0). Sampler interpolates back to full-res UVs.\n  // fragCoord.xy is already at pixel center (e.g. 0.5, 0.5 for first pixel).\n  let bloomUv = fragCoord.xy / max(fullSz, vec2f(1.0));\n  let tint = viewU[1].xyz;\n  let intensity = viewU[1].w;\n  let bloom = textureSampleLevel(bloomTex, bloomSamp, bloomUv, 0.0).rgb * tint * intensity;\n  let combined = straight + bloom;\n  let exposed = combined * exp2(viewU[0].x);\n  let tm = vec3f(filmic(exposed.r), filmic(exposed.g), filmic(exposed.b));\n  var disp = max(tm, vec3f(0.0));\n  if (APPLY_GAMMA) {\n    disp = pow(disp, vec3f(viewU[0].y));\n  }\n  return vec4f(disp * alpha, alpha);\n}\n";
 //# sourceMappingURL=composite.d.ts.map

package/dist/shaders/passes/composite.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"composite.d.ts","sourceRoot":"","sources":["../../../src/shaders/passes/composite.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,qBAAqB,kqFA0DjC,CAAA"}
+{"version":3,"file":"composite.d.ts","sourceRoot":"","sources":["../../../src/shaders/passes/composite.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,qBAAqB,mnGAiEjC,CAAA"}

package/dist/shaders/passes/composite.js

@@ -12,6 +12,12 @@ override APPLY_GAMMA: bool = true;
 @group(0) @binding(1) var bloomTex: texture_2d<f32>;   // bloomUpTexture mip 0 (full pyramid top)
 @group(0) @binding(2) var bloomSamp: sampler;
 @group(0) @binding(3) var<uniform> viewU: array<vec4<f32>, 2>;
+// Aux mask/alpha texture. .r = bloom mask (unused here; bloom blit uses it).
+// .g = accumulated canvas alpha (what hdr.a carried before the HDR format
+// became rg11b10ufloat). We unpremultiply HDR by this alpha for tonemap, then
+// re-premultiply the tonemapped color for output so the premultiplied canvas
+// alphaMode composites the WebGPU surface over the page background correctly.
+@group(0) @binding(4) var maskTex: texture_2d<f32>;
 // viewU[0] = (exposure, invGamma, _, _);  viewU[1] = (tint.rgb, intensity)
 // invGamma = 1/gamma precomputed on CPU — avoids a per-pixel divide.
 
@@ -38,11 +44,12 @@ fn filmic(x: f32) -> f32 {
 }
 
 @fragment fn fs(@builtin(position) fragCoord: vec4f) -> @location(0) vec4f {
-  let hdr = textureLoad(hdrTex, vec2<i32>(fragCoord.xy), 0);
-  let a = max(hdr.a, 1e-6);
+  let coord = vec2<i32>(fragCoord.xy);
+  let hdr = textureLoad(hdrTex, coord, 0);
+  let alpha = textureLoad(maskTex, coord, 0).g;
+  let a = max(alpha, 1e-6);
   let straight = hdr.rgb / a;
   let fullSz = vec2f(textureDimensions(hdrTex));
-  let bloomSz = vec2f(textureDimensions(bloomTex));
   // Bloom is at half-res (pyramid mip 0). Sampler interpolates back to full-res UVs.
   // fragCoord.xy is already at pixel center (e.g. 0.5, 0.5 for first pixel).
   let bloomUv = fragCoord.xy / max(fullSz, vec2f(1.0));
@@ -56,6 +63,6 @@ fn filmic(x: f32) -> f32 {
   if (APPLY_GAMMA) {
     disp = pow(disp, vec3f(viewU[0].y));
   }
-  return vec4f(disp * hdr.a, hdr.a);
+  return vec4f(disp * alpha, alpha);
 }
 `;

package/dist/shaders/passes/ground.d.ts

@@ -1,2 +1,2 @@
-export declare const GROUND_SHADOW_SHADER_WGSL = "\nstruct CameraUniforms { view: mat4x4f, projection: mat4x4f, viewPos: vec3f, _p: f32, };\nstruct Light { direction: vec4f, color: vec4f, };\nstruct LightUniforms { ambientColor: vec4f, lights: array<Light, 4>, };\nstruct GroundShadowMat {\n  diffuseColor: vec3f, fadeStart: f32,\n  fadeEnd: f32, shadowStrength: f32, pcfTexel: f32, gridSpacing: f32,\n  gridLineWidth: f32, gridLineOpacity: f32, noiseStrength: f32, _pad: f32,\n  gridLineColor: vec3f, _pad2: f32,\n};\nstruct LightVP { viewProj: mat4x4f, };\n@group(0) @binding(0) var<uniform> camera: CameraUniforms;\n@group(0) @binding(1) var<uniform> light: LightUniforms;\n@group(0) @binding(2) var shadowMap: texture_depth_2d;\n@group(0) @binding(3) var shadowSampler: sampler_comparison;\n@group(0) @binding(4) var<uniform> material: GroundShadowMat;\n@group(0) @binding(5) var<uniform> lightVP: LightVP;\n\nfn hash2(p: vec2f) -> f32 {\n  var p3 = fract(vec3f(p.x, p.y, p.x) * 0.1031);\n  p3 += dot(p3, vec3f(p3.y + 33.33, p3.z + 33.33, p3.x + 33.33));\n  return fract((p3.x + p3.y) * p3.z);\n}\nfn valueNoise(p: vec2f) -> f32 {\n  let i = floor(p);\n  let f = fract(p);\n  let u = f * f * (3.0 - 2.0 * f);\n  return mix(mix(hash2(i), hash2(i + vec2f(1.0, 0.0)), u.x),\n             mix(hash2(i + vec2f(0.0, 1.0)), hash2(i + vec2f(1.0, 1.0)), u.x), u.y);\n}\nfn fbmNoise(p: vec2f) -> f32 {\n  var v = 0.0;\n  var a = 0.5;\n  var pp = p;\n  for (var i = 0; i < 4; i++) {\n    v += a * valueNoise(pp);\n    pp *= 2.0;\n    a *= 0.5;\n  }\n  return v;\n}\n\nstruct VO { @builtin(position) position: vec4f, @location(0) worldPos: vec3f, @location(1) normal: vec3f, };\n@vertex fn vs(@location(0) position: vec3f, @location(1) normal: vec3f, @location(2) uv: vec2f) -> VO {\n  var o: VO; o.worldPos = position; o.normal = normal;\n  o.position = camera.projection * camera.view * vec4f(position, 1.0); return o;\n}\nstruct FSOut { @location(0) color: vec4f, @location(1) mask: f32 };\n@fragment fn fs(i: VO) -> FSOut {\n  let n = normalize(i.normal);\n  let centerDist = length(i.worldPos.xz);\n  let edgeFade = 1.0 - smoothstep(0.0, 1.0, clamp((centerDist - material.fadeStart) / max(material.fadeEnd - material.fadeStart, 0.001), 0.0, 1.0));\n\n  let lclip = lightVP.viewProj * vec4f(i.worldPos, 1.0);\n  let ndc = lclip.xyz / max(lclip.w, 1e-6);\n  let suv = vec2f(ndc.x * 0.5 + 0.5, 0.5 - ndc.y * 0.5);\n  let suv_c = clamp(suv, vec2f(0.02), vec2f(0.98));\n  let st = material.pcfTexel;\n  let compareZ = ndc.z - 0.0035;\n  var vis = 0.0;\n  for (var y = -2; y <= 2; y++) {\n    for (var x = -2; x <= 2; x++) {\n      vis += textureSampleCompare(shadowMap, shadowSampler, suv_c + vec2f(f32(x), f32(y)) * st, compareZ);\n    }\n  }\n  vis *= 0.04;\n\n  // Frosted/matte micro-texture\n  let noiseVal = fbmNoise(i.worldPos.xz * 3.0);\n  let noiseTint = 1.0 + (noiseVal - 0.5) * material.noiseStrength;\n\n  // Grid lines \u2014 anti-aliased via screen-space derivatives\n  let gp = i.worldPos.xz / material.gridSpacing;\n  let gridFrac = abs(fract(gp - 0.5) - 0.5);\n  let gridDeriv = fwidth(gp);\n  let halfLine = material.gridLineWidth * 0.5;\n  let gridLine = 1.0 - min(\n    smoothstep(halfLine - gridDeriv.x, halfLine + gridDeriv.x, gridFrac.x),\n    smoothstep(halfLine - gridDeriv.y, halfLine + gridDeriv.y, gridFrac.y)\n  );\n  let sun = light.ambientColor.xyz + light.lights[0].color.xyz * light.lights[0].color.w * max(dot(n, -light.lights[0].direction.xyz), 0.0);\n  let dark = (1.0 - vis) * material.shadowStrength;\n  var baseColor = material.diffuseColor * sun * (1.0 - dark * 0.65);\n  baseColor *= noiseTint;\n  let finalColor = mix(baseColor, material.gridLineColor, gridLine * material.gridLineOpacity * edgeFade);\n  var out: FSOut;\n  out.color = vec4f(finalColor * edgeFade, edgeFade);\n  out.mask = 0.0;\n  return out;\n}\n";
+export declare const GROUND_SHADOW_SHADER_WGSL = "\nstruct CameraUniforms { view: mat4x4f, projection: mat4x4f, viewPos: vec3f, _p: f32, };\nstruct Light { direction: vec4f, color: vec4f, };\nstruct LightUniforms { ambientColor: vec4f, lights: array<Light, 4>, };\nstruct GroundShadowMat {\n  diffuseColor: vec3f, fadeStart: f32,\n  fadeEnd: f32, shadowStrength: f32, pcfTexel: f32, gridSpacing: f32,\n  gridLineWidth: f32, gridLineOpacity: f32, noiseStrength: f32, _pad: f32,\n  gridLineColor: vec3f, _pad2: f32,\n};\nstruct LightVP { viewProj: mat4x4f, };\n@group(0) @binding(0) var<uniform> camera: CameraUniforms;\n@group(0) @binding(1) var<uniform> light: LightUniforms;\n@group(0) @binding(2) var shadowMap: texture_depth_2d;\n@group(0) @binding(3) var shadowSampler: sampler_comparison;\n@group(0) @binding(4) var<uniform> material: GroundShadowMat;\n@group(0) @binding(5) var<uniform> lightVP: LightVP;\n\nfn hash2(p: vec2f) -> f32 {\n  var p3 = fract(vec3f(p.x, p.y, p.x) * 0.1031);\n  p3 += dot(p3, vec3f(p3.y + 33.33, p3.z + 33.33, p3.x + 33.33));\n  return fract((p3.x + p3.y) * p3.z);\n}\nfn valueNoise(p: vec2f) -> f32 {\n  let i = floor(p);\n  let f = fract(p);\n  let u = f * f * (3.0 - 2.0 * f);\n  return mix(mix(hash2(i), hash2(i + vec2f(1.0, 0.0)), u.x),\n             mix(hash2(i + vec2f(0.0, 1.0)), hash2(i + vec2f(1.0, 1.0)), u.x), u.y);\n}\nfn fbmNoise(p: vec2f) -> f32 {\n  var v = 0.0;\n  var a = 0.5;\n  var pp = p;\n  for (var i = 0; i < 4; i++) {\n    v += a * valueNoise(pp);\n    pp *= 2.0;\n    a *= 0.5;\n  }\n  return v;\n}\n\nstruct VO { @builtin(position) position: vec4f, @location(0) worldPos: vec3f, @location(1) normal: vec3f, };\n@vertex fn vs(@location(0) position: vec3f, @location(1) normal: vec3f, @location(2) uv: vec2f) -> VO {\n  var o: VO; o.worldPos = position; o.normal = normal;\n  o.position = camera.projection * camera.view * vec4f(position, 1.0); return o;\n}\nstruct FSOut { @location(0) color: vec4f, @location(1) mask: vec4f };\n@fragment fn fs(i: VO) -> FSOut {\n  let n = normalize(i.normal);\n  let centerDist = length(i.worldPos.xz);\n  let edgeFade = 1.0 - smoothstep(0.0, 1.0, clamp((centerDist - material.fadeStart) / max(material.fadeEnd - material.fadeStart, 0.001), 0.0, 1.0));\n\n  let lclip = lightVP.viewProj * vec4f(i.worldPos, 1.0);\n  let ndc = lclip.xyz / max(lclip.w, 1e-6);\n  let suv = vec2f(ndc.x * 0.5 + 0.5, 0.5 - ndc.y * 0.5);\n  let suv_c = clamp(suv, vec2f(0.02), vec2f(0.98));\n  let st = material.pcfTexel;\n  let compareZ = ndc.z - 0.0035;\n  var vis = 0.0;\n  for (var y = -2; y <= 2; y++) {\n    for (var x = -2; x <= 2; x++) {\n      vis += textureSampleCompare(shadowMap, shadowSampler, suv_c + vec2f(f32(x), f32(y)) * st, compareZ);\n    }\n  }\n  vis *= 0.04;\n\n  // Frosted/matte micro-texture\n  let noiseVal = fbmNoise(i.worldPos.xz * 3.0);\n  let noiseTint = 1.0 + (noiseVal - 0.5) * material.noiseStrength;\n\n  // Grid lines \u2014 anti-aliased via screen-space derivatives\n  let gp = i.worldPos.xz / material.gridSpacing;\n  let gridFrac = abs(fract(gp - 0.5) - 0.5);\n  let gridDeriv = fwidth(gp);\n  let halfLine = material.gridLineWidth * 0.5;\n  let gridLine = 1.0 - min(\n    smoothstep(halfLine - gridDeriv.x, halfLine + gridDeriv.x, gridFrac.x),\n    smoothstep(halfLine - gridDeriv.y, halfLine + gridDeriv.y, gridFrac.y)\n  );\n  let sun = light.ambientColor.xyz + light.lights[0].color.xyz * light.lights[0].color.w * max(dot(n, -light.lights[0].direction.xyz), 0.0);\n  let dark = (1.0 - vis) * material.shadowStrength;\n  var baseColor = material.diffuseColor * sun * (1.0 - dark * 0.65);\n  baseColor *= noiseTint;\n  let finalColor = mix(baseColor, material.gridLineColor, gridLine * material.gridLineOpacity * edgeFade);\n  var out: FSOut;\n  out.color = vec4f(finalColor * edgeFade, edgeFade);\n  // mask.r = 0: ground never contributes to bloom. mask.g = 1.0 with src.a =\n  // edgeFade turns the aux blend into alpha-over, so the drawable alpha fades\n  // from edgeFade at the center to 0 at the radial edge \u2014 letting the page\n  // background show through under the premultiplied canvas alphaMode.\n  out.mask = vec4f(0.0, 1.0, 0.0, edgeFade);\n  return out;\n}\n";
 //# sourceMappingURL=ground.d.ts.map

