@forgeax/engine-shader 0.1.20 → 0.1.23

package/src/ibl-sampling.wgsl

@@ -16,8 +16,10 @@
 //   - sampleIblDiffuse(N, irradianceMap, irradianceSampler)
 //   - sampleIblSpecular(N, V, roughness, F0, prefilterMap, prefilterSampler,
 //                       brdfLut, brdfLutSampler)
+//   - box_project(worldPosition, direction, boxCenter, boxExtents)
+//   - sampleReflectionProbeSpecular(..., probeMap, probeSampler, ...)
 
-#import forgeax_pbr::ibl_shared::{PI, fresnelSchlickRoughness, inverseRotateEnvironment}
+#import forgeax_pbr::ibl_shared::{fresnelSchlickRoughness, inverseRotateEnvironment}
 
 // Sample pre-convolved irradiance from the irradiance cubemap.
 // Y is negated to compensate for WebGPU's top-left texture origin vs the
@@ -31,7 +33,7 @@ fn sampleIblDiffuse(
   let rotated = inverseRotateEnvironment(normal, rotation);
   let dir = vec3<f32>(rotated.x, -rotated.y, rotated.z);
   let irradianceEOverPi = textureSample(irradianceMap, irradianceSampler, dir).rgb;
-  return irradianceEOverPi / PI;
+  return irradianceEOverPi;
 }
 
 // Split-sum specular IBL: prefiltered env * (F0 * scale + bias).
@@ -56,3 +58,55 @@ fn sampleIblSpecular(
   let F = fresnelSchlickRoughness(NdotV, F0, roughness);
   return prefilteredColor * (F * envBRDF.r + envBRDF.g);
 }
+
+// Project a reflection ray from a point inside a probe box onto the box
+// boundary. The helper is binding-free so the scene table and the material
+// resource projection remain owned by the host record stage.
+fn box_project(
+  worldPosition: vec3<f32>,
+  direction: vec3<f32>,
+  boxCenter: vec3<f32>,
+  boxExtents: vec3<f32>,
+) -> vec3<f32> {
+  let safeExtents = max(boxExtents, vec3<f32>(0.0001));
+  let safeDirection = select(
+    vec3<f32>(0.0001),
+    direction,
+    abs(direction) >= vec3<f32>(0.0001),
+  );
+  let localPosition = worldPosition - boxCenter;
+  let edgeSign = select(vec3<f32>(-1.0), vec3<f32>(1.0), direction >= vec3<f32>(0.0));
+  let edge = edgeSign * safeExtents;
+  let distances = (edge - localPosition) / safeDirection;
+  let travel = min(distances.x, min(distances.y, distances.z));
+  return normalize(localPosition + direction * max(travel, 0.0));
+}
+
+// Split-sum probe sample. The probe texture is supplied by the caller so a
+// missing scene candidate can keep using sampleIblSpecular with the Skylight
+// resources already present in the Standard material layout.
+fn sampleReflectionProbeSpecular(
+  normal: vec3<f32>,
+  view: vec3<f32>,
+  roughness: f32,
+  F0: vec3<f32>,
+  worldPosition: vec3<f32>,
+  boxCenter: vec3<f32>,
+  boxExtents: vec3<f32>,
+  rotation: vec4<f32>,
+  probeMap: texture_cube<f32>,
+  probeSampler: sampler,
+  brdfLut: texture_2d<f32>,
+  brdfLutSampler: sampler,
+) -> vec3<f32> {
+  let NdotV = max(dot(normal, view), 0.001);
+  let reflection = reflect(-view, normal);
+  let projected = box_project(worldPosition, reflection, boxCenter, boxExtents);
+  let rotated = inverseRotateEnvironment(projected, rotation);
+  let probeDirection = vec3<f32>(rotated.x, -rotated.y, rotated.z);
+  let mip = roughness * 4.0;
+  let prefilteredColor = textureSampleLevel(probeMap, probeSampler, probeDirection, mip).rgb;
+  let envBRDF = textureSample(brdfLut, brdfLutSampler, vec2<f32>(NdotV, roughness)).rg;
+  let F = fresnelSchlickRoughness(NdotV, F0, roughness);
+  return prefilteredColor * (F * envBRDF.r + envBRDF.g);
+}

package/src/index.ts

@@ -39,6 +39,16 @@ export {
   type ShaderErrorDetail,
   shaderNotFound,
 } from './errors.js';
+export {
+  generateLtcTables,
+  hashLtcTable,
+  LTC_SOURCE_PROVENANCE,
+  LTC_TABLE_HASHES,
+  LTC_TABLE_HEIGHT,
+  LTC_TABLE_INPUT_HASHES,
+  LTC_TABLE_WIDTH,
+  LTC_TABLES,
+} from './ltc/tables.js';
 export {
   type MaterialArtifactConflictError,
   type MaterialArtifactInspection,
@@ -46,8 +56,12 @@ export {
   type MaterialRuntimeArtifact,
 } from './material/artifact-registry.js';
 export {
+  createStandardPbrArtifactReceipt,
   isMaterialShaderArtifact,
   type MaterialShaderArtifact,
+  type MaterialShaderArtifactReceipt,
+  type MaterialShaderResourceSlot,
+  type MaterialShaderVertexInput,
 } from './material/artifact-types.js';
 export {
   DEFAULT_MSDF_TEXT_PARAM_SCHEMA,
@@ -55,12 +69,15 @@ export {
   DEFAULT_STANDARD_PBR_PARAM_SCHEMA,
   DEFAULT_UNLIT_PARAM_SCHEMA,
   STANDARD_PBR_ALPHA_CUTOFF_DEFAULT,
+  STANDARD_PBR_ARTIFACT_RECEIPT,
+  STANDARD_PBR_SKIN_ARTIFACT_RECEIPT,
 } from './material-schemas.js';
 export {
   registerDefaultSpriteLit,
   type SpriteLitCaps,
 } from './register-default-sprite-lit.js';
 export { registerDefaultStandardPbrSkin } from './register-default-standard-pbr-skin.js';
+export const RECT_AREA_LTC_SHADER_MODULE = 'forgeax::lighting-rect-area' as const;
 export {
   FORGEAX_RESERVED_PATH_PREFIX,
   type MaterialShaderEntry,
@@ -108,6 +125,12 @@ export const ENGINE_MATERIAL_MODULES = [
   'forgeax_material::sprite-lit',
 ] as const;
 
