reze-engine 0.12.2 → 0.12.3

package/README.md

@@ -14,7 +14,7 @@ npm install reze-engine
 
 - **Anime/MMD-style hybrid renderer** — toon-ramp NPR diffuse mixed with PBR GGX specular (multi-scatter + LTC energy compensation)
 - **Per-material presets** — `face` / `hair` / `body` / `eye` / `stockings` / `metal` / `cloth_smooth` / `cloth_rough` / `default`, assigned by material name
-- **HDR pipeline** with bloom mip pyramid, Filmic tone mapping, 4× MSAA
+- **HDR pipeline** with bloom mip pyramid, Filmic tone mapping, 4× MSAA, tile-memory-friendly on Apple Silicon
 - **Alpha-hashed transparency** (Wyman & McGuire 2017) for self-overlapping transparent meshes like stockings
 - **Screen-space outlines** on opaque + transparent materials
 - **See-through hair over eyes** — stencil-gated MMD post-alpha-eye so eyes read at 50% through hair silhouettes
@@ -249,23 +249,23 @@ let color = eval_principled(
 );
 ```
 
-NPR presets add stage C (and sometimes D) on top, and stage F chooses how NPR-leaning the surface is. The Blender preset this engine targets is authored with `MixShader(NPR, Principled, fac)` throughout — the 7-stage layout is a direct port of that topology.
+NPR presets add stage C (and sometimes D) on top, and stage F chooses how NPR-leaning the surface is.
 
 ### Shared WGSL foundations
 
-- **`nodes.ts`** — WGSL mirrors of the Blender shader nodes the presets use: `hue_sat`, `bright_contrast`, `ramp_constant/linear/cardinal`, `mix_overlay/lighten/linear_light`, `fresnel`, `layer_weight_fresnel/facing`, `tex_noise`, `tex_voronoi`, `mapping_point`, `bump_lh`, `normal_map`. Plus one combined DFG + LTC 64×64 RGBA8 LUT baked at `engine.init()`, one `eval_principled(PrincipledIn, N, L, V, sun, amb, shadow)` EEVEE Principled port, and `principled_sheen`.
-- **`common.ts`** — Uniform structs, bind-group layout (same for every material pipeline), 3×3 PCF shadow sampler, skinning vertex shader, shared `FSOut`.
-- **Per-material files** — constants + NPR stack + optional bump + `eval_principled` call + final mix. That's it.
+- **`nodes.ts`** — WGSL mirrors of the Blender shader nodes the presets use: `hue_sat`, `bright_contrast`, `ramp_constant/linear/cardinal`, `mix_overlay/lighten/linear_light`, `fresnel`, `layer_weight_fresnel/facing`, `tex_noise`, `tex_voronoi`, `mapping_point`, `bump_lh`, `normal_map`. Plus a combined DFG + LTC LUT, `eval_principled(PrincipledIn, N, L, V, sun, amb, shadow)`, and `principled_sheen`.
+- **`common.ts`** — Uniform structs, bind-group layout (same for every material pipeline), PCF shadow sampler, skinning vertex shader, shared `FSOut`.
+- **Per-material files** — constants + NPR stack + optional bump + `eval_principled` call + final mix.