package/dist/shaders/passes/ground.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"ground.d.ts","sourceRoot":"","sources":["../../../src/shaders/passes/ground.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,yBAAyB,2tHA0FrC,CAAA"}
+{"version":3,"file":"ground.d.ts","sourceRoot":"","sources":["../../../src/shaders/passes/ground.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,yBAAyB,ijIA8FrC,CAAA"}

package/dist/shaders/passes/ground.js

@@ -47,7 +47,7 @@ struct VO { @builtin(position) position: vec4f, @location(0) worldPos: vec3f, @l
   var o: VO; o.worldPos = position; o.normal = normal;
   o.position = camera.projection * camera.view * vec4f(position, 1.0); return o;
 }
-struct FSOut { @location(0) color: vec4f, @location(1) mask: f32 };
+struct FSOut { @location(0) color: vec4f, @location(1) mask: vec4f };
 @fragment fn fs(i: VO) -> FSOut {
   let n = normalize(i.normal);
   let centerDist = length(i.worldPos.xz);
@@ -87,7 +87,11 @@ struct FSOut { @location(0) color: vec4f, @location(1) mask: f32 };
   let finalColor = mix(baseColor, material.gridLineColor, gridLine * material.gridLineOpacity * edgeFade);
   var out: FSOut;
   out.color = vec4f(finalColor * edgeFade, edgeFade);
-  out.mask = 0.0;
+  // mask.r = 0: ground never contributes to bloom. mask.g = 1.0 with src.a =
+  // edgeFade turns the aux blend into alpha-over, so the drawable alpha fades
+  // from edgeFade at the center to 0 at the radial edge — letting the page
+  // background show through under the premultiplied canvas alphaMode.
+  out.mask = vec4f(0.0, 1.0, 0.0, edgeFade);
   return out;
 }
 `;

package/dist/shaders/passes/outline.d.ts

@@ -1,2 +1,2 @@
-export declare const OUTLINE_SHADER_WGSL = "\nstruct CameraUniforms {\n  view: mat4x4f,\n  projection: mat4x4f,\n  viewPos: vec3f,\n  _padding: f32,\n};\n\nstruct MaterialUniforms {\n  edgeColor: vec4f,\n  edgeSize: f32,\n  _padding1: f32,\n  _padding2: f32,\n  _padding3: f32,\n};\n\n@group(0) @binding(0) var<uniform> camera: CameraUniforms;\n@group(1) @binding(0) var<storage, read> skinMats: array<mat4x4f>;\n@group(2) @binding(0) var<uniform> material: MaterialUniforms;\n\nstruct VertexOutput {\n  @builtin(position) position: vec4f,\n};\n\n@vertex fn vs(\n  @location(0) position: vec3f,\n  @location(1) normal: vec3f,\n  @location(3) joints0: vec4<u32>,\n  @location(4) weights0: vec4<f32>\n) -> VertexOutput {\n  var output: VertexOutput;\n  let pos4 = vec4f(position, 1.0);\n\n  let weightSum = weights0.x + weights0.y + weights0.z + weights0.w;\n  let invWeightSum = select(1.0, 1.0 / weightSum, weightSum > 0.0001);\n  let normalizedWeights = select(vec4f(1.0, 0.0, 0.0, 0.0), weights0 * invWeightSum, weightSum > 0.0001);\n\n  var skinnedPos = vec4f(0.0, 0.0, 0.0, 0.0);\n  var skinnedNrm = vec3f(0.0, 0.0, 0.0);\n  for (var i = 0u; i < 4u; i++) {\n    let j = joints0[i];\n    let w = normalizedWeights[i];\n    let m = skinMats[j];\n    skinnedPos += (m * pos4) * w;\n    let r3 = mat3x3f(m[0].xyz, m[1].xyz, m[2].xyz);\n    skinnedNrm += (r3 * normal) * w;\n  }\n  let worldPos = skinnedPos.xyz;\n  let worldNormal = normalize(skinnedNrm);\n\n  // Screen-space outline extrusion \u2014 MMD-style pixel-stable edge line.\n  // 1. Project position and normal-as-direction to clip space.\n  // 2. Normalize the 2D clip-space normal, aspect-compensated so \"one pixel horizontally\"\n  //    matches \"one pixel vertically\" (otherwise wide viewports squash the outline in X).\n  // 3. Offset clip-space xy by (normal * edgeSize * edgeScale), then multiply by w\n  //    so the perspective divide cancels out \u2192 offset stays constant in NDC regardless\n  //    of depth, matching how MMD / babylon-mmd style outlines look identical when zooming.\n  // 4. edgeScale is in NDC-y units per PMX edgeSize. \u2248 0.006 gives ~3px at 1080p; it's\n  //    tied to viewport HEIGHT so resizing the window keeps pixel thickness stable.\n  let viewProj = camera.projection * camera.view;\n  let clipPos = viewProj * vec4f(worldPos, 1.0);\n  let clipNormal = (viewProj * vec4f(worldNormal, 0.0)).xy;\n  // projection is column-major: proj[0][0] = 1/(aspect\u00B7tan(fov/2)), proj[1][1] = 1/tan(fov/2).\n  // Ratio proj[1][1]/proj[0][0] recovers the viewport aspect (width/height).\n  let aspect = camera.projection[1][1] / camera.projection[0][0];\n  let pixelDir = normalize(vec2f(clipNormal.x * aspect, clipNormal.y));\n  let ndcDir = vec2f(pixelDir.x / aspect, pixelDir.y);\n  let edgeScale = 0.0016;\n  let offset = ndcDir * material.edgeSize * edgeScale * clipPos.w;\n  output.position = vec4f(clipPos.xy + offset, clipPos.z, clipPos.w);\n  return output;\n}\n\nstruct FSOut { @location(0) color: vec4f, @location(1) mask: f32 };\n@fragment fn fs() -> FSOut {\n  var out: FSOut;\n  out.color = material.edgeColor;\n  out.mask = 1.0;\n  return out;\n}\n";
+export declare const OUTLINE_SHADER_WGSL = "\nstruct CameraUniforms {\n  view: mat4x4f,\n  projection: mat4x4f,\n  viewPos: vec3f,\n  _padding: f32,\n};\n\nstruct MaterialUniforms {\n  edgeColor: vec4f,\n  edgeSize: f32,\n  _padding1: f32,\n  _padding2: f32,\n  _padding3: f32,\n};\n\n@group(0) @binding(0) var<uniform> camera: CameraUniforms;\n@group(1) @binding(0) var<storage, read> skinMats: array<mat4x4f>;\n@group(2) @binding(0) var<uniform> material: MaterialUniforms;\n\nstruct VertexOutput {\n  @builtin(position) position: vec4f,\n};\n\n@vertex fn vs(\n  @location(0) position: vec3f,\n  @location(1) normal: vec3f,\n  @location(3) joints0: vec4<u32>,\n  @location(4) weights0: vec4<f32>\n) -> VertexOutput {\n  var output: VertexOutput;\n  let pos4 = vec4f(position, 1.0);\n\n  let weightSum = weights0.x + weights0.y + weights0.z + weights0.w;\n  let invWeightSum = select(1.0, 1.0 / weightSum, weightSum > 0.0001);\n  let normalizedWeights = select(vec4f(1.0, 0.0, 0.0, 0.0), weights0 * invWeightSum, weightSum > 0.0001);\n\n  var skinnedPos = vec4f(0.0, 0.0, 0.0, 0.0);\n  var skinnedNrm = vec3f(0.0, 0.0, 0.0);\n  for (var i = 0u; i < 4u; i++) {\n    let j = joints0[i];\n    let w = normalizedWeights[i];\n    let m = skinMats[j];\n    skinnedPos += (m * pos4) * w;\n    let r3 = mat3x3f(m[0].xyz, m[1].xyz, m[2].xyz);\n    skinnedNrm += (r3 * normal) * w;\n  }\n  let worldPos = skinnedPos.xyz;\n  let worldNormal = normalize(skinnedNrm);\n\n  // Screen-space outline extrusion \u2014 MMD-style pixel-stable edge line.\n  // 1. Project position and normal-as-direction to clip space.\n  // 2. Normalize the 2D clip-space normal, aspect-compensated so \"one pixel horizontally\"\n  //    matches \"one pixel vertically\" (otherwise wide viewports squash the outline in X).\n  // 3. Offset clip-space xy by (normal * edgeSize * edgeScale), then multiply by w\n  //    so the perspective divide cancels out \u2192 offset stays constant in NDC regardless\n  //    of depth, matching how MMD / babylon-mmd style outlines look identical when zooming.\n  // 4. edgeScale is in NDC-y units per PMX edgeSize. \u2248 0.006 gives ~3px at 1080p; it's\n  //    tied to viewport HEIGHT so resizing the window keeps pixel thickness stable.\n  let viewProj = camera.projection * camera.view;\n  let clipPos = viewProj * vec4f(worldPos, 1.0);\n  let clipNormal = (viewProj * vec4f(worldNormal, 0.0)).xy;\n  // projection is column-major: proj[0][0] = 1/(aspect\u00B7tan(fov/2)), proj[1][1] = 1/tan(fov/2).\n  // Ratio proj[1][1]/proj[0][0] recovers the viewport aspect (width/height).\n  let aspect = camera.projection[1][1] / camera.projection[0][0];\n  let pixelDir = normalize(vec2f(clipNormal.x * aspect, clipNormal.y));\n  let ndcDir = vec2f(pixelDir.x / aspect, pixelDir.y);\n  let edgeScale = 0.0016;\n  let offset = ndcDir * material.edgeSize * edgeScale * clipPos.w;\n  output.position = vec4f(clipPos.xy + offset, clipPos.z, clipPos.w);\n  return output;\n}\n\nstruct FSOut { @location(0) color: vec4f, @location(1) mask: vec4f };\n@fragment fn fs() -> FSOut {\n  var out: FSOut;\n  out.color = material.edgeColor;\n  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);\n  return out;\n}\n";
 //# sourceMappingURL=outline.d.ts.map

package/dist/shaders/passes/outline.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"outline.d.ts","sourceRoot":"","sources":["../../../src/shaders/passes/outline.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,mBAAmB,kjGAgF/B,CAAA"}
+{"version":3,"file":"outline.d.ts","sourceRoot":"","sources":["../../../src/shaders/passes/outline.ts"],"names":[],"mappings":"AAGA,eAAO,MAAM,mBAAmB,klGAgF/B,CAAA"}

package/dist/shaders/passes/outline.js

@@ -73,11 +73,11 @@ struct VertexOutput {
   return output;
 }
 
-struct FSOut { @location(0) color: vec4f, @location(1) mask: f32 };
+struct FSOut { @location(0) color: vec4f, @location(1) mask: vec4f };
 @fragment fn fs() -> FSOut {
   var out: FSOut;
   out.color = material.edgeColor;
-  out.mask = 1.0;
+  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);
   return out;
 }
 `;