+/** Names of the zero-binding Standard reflection-probe shader helpers. */
+export const REFLECTION_PROBE_SHADER_HELPERS = Object.freeze([
+  'box_project',
+  'sampleReflectionProbeSpecular',
+] as const);
+
 export function isEngineMaterialModule(module: string): boolean {
   return (ENGINE_MATERIAL_MODULES as readonly string[]).includes(module);
 }

package/src/lighting-directional.wgsl

@@ -46,128 +46,43 @@
 //     layered BRDFs such as base + clearcoat)
 
 #import forgeax_view::common::{view, shadowMap, shadowSampler}
-#import forgeax_pbr::brdf::{f_schlick, v_smith, d_ggx}
+#import forgeax_pbr::brdf::{f_schlick, v_smith, d_ggx, threeR184DirectMultiScatter}
 #import forgeax_pbr::shadow_pcf::{sample_shadow_2d}
+#ifdef DIRECTIONAL_PCSS_AVAILABLE
+#import forgeax_pbr::shadow_pcf::{shadow_biased_receiver_depth, shadow_clamp_texel_to_tile, shadow_load_raw_depth, shadow_sample_compare}
+#endif
 
 // feat-20260621-learn-render-5-3-production-shadow-demos M0 / AC-14:
 // compile-time upper bound on the PCF half-extent so the WGSL tap loops keep
 // a constant trip count (no dynamic loop bounds / shader variants). half=2
-// covers pcfKernelSize in {1,3,5} -> {1,9,25} taps; view.pcfKernelSize selects
-// the runtime radius via a per-iteration clip (plan-strategy D-1).
+// covers the PCF1/3/5 profiles -> {1,9,25} taps; the profile carrier selects
+// the runtime radius while keeping the receiver variant-free.
 const MAX_PCF_HALF : u32 = 2u;
 