 
 ### PBR specular core
 
 Inside `eval_principled`:
 
-- GGX microfacet specular with Schlick Fresnel, Walter–Smith G1 (reciprocal form collapses the `4·NL·NV` denominator)
-- **Multi-scatter compensation** (Fdez-Agüera 2019, `F_brdf_multi_scatter`) — restores energy at high roughness so metals don't darken
-- **Split-sum DFG LUT** (Karis 2013) — drives indirect specular, packed into the `.rg` channels of the baked LUT
-- **LTC direct-spec scale** (Heitz 2016 magnitude LUT) — keeps analytic-light specular in the same energy budget as image-based lighting; packed into `.ba` so a single LUT tap serves both paths
-- **Sheen coarse curve** (`f³·0.077 + f·0.01 + 0.00026`) gated by the `sheen` field on `PrincipledIn` — used by `stockings`
+- GGX microfacet specular with Schlick Fresnel and Walter–Smith G1
+- **Multi-scatter compensation** (Fdez-Agüera 2019) — restores energy at high roughness so metals don't darken
+- **Split-sum DFG LUT** (Karis 2013) — drives indirect specular
+- **LTC direct-spec scale** (Heitz 2016) — keeps analytic-light specular in the same energy budget as image-based lighting
+- **Sheen coarse curve** gated by the `sheen` field on `PrincipledIn`
 
 ### NPR toolbox
 
@@ -297,26 +297,26 @@ Assign presets per-model with `engine.setMaterialPresets(name, map)` (see the [U
 
 ### Shadows, post, output
 
-- Directional shadow map (2048², depth32float, 3×3 PCF unrolled, normal + depth bias)
-- HDR (rgba16f) main pass with 4× MSAA, resolved before tonemap
-- Bloom via threshold + downsample/upsample mip pyramid, gated by an MRT mask channel from emissive presets
-- Filmic tone mapping (LUT extracted from Blender 3.6 OCIO "Filmic / Medium High Contrast", exposure baked in)
+- Directional shadow map (2048², depth32float, PCF, normal + depth bias)
+- HDR main pass with 4× MSAA. Color is `rg11b10ufloat` paired with an `rg8unorm` aux MRT carrying bloom mask (`.r`) and accumulated alpha (`.g`). The combined footprint fits Apple Silicon TBDR tile memory, so the 4× MSAA buffer resolves in-tile rather than spilling to system memory every frame. Falls back to `rgba16float` when the device does not expose `rg11b10ufloat-renderable`.
+- Bloom via threshold + downsample/upsample mip pyramid, gated by the aux bloom-mask channel
+- Filmic tone mapping (LUT extracted from Blender 3.6 OCIO "Filmic / Medium High Contrast")
 - Screen-space outline pass (inverted-hull) on opaque and transparent materials
 
 ### Alpha-hashed transparency
 
-`stockings` uses the Wyman & McGuire 2017 derivative-aware stochastic discard instead of alpha blend. Self-overlapping transparent meshes (stockings wrap the leg — front and back surfaces share screen pixels) can't be sorted per-fragment in one draw call, so blend produces "cracks". Hashed alpha keeps opaque-style depth writes and resolves cleanly under MSAA. The hash is derived from world-space position, not screen-space, so the dither pattern doesn't swim when the camera moves.
+`stockings` uses the Wyman & McGuire 2017 derivative-aware stochastic discard so self-overlapping transparent meshes (e.g. the front and back of a stocking wrapped around a leg) resolve cleanly under MSAA with opaque-style depth writes. The hash is derived from world-space position, so the dither pattern does not swim when the camera moves.
 
 ### See-through hair over eyes (MMD post-alpha-eye)
 
 The classic MMD effect where hair strands covering the eye are rendered at 50% so the iris stays readable — implemented as a single extra pass driven by the stencil buffer, not a two-texture composite.
 
-- **Eye pipeline** stamps `stencil = EYE_VALUE` on every fragment it writes, and uses `cullMode: "front"` + a small negative `depthBias` so only the back half of the eye mesh renders (the MMD trick that prevents eyes from leaking through the back of the head without a per-model skull occluder).
-- **Main hair pipeline** stencil-tests `not-equal EYE_VALUE` and skips those fragments entirely — depth and color stay as the eye wrote them.
-- **Hair-over-eyes pipeline** re-issues the hair draws through the same shader compiled with `override IS_OVER_EYES = true` (pipeline-override constant, so the dead branch is dropped at compile time). Stencil-tests `equal EYE_VALUE`, `depthWriteEnabled: false`, and alpha blends at 50% — so eye-stamped pixels end up `0.5·hair + 0.5·eye` in linear HDR before tonemap.
-- **Outline pipeline** also stencil-tests `not-equal EYE_VALUE` so the edge color doesn't overwrite the freshly blended eye pixels (otherwise the near-black outline paints a dark almond shape over the see-through area).
+- **Eye pipeline** stamps `stencil = EYE_VALUE` on every fragment it writes, with `cullMode: "front"` and a small negative `depthBias` so only the back half of the eye mesh renders (the MMD trick that keeps eyes from leaking through the back of the head).
+- **Main hair pipeline** stencil-tests `not-equal EYE_VALUE` and skips those fragments.
+- **Hair-over-eyes pipeline** re-issues the hair draws with `IS_OVER_EYES = true`, stencil-tests `equal EYE_VALUE`, disables depth writes, and alpha blends at 50% — eye-stamped pixels end up `0.5·hair + 0.5·eye` in linear HDR before tonemap.
+- **Outline pipeline** stencil-tests `not-equal EYE_VALUE` so edge color does not overwrite the see-through region.