package/dist/shaders/stockings.d.ts

@@ -1,2 +1,2 @@
-export declare const STOCKINGS_SHADER_WGSL = "\n\n\n\n// Baked 64\u00D764 rgba8unorm combined BRDF LUT \u2014 created once at engine init by dfg_lut.ts.\n//   .rg = split-sum DFG (Karis: tint = f0\u00B7x + f90\u00B7y)  \u2192 F_brdf_*_scatter\n//   .ba = Heitz 2016 LTC magnitude (ltc_mag_ggx)       \u2192 ltc_brdf_scale_from_lut\n// Paired with group(0) binding(2) diffuseSampler (linear filter). Sample once per\n// fragment via brdf_lut_sample() \u2014 callers feed .rg and the whole vec4 into the\n// helpers below, halving LUT taps on the default Principled path.\n@group(0) @binding(9) var brdfLut: texture_2d<f32>;\n\n// \u2500\u2500\u2500 RGB \u2194 HSV \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn rgb_to_hsv(rgb: vec3f) -> vec3f {\n  let c_max = max(rgb.r, max(rgb.g, rgb.b));\n  let c_min = min(rgb.r, min(rgb.g, rgb.b));\n  let delta = c_max - c_min;\n\n  var h = 0.0;\n  if (delta > 1e-6) {\n    if (c_max == rgb.r) {\n      h = (rgb.g - rgb.b) / delta;\n      if (h < 0.0) { h += 6.0; }\n    } else if (c_max == rgb.g) {\n      h = 2.0 + (rgb.b - rgb.r) / delta;\n    } else {\n      h = 4.0 + (rgb.r - rgb.g) / delta;\n    }\n    h /= 6.0;\n  }\n  let s = select(0.0, delta / c_max, c_max > 1e-6);\n  return vec3f(h, s, c_max);\n}\n\nfn hsv_to_rgb(hsv: vec3f) -> vec3f {\n  let h = hsv.x;\n  let s = hsv.y;\n  let v = hsv.z;\n  if (s < 1e-6) { return vec3f(v); }\n\n  let hh = fract(h) * 6.0;\n  let sector = u32(hh);\n  let f = hh - f32(sector);\n  let p = v * (1.0 - s);\n  let q = v * (1.0 - s * f);\n  let t = v * (1.0 - s * (1.0 - f));\n\n  switch (sector) {\n    case 0u: { return vec3f(v, t, p); }\n    case 1u: { return vec3f(q, v, p); }\n    case 2u: { return vec3f(p, v, t); }\n    case 3u: { return vec3f(p, q, v); }\n    case 4u: { return vec3f(t, p, v); }\n    default: { return vec3f(v, p, q); }\n  }\n}\n\n// \u2500\u2500\u2500 HUE_SAT node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn hue_sat(hue: f32, saturation: f32, value: f32, fac: f32, color: vec3f) -> vec3f {\n  var hsv = rgb_to_hsv(color);\n  hsv.x = fract(hsv.x + hue - 0.5);\n  hsv.y = clamp(hsv.y * saturation, 0.0, 1.0);\n  hsv.z *= value;\n  return mix(color, hsv_to_rgb(hsv), fac);\n}\n\n// hue_sat specialization for hue=0.5 (identity hue shift \u2014 fract(h + 0.5 - 0.5) = h).\n// Branchless equivalent that skips the rgb_to_hsv \u2192 hsv_to_rgb roundtrip: WebKit's\n// Metal backend serializes the 3-way if chain in rgb_to_hsv and the 6-way switch in\n// hsv_to_rgb, where this form compiles to linear SIMD ops + a single select.\nfn hue_sat_id(saturation: f32, value: f32, fac: f32, color: vec3f) -> vec3f {\n  let m = max(max(color.r, color.g), color.b);\n  let n = min(min(color.r, color.g), color.b);\n  // Unclamped (sat*old_s \u2264 1): reproj = mix(vec3f(m), color, saturation).\n  // Clamped (saturated to 1):   reproj = (color - n) * m / (m - n).\n  let range = max(m - n, 1e-6);\n  let unclamped = mix(vec3f(m), color, saturation);\n  let clamped = (color - vec3f(n)) * m / range;\n  let needs_clamp = (m - n) * saturation >= m;\n  let reproj = select(unclamped, clamped, needs_clamp);\n  return mix(color, reproj * value, fac);\n}\n\n// \u2500\u2500\u2500 BRIGHTCONTRAST node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn bright_contrast(color: vec3f, bright: f32, contrast: f32) -> vec3f {\n  let a = 1.0 + contrast;\n  let b = bright - contrast * 0.5;\n  return max(vec3f(0.0), color * a + vec3f(b));\n}\n\n// \u2500\u2500\u2500 INVERT node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn invert(fac: f32, color: vec3f) -> vec3f {\n  return mix(color, vec3f(1.0) - color, fac);\n}\n\nfn invert_f(fac: f32, val: f32) -> f32 {\n  return mix(val, 1.0 - val, fac);\n}\n\n// \u2500\u2500\u2500 Color ramp (VALTORGB) \u2014 2-stop variants \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// All 7 presets use exclusively 2-stop ramps.\n\nfn ramp_constant(f: f32, p0: f32, c0: vec4f, p1: f32, c1: vec4f) -> vec4f {\n  return select(c0, c1, f >= p1);\n}\n\n// CONSTANT ramp with screen-space edge AA \u2014 kills sparkle where fwidth(f) straddles a hard step (NPR terminator)\nfn ramp_constant_edge_aa(f: f32, edge: f32, c0: vec4f, c1: vec4f) -> vec4f {\n  let w = max(fwidth(f) * 1.75, 6e-6);\n  let t = smoothstep(edge - w, edge + w, f);\n  return mix(c0, c1, t);\n}\n\nfn ramp_linear(f: f32, p0: f32, c0: vec4f, p1: f32, c1: vec4f) -> vec4f {\n  let t = saturate((f - p0) / max(p1 - p0, 1e-6));\n  return mix(c0, c1, t);\n}\n\nfn ramp_cardinal(f: f32, p0: f32, c0: vec4f, p1: f32, c1: vec4f) -> vec4f {\n  // cardinal spline with 2 stops degrades to smoothstep\n  let t = saturate((f - p0) / max(p1 - p0, 1e-6));\n  let ss = t * t * (3.0 - 2.0 * t);\n  return mix(c0, c1, ss);\n}\n\n// \u2500\u2500\u2500 MATH node operations \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn math_add(a: f32, b: f32) -> f32 { return a + b; }\nfn math_multiply(a: f32, b: f32) -> f32 { return a * b; }\nfn math_power(a: f32, b: f32) -> f32 { return pow(max(a, 0.0), b); }\nfn math_greater_than(a: f32, b: f32) -> f32 { return select(0.0, 1.0, a > b); }\n\n// Blender's implicit Color \u2192 Float socket conversion uses BT.601 grayscale\n// (rgb_to_grayscale in blenkernel/intern/node.cc). When a material graph plugs a\n// Color output into a Math node's Value input, this is the scalar it actually sees.\nfn color_to_value(c: vec3f) -> f32 {\n  return 0.299 * c.r + 0.587 * c.g + 0.114 * c.b;\n}\n\n// \u2500\u2500\u2500 MIX node (blend_type variants) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn mix_blend(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  return mix(a, b, fac);\n}\n\nfn mix_overlay(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  let lo = 2.0 * a * b;\n  let hi = vec3f(1.0) - 2.0 * (vec3f(1.0) - a) * (vec3f(1.0) - b);\n  let overlay = select(hi, lo, a < vec3f(0.5));\n  return mix(a, overlay, fac);\n}\n\nfn mix_multiply(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  return mix(a, a * b, fac);\n}\n\nfn mix_lighten(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  return mix(a, max(a, b), fac);\n}\n\n// Blender Mix (Color) blend LINEAR_LIGHT: result = mix(A, A + 2*B - 1, Fac)\nfn mix_linear_light(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  return mix(a, a + 2.0 * b - vec3f(1.0), fac);\n}\n\n// Luminance for Shader\u2192RGB scalar gates (linear RGB, Rec.709 weights)\nfn luminance_rec709_linear(c: vec3f) -> f32 {\n  return dot(max(c, vec3f(0.0)), vec3f(0.2126, 0.7152, 0.0722));\n}\n\n// \u2500\u2500\u2500 FRESNEL node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Schlick approximation matching Blender's Fresnel node\n\nfn fresnel(ior: f32, n: vec3f, v: vec3f) -> f32 {\n  let r = (ior - 1.0) / (ior + 1.0);\n  let f0 = r * r;\n  let cos_theta = clamp(dot(n, v), 0.0, 1.0);\n  let m = 1.0 - cos_theta;\n  let m2 = m * m;\n  let m5 = m2 * m2 * m;\n  return f0 + (1.0 - f0) * m5;\n}\n\n// \u2500\u2500\u2500 LAYER_WEIGHT node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn layer_weight_fresnel(blend: f32, n: vec3f, v: vec3f) -> f32 {\n  let eta = max(1.0 - blend, 1e-4);\n  let r = (1.0 - eta) / (1.0 + eta);\n  let f0 = r * r;\n  let cos_theta = clamp(abs(dot(n, v)), 0.0, 1.0);\n  let m = 1.0 - cos_theta;\n  let m2 = m * m;\n  let m5 = m2 * m2 * m;\n  return f0 + (1.0 - f0) * m5;\n}\n\nfn layer_weight_facing(blend: f32, n: vec3f, v: vec3f) -> f32 {\n  var facing = abs(dot(n, v));\n  let b = clamp(blend, 0.0, 0.99999);\n  if (b != 0.5) {\n    let exponent = select(2.0 * b, 0.5 / (1.0 - b), b >= 0.5);\n    facing = pow(facing, exponent);\n  }\n  return 1.0 - facing;\n}\n\n// \u2500\u2500\u2500 SHADER_TO_RGB (white DiffuseBSDF) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Eevee captures lit diffuse: (albedo/\u03C0)*sun*N\u00B7L*shadow + ambient (linear). Albedo=1.\n// Matches default.ts direct term scale so VALTORGB thresholds from Blender JSON stay valid.\n\nfn shader_to_rgb_diffuse(n: vec3f, l: vec3f, sun_rgb: vec3f, ambient_rgb: vec3f, shadow: f32) -> f32 {\n  const PI_S: f32 = 3.141592653589793;\n  let ndotl = max(dot(n, l), 0.0);\n  let rgb = sun_rgb * (ndotl * shadow / PI_S) + ambient_rgb;\n  return luminance_rec709_linear(rgb);\n}\n\n// \u2500\u2500\u2500 AMBIENT_OCCLUSION node (faked) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Real SSAO is a non-goal. We approximate: use the \"inside\" value from\n// concavity heuristic: 1.0 = fully lit, lower = occluded.\n// For now returns 1.0 (no darkening). Individual presets can override.\n\nfn ao_fake(n: vec3f, v: vec3f) -> f32 {\n  return 1.0;\n}\n\n// \u2500\u2500\u2500 BUMP node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Screen-space bump from a scalar height field. Needs dFdx/dFdy which\n// WGSL provides as dpdx/dpdy.\n\nfn bump(strength: f32, height: f32, normal: vec3f, world_pos: vec3f) -> vec3f {\n  let dhdx = dpdx(height);\n  let dhdy = dpdy(height);\n  let dpdx_pos = dpdx(world_pos);\n  let dpdy_pos = dpdy(world_pos);\n  let perturbed = normalize(normal) - strength * (dhdx * normalize(cross(dpdy_pos, normal)) + dhdy * normalize(cross(normal, dpdx_pos)));\n  return normalize(perturbed);\n}\n\n// LH engine + WebGPU fragment Y: flip dhdy contribution so height peaks read as outward bumps vs Blender reference\nfn bump_lh(strength: f32, height: f32, normal: vec3f, world_pos: vec3f) -> vec3f {\n  let dhdx = dpdx(height);\n  let dhdy = dpdy(height);\n  let dpdx_pos = dpdx(world_pos);\n  let dpdy_pos = dpdy(world_pos);\n  let perturbed = normalize(normal) - strength * (dhdx * normalize(cross(dpdy_pos, normal)) - dhdy * normalize(cross(normal, dpdx_pos)));\n  return normalize(perturbed);\n}\n\n// \u2500\u2500\u2500 NOISE texture (Perlin-style) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Simplified gradient noise matching Blender's default noise output.\n\n// PCG-style integer hash. Replaces the classic 'fract(sin(q) * LARGE)' trick because\n// WebKit's Metal backend compiles 'sin' to a full transcendental op (slow), while\n// Safari's Apple-GPU scalar ALU handles int muls/xors near free. Inputs arrive as\n// integer-valued floats (floor(p) + unit offsets) from _noise3, so vec3i cast is exact.\nfn _hash33(p: vec3f) -> vec3f {\n  var h = vec3u(vec3i(p) + vec3i(32768));\n  h = h * vec3u(1664525u, 1013904223u, 2654435761u);\n  h = (h.yzx ^ h) * vec3u(2246822519u, 3266489917u, 668265263u);\n  h = h ^ (h >> vec3u(16u));\n  // Mask to 24 bits \u2014 above that f32 loses precision on the u32\u2192f32 convert.\n  let hm = h & vec3u(16777215u);\n  return vec3f(hm) * (2.0 / 16777216.0) - 1.0;\n}\n\nfn _noise3(p: vec3f) -> f32 {\n  let i = floor(p);\n  let f = fract(p);\n  let u = f * f * (3.0 - 2.0 * f);\n\n  return mix(\n    mix(\n      mix(dot(_hash33(i + vec3f(0,0,0)), f - vec3f(0,0,0)),\n          dot(_hash33(i + vec3f(1,0,0)), f - vec3f(1,0,0)), u.x),\n      mix(dot(_hash33(i + vec3f(0,1,0)), f - vec3f(0,1,0)),\n          dot(_hash33(i + vec3f(1,1,0)), f - vec3f(1,1,0)), u.x), u.y),\n    mix(\n      mix(dot(_hash33(i + vec3f(0,0,1)), f - vec3f(0,0,1)),\n          dot(_hash33(i + vec3f(1,0,1)), f - vec3f(1,0,1)), u.x),\n      mix(dot(_hash33(i + vec3f(0,1,1)), f - vec3f(0,1,1)),\n          dot(_hash33(i + vec3f(1,1,1)), f - vec3f(1,1,1)), u.x), u.y),\n    u.z);\n}\n\nfn tex_noise(p: vec3f, scale: f32, detail: f32, roughness: f32, distortion: f32) -> f32 {\n  var q = p;\n  if (abs(distortion) > 1e-6) {\n    let w = _noise3(p * scale * 1.37 + vec3f(2.31, 5.17, 8.09));\n    q = p + (w * 2.0 - 1.0) * distortion;\n  }\n  let coords = q * scale;\n  var value = 0.0;\n  var amplitude = 1.0;\n  var frequency = 1.0;\n  var total_amp = 0.0;\n  let octaves = i32(clamp(detail, 0.0, 15.0)) + 1;\n  for (var i = 0; i < octaves; i++) {\n    value += amplitude * _noise3(coords * frequency);\n    total_amp += amplitude;\n    amplitude *= roughness;\n    frequency *= 2.0;\n  }\n  return value / max(total_amp, 1e-6) * 0.5 + 0.5;\n}\n\n// tex_noise specialization: detail=2.0 (3 octaves), roughness=0.5, distortion=0.