-// Source: Three.js r184 src/nodes/functions/BSDF/DFGLUT.js. The full 16x16
-// RG16F table is decoded to exact f32 constants here; the lookup below mirrors
-// the source DataTexture's linear filtering and clamp-to-edge sampling without
-// adding a host binding or coupling direct light to the engine IBL LUT.
-const THREE_R184_DFG_LUT_SIZE : u32 = 16u;
-const THREE_R184_DFG_LUT : array<vec2<f32>, 256> = array<vec2<f32>, 256>(
-  vec2<f32>(0.1470947265625, 0.85205078125), vec2<f32>(0.16552734375, 0.78759765625), vec2<f32>(0.244384765625, 0.638671875), vec2<f32>(0.370849609375, 0.51953125),
-  vec2<f32>(0.496826171875, 0.41552734375), vec2<f32>(0.60205078125, 0.326416015625), vec2<f32>(0.68408203125, 0.25390625), vec2<f32>(0.74609375, 0.197509765625),
-  vec2<f32>(0.79052734375, 0.154296875), vec2<f32>(0.822265625, 0.12164306640625), vec2<f32>(0.84326171875, 0.09698486328125), vec2<f32>(0.8564453125, 0.07843017578125),
-  vec2<f32>(0.86328125, 0.06439208984375), vec2<f32>(0.86572265625, 0.0537109375), vec2<f32>(0.8642578125, 0.045440673828125), vec2<f32>(0.85986328125, 0.039031982421875),
-  vec2<f32>(0.388671875, 0.611328125), vec2<f32>(0.39306640625, 0.6005859375), vec2<f32>(0.412353515625, 0.5458984375), vec2<f32>(0.45654296875, 0.4482421875),
-  vec2<f32>(0.52783203125, 0.3525390625), vec2<f32>(0.607421875, 0.27392578125), vec2<f32>(0.6787109375, 0.21142578125), vec2<f32>(0.7333984375, 0.16259765625),
-  vec2<f32>(0.77099609375, 0.1253662109375), vec2<f32>(0.79345703125, 0.0972900390625), vec2<f32>(0.80322265625, 0.07623291015625), vec2<f32>(0.8037109375, 0.06036376953125),
-  vec2<f32>(0.79638671875, 0.048431396484375), vec2<f32>(0.78564453125, 0.03936767578125), vec2<f32>(0.77197265625, 0.03240966796875), vec2<f32>(0.7548828125, 0.0269775390625),
-  vec2<f32>(0.572265625, 0.427490234375), vec2<f32>(0.57373046875, 0.424072265625), vec2<f32>(0.57958984375, 0.403564453125), vec2<f32>(0.591796875, 0.3544921875),
-  vec2<f32>(0.6162109375, 0.2880859375), vec2<f32>(0.6552734375, 0.224853515625), vec2<f32>(0.69873046875, 0.172607421875), vec2<f32>(0.7353515625, 0.1318359375),
-  vec2<f32>(0.75927734375, 0.10089111328125), vec2<f32>(0.77001953125, 0.07745361328125), vec2<f32>(0.77197265625, 0.0599365234375), vec2<f32>(0.76611328125, 0.046844482421875),
-  vec2<f32>(0.751953125, 0.0369873046875), vec2<f32>(0.732421875, 0.029541015625), vec2<f32>(0.70947265625, 0.023834228515625), vec2<f32>(0.68359375, 0.019439697265625),
-  vec2<f32>(0.708984375, 0.291015625), vec2<f32>(0.708984375, 0.289794921875), vec2<f32>(0.7099609375, 0.28125), vec2<f32>(0.7099609375, 0.258544921875),
-  vec2<f32>(0.71142578125, 0.220458984375), vec2<f32>(0.7197265625, 0.1768798828125), vec2<f32>(0.734375, 0.1370849609375), vec2<f32>(0.748046875, 0.10479736328125),
-  vec2<f32>(0.755859375, 0.07989501953125), vec2<f32>(0.759765625, 0.061004638671875), vec2<f32>(0.75341796875, 0.046844482421875), vec2<f32>(0.73876953125, 0.036224365234375),
-  vec2<f32>(0.7177734375, 0.02825927734375), vec2<f32>(0.69189453125, 0.0222625732421875), vec2<f32>(0.6611328125, 0.0177001953125), vec2<f32>(0.62841796875, 0.01419830322265625),
-  vec2<f32>(0.80810546875, 0.1917724609375), vec2<f32>(0.8076171875, 0.1912841796875), vec2<f32>(0.80615234375, 0.18798828125), vec2<f32>(0.8017578125, 0.1781005859375),
-  vec2<f32>(0.79296875, 0.15869140625), vec2<f32>(0.78369140625, 0.1322021484375), vec2<f32>(0.775390625, 0.1048583984375), vec2<f32>(0.7685546875, 0.08111572265625),
-  vec2<f32>(0.7646484375, 0.061981201171875), vec2<f32>(0.75537109375, 0.047271728515625), vec2<f32>(0.740234375, 0.036163330078125), vec2<f32>(0.71875, 0.0277557373046875),
-  vec2<f32>(0.69091796875, 0.021484375), vec2<f32>(0.658203125, 0.0167388916015625), vec2<f32>(0.6220703125, 0.01316070556640625), vec2<f32>(0.583984375, 0.01041412353515625),
-  vec2<f32>(0.87841796875, 0.1217041015625), vec2<f32>(0.8779296875, 0.1217041015625), vec2<f32>(0.875, 0.12060546875), vec2<f32>(0.869140625, 0.116943359375),
-  vec2<f32>(0.85693359375, 0.1080322265625), vec2<f32>(0.837890625, 0.09375), vec2<f32>(0.81494140625, 0.076904296875), vec2<f32>(0.7958984375, 0.060699462890625),