 
-Draw order within a model: non-eye/non-hair opaque → eye (stamp) → hair (skip stamp) → outlines (skip stamp) → hair-over-eyes (match stamp). One `setStencilReference(EYE_VALUE)` per model covers all four stencil-aware pipelines.
+Draw order within a model: non-eye/non-hair opaque → eye (stamp) → hair (skip stamp) → outlines (skip stamp) → hair-over-eyes (match stamp).
 
 ## Projects Using This Engine
 

package/dist/engine.d.ts

@@ -128,13 +128,18 @@ export declare class Engine {
     private multisampleTexture;
     private hdrResolveTexture;
     private static readonly MULTISAMPLE_COUNT;
-    private static readonly HDR_FORMAT;
+    private hdrFormat;
     /** 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;
-    /** 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. */
+    /** 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;
     private multisampleMaskTexture;
     private maskResolveTexture;

package/dist/engine.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"engine.d.ts","sourceRoot":"","sources":["../src/engine.ts"],"names":[],"mappings":"AACA,OAAO,EAAQ,IAAI,EAAE,MAAM,QAAQ,CAAA;AACnC,OAAO,EAAE,KAAK,EAAE,MAAM,SAAS,CAAA;AAE/B,OAAO,EAAW,KAAK,cAAc,EAAE,MAAM,WAAW,CAAA;AACxD,OAAO,EAQL,KAAK,WAAW,EACjB,MAAM,gBAAgB,CAAA;AA0BvB,MAAM,MAAM,cAAc,GACtB,SAAS,GACT,MAAM,GACN,MAAM,GACN,MAAM,GACN,KAAK,GACL,WAAW,GACX,OAAO,GACP,cAAc,GACd,aAAa,CAAA;AAEjB,MAAM,MAAM,iBAAiB,GAAG,OAAO,CAAC,MAAM,CAAC,cAAc,EAAE,MAAM,EAAE,CAAC,CAAC,CAAA;AAUzE,MAAM,MAAM,eAAe,GAAG,CAAC,SAAS,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,IAAI,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,KAAK,IAAI,CAAA;AAEpH,kHAAkH;AAClH,MAAM,MAAM,yBAAyB,GAAG;IACtC,KAAK,EAAE,QAAQ,GAAG,IAAI,EAAE,CAAA;IACxB,OAAO,CAAC,EAAE,IAAI,CAAA;CACf,CAAA;AAID,MAAM,MAAM,YAAY,GAAG;IACzB,8FAA8F;IAC9F,KAAK,CAAC,EAAE,IAAI,CAAA;IACZ,uEAAuE;IACvE,QAAQ,CAAC,EAAE,MAAM,CAAA;CAClB,CAAA;AAED,MAAM,MAAM,UAAU,GAAG;IACvB,6DAA6D;IAC7D,KAAK,CAAC,EAAE,IAAI,CAAA;IACZ,+DAA+D;IAC/D,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,yFAAyF;IACzF,SAAS,CAAC,EAAE,IAAI,CAAA;CACjB,CAAA;AAED,MAAM,MAAM,aAAa,GAAG;IAC1B,kCAAkC;IAClC,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,gCAAgC;IAChC,MAAM,CAAC,EAAE,IAAI,CAAA;IACb,yCAAyC;IACzC,GAAG,CAAC,EAAE,MAAM,CAAA;CACb,CAAA;AAED,wFAAwF;AACxF,MAAM,MAAM,YAAY,GAAG;IACzB,OAAO,EAAE,OAAO,CAAA;IAChB,SAAS,EAAE,MAAM,CAAA;IACjB,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,MAAM,CAAA;IACd,KAAK,EAAE,IAAI,CAAA;IACX,SAAS,EAAE,MAAM,CAAA;IACjB,KAAK,EAAE,MAAM,CAAA;CACd,CAAA;AAED,eAAO,MAAM,qBAAqB,EAAE,YAQnC,CAAA;AAED,4HAA4H;AAC5H,MAAM,MAAM,oBAAoB,GAAG;IACjC,2DAA2D;IAC3D,QAAQ,EAAE,MAAM,CAAA;IAChB,yDAAyD;IACzD,KAAK,EAAE,MAAM,CAAA;IACb,IAAI,EAAE,SAAS,GAAG,sBAAsB,CAAA;CACzC,CAAA;AAID,eAAO,MAAM,sBAAsB,EAAE,oBAIpC,CAAA;AAED,MAAM,MAAM,aAAa,GAAG;IAC1B,KAAK,CAAC,EAAE,YAAY,CAAA;IACpB,GAAG,CAAC,EAAE,UAAU,CAAA;IAChB,MAAM,CAAC,EAAE,aAAa,CAAA;IACtB,yEAAyE;IACzE,KAAK,CAAC,EAAE,OAAO,CAAC,YAAY,CAAC,CAAA;IAC7B,gFAAgF;IAChF,IAAI,CAAC,EAAE,OAAO,CAAC,oBAAoB,CAAC,CAAA;IACpC,SAAS,CAAC,EAAE,eAAe,CAAA;IAC3B,cAAc,CAAC,EAAE,cAAc,CAAA;CAChC,CAAA;AAED,eAAO,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;CAMlC,CAAA;AAED,MAAM,WAAW,WAAW;IAC1B,GAAG,EAAE,MAAM,CAAA;IACX,SAAS,EAAE,MAAM,CAAA;CAClB;AA2CD,qBAAa,MAAM;IACjB,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAsB;IAE7C,MAAM,CAAC,WAAW,IAAI,MAAM;IAO5B,OAAO,CAAC,MAAM,CAAmB;IACjC,OAAO,CAAC,MAAM,CAAY;IAC1B,OAAO,CAAC,OAAO,CAAmB;IAClC,OAAO,CAAC,kBAAkB,CAAmB;IAC7C,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,mBAAmB,CAAY;IACvC,OAAO,CAAC,gBAAgB,CAAuB;IAE/C,OAAO,CAAC,KAAK,CAAoC;IACjD,OAAO,CAAC,GAAG,CAAqD;IAChE,OAAO,CAAC,YAAY,CAAkD;IACtE,OAAO,CAAC,kBAAkB,CAAY;IACtC,OAAO,CAAC,SAAS,CAAuB;IACxC,OAAO,CAAC,UAAU,CAAI;IACtB,OAAO,CAAC,cAAc,CAA8B;IACpD,OAAO,CAAC,YAAY,CAAa;IACjC,OAAO,CAAC,aAAa,CAAoB;IACzC,OAAO,CAAC,YAAY,CAAoB;IACxC,OAAO,CAAC,YAAY,CAAoB;IACxC,OAAO,CAAC,mBAAmB,CAAoB;IAC/C,OAAO,CAAC,kBAAkB,CAAoB;IAC9C,OAAO,CAAC,aAAa,CAAoB;IACzC,OAAO,CAAC,YAAY,CAAoB;IACxC,OAAO,CAAC,WAAW,CAAoB;IACvC,OAAO,CAAC,oBAAoB,CAAoB;IAChD,OAAO,CAAC,iBAAiB,CAAoB;IAC7C,OAAO,CAAC,oBAAoB,CAAoB;IAChD,OAAO,CAAC,2BAA2B,CAAqB;IACxD,OAAO,CAAC,eAAe,CAAoB;IAC3C,OAAO,CAAC,2BAA2B,CAAqB;IACxD,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,iCAAiC,CAAqB;IAC9D,OAAO,CAAC,iBAAiB,CAAe;IACxC,OAAO,CAAC,wBAAwB,CAAe;IAC/C,OAAO,CAAC,kBAAkB,CAAa;IACvC,OAAO,CAAC,iBAAiB,CAAa;IACtC,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,iBAAiB,CAAI;IAC7C,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,UAAU,CAAkC;IACpE;0FACsF;IACtF,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,iBAAiB,CAAI;IAC7C;;+DAE2D;IAC3D,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