\n// WebKit can't unroll tex_noise's for-loop because 'octaves' is a runtime value;\n// this variant is fully unrolled with constants folded (total_amp = 1.75).\nfn tex_noise_d2(p: vec3f, scale: f32) -> f32 {\n  let c = p * scale;\n  let v = _noise3(c) + 0.5 * _noise3(c * 2.0) + 0.25 * _noise3(c * 4.0);\n  return v * (1.0 / 1.75) * 0.5 + 0.5;\n}\n\n// \u2500\u2500\u2500 TEX_GRADIENT (linear) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Used by Stockings preset. Maps the input vector's X to a 0\u20131 gradient.\n\nfn tex_gradient_linear(uv: vec3f) -> f32 {\n  return clamp(uv.x, 0.0, 1.0);\n}\n\n// \u2500\u2500\u2500 TEX_VORONOI (distance only) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Used by Metal preset. Simplified F1 cell noise.\n\nfn tex_voronoi_f1(p: vec3f, scale: f32) -> f32 {\n  let coords = p * scale;\n  let i = floor(coords);\n  let f = fract(coords);\n  var min_dist = 1e10;\n  for (var z = -1; z <= 1; z++) {\n    for (var y = -1; y <= 1; y++) {\n      for (var x = -1; x <= 1; x++) {\n        let neighbor = vec3f(f32(x), f32(y), f32(z));\n        let point = _hash33(i + neighbor) * 0.5 + 0.5;\n        let diff = neighbor + point - f;\n        min_dist = min(min_dist, dot(diff, diff));\n      }\n    }\n  }\n  return sqrt(min_dist);\n}\n\n// \u2500\u2500\u2500 SEPXYZ node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn separate_xyz(v: vec3f) -> vec3f { return v; }\n\n// \u2500\u2500\u2500 VECT_MATH (cross product) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn vect_math_cross(a: vec3f, b: vec3f) -> vec3f { return cross(a, b); }\n\n// \u2500\u2500\u2500 MAPPING node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Point-type mapping: scale, rotate (euler XYZ), translate.\n\nfn mapping_point(v: vec3f, loc: vec3f, rot: vec3f, scl: vec3f) -> vec3f {\n  var p = v * scl;\n  // simplified: skip rotation when all angles are zero (common case)\n  if (abs(rot.x) + abs(rot.y) + abs(rot.z) > 1e-6) {\n    let cx = cos(rot.x); let sx = sin(rot.x);\n    let cy = cos(rot.y); let sy = sin(rot.y);\n    let cz = cos(rot.z); let sz = sin(rot.z);\n    let rx = vec3f(p.x, cx*p.y - sx*p.z, sx*p.y + cx*p.z);\n    let ry = vec3f(cy*rx.x + sy*rx.z, rx.y, -sy*rx.x + cy*rx.z);\n    p = vec3f(cz*ry.x - sz*ry.y, sz*ry.x + cz*ry.y, ry.z);\n  }\n  return p + loc;\n}\n\n// \u2500\u2500\u2500 NORMAL_MAP node (tangent-space) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Applies a tangent-space normal map. Requires TBN from vertex stage.\n\nfn normal_map(strength: f32, map_color: vec3f, normal: vec3f, tangent: vec3f, bitangent: vec3f) -> vec3f {\n  let ts = map_color * 2.0 - 1.0;\n  let perturbed = normalize(tangent * ts.x + bitangent * ts.y + normal * ts.z);\n  return normalize(mix(normal, perturbed, strength));\n}\n\n// \u2500\u2500\u2500 EEVEE Principled BSDF primitives \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Ports from Blender 3.6 source/blender/draw/engines/eevee/shaders/\n//   bsdf_common_lib.glsl + gpu_shader_material_principled.glsl.\n// Usage pattern (see material shaders): direct spec = bsdf_ggx \u00D7 sun \u00D7 shadow\n// (NL baked in, no F yet); ambient spec = probe_radiance; tint both with\n// reflection_color = F_brdf_multi_scatter(f0, f90, split_sum) AFTER summing.\n\nconst EEVEE_PI: f32 = 3.141592653589793;\n\n// Fused analytic GGX specular (direct lights). Returns BRDF \u00D7 NL.\n// 4\u00B7NL\u00B7NV is cancelled via G1_Smith reciprocal form \u2014 see bsdf_common_lib.glsl:115.\n// Caller passes NL, NV (already computed for diffuse + brdf_lut_sample) so WebKit\n// can reuse them instead of recomputing dot products across the function boundary.\nfn bsdf_ggx(N: vec3f, L: vec3f, V: vec3f, NL_in: f32, NV_in: f32, roughness: f32) -> f32 {\n  let a = max(roughness, 1e-4);\n  let a2 = a * a;\n  let H = normalize(L + V);\n  let NH = max(dot(N, H), 1e-8);\n  let NL = max(NL_in, 1e-8);\n  let NV = max(NV_in, 1e-8);\n  // G1_Smith_GGX_opti reciprocal form \u2014 denominator piece only.\n  let G1L = NL + sqrt(NL * (NL - NL * a2) + a2);\n  let G1V = NV + sqrt(NV * (NV - NV * a2) + a2);\n  let G = G1L * G1V;\n  // D_ggx_opti = pi * denom\u00B2 \u2014 reciprocal D \u00D7 a\u00B2.\n  let tmp = (NH * a2 - NH) * NH + 1.0;\n  let D_opti = EEVEE_PI * tmp * tmp;\n  return NL * a2 / (D_opti * G);\n}\n\n// Split-sum DFG LUT \u2014 Karis 2013 curve fit stand-in for the 64\u00D764 baked LUT.\n// Returns (lut.x, lut.y) in Blender convention: tint = f0\u00B7lut.x + f90\u00B7lut.y.\nfn brdf_lut_approx(NV: f32, roughness: f32) -> vec2f {\n  let c0 = vec4f(-1.0, -0.0275, -0.572, 0.022);\n  let c1 = vec4f(1.0, 0.0425, 1.04, -0.04);\n  let r = roughness * c0 + c1;\n  let a004 = min(r.x * r.x, exp2(-9.28 * NV)) * r.x + r.y;\n  return vec2f(-1.04, 1.04) * a004 + r.zw;\n}\n\n// Baked combined BRDF LUT \u2014 exact port of Blender bsdf_lut_frag.glsl packed with\n// ltc_mag_ggx from eevee_lut.c. Single sample returns DFG (.rg) and LTC mag (.ba).\n// Addressed as Blender's common_utiltex_lib.glsl:lut_coords:\n//   coords = (roughness, sqrt(1 - NV)), then half-texel bias for filtering.\n// Requires group(0) binding(9) brdfLut + binding(2) diffuseSampler in the host shader.\nfn brdf_lut_sample(NV: f32, roughness: f32) -> vec4f {\n  let LUT_SIZE: f32 = 64.0;\n  var uv = vec2f(saturate(roughness), sqrt(saturate(1.0 - NV)));\n  uv = uv * ((LUT_SIZE - 1.0) / LUT_SIZE) + 0.5 / LUT_SIZE;\n  return textureSampleLevel(brdfLut, diffuseSampler, uv, 0.0);\n}\n\nfn F_brdf_single_scatter(f0: vec3f, f90: vec3f, lut: vec2f) -> vec3f {\n  return lut.y * f90 + lut.x * f0;\n}\n\n// Fdez-Ag\u00FCera 2019 multi-scatter compensation (EEVEE do_multiscatter=1).\nfn F_brdf_multi_scatter(f0: vec3f, f90: vec3f, lut: vec2f) -> vec3f {\n  let FssEss = lut.y * f90 + lut.x * f0;\n  let Ess = lut.x + lut.y;\n  let Ems = 1.0 - Ess;\n  let Favg = f0 + (1.0 - f0) / 21.0;\n  let Fms = FssEss * Favg / (1.0 - (1.0 - Ess) * Favg);\n  return FssEss + Fms * Ems;\n}\n\n// EEVEE direct-specular energy compensation factor \u2014 closure_eval_glossy_lib.glsl:79-81:\n//   ltc_brdf_scale = (ltc.x + ltc.y) / (split_sum.x + split_sum.y)\n// Blender evaluates direct lights via LTC (Heitz 2016) but indirect via split-sum;\n// direct radiance is rescaled so total-energy matches the split-sum LUT.\n// Takes a pre-sampled vec4f from brdf_lut_sample() to share the fetch with\n// F_brdf_multi_scatter on the same fragment.\nfn ltc_brdf_scale_from_lut(lut: vec4f) -> f32 {\n  return (lut.z + lut.w) / max(lut.x + lut.y, 1e-6);\n}\n\n// Luminance-normalized hue extraction \u2014 Blender tint_from_color (isolates hue+sat).\nfn tint_from_color(color: vec3f) -> vec3f {\n  let lum = dot(color, vec3f(0.3, 0.6, 0.1));\n  return select(vec3f(1.0), color / lum, lum > 0.0);\n}\n\n\n\nstruct CameraUniforms {\n  view: mat4x4f,\n  projection: mat4x4f,\n  viewPos: vec3f,\n  _padding: f32,\n};\n\nstruct Light {\n  direction: vec4f,\n  color: vec4f,\n};\n\nstruct LightUniforms {\n  ambientColor: vec4f,\n  lights: array<Light, 4>,\n};\n\nstruct MaterialUniforms {\n  diffuseColor: vec3f,\n  alpha: f32,\n};\n\nstruct VertexOutput {\n  @builtin(position) position: vec4f,\n  @location(0) normal: vec3f,\n  @location(1) uv: vec2f,\n  @location(2) worldPos: vec3f,\n};\n\nstruct LightVP { viewProj: mat4x4f, };\n\n@group(0) @binding(0) var<uniform> camera: CameraUniforms;\n@group(0) @binding(1) var<uniform> light: LightUniforms;\n@group(0) @binding(2) var diffuseSampler: sampler;\n@group(0) @binding(3) var shadowMap: texture_depth_2d;\n@group(0) @binding(4) var shadowSampler: sampler_comparison;\n@group(0) @binding(5) var<uniform> lightVP: LightVP;\n@group(1) @binding(0) var<storage, read> skinMats: array<mat4x4f>;\n@group(2) @binding(0) var diffuseTexture: texture_2d<f32>;\n@group(2) @binding(1) var<uniform> material: MaterialUniforms;\n\nfn sampleShadow(worldPos: vec3f, n: vec3f) -> f32 {\n  // Back-facing to key light: direct contribution is zero anyway, skip 9 texture samples.\n  if (dot(n, -light.lights[0].direction.xyz) <= 0.0) { return 0.0; }\n  let biasedPos = worldPos + n * 0.08;\n  let lclip = lightVP.viewProj * vec4f(biasedPos, 1.0);\n  let ndc = lclip.xyz / max(lclip.w, 1e-6);\n  let suv = vec2f(ndc.x * 0.5 + 0.5, 0.5 - ndc.y * 0.5);\n  let cmpZ = ndc.z - 0.001;\n  let ts = 1.0 / 2048.0;\n  // 3x3 PCF unrolled \u2014 Safari's Metal backend doesn't unroll nested shadow loops reliably.\n  let s00 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, -ts), cmpZ);\n  let s10 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0, -ts), cmpZ);\n  let s20 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, -ts), cmpZ);\n  let s01 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, 0.0), cmpZ);\n  let s11 = textureSampleCompareLevel(shadowMap, shadowSampler, suv, cmpZ);\n  let s21 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, 0.0), cmpZ);\n  let s02 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts,  ts), cmpZ);\n  let s12 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0,  ts), cmpZ);\n  let s22 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts,  ts), cmpZ);\n  return (s00 + s10 + s20 + s01 + s11 + s21 + s02 + s12 + s22) * (1.0 / 9.0);\n}\n\nconst PI_S: f32 = 3.141592653589793;\n// Principled BSDF params from dump (Alpha=0.95 is intentionally dropped \u2014 see alpha-hash note below)\nconst STOCK_METALLIC: f32 = 0.1;\nconst STOCK_SPECULAR: f32 = 1.0;\nconst STOCK_ROUGHNESS: f32 = 0.5;\nconst STOCK_SHEEN: f32 = 0.7017999887466431;\nconst STOCK_SHEEN_TINT: f32 = 0.5;\n// NPR mask ramps\nconst STOCK_RAMP002_P1: f32 = 0.9565;  // EASE [0\u2192black, 0.9565\u2192white]\nconst STOCK_RAMPFACE_P1: f32 = 0.5435; // EASE [0\u2192black, 0.5435\u2192white]\nconst STOCK_LW_BLEND: f32 = 0.4;       // Layer Weight Blend\n\n// principled_sheen (gpu_shader_material_principled.glsl:8-14) \u2014 empirical NV curve\nfn principled_sheen(NV: f32) -> f32 {\n  let f = 1.0 - NV;\n  return f * f * f * 0.077 + f * 0.01 + 0.00026;\n}\n\n// Wyman & McGuire \"Hashed Alpha Testing\" (2017) \u2014 world-space hash with derivative-aware\n// pixel-scale selection, matches Blender EEVEE prepass_frag.glsl::hashed_alpha_threshold.\n// Key property: dither pattern is stable in object/world space (doesn't swim) and stays\n// at one-pixel frequency regardless of view distance, which makes it tolerable without TAA.