-  vec2<f32>(0.7763671875, 0.046875), vec2<f32>(0.75634765625, 0.035919189453125), vec2<f32>(0.732421875, 0.0274505615234375), vec2<f32>(0.703125, 0.021026611328125),
-  vec2<f32>(0.6689453125, 0.01617431640625), vec2<f32>(0.63037109375, 0.01251220703125), vec2<f32>(0.58935546875, 0.00974273681640625), vec2<f32>(0.54638671875, 0.007633209228515625),
-  vec2<f32>(0.92626953125, 0.07379150390625), vec2<f32>(0.92578125, 0.07391357421875), vec2<f32>(0.9228515625, 0.07391357421875), vec2<f32>(0.9169921875, 0.0731201171875),
-  vec2<f32>(0.90380859375, 0.06982421875), vec2<f32>(0.88134765625, 0.0631103515625), vec2<f32>(0.85107421875, 0.053802490234375), vec2<f32>(0.8212890625, 0.043701171875),
-  vec2<f32>(0.791015625, 0.034393310546875), vec2<f32>(0.76123046875, 0.026611328125), vec2<f32>(0.72802734375, 0.02044677734375), vec2<f32>(0.69189453125, 0.015655517578125),
-  vec2<f32>(0.65087890625, 0.0120086669921875), vec2<f32>(0.607421875, 0.00925445556640625), vec2<f32>(0.56103515625, 0.0071563720703125), vec2<f32>(0.51416015625, 0.005565643310546875),
-  vec2<f32>(0.95751953125, 0.042327880859375), vec2<f32>(0.95703125, 0.042449951171875), vec2<f32>(0.95458984375, 0.042816162109375), vec2<f32>(0.94921875, 0.043182373046875),
-  vec2<f32>(0.93701171875, 0.04266357421875), vec2<f32>(0.9130859375, 0.040252685546875), vec2<f32>(0.8818359375, 0.035797119140625), vec2<f32>(0.8447265625, 0.0301361083984375),
-  vec2<f32>(0.80615234375, 0.0243377685546875), vec2<f32>(0.76708984375, 0.0191497802734375), vec2<f32>(0.7265625, 0.01485443115234375), vec2<f32>(0.6826171875, 0.01142120361328125),
-  vec2<f32>(0.63623046875, 0.0087738037109375), vec2<f32>(0.5869140625, 0.006740570068359375), vec2<f32>(0.53662109375, 0.005199432373046875), vec2<f32>(0.486083984375, 0.00402069091796875),
-  vec2<f32>(0.9775390625, 0.0226287841796875), vec2<f32>(0.97705078125, 0.0227508544921875), vec2<f32>(0.97509765625, 0.023193359375), vec2<f32>(0.9697265625, 0.0239105224609375),
-  vec2<f32>(0.95947265625, 0.0244903564453125), vec2<f32>(0.93603515625, 0.0241241455078125), vec2<f32>(0.9052734375, 0.02252197265625), vec2<f32>(0.865234375, 0.0197601318359375),
-  vec2<f32>(0.82177734375, 0.016510009765625), vec2<f32>(0.7744140625, 0.0132904052734375), vec2<f32>(0.7265625, 0.01045989990234375), vec2<f32>(0.67626953125, 0.0081329345703125),
-  vec2<f32>(0.6240234375, 0.00627899169921875), vec2<f32>(0.56982421875, 0.004825592041015625), vec2<f32>(0.51513671875, 0.0037136077880859375), vec2<f32>(0.46142578125, 0.0028629302978515625),
-  vec2<f32>(0.98876953125, 0.01107025146484375), vec2<f32>(0.98876953125, 0.01116180419921875), vec2<f32>(0.98681640625, 0.01151275634765625), vec2<f32>(0.98291015625, 0.01221466064453125),
-  vec2<f32>(0.97314453125, 0.01299285888671875), vec2<f32>(0.953125, 0.013519287109375), vec2<f32>(0.9228515625, 0.01328277587890625), vec2<f32>(0.88232421875, 0.01226043701171875),
-  vec2<f32>(0.8349609375, 0.01065826416015625), vec2<f32>(0.783203125, 0.00885009765625), vec2<f32>(0.728515625, 0.007114410400390625), vec2<f32>(0.671875, 0.00560760498046875),
-  vec2<f32>(0.61376953125, 0.004360198974609375), vec2<f32>(0.5546875, 0.0033721923828125), vec2<f32>(0.496337890625, 0.0025997161865234375), vec2<f32>(0.439453125, 0.0020046234130859375),
-  vec2<f32>(0.9951171875, 0.00479888916015625), vec2<f32>(0.9951171875, 0.004863739013671875), vec2<f32>(0.99365234375, 0.005096435546875), vec2<f32>(0.990234375, 0.005603790283203125),
-  vec2<f32>(0.9814453125, 0.0063323974609375), vec2<f32>(0.96435546875, 0.006977081298828125), vec2<f32>(0.93603515625, 0.007297515869140625), vec2<f32>(0.896484375, 0.0071258544921875),
-  vec2<f32>(0.8466796875, 0.00649261474609375), vec2<f32>(0.79150390625, 0.005596160888671875), vec2<f32>(0.7314453125, 0.004627227783203125), vec2<f32>(0.6689453125, 0.0037174224853515625),
-  vec2<f32>(0.60546875, 0.0029296875), vec2<f32>(0.54150390625, 0.00228118896484375), vec2<f32>(0.479248046875, 0.001766204833984375), vec2<f32>(0.419677734375, 0.00136566162109375),