,iBAAiB,CAA8B;IACvE,OAAO,CAAC,sBAAsB,CAAa;IAC3C,OAAO,CAAC,kBAAkB,CAAa;IACvC,OAAO,CAAC,eAAe,CAAiB;IACxC,OAAO,CAAC,oBAAoB,CAA0B;IACtD,OAAO,CAAC,uBAAuB,CAA0B;IAIzD,OAAO,CAAC,yBAAyB,CAAoB;IACrD,OAAO,CAAC,sBAAsB,CAAoB;IAClD,OAAO,CAAC,wBAAwB,CAAqB;IACrD,OAAO,CAAC,kBAAkB,CAAe;IACzC,OAAO,CAAC,sBAAsB,CAAY;IAE1C,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAsB;IAQ3D,OAAO,CAAC,YAAY,CAAa;IACjC,OAAO,CAAC,sBAAsB,CAAY;IAC1C,OAAO,CAAC,0BAA0B,CAAY;IAC9C,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAsB;IAC3D,OAAO,CAAC,QAAQ,CAAC,wBAAwB,CAAsB;IAC/D,OAAO,CAAC,iBAAiB,CAAoB;IAC7C,OAAO,CAAC,uBAAuB,CAAoB;IACnD,OAAO,CAAC,qBAAqB,CAAoB;IACjD,OAAO,CAAC,wBAAwB,CAAqB;IACrD,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,4BAA4B,CAAqB;IACzD,OAAO,CAAC,gBAAgB,CAAa;IACrC,OAAO,CAAC,cAAc,CAAa;IACnC,OAAO,CAAC,aAAa,CAAI;IACzB,OAAO,CAAC,iBAAiB,CAAuB;IAChD,OAAO,CAAC,eAAe,CAAuB;IAC9C,OAAO,CAAC,kBAAkB,CAAe;IACzC,OAAO,CAAC,yBAAyB,CAAqB;IACtD,OAAO,CAAC,uBAAuB,CAAqB;IACpD,2EAA2E;IAC3E,OAAO,CAAC,mBAAmB,CAA0B;IACrD,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,gBAAgB,CAAI;IAG5C,OAAO,CAAC,kBAAkB,CAAC,CAAW;IACtC,OAAO,CAAC,iBAAiB,CAAC,CAAW;IACrC,OAAO,CAAC,SAAS,CAAQ;IACzB,OAAO,CAAC,gBAAgB,CAAa;IACrC,OAAO,CAAC,kBAAkB,CAAiB;IAC3C,OAAO,CAAC,cAAc,CAAa;IACnC,OAAO,CAAC,WAAW,CAAiB;IACpC,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,eAAe,CAAO;IAC9C,OAAO,CAAC,mBAAmB,CAAoB;IAC/C,OAAO,CAAC,mBAAmB,CAAY;IACvC,OAAO,CAAC,mBAAmB,CAAuB;IAClD,OAAO,CAAC,qBAAqB,CAAC,CAAc;IAC5C,OAAO,CAAC,uBAAuB,CAAa;IAC5C,OAAO,CAAC,0BAA0B,CAAC,CAAW;IAC9C,OAAO,CAAC,cAAc,CAAwB;IAE9C,OAAO,CAAC,SAAS,CAAC,CAAiB;IACnC,OAAO,CAAC,cAAc,CAAwD;IAC9E,OAAO,CAAC,aAAa,CAAI;IACzB,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAM;IAEvC,OAAO,CAAC,YAAY,CAAoB;IACxC,OAAO,CAAC,2BAA2B,CAAqB;IACxD,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,qBAAqB,CAAe;IAC5C,OAAO,CAAC,WAAW,CAAa;IAChC,OAAO,CAAC,gBAAgB,CAAa;IACrC,OAAO,CAAC,kBAAkB,CAAY;IACtC,OAAO,CAAC,WAAW,CAAwC;IAE3D,OAAO,CAAC,cAAc,CAAmC;IACzD,OAAO,CAAC,eAAe,CAAa;IACpC,OAAO,CAAC,YAAY,CAAgC;IACpD,OAAO,CAAC,eAAe,CAAiC;IACxD,OAAO,CAAC,cAAc,CAA0B;IAChD,OAAO,CAAC,mBAAmB,CAAI;IAG/B,OAAO,CAAC,SAAS,CAAO;IACxB,OAAO,CAAC,cAAc,CAAO;IAG7B,OAAO,CAAC,iBAAiB,CAAqB;IAC9C,OAAO,CAAC,oBAAoB,CAAS;IACrC,OAAO,CAAC,kBAAkB,CAA0B;IAEpD,OAAO,CAAC,aAAa,CAAoB;IACzC,OAAO,CAAC,qBAAqB,CAAI;IACjC,OAAO,