\nfn _hash_wm(a: vec2f) -> f32 {\n  return fract(1e4 * sin(17.0 * a.x + 0.1 * a.y) * (0.1 + abs(sin(13.0 * a.y + a.x))));\n}\nfn _hash3d_wm(a: vec3f) -> f32 {\n  return _hash_wm(vec2f(_hash_wm(a.xy), a.z));\n}\nfn hashed_alpha_threshold(co: vec3f) -> f32 {\n  let alphaHashScale: f32 = 1.0;\n  let max_deriv = max(length(dpdx(co)), length(dpdy(co)));\n  let pix_scale = 1.0 / max(alphaHashScale * max_deriv, 1e-6);\n  let pix_scale_log = log2(pix_scale);\n  let px_lo = exp2(floor(pix_scale_log));\n  let px_hi = exp2(ceil(pix_scale_log));\n  let a_lo = _hash3d_wm(floor(px_lo * co));\n  let a_hi = _hash3d_wm(floor(px_hi * co));\n  let fac = fract(pix_scale_log);\n  let x = mix(a_lo, a_hi, fac);\n  // CDF remap so that discard-probability = (1 - alpha) uniformly across scale transitions\n  let a = min(fac, 1.0 - fac);\n  let one_a = 1.0 - a;\n  let denom = 1.0 / max(2.0 * a * one_a, 1e-6);\n  let one_x = 1.0 - x;\n  let case_lo = (x * x) * denom;\n  let case_mid = (x - 0.5 * a) / max(one_a, 1e-6);\n  let case_hi = 1.0 - (one_x * one_x) * denom;\n  var threshold = select(case_hi, select(case_lo, case_mid, x >= a), x < one_a);\n  return clamp(threshold, 1e-6, 1.0);\n}\n\n// Smoothstep-based EASE ramp (Blender VALTORGB EASE) \u2014 2 stops, saturate+smoothstep between\nfn ramp_ease_s(f: f32, p0: f32, p1: f32) -> f32 {\n  let t = saturate((f - p0) / max(p1 - p0, 1e-6));\n  return t * t * (3.0 - 2.0 * t);\n}\n\n@vertex fn vs(\n  @location(0) position: vec3f,\n  @location(1) normal: vec3f,\n  @location(2) uv: vec2f,\n  @location(3) joints0: vec4<u32>,\n  @location(4) weights0: vec4<f32>\n) -> VertexOutput {\n  var output: VertexOutput;\n  let pos4 = vec4f(position, 1.0);\n  let weightSum = weights0.x + weights0.y + weights0.z + weights0.w;\n  let invWeightSum = select(1.0, 1.0 / weightSum, weightSum > 0.0001);\n  let nw = select(vec4f(1.0, 0.0, 0.0, 0.0), weights0 * invWeightSum, weightSum > 0.0001);\n  var skinnedPos = vec4f(0.0);\n  var skinnedNrm = vec3f(0.0);\n  for (var i = 0u; i < 4u; i++) {\n    let m = skinMats[joints0[i]];\n    let w = nw[i];\n    skinnedPos += (m * pos4) * w;\n    skinnedNrm += (mat3x3f(m[0].xyz, m[1].xyz, m[2].xyz) * normal) * w;\n  }\n  output.position = camera.projection * camera.view * vec4f(skinnedPos.xyz, 1.0);\n  // Skip VS normalize \u2014 interpolation denormalizes anyway, and FS always does normalize(input.normal).\n  output.normal = skinnedNrm;\n  output.uv = uv;\n  output.worldPos = skinnedPos.xyz;\n  return output;\n}\n\nstruct FSOut {\n  @location(0) color: vec4f,\n  @location(1) bloom_mask: f32,\n};\n\n@fragment fn fs(input: VertexOutput) -> FSOut {\n  let n = normalize(input.normal);\n  let v = normalize(camera.viewPos - input.worldPos);\n  let l = -light.lights[0].direction.xyz;\n  let sun = light.lights[0].color.xyz * light.lights[0].color.w;\n  let amb = light.ambientColor.xyz;\n  let shadow = sampleShadow(input.worldPos, n);\n\n  let tex_s = textureSample(diffuseTexture, diffuseSampler, input.uv);\n  let tex_rgb = tex_s.rgb;\n  // Alpha HASHED (Blender EEVEE \"Hashed\" blend mode) per preset author's note \u2014\n  // self-overlap on the stockings produces sort cracks under alpha blend. Wyman-style\n  // worldPos hash + depth-write is sort-independent. NOTE: Principled.Alpha=0.95 from\n  // the dump is DROPPED here \u2014 it relies on TAA to smooth the resulting 5%-everywhere\n  // dither, and without TAA it shows as a pervasive dot pattern. Hash now gates only\n  // on texture/material alpha, so solid stockings regions stay fully opaque.\n  let combined_alpha = material.alpha * tex_s.a;\n  if (combined_alpha < hashed_alpha_threshold(input.worldPos)) { discard; }\n  let out_alpha = 1.0;\n\n  // \u2550\u2550\u2550 NPR MASK: TEX_COORD.Generated \u2192 Mapping(Rot=0,\u03C0/2,\u03C0/2, Loc=(1,1,1)) \u2192 Gradient Texture\n  // The Blender mapping reduces to gradient.x = 1 - input.y (rot swaps axes, loc offsets by 1).\n  // We approximate Generated with UV since Y-up PMX has no object bbox in pipeline state.\n  let gen_coord = vec3f(input.uv, 0.0);\n  let mapped = mapping_point(gen_coord, vec3f(1.0), vec3f(0.0, 1.5708, 1.5708), vec3f(1.0));\n  let gradient = tex_gradient_linear(mapped);\n\n  // Ramp.001 LINEAR [0\u2192black, 0.5\u2192white, 1.0\u2192black] \u2014 triangular peak at 0.5\n  let ramp001 = 1.0 - abs(2.0 * gradient - 1.0);\n  // Ramp.002 EASE [0\u2192black, 0.9565\u2192white]\n  let ramp002 = ramp_ease_s(ramp001, 0.0, STOCK_RAMP002_P1);\n\n  // Layer Weight.Facing (Blend=0.4) \u2192 Ramp EASE [0\u2192black, 0.5435\u2192white]\n  let facing = layer_weight_facing(STOCK_LW_BLEND, n, v);\n  let ramp_face = ramp_ease_s(facing, 0.0, STOCK_RAMPFACE_P1);\n\n  // Mix.001: MIX blend Fac=0.5, A=white, B=ramp_face \u2192 (A,B) averaged 50/50\n  let mix001 = mix(1.0, ramp_face, 0.5);\n  // Mix: LIGHTEN blend Fac=0.5, A=mix001, B=ramp002 \u2192 A smoothly lightens toward max(A,B)\n  let lighten = max(mix001, ramp002);\n  let mask = mix(mix001, lighten, 0.5);\n\n  // \u2550\u2550\u2550 EMISSION SHADER \u2550\u2550\u2550\n  // Hue=0.5 (identity rotation), Sat=1.0, Val=5.0 (5\u00D7 brightness boost), Fac=1; Strength=1\n  let emission = hue_sat_id(1.0, 5.0, 1.0, tex_rgb);\n\n  // \u2550\u2550\u2550 PRINCIPLED BSDF (EEVEE port) \u2550\u2550\u2550\n  // base_color_tint, metallic f0, sheen coarse approx (scales diffuse radiance).\n  let NL = max(dot(n, l), 0.0);\n  let NV = max(dot(n, v), 1e-4);\n\n  // f0 = mix((0.08*spec)*dielectric_tint, base, metallic); dielectric_tint=1 since specular_tint=0.\n  let dielectric_f0 = vec3f(0.08 * STOCK_SPECULAR);\n  let f0 = mix(dielectric_f0, tex_rgb, STOCK_METALLIC);\n  let f90 = mix(f0, vec3f(1.0), sqrt(STOCK_SPECULAR));\n  let brdf_lut = brdf_lut_sample(NV, STOCK_ROUGHNESS);\n  let reflection_color = F_brdf_multi_scatter(f0, f90, brdf_lut.xy);\n\n  let spec_direct = bsdf_ggx(n, l, v, NL, NV, STOCK_ROUGHNESS) * sun * shadow * ltc_brdf_scale_from_lut(brdf_lut);\n  let spec_indirect = amb;\n  let spec_radiance = (spec_direct + spec_indirect) * reflection_color;\n\n  // Sheen coarse: diffuse_color += sheen * sheen_color * principled_sheen(NV).\n  let base_tint = tint_from_color(tex_rgb);\n  let sheen_color = mix(vec3f(1.0), base_tint, STOCK_SHEEN_TINT);\n  let diffuse_color = tex_rgb + STOCK_SHEEN * sheen_color * principled_sheen(NV);\n\n  // diffuse_weight = (1 - metallic). Indirect diffuse uses L_w (no \u03C0; see closure_eval_surface_lib:302).\n  let diffuse_weight = 1.0 - STOCK_METALLIC;\n  let diffuse_radiance = diffuse_color * (sun * NL * shadow / PI_S + amb) * diffuse_weight;\n  let principled = diffuse_radiance + spec_radiance;\n\n  // \u2550\u2550\u2550 MIX SHADER: Shader=Emission, Shader_001=Principled, Fac=mask \u2550\u2550\u2550\n  let final_color = mix(emission, principled, mask);\n\n  var out: FSOut;\n  out.color = vec4f(final_color, out_alpha);\n  out.bloom_mask = 1.0;\n  return out;\n}\n\n";
+export declare const STOCKINGS_SHADER_WGSL = "\n\n\n\n// Baked 64\u00D764 rgba8unorm combined BRDF LUT \u2014 created once at engine init by dfg_lut.ts.\n//   .rg = split-sum DFG (Karis: tint = f0\u00B7x + f90\u00B7y)  \u2192 F_brdf_*_scatter\n//   .ba = Heitz 2016 LTC magnitude (ltc_mag_ggx)       \u2192 ltc_brdf_scale_from_lut\n// Paired with group(0) binding(2) diffuseSampler (linear filter). Sample once per\n// fragment via brdf_lut_sample() \u2014 callers feed .rg and the whole vec4 into the\n// helpers below, halving LUT taps on the default Principled path.\n@group(0) @binding(9) var brdfLut: texture_2d<f32>;\n\n// \u2500\u2500\u2500 RGB \u2194 HSV \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn rgb_to_hsv(rgb: vec3f) -> vec3f {\n  let c_max = max(rgb.r, max(rgb.g, rgb.b));\n  let c_min = min(rgb.r, min(rgb.g, rgb.b));\n  let delta = c_max - c_min;\n\n  var h = 0.0;\n  if (delta > 1e-6) {\n    if (c_max == rgb.r) {\n      h = (rgb.g - rgb.b) / delta;\n      if (h < 0.0) { h += 6.0; }\n    } else if (c_max == rgb.g) {\n      h = 2.0 + (rgb.b - rgb.r) / delta;\n    } else {\n      h = 4.0 + (rgb.r - rgb.g) / delta;\n    }\n    h /= 6.0;\n  }\n  let s = select(0.0, delta / c_max, c_max > 1e-6);\n  return vec3f(h, s, c_max);\n}\n\nfn hsv_to_rgb(hsv: vec3f) -> vec3f {\n  let h = hsv.x;\n  let s = hsv.y;\n  let v = hsv.z;\n  if (s < 1e-6) { return vec3f(v); }\n\n  let hh = fract(h) * 6.0;\n  let sector = u32(hh);\n  let f = hh - f32(sector);\n  let p = v * (1.0 - s);\n  let q = v * (1.0 - s * f);\n  let t = v * (1.0 - s * (1.0 - f));\n\n  switch (sector) {\n    case 0u: { return vec3f(v, t, p); }\n    case 1u: { return vec3f(q, v, p); }\n    case 2u: { return vec3f(p, v, t); }\n    case 3u: { return vec3f(p, q, v); }\n    case 4u: { return vec3f(t, p, v); }\n    default: { return vec3f(v, p, q); }\n  }\n}\n\n// \u2500\u2500\u2500 HUE_SAT node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn hue_sat(hue: f32, saturation: f32, value: f32, fac: f32, color: vec3f) -> vec3f {\n  var hsv = rgb_to_hsv(color);\n  hsv.x = fract(hsv.x + hue - 0.5);\n  hsv.y = clamp(hsv.y * saturation, 0.0, 1.0);\n  hsv.z *= value;\n  return mix(color, hsv_to_rgb(hsv), fac);\n}\n\n// hue_sat specialization for hue=0.5 (identity hue shift \u2014 fract(h + 0.5 - 0.5) = h).\n// Branchless equivalent that skips the rgb_to_hsv \u2192 hsv_to_rgb roundtrip: WebKit's\n// Metal backend serializes the 3-way if chain in rgb_to_hsv and the 6-way switch in\n// hsv_to_rgb, where this form compiles to linear SIMD ops + a single select.\nfn hue_sat_id(saturation: f32, value: f32, fac: f32, color: vec3f) -> vec3f {\n  let m = max(max(color.r, color.g), color.b);\n  let n = min(min(color.r, color.g), color.b);\n  // Unclamped (sat*old_s \u2264 1): reproj = mix(vec3f(m), color, saturation).\n  // Clamped (saturated to 1):   reproj = (color - n) * m / (m - n).\n  let range = max(m - n, 1e-6);\n  let unclamped = mix(vec3f(m), color, saturation);\n  let clamped = (color - vec3f(n)) * m / range;\n  let needs_clamp = (m - n) * saturation >= m;\n  let reproj = select(unclamped, clamped, needs_clamp);\n  return mix(color, reproj * value, fac);\n}\n\n// \u2500\u2500\u2500 BRIGHTCONTRAST node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn bright_contrast(color: vec3f, bright: f32, contrast: f32) -> vec3f {\n  let a = 1.0 + contrast;\n  let b = bright - contrast * 0.5;\n  return max(vec3f(0.0), color * a + vec3f(b));\n}\n\n// \u2500\u2500\u2500 INVERT node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn invert(fac: f32, color: vec3f) -> vec3f {\n  return mix(color, vec3f(1.0) - color, fac);\n}\n\nfn invert_f(fac: f32, val: f32) -> f32 {\n  return mix(val, 1.0 - val, fac);\n}\n\n// \u2500\u2500\u2500 Color ramp (VALTORGB) \u2014 2-stop variants \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// All 7 presets use exclusively 2-stop ramps.\n\nfn ramp_constant(f: f32, p0: f32, c0: vec4f, p1: f32, c1: vec4f) -> vec4f {\n  return select(c0, c1, f >= p1);\n}\n\n// CONSTANT ramp with screen-space edge AA \u2014 kills sparkle where fwidth(f) straddles a hard step (NPR terminator)\nfn ramp_constant_edge_aa(f: f32, edge: f32, c0: vec4f, c1: vec4f) -> vec4f {\n  let w = max(fwidth(f) * 1.75, 6e-6);\n  let t = smoothstep(edge - w, edge + w, f);\n  return mix(c0, c1, t);\n}\n\nfn ramp_linear(f: f32, p0: f32, c0: vec4f, p1: f32, c1: vec4f) -> vec4f {\n  let t = saturate((f - p0) / max(p1 - p0, 1e-6));\n  return mix(c0, c1, t);\n}\n\nfn ramp_cardinal(f: f32, p0: f32, c0: vec4f, p1: f32, c1: vec4f) -> vec4f {\n  // cardinal spline with 2 stops degrades to smoothstep\n  let t = saturate((f - p0) / max(p1 - p0, 1e-6));\n  let ss = t * t * (3.0 - 2.0 * t);\n  return mix(c0, c1, ss);\n}\n\n// \u2500\u2500\u2500 MATH node operations \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn math_add(a: f32, b: f32) -> f32 { return a + b; }\nfn math_multiply(a: f32, b: f32) -> f32 { return a * b; }\nfn math_power(a: f32, b: f32) -> f32 { return pow(max(a, 0.0), b); }\nfn math_greater_than(a: f32, b: f32) -> f32 { return select(0.0, 1.0, a > b); }\n\n// Blender's implicit Color \u2192 Float socket conversion uses BT.601 grayscale\n// (rgb_to_grayscale in blenkernel/intern/node.cc). When a material graph plugs a\n// Color output into a Math node's Value input, this is the scalar it actually sees.