-  vec2<f32>(0.998046875, 0.0017604827880859375), vec2<f32>(0.998046875, 0.0017938613891601562), vec2<f32>(0.99658203125, 0.0019292831420898438), vec2<f32>(0.994140625, 0.002239227294921875),
-  vec2<f32>(0.986328125, 0.0027294158935546875), vec2<f32>(0.9716796875, 0.0032501220703125), vec2<f32>(0.94580078125, 0.00366973876953125), vec2<f32>(0.90771484375, 0.003818511962890625),
-  vec2<f32>(0.8583984375, 0.003681182861328125), vec2<f32>(0.79931640625, 0.0033206939697265625), vec2<f32>(0.7353515625, 0.0028438568115234375), vec2<f32>(0.66748046875, 0.0023441314697265625),
-  vec2<f32>(0.5986328125, 0.0018796920776367188), vec2<f32>(0.5302734375, 0.0014829635620117188), vec2<f32>(0.464111328125, 0.0011587142944335938), vec2<f32>(0.402099609375, 0.0008993148803710938),
-  vec2<f32>(0.99951171875, 0.0005011558532714844), vec2<f32>(0.99951171875, 0.0005154609680175781), vec2<f32>(0.998046875, 0.000583648681640625), vec2<f32>(0.99658203125, 0.0007505416870117188),
-  vec2<f32>(0.9892578125, 0.00101470947265625), vec2<f32>(0.97607421875, 0.001346588134765625), vec2<f32>(0.95263671875, 0.0016527175903320312), vec2<f32>(0.916015625, 0.0018558502197265625),
-  vec2<f32>(0.8671875, 0.0019054412841796875), vec2<f32>(0.8076171875, 0.0018167495727539062), vec2<f32>(0.7392578125, 0.001621246337890625), vec2<f32>(0.6669921875, 0.0013799667358398438),
-  vec2<f32>(0.59326171875, 0.0011358261108398438), vec2<f32>(0.52001953125, 0.0009121894836425781), vec2<f32>(0.450439453125, 0.0007219314575195312), vec2<f32>(0.385986328125, 0.0005650520324707031),
-  vec2<f32>(1.0, 0.00009316205978393555), vec2<f32>(1.0, 0.0000978708267211914), vec2<f32>(0.9990234375, 0.00012540817260742188), vec2<f32>(0.9970703125, 0.00019216537475585938),
-  vec2<f32>(0.99072265625, 0.0003104209899902344), vec2<f32>(0.97900390625, 0.0004699230194091797), vec2<f32>(0.95751953125, 0.0006470680236816406), vec2<f32>(0.92333984375, 0.0007915496826171875),
-  vec2<f32>(0.87548828125, 0.0008769035339355469), vec2<f32>(0.81494140625, 0.0008869171142578125), vec2<f32>(0.744140625, 0.0008325576782226562), vec2<f32>(0.66748046875, 0.0007386207580566406),
-  vec2<f32>(0.58837890625, 0.0006265640258789062), vec2<f32>(0.51123046875, 0.0005168914794921875), vec2<f32>(0.438232421875, 0.0004165172576904297), vec2<f32>(0.37109375, 0.0003311634063720703),
-  vec2<f32>(1.0, 0.000007271766662597656), vec2<f32>(1.0, 0.000008165836334228516), vec2<f32>(0.9990234375, 0.000016987323760986328), vec2<f32>(0.99755859375, 0.00003790855407714844),
-  vec2<f32>(0.9921875, 0.00007593631744384766), vec2<f32>(0.9814453125, 0.00013744831085205078), vec2<f32>(0.96142578125, 0.00020754337310791016), vec2<f32>(0.92919921875, 0.00028014183044433594),
-  vec2<f32>(0.8828125, 0.0003345012664794922), vec2<f32>(0.82177734375, 0.00036263465881347656), vec2<f32>(0.7490234375, 0.00036215782165527344), vec2<f32>(0.66845703125, 0.0003380775451660156),
-  vec2<f32>(0.58544921875, 0.00029969215393066406), vec2<f32>(0.50390625, 0.00025582313537597656), vec2<f32>(0.427001953125, 0.0002117156982421875), vec2<f32>(0.357666015625, 0.00017201900482177734),
-  vec2<f32>(1.0, 0.0), vec2<f32>(1.0, 5.960464477539063e-8), vec2<f32>(0.99951171875, 0.0000012516975402832031), vec2<f32>(0.99755859375, 0.000005304813385009766),
-  vec2<f32>(0.9931640625, 0.000015079975128173828), vec2<f32>(0.98291015625, 0.00002855062484741211), vec2<f32>(0.96435546875, 0.00004744529724121094), vec2<f32>(0.93408203125, 0.00006842613220214844),
-  vec2<f32>(0.88916015625, 0.00008893013000488281), vec2<f32>(0.828125, 0.0001042485237121582), vec2<f32>(0.75390625, 0.00011217594146728516), vec2<f32>(0.67041015625, 0.00011241436004638672),
-  vec2<f32>(0.5830078125, 0.00010627508163452148), vec2<f32>(0.4970703125, 0.00009584426879882812), vec2<f32>(0.4169921875, 0.0000833272933959961), vec2<f32>(0.34521484375, 0.00007051229476928711),
-);
+#ifdef DIRECTIONAL_PCSS_AVAILABLE
+const PCSS_MEDIUM_RAW_TAPS : u32 = 8u;
+const PCSS_MEDIUM_COMPARE_TAPS : u32 = 16u;
+const PCSS_HIGH_RAW_TAPS : u32 = 16u;
+const PCSS_HIGH_COMPARE_TAPS : u32 = 32u;
 