CAAC,aAAa,CAAoB;IACzC,OAAO,CAAC,YAAY,CAAI;IACxB,OAAO,CAAC,cAAc,CAAI;IAC1B,OAAO,CAAC,KAAK,CAGZ;IACD,OAAO,CAAC,gBAAgB,CAAsB;IAC9C,OAAO,CAAC,kBAAkB,CAA4B;IACtD,OAAO,CAAC,aAAa,CAAe;IACpC,OAAO,CAAC,aAAa,CAAuB;gBAEhC,MAAM,EAAE,iBAAiB,EAAE,OAAO,CAAC,EAAE,aAAa;IAuB9D,qEAAqE;IACrE,MAAM,CAAC,kBAAkB,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,YAAY;IAcxE,MAAM,CAAC,0BAA0B,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC,oBAAoB,CAAC,GAAG,oBAAoB;IAShG,uEAAuE;IACvE,eAAe,IAAI,YAAY;IAa/B,uBAAuB,IAAI,oBAAoB;IAK/C,uBAAuB,CAAC,KAAK,EAAE,OAAO,CAAC,oBAAoB,CAAC,GAAG,IAAI;IAUnE,OAAO,CAAC,0BAA0B;IAmBlC,wEAAwE;IACxE,eAAe,CAAC,KAAK,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,IAAI;IAoBnD,OAAO,CAAC,kBAAkB;IAsBpB,IAAI;IAmCV,OAAO,CAAC,WAAW;IA2EnB,OAAO,CAAC,oBAAoB;IAiC5B,OAAO,CAAC,eAAe;IA8pBvB,OAAO,CAAC,WAAW;IAYnB,OAAO,CAAC,YAAY;IAsNpB,OAAO,CAAC,WAAW;IAmBnB,iFAAiF;IACjF,eAAe,CAAC,CAAC,EAAE,IAAI,GAAG,IAAI;IAC9B,gGAAgG;IAChG,eAAe,CAAC,KAAK,EAAE,KAAK,GAAG,IAAI,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,IAAI,GAAG,IAAI;IAoB3E,mIAAmI;IACnI,eAAe,CAAC,KAAK,EAAE,KAAK,GAAG,IAAI,EAAE,QAAQ,CAAC,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,IAAI,GAAG,IAAI;IAY5E,iBAAiB,IAAI,MAAM;IAG3B,iBAAiB,CAAC,CAAC,EAAE,MAAM,GAAG,IAAI;IAGlC,cAAc,IAAI,MAAM;IAGxB,cAAc,CAAC,CAAC,EAAE,MAAM,GAAG,IAAI;IAG/B,aAAa,IAAI,MAAM;IAGvB,aAAa,CAAC,CAAC,EAAE,MAAM,GAAG,IAAI;IAK9B,OAAO,CAAC,aAAa;IAYrB;;;;OAIG;IACH,OAAO,CAAC,UAAU;IASlB,qFAAqF;IACrF,OAAO,CAAC,QAAQ;IAgBhB,gGAAgG;IAChG,QAAQ,CAAC,OAAO,EAAE,YAAY,GAAG,IAAI;IAMrC,0FAA0F;IAC1F,MAAM,CAAC,OAAO,EAAE,UAAU,GAAG,IAAI;IAUjC,QAAQ,IAAI,QAAQ,CAAC;QAAE,KAAK,EAAE,IAAI,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAC;IAGvD,MAAM,IAAI,QAAQ,CAAC;QAAE,KAAK,EAAE,IAAI,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,IAAI,CAAA;KAAE,CAAC;IAItE,SAAS,CAAC,OAAO,CAAC,EAAE;QAClB,KAAK,CAAC,EAAE,MAAM,CAAA;QACd,MAAM,CAAC,EAAE,MAAM,CAAA;QACf,YAAY,CAAC,EAAE,IAAI,CAAA;QACnB,SAAS,CAAC,EAAE,MAAM,CAAA;QAClB,OAAO,CAAC,EAAE,MAAM,CAAA;QAChB,cAAc,CAAC,EAAE,MAAM,CAAA;QACvB,WAAW,CAAC,EAAE,MAAM,CAAA;QACpB,aAAa,CAAC,EAAE,MAAM,CAAA;QACtB,eAAe,CAAC,EAAE,MAAM,CAAA;QACxB,aAAa,CAAC,EAAE,IAAI,CAAA;QACpB,aAAa,CAAC,EAAE,MAAM,CAAA;KACvB,GAAG,IAAI;IA4BR,OAAO,CAAC,iBAAiB;IAIzB,QAAQ,IAAI,WAAW;IAIvB,aAAa,CAAC,QAAQ,CAAC,EAAE,MAAM,IAAI;IAgBnC,cAAc;IAQd,OAAO;IAkBD,SAAS,