\nfn color_to_value(c: vec3f) -> f32 {\n  return 0.299 * c.r + 0.587 * c.g + 0.114 * c.b;\n}\n\n// \u2500\u2500\u2500 MIX node (blend_type variants) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn mix_blend(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  return mix(a, b, fac);\n}\n\nfn mix_overlay(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  let lo = 2.0 * a * b;\n  let hi = vec3f(1.0) - 2.0 * (vec3f(1.0) - a) * (vec3f(1.0) - b);\n  let overlay = select(hi, lo, a < vec3f(0.5));\n  return mix(a, overlay, fac);\n}\n\nfn mix_multiply(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  return mix(a, a * b, fac);\n}\n\nfn mix_lighten(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  return mix(a, max(a, b), fac);\n}\n\n// Blender Mix (Color) blend LINEAR_LIGHT: result = mix(A, A + 2*B - 1, Fac)\nfn mix_linear_light(fac: f32, a: vec3f, b: vec3f) -> vec3f {\n  return mix(a, a + 2.0 * b - vec3f(1.0), fac);\n}\n\n// Luminance for Shader\u2192RGB scalar gates (linear RGB, Rec.709 weights)\nfn luminance_rec709_linear(c: vec3f) -> f32 {\n  return dot(max(c, vec3f(0.0)), vec3f(0.2126, 0.7152, 0.0722));\n}\n\n// \u2500\u2500\u2500 FRESNEL node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Schlick approximation matching Blender's Fresnel node\n\nfn fresnel(ior: f32, n: vec3f, v: vec3f) -> f32 {\n  let r = (ior - 1.0) / (ior + 1.0);\n  let f0 = r * r;\n  let cos_theta = clamp(dot(n, v), 0.0, 1.0);\n  let m = 1.0 - cos_theta;\n  let m2 = m * m;\n  let m5 = m2 * m2 * m;\n  return f0 + (1.0 - f0) * m5;\n}\n\n// \u2500\u2500\u2500 LAYER_WEIGHT node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn layer_weight_fresnel(blend: f32, n: vec3f, v: vec3f) -> f32 {\n  let eta = max(1.0 - blend, 1e-4);\n  let r = (1.0 - eta) / (1.0 + eta);\n  let f0 = r * r;\n  let cos_theta = clamp(abs(dot(n, v)), 0.0, 1.0);\n  let m = 1.0 - cos_theta;\n  let m2 = m * m;\n  let m5 = m2 * m2 * m;\n  return f0 + (1.0 - f0) * m5;\n}\n\nfn layer_weight_facing(blend: f32, n: vec3f, v: vec3f) -> f32 {\n  var facing = abs(dot(n, v));\n  let b = clamp(blend, 0.0, 0.99999);\n  if (b != 0.5) {\n    let exponent = select(2.0 * b, 0.5 / (1.0 - b), b >= 0.5);\n    facing = pow(facing, exponent);\n  }\n  return 1.0 - facing;\n}\n\n// \u2500\u2500\u2500 SHADER_TO_RGB (white DiffuseBSDF) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Eevee captures lit diffuse: (albedo/\u03C0)*sun*N\u00B7L*shadow + ambient (linear). Albedo=1.\n// Matches default.ts direct term scale so VALTORGB thresholds from Blender JSON stay valid.\n\nfn shader_to_rgb_diffuse(n: vec3f, l: vec3f, sun_rgb: vec3f, ambient_rgb: vec3f, shadow: f32) -> f32 {\n  const PI_S: f32 = 3.141592653589793;\n  let ndotl = max(dot(n, l), 0.0);\n  let rgb = sun_rgb * (ndotl * shadow / PI_S) + ambient_rgb;\n  return luminance_rec709_linear(rgb);\n}\n\n// \u2500\u2500\u2500 BUMP node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Screen-space bump from a scalar height field. Needs dFdx/dFdy which\n// WGSL provides as dpdx/dpdy.\n\nfn bump(strength: f32, height: f32, normal: vec3f, world_pos: vec3f) -> vec3f {\n  let dhdx = dpdx(height);\n  let dhdy = dpdy(height);\n  let dpdx_pos = dpdx(world_pos);\n  let dpdy_pos = dpdy(world_pos);\n  let perturbed = normalize(normal) - strength * (dhdx * normalize(cross(dpdy_pos, normal)) + dhdy * normalize(cross(normal, dpdx_pos)));\n  return normalize(perturbed);\n}\n\n// LH engine + WebGPU fragment Y: flip dhdy contribution so height peaks read as outward bumps vs Blender reference\nfn bump_lh(strength: f32, height: f32, normal: vec3f, world_pos: vec3f) -> vec3f {\n  let dhdx = dpdx(height);\n  let dhdy = dpdy(height);\n  let dpdx_pos = dpdx(world_pos);\n  let dpdy_pos = dpdy(world_pos);\n  let perturbed = normalize(normal) - strength * (dhdx * normalize(cross(dpdy_pos, normal)) - dhdy * normalize(cross(normal, dpdx_pos)));\n  return normalize(perturbed);\n}\n\n// \u2500\u2500\u2500 NOISE texture (Perlin-style) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Simplified gradient noise matching Blender's default noise output.\n\n// PCG-style integer hash. Replaces the classic 'fract(sin(q) * LARGE)' trick because\n// WebKit's Metal backend compiles 'sin' to a full transcendental op (slow), while\n// Safari's Apple-GPU scalar ALU handles int muls/xors near free. Inputs arrive as\n// integer-valued floats (floor(p) + unit offsets) from _noise3, so vec3i cast is exact.\nfn _hash33(p: vec3f) -> vec3f {\n  var h = vec3u(vec3i(p) + vec3i(32768));\n  h = h * vec3u(1664525u, 1013904223u, 2654435761u);\n  h = (h.yzx ^ h) * vec3u(2246822519u, 3266489917u, 668265263u);\n  h = h ^ (h >> vec3u(16u));\n  // Mask to 24 bits \u2014 above that f32 loses precision on the u32\u2192f32 convert.\n  let hm = h & vec3u(16777215u);\n  return vec3f(hm) * (2.0 / 16777216.0) - 1.0;\n}\n\nfn _noise3(p: vec3f) -> f32 {\n  let i = floor(p);\n  let f = fract(p);\n  let u = f * f * (3.0 - 2.0 * f);\n\n  return mix(\n    mix(\n      mix(dot(_hash33(i + vec3f(0,0,0)), f - vec3f(0,0,0)),\n          dot(_hash33(i + vec3f(1,0,0)), f - vec3f(1,0,0)), u.x),\n      mix(dot(_hash33(i + vec3f(0,1,0)), f - vec3f(0,1,0)),\n          dot(_hash33(i + vec3f(1,1,0)), f - vec3f(1,1,0)), u.x), u.y),\n    mix(\n      mix(dot(_hash33(i + vec3f(0,0,1)), f - vec3f(0,0,1)),\n          dot(_hash33(i + vec3f(1,0,1)), f - vec3f(1,0,1)), u.x),\n      mix(dot(_hash33(i + vec3f(0,1,1)), f - vec3f(0,1,1)),\n          dot(_hash33(i + vec3f(1,1,1)), f - vec3f(1,1,1)), u.x), u.y),\n    u.z);\n}\n\nfn tex_noise(p: vec3f, scale: f32, detail: f32, roughness: f32, distortion: f32) -> f32 {\n  var q = p;\n  if (abs(distortion) > 1e-6) {\n    let w = _noise3(p * scale * 1.37 + vec3f(2.31, 5.17, 8.09));\n    q = p + (w * 2.0 - 1.0) * distortion;\n  }\n  let coords = q * scale;\n  var value = 0.0;\n  var amplitude = 1.0;\n  var frequency = 1.0;\n  var total_amp = 0.0;\n  let octaves = i32(clamp(detail, 0.0, 15.0)) + 1;\n  for (var i = 0; i < octaves; i++) {\n    value += amplitude * _noise3(coords * frequency);\n    total_amp += amplitude;\n    amplitude *= roughness;\n    frequency *= 2.0;\n  }\n  return value / max(total_amp, 1e-6) * 0.5 + 0.5;\n}\n\n// tex_noise specialization: detail=2.0 (3 octaves), roughness=0.5, distortion=0.\n// WebKit can't unroll tex_noise's for-loop because 'octaves' is a runtime value;\n// this variant is fully unrolled with constants folded (total_amp = 1.75).\nfn tex_noise_d2(p: vec3f, scale: f32) -> f32 {\n  let c = p * scale;\n  let v = _noise3(c) + 0.5 * _noise3(c * 2.0) + 0.25 * _noise3(c * 4.0);\n  return v * (1.0 / 1.75) * 0.5 + 0.5;\n}\n\n// \u2500\u2500\u2500 TEX_GRADIENT (linear) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Used by Stockings preset. Maps the input vector's X to a 0\u20131 gradient.\n\nfn tex_gradient_linear(uv: vec3f) -> f32 {\n  return clamp(uv.x, 0.0, 1.0);\n}\n\n// \u2500\u2500\u2500 TEX_VORONOI (distance only) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Used by Metal preset. Simplified F1 cell noise.\n\nfn tex_voronoi_f1(p: vec3f, scale: f32) -> f32 {\n  let coords = p * scale;\n  let i = floor(coords);\n  let f = fract(coords);\n  var min_dist = 1e10;\n  for (var z = -1; z <= 1; z++) {\n    for (var y = -1; y <= 1; y++) {\n      for (var x = -1; x <= 1; x++) {\n        let neighbor = vec3f(f32(x), f32(y), f32(z));\n        let point = _hash33(i + neighbor) * 0.5 + 0.5;\n        let diff = neighbor + point - f;\n        min_dist = min(min_dist, dot(diff, diff));\n      }\n    }\n  }\n  return sqrt(min_dist);\n}\n\n// \u2500\u2500\u2500 SEPXYZ node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn separate_xyz(v: vec3f) -> vec3f { return v; }\n\n// \u2500\u2500\u2500 VECT_MATH (cross product) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn vect_math_cross(a: vec3f, b: vec3f) -> vec3f { return cross(a, b); }\n\n// \u2500\u2500\u2500 MAPPING node \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Point-type mapping: scale, rotate (euler XYZ), translate.\n\nfn mapping_point(v: vec3f, loc: vec3f, rot: vec3f, scl: vec3f) -> vec3f {\n  var p = v * scl;\n  // simplified: skip rotation when all angles are zero (common case)\n  if (abs(rot.x) + abs(rot.y) + abs(rot.z) > 1e-6) {\n    let cx = cos(rot.x); let sx = sin(rot.x);\n    let cy = cos(rot.y); let sy = sin(rot.y);\n    let cz = cos(rot.z); let sz = sin(rot.z);\n    let rx = vec3f(p.x, cx*p.y - sx*p.z, sx*p.y + cx*p.z);\n    let ry = vec3f(cy*rx.x + sy*rx.z, rx.y, -sy*rx.x + cy*rx.z);\n    p = vec3f(cz*ry.x - sz*ry.y, sz*ry.x + cz*ry.y, ry.z);\n  }\n  return p + loc;\n}\n\n// \u2500\u2500\u2500 NORMAL_MAP node (tangent-space) \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Applies a tangent-space normal map. Requires TBN from vertex stage.\n\nfn normal_map(strength: f32, map_color: vec3f, normal: vec3f, tangent: vec3f, bitangent: vec3f) -> vec3f {\n  let ts = map_color * 2.0 - 1.0;\n  let perturbed = normalize(tangent * ts.x + bitangent * ts.y + normal * ts.z);\n  return normalize(mix(normal, perturbed, strength));\n}\n\n// \u2500\u2500\u2500 EEVEE Principled BSDF primitives \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n// Ports from Blender 3.6 source/blender/draw/engines/eevee/shaders/\n//   bsdf_common_lib.glsl + gpu_shader_material_principled.glsl.\n// Usage pattern (see material shaders): direct spec = bsdf_ggx \u00D7 sun \u00D7 shadow\n// (NL baked in, no F yet); ambient spec = probe_radiance; tint both with\n// reflection_color = F_brdf_multi_scatter(f0, f90, split_sum) AFTER summing.\n\nconst EEVEE_PI: f32 = 3.141592653589793;\n\n// Fused analytic GGX specular (direct lights). Returns BRDF \u00D7 NL.\n// 4\u00B7NL\u00B7NV is cancelled via G1_Smith reciprocal form \u2014 see bsdf_common_lib.glsl:115.\n// Caller passes NL, NV (already computed for diffuse + brdf_lut_sample) so WebKit\n// can reuse them instead of recomputing dot products across the function boundary.\nfn bsdf_ggx(N: vec3f, L: vec3f, V: vec3f, NL_in: f32, NV_in: f32, roughness: f32) -> f32 {\n  let a = max(roughness, 1e-4);\n  let a2 = a * a;\n  let H = normalize(L + V);\n  let NH = max(dot(N, H), 1e-8);\n  let NL = max(NL_in, 1e-8);\n  let NV = max(NV_in, 1e-8);\n  // G1_Smith_GGX_opti reciprocal form \u2014 denominator piece only.\n  let G1L = NL + sqrt(NL * (NL - NL * a2) + a2);\n  let G1V = NV + sqrt(NV * (NV - NV * a2) + a2);\n  let G = G1L * G1V;\n  // D_ggx_opti = pi * denom\u00B2 \u2014 reciprocal D \u00D7 a\u00B2.\n  let tmp = (NH * a2 - NH) * NH + 1.0;\n  let D_opti = EEVEE_PI * tmp * tmp;\n  return NL * a2 / (D_opti * G);\n}\n\n// Split-sum DFG LUT \u2014 Karis 2013 curve fit stand-in for the 64\u00D764 baked LUT.\n// Returns (lut.x, lut.y) in Blender convention: tint = f0\u00B7lut.x + f90\u00B7lut.y.\nfn brdf_lut_approx(NV: f32, roughness: f32) -> vec2f {\n  let c0 = vec4f(-1.0, -0.0275, -0.572, 0.022);\n  let c1 = vec4f(1.0, 0.0425, 1.04, -0.04);\n  let r = roughness * c0 + c1;\n  let a004 = min(r.x * r.x, exp2(-9.28 * NV)) * r.x + r.y;\n  return vec2f(-1.04, 1.04) * a004 + r.zw;\n}\n\n// Baked combined BRDF LUT \u2014 exact port of Blender bsdf_lut_frag.glsl packed with\n// ltc_mag_ggx from eevee_lut.c. Single sample returns DFG (.rg) and LTC mag (.ba).\n// Addressed as Blender's common_utiltex_lib.glsl:lut_coords:\n//   coords = (roughness, sqrt(1 - NV)), then half-texel bias for filtering.\n// Requires group(0) binding(9) brdfLut + binding(2) diffuseSampler in the host shader.\nfn brdf_lut_sample(NV: f32, roughness: f32) -> vec4f {\n  let LUT_SIZE: f32 = 64.0;\n  var uv = vec2f(saturate(roughness), sqrt(saturate(1.0 - NV)));\n  uv = uv * ((LUT_SIZE - 1.0) / LUT_SIZE) + 0.5 / LUT_SIZE;\n  return textureSampleLevel(brdfLut, diffuseSampler, uv, 0.0);\n}\n\nfn F_brdf_single_scatter(f0: vec3f, f90: vec3f, lut: vec2f) -> vec3f {\n  return lut.y * f90 + lut.x * f0;\n}\n\n// Fdez-Ag\u00FCera 2019 multi-scatter compensation (EEVEE do_multiscatter=1).