-fn _sampleThreeR184DfgLut(roughness : f32, dotNV : f32) -> vec2<f32> {
-  let uv = clamp(vec2<f32>(roughness, dotNV), vec2<f32>(0.0), vec2<f32>(1.0));
-  let samplePosition = uv * f32(THREE_R184_DFG_LUT_SIZE) - vec2<f32>(0.5);
-  let base = vec2<i32>(floor(samplePosition));
-  let weight = fract(samplePosition);
-  let last = i32(THREE_R184_DFG_LUT_SIZE - 1u);
-  let lo = clamp(base, vec2<i32>(0), vec2<i32>(last));
-  let hi = clamp(base + vec2<i32>(1), vec2<i32>(0), vec2<i32>(last));
-  let rowLo = mix(
-    THREE_R184_DFG_LUT[u32(lo.y) * THREE_R184_DFG_LUT_SIZE + u32(lo.x)],
-    THREE_R184_DFG_LUT[u32(lo.y) * THREE_R184_DFG_LUT_SIZE + u32(hi.x)],
-    weight.x,
-  );
-  let rowHi = mix(
-    THREE_R184_DFG_LUT[u32(hi.y) * THREE_R184_DFG_LUT_SIZE + u32(lo.x)],
-    THREE_R184_DFG_LUT[u32(hi.y) * THREE_R184_DFG_LUT_SIZE + u32(hi.x)],
-    weight.x,
-  );
-  return mix(rowLo, rowHi, weight.y);
-}
-
-fn _threeR184DirectMultiScatter(
-  roughness : f32,
-  nDotV     : f32,
-  nDotL     : f32,
-  F0        : vec3<f32>,
-) -> vec3<f32> {
-  let dfgV = _sampleThreeR184DfgLut(roughness, nDotV);
-  let dfgL = _sampleThreeR184DfgLut(roughness, nDotL);
-  let fssEssV = F0 * dfgV.x + vec3<f32>(dfgV.y);
-  let fssEssL = F0 * dfgL.x + vec3<f32>(dfgL.y);
-  let emsV = 1.0 - dfgV.x - dfgV.y;
-  let emsL = 1.0 - dfgL.x - dfgL.y;
-  let favg = F0 + (vec3<f32>(1.0) - F0) * 0.047619;
-  let energyLoss = emsV * emsL;
-  let fms = fssEssV * fssEssL * favg
-    / (vec3<f32>(1.0) - energyLoss * favg * favg + vec3<f32>(1e-6));
-  return fms * energyLoss;
-}
+// Cascade-local disk points are fixed, bounded, and shared by raw and compare
+// phases. A rotation derived from the integer receiver texel and cascade keeps
+// the pattern stable without a temporal input.
+const PCSS_DISK_OFFSETS : array<vec2<f32>, 32> = array<vec2<f32>, 32>(
+  vec2<f32>(-0.326, -0.945), vec2<f32>(0.236, -0.873), vec2<f32>(0.891, -0.404), vec2<f32>(-0.761, -0.581),
+  vec2<f32>(0.612, 0.146), vec2<f32>(-0.148, 0.514), vec2<f32>(-0.527, -0.109), vec2<f32>(0.074, 0.911),
+  vec2<f32>(-0.944, 0.238), vec2<f32>(0.444, -0.736), vec2<f32>(0.707, 0.641), vec2<f32>(-0.184, -0.342),
+  vec2<f32>(0.318, 0.382), vec2<f32>(-0.638, 0.526), vec2<f32>(0.955, -0.083), vec2<f32>(-0.401, 0.816),
+  vec2<f32>(-0.083, -0.632), vec2<f32>(0.539, -0.262), vec2<f32>(-0.735, -0.168), vec2<f32>(0.162, 0.719),
+  vec2<f32>(-0.841, 0.003), vec2<f32>(0.791, 0.332), vec2<f32>(-0.291, -0.791), vec2<f32>(0.021, -0.224),
+  vec2<f32>(0.386, 0.799), vec2<f32>(-0.558, 0.134), vec2<f32>(0.638, -0.555), vec2<f32>(-0.189, 0.957),
+  vec2<f32>(-0.977, -0.117), vec2<f32>(0.271, 0.589), vec2<f32>(0.819, -0.719), vec2<f32>(-0.472, -0.409),
+);
+#endif
 
+// Direct multiscatter is owned by forgeax_pbr::brdf. Keep directional
+// lighting as a consumer of that shared function so its LUT and energy
+// compensation cannot diverge from punctual and cluster lighting.
 // Pick the cascade layer for a positive view-space depth -- walks
 // splitPlanes in order, returns the first split the depth falls below. Last
 // layer (count - 1) catches everything beyond splits[count-2]. cascadeCount=1
@@ -226,8 +141,128 @@ fn _atlasTileOrigin(layer : u32, count : u32) -> vec2<f32> {
   return vec2<f32>(tile) / vec2<f32>(grid);
 }
 
+#ifdef DIRECTIONAL_PCSS_AVAILABLE
+fn _pcssDiskRotation(texel : vec2<i32>, layer : u32) -> f32 {
+  let hash = u32(texel.x) * 1664525u + u32(texel.y) * 1013904223u + (layer + 1u) * 374761393u;
+  return f32(hash % 6283u) * 0.001;
+}
+
+fn _pcssDiskPoint(index : u32, angle : f32, radius : f32) -> vec2<f32> {
+  let point = PCSS_DISK_OFFSETS[index];
+  let cs = cos(angle);
+  let sn = sin(angle);
+  return vec2<f32>(point.x * cs - point.y * sn, point.x * sn + point.y * cs) * radius;
+}
+
+fn _pcssProjectedSample(
+  shadowMapSize : vec2<u32>,
+  tileOrigin   : vec2<u32>,
+  tileSize     : vec2<u32>,
+  baseTexel    : vec2<i32>,
+  offset       : vec2<f32>,
+) -> vec2<f32> {
+  let texel = shadow_clamp_texel_to_tile(
+    baseTexel + vec2<i32>(round(offset)),
+    vec2<i32>(tileOrigin),
+    vec2<i32>(tileSize),
+    1,
+  );
+  return (vec2<f32>(texel) + vec2<f32>(0.5)) / vec2<f32>(shadowMapSize);
+}
+
+fn _samplePcssForCascade(
+  worldPos : vec3<f32>,
+  layer    : u32,
+  count    : u32,
+  normal   : vec3<f32>,
+  l        : vec3<f32>,
+  profile  : u32,
+) -> f32 {
+  let lightClip = _cascadeLightViewProj(layer) * vec4<f32>(worldPos, 1.0);
+  if (!(lightClip.w > 0.0 || lightClip.w < 0.0)) {
+    return 1.0;
+  }
+  let projCoords = lightClip.xyz / lightClip.w;
+  let tileUv = vec2<f32>(projCoords.x * 0.5 + 0.5, -projCoords.y * 0.5 + 0.5);
+  if (!(tileUv.x >= 0.0 && tileUv.x <= 1.0 && tileUv.y >= 0.0 && tileUv.y <= 1.0 && projCoords.z >= 0.0 && projCoords.z <= 1.0)) {
+    return 1.0;
+  }
+  let nDotL = dot(normal, l);
+  if (!(nDotL >= -1.0 && nDotL <= 1.0)) {
+    return 1.0;
+  }
+  let shadowMapSize = textureDimensions(shadowMap, 0);
+  let grid = _atlasTileGrid(count);
+  let tileSize = shadowMapSize / grid;
+  let tileOrigin = vec2<u32>(layer % grid.x, layer / grid.x) * tileSize;
+  let baseTexel = vec2<i32>(tileOrigin) + vec2<i32>(floor(tileUv * vec2<f32>(tileSize)));
+  let angle = _pcssDiskRotation(baseTexel, layer);
+  let biasedDepth = shadow_biased_receiver_depth(
+    projCoords.z,
+    view.normalBias,
+    view.depthBias,
+    nDotL,
+  );
+
+  let rawRadius = select(2.0, 3.0, profile == 5u);
+  var blockerDepth = 0.0;
+  var blockerCount = 0u;
+  if (profile == 4u) {
+    for (var i = 0u; i < PCSS_MEDIUM_RAW_TAPS; i = i + 1u) {
+      let rawUv = _pcssProjectedSample(shadowMapSize, tileOrigin, tileSize, baseTexel, _pcssDiskPoint(i, angle, rawRadius));
+      let rawTexel = vec2<i32>(rawUv * vec2<f32>(shadowMapSize) - vec2<f32>(0.5));
+      let rawDepth = shadow_load_raw_depth(shadowMap, rawTexel);
+      if (rawDepth >= 0.0 && rawDepth < biasedDepth) {
+        blockerDepth = blockerDepth + rawDepth;
+        blockerCount = blockerCount + 1u;
+      }
+    }
+  } else {
+    for (var i = 0u; i < PCSS_HIGH_RAW_TAPS; i = i + 1u) {
+      let rawUv = _pcssProjectedSample(shadowMapSize, tileOrigin, tileSize, baseTexel, _pcssDiskPoint(i, angle, rawRadius));
+      let rawTexel = vec2<i32>(rawUv * vec2<f32>(shadowMapSize) - vec2<f32>(0.5));
+      let rawDepth = shadow_load_raw_depth(shadowMap, rawTexel);
+      if (rawDepth >= 0.0 && rawDepth < biasedDepth) {
+        blockerDepth = blockerDepth + rawDepth;
+        blockerCount = blockerCount + 1u;
+      }
+    }
+  }
+  if (blockerCount == 0u) {
+    return 1.0;
+  }
+
+  let averageBlockerDepth = blockerDepth / f32(blockerCount);
+  let lightDepthWorldSpan = max(view.splitPlanes[layer].z, 0.0);
+  let worldUnitsPerTexel = max(view.splitPlanes[layer].y, 0.0);
+  if (!(lightDepthWorldSpan > 0.0 && worldUnitsPerTexel > 0.0 && view.directionalShadowFilter.y >= 0.0 && view.directionalShadowFilter.z >= 0.0)) {
+    return 1.0;
+  }
+  let worldDistance = max(0.0, biasedDepth - averageBlockerDepth) * lightDepthWorldSpan;
+  let penumbraTexels = clamp(
+    worldDistance * tan(view.directionalShadowFilter.y) / worldUnitsPerTexel,
+    0.0,
+    view.directionalShadowFilter.z,
+  );
+  let compareRadius = max(0.5, penumbraTexels);
+  var litSum = 0.0;
+  if (profile == 4u) {
+    for (var i = 0u; i < PCSS_MEDIUM_COMPARE_TAPS; i = i + 1u) {
+      let compareUv = _pcssProjectedSample(shadowMapSize, tileOrigin, tileSize, baseTexel, _pcssDiskPoint(i, angle, compareRadius));
+      litSum = litSum + shadow_sample_compare(shadowMap, shadowSampler, compareUv, biasedDepth);
+    }
+    return litSum / f32(PCSS_MEDIUM_COMPARE_TAPS);
+  }
+  for (var i = 0u; i < PCSS_HIGH_COMPARE_TAPS; i = i + 1u) {
+    let compareUv = _pcssProjectedSample(shadowMapSize, tileOrigin, tileSize, baseTexel, _pcssDiskPoint(i, angle, compareRadius));
+    litSum = litSum + shadow_sample_compare(shadowMap, shadowSampler, compareUv, biasedDepth);
+  }
+  return litSum / f32(PCSS_HIGH_COMPARE_TAPS);
+}
+#endif
+
 // Sample the shadow atlas with the LO 3.1.3 slope-scaled bias + dynamic PCF
-// kernel (driven by view.pcfKernelSize, MAX_PCF_HALF=2),
+// kernel (driven by the directional filter profile, MAX_PCF_HALF=2),
 // against the lightViewProj for the chosen cascade. The shader maps NDC
 // xy to that cascade's atlas tile in fragment space (matrix carries
 // clip-space; tile placement happens here so shadow_caster.gl_Position
@@ -275,21 +310,16 @@ fn _sampleShadowForCascade(
   // [tileOrigin, tileOrigin+tileScale) (one texel inset) so taps stay in-tile.
   let tileLo = tileOrigin + texel;
   let tileHi = tileOrigin + tileScale - texel;
-  // Variable-width PCF kernel driven by view.pcfKernelSize (feat-20260621
-  // 5.3-production-shadow-demos AC-14 merged with the DirectionalLightShadow
-  // merge). Constant trip count to MAX_PCF_HALF with a per-iteration clip to the
-  // runtime radius keeps the shader variant-free (no dynamic loop bound, legal
-  // for textureSampleCompareLevel uniform control flow). Host clamps
-  // view.pcfKernelSize to {1,3,5}; divisor = actual tap count, so pcfKernelSize=3
-  // -> half=1 -> 9 taps / 9.0 (result-identical to the prior hard-coded 3x3);
-  // pcfKernelSize=1 -> half=0 -> single centre tap (hard edge); pcfKernelSize=5
-  // -> half=2 -> 25-tap soft penumbra.
-  let requestedKernel = clamp(u32(round(view.pcfKernelSize)), 1u, 2u * MAX_PCF_HALF + 1u);
-  // Preserve the previous radius mapping for malformed even values:
-  // 1 -> 1x1, 2/3/4 -> 3x3, 5 -> 5x5. The authored component accepts only
-  // {1,3,5}, but keeping this normalization makes the refactor output-stable
-  // for raw UBO callers too.
-  let kernel = select(select(5u, 3u, requestedKernel <= 4u), 1u, requestedKernel == 1u);
+  // PCF1/3/5 profiles 1/2/3 retain their existing receiver. PCSS profiles 4/5
+  // use the bounded three-stage receiver below without a material or variant
+  // axis; the carrier y/z lanes provide angular radius and penumbra limit.
+  let filterProfile = clamp(u32(round(view.directionalShadowFilter.x)), 1u, 5u);
+  #ifdef DIRECTIONAL_PCSS_AVAILABLE
+  if (filterProfile >= 4u) {
+    return _samplePcssForCascade(worldPos, layer, count, normal, l, filterProfile);
+  }
+  #endif
+  let kernel = select(select(3u, 5u, filterProfile == 3u), 1u, filterProfile == 1u);
   if (kernel == 1u) {
     let lit = textureSampleCompareLevel(shadowMap, shadowSampler, clamp(uv, tileLo, tileHi), adjustedDepth);
     return lit;
@@ -383,10 +413,9 @@ fn evalDirectionalNoShadow(
   let nDotV = max(dot(normal, viewDir), 1e-5);
   let nDotH = max(dot(normal, h), 0.0);
   let vDotH = max(dot(viewDir, h), 0.0);
-  let fresnel = exp2((-5.55473 * vDotH - 6.98316) * vDotH);
-  let f = F0 * (vec3<f32>(1.0) - fresnel) + vec3<f32>(fresnel);
+  let f = f_schlick(vDotH, F0);
   let roughness = sqrt(max(alphaSq, 0.0));
-  let multiScatter = _threeR184DirectMultiScatter(roughness, nDotV, nDotL, F0);
+  let multiScatter = threeR184DirectMultiScatter(roughness, nDotV, nDotL, F0);
   let specular = d_ggx(nDotH, alphaSq) * v_smith(nDotV, nDotL, alphaSq) * f + multiScatter;
   let diffuse = (1.0 - metallic) * baseColor / 3.14159265;
   return (diffuse + specular) * view.lightColor * nDotL;