\nfn F_brdf_multi_scatter(f0: vec3f, f90: vec3f, lut: vec2f) -> vec3f {\n  let FssEss = lut.y * f90 + lut.x * f0;\n  let Ess = lut.x + lut.y;\n  let Ems = 1.0 - Ess;\n  let Favg = f0 + (1.0 - f0) / 21.0;\n  let Fms = FssEss * Favg / (1.0 - (1.0 - Ess) * Favg);\n  return FssEss + Fms * Ems;\n}\n\n// EEVEE direct-specular energy compensation factor \u2014 closure_eval_glossy_lib.glsl:79-81:\n//   ltc_brdf_scale = (ltc.x + ltc.y) / (split_sum.x + split_sum.y)\n// Blender evaluates direct lights via LTC (Heitz 2016) but indirect via split-sum;\n// direct radiance is rescaled so total-energy matches the split-sum LUT.\n// Takes a pre-sampled vec4f from brdf_lut_sample() to share the fetch with\n// F_brdf_multi_scatter on the same fragment.\nfn ltc_brdf_scale_from_lut(lut: vec4f) -> f32 {\n  return (lut.z + lut.w) / max(lut.x + lut.y, 1e-6);\n}\n\n// Luminance-normalized hue extraction \u2014 Blender tint_from_color (isolates hue+sat).\nfn tint_from_color(color: vec3f) -> vec3f {\n  let lum = dot(color, vec3f(0.3, 0.6, 0.1));\n  return select(vec3f(1.0), color / lum, lum > 0.0);\n}\n\n\n\nstruct CameraUniforms {\n  view: mat4x4f,\n  projection: mat4x4f,\n  viewPos: vec3f,\n  _padding: f32,\n};\n\nstruct Light {\n  direction: vec4f,\n  color: vec4f,\n};\n\nstruct LightUniforms {\n  ambientColor: vec4f,\n  lights: array<Light, 4>,\n};\n\nstruct MaterialUniforms {\n  diffuseColor: vec3f,\n  alpha: f32,\n};\n\nstruct VertexOutput {\n  @builtin(position) position: vec4f,\n  @location(0) normal: vec3f,\n  @location(1) uv: vec2f,\n  @location(2) worldPos: vec3f,\n};\n\nstruct LightVP { viewProj: mat4x4f, };\n\n@group(0) @binding(0) var<uniform> camera: CameraUniforms;\n@group(0) @binding(1) var<uniform> light: LightUniforms;\n@group(0) @binding(2) var diffuseSampler: sampler;\n@group(0) @binding(3) var shadowMap: texture_depth_2d;\n@group(0) @binding(4) var shadowSampler: sampler_comparison;\n@group(0) @binding(5) var<uniform> lightVP: LightVP;\n@group(1) @binding(0) var<storage, read> skinMats: array<mat4x4f>;\n@group(2) @binding(0) var diffuseTexture: texture_2d<f32>;\n@group(2) @binding(1) var<uniform> material: MaterialUniforms;\n\nfn sampleShadow(worldPos: vec3f, n: vec3f) -> f32 {\n  // Back-facing to key light: direct contribution is zero anyway, skip 9 texture samples.\n  if (dot(n, -light.lights[0].direction.xyz) <= 0.0) { return 0.0; }\n  let biasedPos = worldPos + n * 0.08;\n  let lclip = lightVP.viewProj * vec4f(biasedPos, 1.0);\n  let ndc = lclip.xyz / max(lclip.w, 1e-6);\n  let suv = vec2f(ndc.x * 0.5 + 0.5, 0.5 - ndc.y * 0.5);\n  let cmpZ = ndc.z - 0.001;\n  let ts = 1.0 / 2048.0;\n  // 3x3 PCF unrolled \u2014 Safari's Metal backend doesn't unroll nested shadow loops reliably.\n  let s00 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, -ts), cmpZ);\n  let s10 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0, -ts), cmpZ);\n  let s20 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, -ts), cmpZ);\n  let s01 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, 0.0), cmpZ);\n  let s11 = textureSampleCompareLevel(shadowMap, shadowSampler, suv, cmpZ);\n  let s21 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, 0.0), cmpZ);\n  let s02 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts,  ts), cmpZ);\n  let s12 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0,  ts), cmpZ);\n  let s22 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts,  ts), cmpZ);\n  return (s00 + s10 + s20 + s01 + s11 + s21 + s02 + s12 + s22) * (1.0 / 9.0);\n}\n\nconst PI_S: f32 = 3.141592653589793;\n// Principled BSDF params from dump (Alpha=0.95 is intentionally dropped \u2014 see alpha-hash note below)\nconst STOCK_METALLIC: f32 = 0.1;\nconst STOCK_SPECULAR: f32 = 1.0;\nconst STOCK_ROUGHNESS: f32 = 0.5;\nconst STOCK_SHEEN: f32 = 0.7017999887466431;\nconst STOCK_SHEEN_TINT: f32 = 0.5;\n// NPR mask ramps\nconst STOCK_RAMP002_P1: f32 = 0.9565;  // EASE [0\u2192black, 0.9565\u2192white]\nconst STOCK_RAMPFACE_P1: f32 = 0.5435; // EASE [0\u2192black, 0.5435\u2192white]\nconst STOCK_LW_BLEND: f32 = 0.4;       // Layer Weight Blend\n\n// principled_sheen (gpu_shader_material_principled.glsl:8-14) \u2014 empirical NV curve\nfn principled_sheen(NV: f32) -> f32 {\n  let f = 1.0 - NV;\n  return f * f * f * 0.077 + f * 0.01 + 0.00026;\n}\n\n// Wyman & McGuire \"Hashed Alpha Testing\" (2017) \u2014 world-space hash with derivative-aware\n// pixel-scale selection, matches Blender EEVEE prepass_frag.glsl::hashed_alpha_threshold.\n// Key property: dither pattern is stable in object/world space (doesn't swim) and stays\n// at one-pixel frequency regardless of view distance, which makes it tolerable without TAA.\nfn _hash_wm(a: vec2f) -> f32 {\n  return fract(1e4 * sin(17.0 * a.x + 0.1 * a.y) * (0.1 + abs(sin(13.0 * a.y + a.x))));\n}\nfn _hash3d_wm(a: vec3f) -> f32 {\n  return _hash_wm(vec2f(_hash_wm(a.xy), a.z));\n}\nfn hashed_alpha_threshold(co: vec3f) -> f32 {\n  let alphaHashScale: f32 = 1.0;\n  let max_deriv = max(length(dpdx(co)), length(dpdy(co)));\n  let pix_scale = 1.0 / max(alphaHashScale * max_deriv, 1e-6);\n  let pix_scale_log = log2(pix_scale);\n  let px_lo = exp2(floor(pix_scale_log));\n  let px_hi = exp2(ceil(pix_scale_log));\n  let a_lo = _hash3d_wm(floor(px_lo * co));\n  let a_hi = _hash3d_wm(floor(px_hi * co));\n  let fac = fract(pix_scale_log);\n  let x = mix(a_lo, a_hi, fac);\n  // CDF remap so that discard-probability = (1 - alpha) uniformly across scale transitions\n  let a = min(fac, 1.0 - fac);\n  let one_a = 1.0 - a;\n  let denom = 1.0 / max(2.0 * a * one_a, 1e-6);\n  let one_x = 1.0 - x;\n  let case_lo = (x * x) * denom;\n  let case_mid = (x - 0.5 * a) / max(one_a, 1e-6);\n  let case_hi = 1.0 - (one_x * one_x) * denom;\n  var threshold = select(case_hi, select(case_lo, case_mid, x >= a), x < one_a);\n  return clamp(threshold, 1e-6, 1.0);\n}\n\n// Smoothstep-based EASE ramp (Blender VALTORGB EASE) \u2014 2 stops, saturate+smoothstep between\nfn ramp_ease_s(f: f32, p0: f32, p1: f32) -> f32 {\n  let t = saturate((f - p0) / max(p1 - p0, 1e-6));\n  return t * t * (3.0 - 2.0 * t);\n}\n\n@vertex fn vs(\n  @location(0) position: vec3f,\n  @location(1) normal: vec3f,\n  @location(2) uv: vec2f,\n  @location(3) joints0: vec4<u32>,\n  @location(4) weights0: vec4<f32>\n) -> VertexOutput {\n  var output: VertexOutput;\n  let pos4 = vec4f(position, 1.0);\n  let weightSum = weights0.x + weights0.y + weights0.z + weights0.w;\n  let invWeightSum = select(1.0, 1.0 / weightSum, weightSum > 0.0001);\n  let nw = select(vec4f(1.0, 0.0, 0.0, 0.0), weights0 * invWeightSum, weightSum > 0.0001);\n  var skinnedPos = vec4f(0.0);\n  var skinnedNrm = vec3f(0.0);\n  for (var i = 0u; i < 4u; i++) {\n    let m = skinMats[joints0[i]];\n    let w = nw[i];\n    skinnedPos += (m * pos4) * w;\n    skinnedNrm += (mat3x3f(m[0].xyz, m[1].xyz, m[2].xyz) * normal) * w;\n  }\n  output.position = camera.projection * camera.view * vec4f(skinnedPos.xyz, 1.0);\n  // Skip VS normalize \u2014 interpolation denormalizes anyway, and FS always does normalize(input.normal).\n  output.normal = skinnedNrm;\n  output.uv = uv;\n  output.worldPos = skinnedPos.xyz;\n  return output;\n}\n\nstruct FSOut {\n  @location(0) color: vec4f,\n  @location(1) bloom_mask: f32,\n};\n\n@fragment fn fs(input: VertexOutput) -> FSOut {\n  let n = normalize(input.normal);\n  let v = normalize(camera.viewPos - input.worldPos);\n  let l = -light.lights[0].direction.xyz;\n  let sun = light.lights[0].color.xyz * light.lights[0].color.w;\n  let amb = light.ambientColor.xyz;\n  let shadow = sampleShadow(input.worldPos, n);\n\n  let tex_s = textureSample(diffuseTexture, diffuseSampler, input.uv);\n  let tex_rgb = tex_s.rgb;\n  // Alpha HASHED (Blender EEVEE \"Hashed\" blend mode) per preset author's note \u2014\n  // self-overlap on the stockings produces sort cracks under alpha blend. Wyman-style\n  // worldPos hash + depth-write is sort-independent. NOTE: Principled.Alpha=0.95 from\n  // the dump is DROPPED here \u2014 it relies on TAA to smooth the resulting 5%-everywhere\n  // dither, and without TAA it shows as a pervasive dot pattern. Hash now gates only\n  // on texture/material alpha, so solid stockings regions stay fully opaque.\n  let combined_alpha = material.alpha * tex_s.a;\n  if (combined_alpha < hashed_alpha_threshold(input.worldPos)) { discard; }\n  let out_alpha = 1.0;\n\n  // \u2550\u2550\u2550 NPR MASK: TEX_COORD.Generated \u2192 Mapping(Rot=0,\u03C0/2,\u03C0/2, Loc=(1,1,1)) \u2192 Gradient Texture\n  // The Blender mapping reduces to gradient.x = 1 - input.y (rot swaps axes, loc offsets by 1).\n  // We approximate Generated with UV since Y-up PMX has no object bbox in pipeline state.\n  let gen_coord = vec3f(input.uv, 0.0);\n  let mapped = mapping_point(gen_coord, vec3f(1.0), vec3f(0.0, 1.5708, 1.5708), vec3f(1.0));\n  let gradient = tex_gradient_linear(mapped);\n\n  // Ramp.001 LINEAR [0\u2192black, 0.5\u2192white, 1.0\u2192black] \u2014 triangular peak at 0.5\n  let ramp001 = 1.0 - abs(2.0 * gradient - 1.0);\n  // Ramp.002 EASE [0\u2192black, 0.9565\u2192white]\n  let ramp002 = ramp_ease_s(ramp001, 0.0, STOCK_RAMP002_P1);\n\n  // Layer Weight.Facing (Blend=0.4) \u2192 Ramp EASE [0\u2192black, 0.5435\u2192white]\n  let facing = layer_weight_facing(STOCK_LW_BLEND, n, v);\n  let ramp_face = ramp_ease_s(facing, 0.0, STOCK_RAMPFACE_P1);\n\n  // Mix.001: MIX blend Fac=0.5, A=white, B=ramp_face \u2192 (A,B) averaged 50/50\n  let mix001 = mix(1.0, ramp_face, 0.5);\n  // Mix: LIGHTEN blend Fac=0.5, A=mix001, B=ramp002 \u2192 A smoothly lightens toward max(A,B)\n  let lighten = max(mix001, ramp002);\n  let mask = mix(mix001, lighten, 0.5);\n\n  // \u2550\u2550\u2550 EMISSION SHADER \u2550\u2550\u2550\n  // Hue=0.5 (identity rotation), Sat=1.0, Val=5.0 (5\u00D7 brightness boost), Fac=1; Strength=1\n  let emission = hue_sat_id(1.0, 5.0, 1.0, tex_rgb);\n\n  // \u2550\u2550\u2550 PRINCIPLED BSDF (EEVEE port) \u2550\u2550\u2550\n  // base_color_tint, metallic f0, sheen coarse approx (scales diffuse radiance).\n  let NL = max(dot(n, l), 0.0);\n  let NV = max(dot(n, v), 1e-4);\n\n  // f0 = mix((0.08*spec)*dielectric_tint, base, metallic); dielectric_tint=1 since specular_tint=0.\n  let dielectric_f0 = vec3f(0.08 * STOCK_SPECULAR);\n  let f0 = mix(dielectric_f0, tex_rgb, STOCK_METALLIC);\n  let f90 = mix(f0, vec3f(1.0), sqrt(STOCK_SPECULAR));\n  let brdf_lut = brdf_lut_sample(NV, STOCK_ROUGHNESS);\n  let reflection_color = F_brdf_multi_scatter(f0, f90, brdf_lut.xy);\n\n  let spec_direct = bsdf_ggx(n, l, v, NL, NV, STOCK_ROUGHNESS) * sun * shadow * ltc_brdf_scale_from_lut(brdf_lut);\n  let spec_indirect = amb;\n  let spec_radiance = (spec_direct + spec_indirect) * reflection_color;\n\n  // Sheen coarse: diffuse_color += sheen * sheen_color * principled_sheen(NV).\n  let base_tint = tint_from_color(tex_rgb);\n  let sheen_color = mix(vec3f(1.0), base_tint, STOCK_SHEEN_TINT);\n  let diffuse_color = tex_rgb + STOCK_SHEEN * sheen_color * principled_sheen(NV);\n\n  // diffuse_weight = (1 - metallic). Indirect diffuse uses L_w (no \u03C0; see closure_eval_surface_lib:302).\n  let diffuse_weight = 1.0 - STOCK_METALLIC;\n  let diffuse_radiance = diffuse_color * (sun * NL * shadow / PI_S + amb) * diffuse_weight;\n  let principled = diffuse_radiance + spec_radiance;\n\n  // \u2550\u2550\u2550 MIX SHADER: Shader=Emission, Shader_001=Principled, Fac=mask \u2550\u2550\u2550\n  let final_color = mix(emission, principled, mask);\n\n  var out: FSOut;\n  out.color = vec4f(final_color, out_alpha);\n  out.bloom_mask = 1.0;\n  return out;\n}\n\n";
 //# sourceMappingURL=stockings.d.ts.map