package/src/lighting-probe.wgsl

@@ -0,0 +1,51 @@
+// @forgeax/engine-shader - lighting-probe.wgsl
+// Local diffuse LightProbe composition. The record is supplied by the
+// renderer's per-object retained projection; Skylight remains the sole Sky
+// and specular owner.
+
+#define_import_path forgeax_pbr::lighting_probe
+
+const PROBE_INV_PI : f32 = 0.3183098861837907;
+
+fn probe_sh9(shPreblend : array<vec4<f32>, 9>, normal : vec3<f32>) -> vec3<f32> {
+  let x = normal.x;
+  let y = normal.y;
+  let z = normal.z;
+  let basis = array<f32, 9>(
+    0.2820947918,
+    0.4886025119 * y,
+    0.4886025119 * z,
+    0.4886025119 * x,
+    1.0925484306 * x * y,
+    1.0925484306 * y * z,
+    0.3153915653 * (3.0 * z * z - 1.0),
+    1.0925484306 * x * z,
+    0.5462742153 * (x * x - y * y),
+  );
+  var result = vec3<f32>(0.0);
+  for (var band = 0u; band < 9u; band = band + 1u) {
+    result += shPreblend[band].xyz * basis[band];
+  }
+  return result;
+}
+
+// One SH9 evaluation, then the real Skylight residual. No Sky q, probe loop,
+// hidden gain, smoothstep, or specular probe path is present here.
+fn evaluateProbeDiffuse(
+  shPreblend : array<vec4<f32>, 9>,
+  localBlendFraction : f32,
+  normal : vec3<f32>,
+  e_sky : vec3<f32>,
+  k_d : vec3<f32>,
+  albedo : vec3<f32>,
+  metallic : f32,
+) -> vec3<f32> {
+  // `shPreblend` stores cosine-convolved irradiance. Match the existing
+  // Standard-PBR Skylight path and Three's BRDF_Lambert: irradiance becomes
+  // reflected diffuse radiance only after the single 1/pi factor. The Sky
+  // argument is already the Skylight path's E/pi * color * intensity value,
+  // so it must not be divided a second time below.
+  let local = max(probe_sh9(shPreblend, normal), vec3<f32>(0.0)) * PROBE_INV_PI;
+  let skyResidualFraction = 1.0 - localBlendFraction;
+  return (local + skyResidualFraction * e_sky) * k_d * albedo * (1.0 - metallic);
+}