package/dist/shaders/stockings.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stockings.d.ts","sourceRoot":"","sources":["../../src/shaders/stockings.ts"],"names":[],"mappings":"AAOA,eAAO,MAAM,qBAAqB,g/gCA8OjC,CAAA"}
+{"version":3,"file":"stockings.d.ts","sourceRoot":"","sources":["../../src/shaders/stockings.ts"],"names":[],"mappings":"AAOA,eAAO,MAAM,qBAAqB,o+/BA8OjC,CAAA"}

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "reze-engine",
-    "version": "0.12.2",
+    "version": "0.12.3",
     "description": "A lightweight WebGPU engine for real-time 3D MMD/PMX model rendering",
     "main": "./dist/index.js",
     "types": "./dist/index.d.ts",

package/src/engine.ts

@@ -247,14 +247,35 @@ export class Engine {
   private multisampleTexture!: GPUTexture
   private hdrResolveTexture!: GPUTexture
   private static readonly MULTISAMPLE_COUNT = 4
-  private static readonly HDR_FORMAT: GPUTextureFormat = "rgba16float"
+  // HDR intermediate format. rg11b10ufloat when the adapter exposes the
+  // `rg11b10ufloat-renderable` feature (Chrome + Safari on Apple Silicon both
+  // do), else fall back to rgba16float.
+  //
+  // Why it matters — Apple TBDR tile memory: rgba16float is 8 bytes/texel, so
+  // 4× MSAA is 32 bytes/texel and does not fit Apple Silicon's tile memory at
+  // useful tile sizes. The driver then stores the full MSAA buffer to system
+  // memory every frame and resolves from there — ~300 MB/frame of extra
+  // bandwidth at 1920×1200 DPR=2, which is the dominant frame-pacing hit on
+  // Safari (visibly: shrinking the window made Safari smooth; Chrome was
+  // always smooth because Dawn apparently amortizes it). rg11b10ufloat at
+  // 4 bytes/texel → 16 bytes/texel at 4× MSAA → fits tile memory like
+  // rgba8unorm does, resolves in-tile, no system-memory round-trip. No alpha
+  // channel (the HDR path never needed one — alpha blending reads src.a from
+  // the fragment shader and treats missing dst.a as 1, so the blend math is
+  // unchanged).
+  private hdrFormat: GPUTextureFormat = "rgba16float"
   /** Stencil value stamped by eye draws so hair can stencil-test against it and
    *  alpha-blend a second pass over eye silhouette pixels (see-through-hair effect). */
   private static readonly STENCIL_EYE_VALUE = 1
-  /** Single-channel mask written alongside HDR color — 1 = model geometry (contributes
-   *  to bloom), 0 = ground (never blooms). Sampled by the bloom blit pass to gate the
-   *  prefilter so ground brightness can't halo the scene. */
-  private static readonly BLOOM_MASK_FORMAT: GPUTextureFormat = "r8unorm"
+  /** Aux MRT alongside HDR color. Two channels:
+   *   .r — bloom mask (1 = model geometry, 0 = ground; sampled by bloom blit to gate prefilter).
+   *   .g — accumulated alpha (the channel that used to live in hdr.a before the HDR format
+   *        switched to rg11b10ufloat, which has no alpha). Sampled by composite/bloom to
+   *        un-premultiply color for tonemap and to produce the canvas-drawable alpha used by
+   *        the premultiplied alphaMode compositor (so the page background still shows through
+   *        cleared / edge-faded regions like before).
+   *  rg8unorm at 4× MSAA is 8 bytes/texel — still fits Apple TBDR tile memory comfortably. */
+  private static readonly BLOOM_MASK_FORMAT: GPUTextureFormat = "rg8unorm"
   private multisampleMaskTexture!: GPUTexture
   private maskResolveTexture!: GPUTexture
   private maskResolveView!: GPUTextureView
@@ -502,11 +523,17 @@ export class Engine {
   // Step 1: Get WebGPU device and context
   async init() {
     const adapter = await navigator.gpu?.requestAdapter()
-    const device = await adapter?.requestDevice()
+    if (!adapter) throw new Error("WebGPU is not supported in this browser.")
+    const wantFeature: GPUFeatureName = "rg11b10ufloat-renderable"
+    const hasRg11b10 = adapter.features.has(wantFeature)
+    const device = await adapter.requestDevice({
+      requiredFeatures: hasRg11b10 ? [wantFeature] : [],
+    })
     if (!device) {
       throw new Error("WebGPU is not supported in this browser.")
     }
     this.device = device
+    if (hasRg11b10) this.hdrFormat = "rg11b10ufloat"
 
     const context = this.canvas.getContext("webgpu")
     if (!context) {
@@ -694,7 +721,7 @@ export class Engine {
     // composite pass writes the swapchain. Tonemap moved to composite so bloom
     // (added next) can run on linear HDR.
     const standardBlend: GPUColorTargetState = {
-      format: Engine.HDR_FORMAT,
+      format: this.hdrFormat,
       blend: {
         color: {
           srcFactor: "src-alpha",
@@ -709,10 +736,21 @@ export class Engine {
       },
     }
 
-    // Bloom mask target — r8unorm has no alpha channel, so src-alpha blending is invalid.
-    // Use replace mode: depth test already rejects occluded fragments, so last-writer-wins
-    // on surviving pixels gives the right result (ground writes 0; models/outlines write 1).
-    const maskBlend: GPUColorTargetState = { format: Engine.BLOOM_MASK_FORMAT }
+    // Aux target carrying (bloom mask, alpha). Src-alpha blend so the .g channel
+    // accumulates proper alpha-over (same semantic the old rgba16f hdr.a had).
+    // Materials write vec2f(mask, 1.0); ground writes vec2f(0.0, 1.0). With src.a
+    // coming from the fragment color.a, the blend equation produces
+    //   out.g = 1·src.a + dst.g·(1-src.a)  →  premultiplied over operator on alpha.
+    // .r gets weighted by src.a too, which is fine: opaque pixels (α=1) give full
+    // mask, partially translucent fragments dilute mask proportionally — acceptable
+    // for the bloom-gate use.
+    const maskBlend: GPUColorTargetState = {
+      format: Engine.BLOOM_MASK_FORMAT,
+      blend: {
+        color: { srcFactor: "src-alpha", dstFactor: "one-minus-src-alpha", operation: "add" },
+        alpha: { srcFactor: "one", dstFactor: "one-minus-src-alpha", operation: "add" },
+      },
+    }
     const sceneTargets: GPUColorTargetState[] = [standardBlend, maskBlend]
 
     const shaderModule = this.device.createShaderModule({
@@ -1184,21 +1222,21 @@ export class Engine {
       label: "bloom blit pipeline",
       layout: bloomBlitLayout,
       vertex: { module: bloomBlitShader, entryPoint: "vs" },
-      fragment: { module: bloomBlitShader, entryPoint: "fs", targets: [{ format: Engine.HDR_FORMAT }] },
+      fragment: { module: bloomBlitShader, entryPoint: "fs", targets: [{ format: this.hdrFormat }] },
       primitive: { topology: "triangle-list" },
     })
     this.bloomDownsamplePipeline = this.device.createRenderPipeline({
       label: "bloom downsample pipeline",
       layout: bloomDownLayout,
       vertex: { module: bloomDownsampleShader, entryPoint: "vs" },
-      fragment: { module: bloomDownsampleShader, entryPoint: "fs", targets: [{ format: Engine.HDR_FORMAT }] },
+      fragment: { module: bloomDownsampleShader, entryPoint: "fs", targets: [{ format: this.hdrFormat }] },
       primitive: { topology: "triangle-list" },
     })
     this.bloomUpsamplePipeline = this.device.createRenderPipeline({
       label: "bloom upsample pipeline",
       layout: bloomUpLayout,
       vertex: { module: bloomUpsampleShader, entryPoint: "vs" },
-      fragment: { module: bloomUpsampleShader, entryPoint: "fs", targets: [{ format: Engine.HDR_FORMAT }] },
+      fragment: { module: bloomUpsampleShader, entryPoint: "fs", targets: [{ format: this.hdrFormat }] },
       primitive: { topology: "triangle-list" },
     })
 
@@ -1217,6 +1255,9 @@ export class Engine {
         { binding: 1, visibility: GPUShaderStage.FRAGMENT, texture: {} },
         { binding: 2, visibility: GPUShaderStage.FRAGMENT, sampler: {} },
         { binding: 3, visibility: GPUShaderStage.FRAGMENT, buffer: { type: "uniform" } },
+        // Aux mask/alpha texture — composite reads .g to reconstruct the alpha that
+        // used to live in the HDR target before the rg11b10ufloat switch.
+        { binding: 4, visibility: GPUShaderStage.FRAGMENT, texture: {} },
       ],
     })
 
@@ -1341,14 +1382,14 @@ export class Engine {
         label: "multisample HDR render target",
         size: [width, height],
         sampleCount: Engine.MULTISAMPLE_COUNT,
-        format: Engine.HDR_FORMAT,
+        format: this.hdrFormat,
         usage: GPUTextureUsage.RENDER_ATTACHMENT,
       })
 
       this.hdrResolveTexture = this.device.createTexture({
         label: "HDR resolve target",
         size: [width, height],
-        format: Engine.HDR_FORMAT,
+        format: this.hdrFormat,
         usage: GPUTextureUsage.RENDER_ATTACHMENT | GPUTextureUsage.TEXTURE_BINDING,
       })
 
@@ -1379,14 +1420,14 @@ export class Engine {
         label: "bloom down pyramid",
         size: [bw, bh],
         mipLevelCount: this.bloomMipCount,
-        format: Engine.HDR_FORMAT,
+        format: this.hdrFormat,
         usage: GPUTextureUsage.RENDER_ATTACHMENT | GPUTextureUsage.TEXTURE_BINDING,
       })
       this.bloomUpTexture = this.device.createTexture({
         label: "bloom up pyramid",
         size: [bw, bh],
         mipLevelCount: Math.max(1, this.bloomMipCount - 1),
-        format: Engine.HDR_FORMAT,
+        format: this.hdrFormat,
         usage: GPUTextureUsage.RENDER_ATTACHMENT | GPUTextureUsage.TEXTURE_BINDING,
       })
       this.bloomDownMipViews = []
@@ -1513,6 +1554,7 @@ export class Engine {
             { binding: 1, resource: compositeBloomView },
             { binding: 2, resource: this.bloomSampler },
             { binding: 3, resource: { buffer: this.compositeUniformBuffer } },
+            { binding: 4, resource: this.maskResolveView },
           ],
         })
       }

package/src/shaders/materials/body.ts

@@ -83,7 +83,7 @@ fn ramp_ease(f: f32, p0: f32, c0: vec4f, p1: f32, c1: vec4f) -> vec4f {
 
   var out: FSOut;
   out.color = vec4f(final_color, alpha);
-  out.mask = 1.0;
+  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);
   return out;
 }
 

package/src/shaders/materials/cloth_rough.ts

@@ -62,7 +62,7 @@ const CLOTH_R_SPEC_CLAMP: f32 = 10.0;
 
   var out: FSOut;
   out.color = vec4f(final_color, out_alpha);
-  out.mask = 1.0;
+  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);
   return out;
 }
 

package/src/shaders/materials/cloth_smooth.ts

@@ -56,7 +56,7 @@ const NPR_MIX_SHADER_FAC: f32 = 0.8999999761581421;
 
   var out: FSOut;
   out.color = vec4f(final_color, out_alpha);
-  out.mask = 1.0;
+  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);
   return out;
 }
 

package/src/shaders/materials/common.ts

@@ -135,14 +135,30 @@ export const COMMON_VS_WGSL = /* wgsl */ `
 `;
 
 // ─── FS output struct ───────────────────────────────────────────────
-// Location 0: final radiance+alpha. Location 1: bloom/highlight mask (single f32,
-// currently 1.0 everywhere — slot is reserved for future per-material bloom control).
+// Location 0: final radiance+alpha (blended into rg11b10ufloat; the HDR target
+// has no alpha channel, but the blend equation still uses the .a you write here
+// as the src-alpha factor that premultiplies rgb into the HDR target).
+// Location 1: auxiliary rg8unorm carrying
+//   .r = bloom mask (1 = contributes to bloom, 0 = skip — e.g. ground).
+//   .g = accumulated canvas alpha — the channel that used to live in hdr.a
+//        before the switch to rg11b10ufloat. Sampled by composite to
+//        un-premultiply color for tonemap and to set the final drawable alpha
+//        (needed for the `premultiplied` canvas alphaMode that blends the
+//        WebGPU surface over the page background).
+// FS output at location 1 must be vec4f — the blend state references src.a, and
+// WebGPU requires the fragment output to provide an alpha component even though
+// the rg8unorm target only stores .r and .g (extra components are discarded).
+// Materials write mask = vec4f(1.0, 1.0, 0.0, color.a); ground writes
+// vec4f(0.0, 1.0, 0.0, edgeFade). With src.a coming from the 4th component and
+// src-alpha blending enabled:
+//   out.r = mask_r · src.a + dst.r · (1-src.a)   (bloom mask, weighted by alpha)
+//   out.g = 1.0    · src.a + dst.g · (1-src.a)   (canonical premultiplied alpha-over)
 
 export const COMMON_FS_OUT_WGSL = /* wgsl */ `
 
 struct FSOut {
   @location(0) color: vec4f,
-  @location(1) mask: f32,
+  @location(1) mask: vec4f,
 };
 
 `;

package/src/shaders/materials/default.ts

@@ -33,7 +33,7 @@ const DEFAULT_ROUGHNESS: f32 = 0.5;
 
   var out: FSOut;
   out.color = vec4f(color, alpha);
-  out.mask = 1.0;
+  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);
   return out;
 }
 

package/src/shaders/materials/eye.ts

@@ -36,7 +36,7 @@ const EYE_EMISSION_STRENGTH: f32 = 1.5;
 
   var out: FSOut;
   out.color = vec4f(shaded + emission, alpha);
-  out.mask = 1.0;
+  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);
   return out;
 }
 

package/src/shaders/materials/face.ts

@@ -82,7 +82,7 @@ const FACE_BRIGHT_TEX_THRESH: f32 = 0.9300000071525574;
 
   var out: FSOut;
   out.color = vec4f(final_color, alpha);
-  out.mask = 1.0;
+  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);
   return out;
 }
 

package/src/shaders/materials/hair.ts

@@ -75,7 +75,7 @@ const HAIR_MIX_NPR: f32 = 0.2;
 
   var out: FSOut;
   out.color = vec4f(final_color, outAlpha);
-  out.mask = 1.0;
+  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);
   return out;
 }
 

package/src/shaders/materials/metal.ts

@@ -72,7 +72,7 @@ const METAL_VORONOI_SCALE: f32 = 4.3;
 
   var out: FSOut;
   out.color = vec4f(final_color, out_alpha);
-  out.mask = 1.0;
+  out.mask = vec4f(1.0, 1.0, 0.0, out.color.a);
   return out;
 }