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,KAAK,CAAC;IACvC,SAAS,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,KAAK,CAAC;IACrD,SAAS,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,yBAAyB,GAAG,OAAO,CAAC,KAAK,CAAC;IAyB3E,QAAQ,CAAC,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,MAAM,EAAE,WAAW,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,MAAM,CAAC;IAcxG,WAAW,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI;IAiB/B,aAAa,IAAI,MAAM,EAAE;IAIzB,QAAQ,CAAC,IAAI,EAAE,MAAM,GAAG,KAAK,GAAG,IAAI;IAIpC,qBAAqB,CAAC,gBAAgB,CAAC,EAAE,MAAM,GAAG,KAAK,GAAG,IAAI;IAe9D,kBAAkB,CAAC,SAAS,EAAE,MAAM,EAAE,OAAO,EAAE,iBAAiB,GAAG,IAAI;IASvE,kBAAkB,CAAC,SAAS,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,IAAI;IAOnF,qBAAqB,CAAC,SAAS,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,GAAG,IAAI;IAOpE,iBAAiB,CAAC,SAAS,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,GAAG,OAAO;IAKnE,YAAY,CAAC,OAAO,EAAE,OAAO,GAAG,IAAI;IAIpC,YAAY,IAAI,OAAO;IAIvB,iBAAiB,CAAC,OAAO,EAAE,OAAO,GAAG,IAAI;IAIzC,iBAAiB,IAAI,OAAO;IAI5B,OAAO,CAAC,eAAe;IAIvB,OAAO,CAAC,eAAe;IAWvB,OAAO,CAAC,kBAAkB;YAOZ,kBAAkB;IAiHhC,OAAO,CAAC,oBAAoB;IAwE5B,OAAO,CAAC,2BAA2B;IA4DnC,OAAO,CAAC,kBAAkB,CAAO;IACjC,OAAO,CAAC,mBAAmB;YAeb,yBAAyB;IA6HvC,OAAO,CAAC,2BAA2B;IAUnC,OAAO,CAAC,mBAAmB;IAU3B,OAAO,CAAC,oBAAoB;YAId,4BAA4B;IAuC1C,OAAO,CAAC,eAAe;IA6CvB,OAAO,CAAC,YAAY;IASpB,OAAO,CAAC,uBAAuB,CAI9B;IAED,OAAO,CAAC,iBAAiB,CA0BxB;IAED,OAAO,CAAC,cAAc;IAStB,OAAO,CAAC,cAAc;YA+CR,iBAAiB;IA6C/B,MAAM;IAiHN,OAAO,CAAC,kBAAkB;IAW1B,OAAO,CAAC,iBAAiB;IAYzB;;;;;OAKG;IACH,OAAO,CAAC,aAAa;IAoBrB;;;;;OAKG;IACH,OAAO,CAAC,YAAY;IAepB;;;;;OAKG;IACH,OAAO,CAAC,cAAc;IAiBtB;;;;;OAKG;IACH,OAAO,CAAC,gBAAgB;IAexB,OAAO,CAAC,oBAAoB;IAY5B,OAAO,CAAC,kBAAkB;IAa1B,OAAO,CAAC,WAAW;CAwBpB"}
+{"version":3,"file":"engine.d.ts","sourceRoot":"","sources":["../src/engine.ts"],"names":[],"mappings":"AACA,OAAO,EAAQ,IAAI,EAAE,MAAM,QAAQ,CAAA;AACnC,OAAO,EAAE,KAAK,EAAE,MAAM,SAAS,CAAA;AAE/B,OAAO,EAAW,KAAK,cAAc,EAAE,MAAM,WAAW,CAAA;AACxD,OAAO,EAQL,KAAK,WAAW,EACjB,MAAM,gBAAgB,CAAA;AA0BvB,MAAM,MAAM,cAAc,GACtB,SAAS,GACT,MAAM,GACN,MAAM,GACN,MAAM,GACN,KAAK,GACL,WAAW,GACX,OAAO,GACP,cAAc,GACd,aAAa,CAAA;AAEjB,MAAM,MAAM,iBAAiB,GAAG,OAAO,CAAC,MAAM,CAAC,cAAc,EAAE,MAAM,EAAE,CAAC,CAAC,CAAA;AAUzE,MAAM,MAAM,eAAe,GAAG,CAAC,SAAS,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,IAAI,EAAE,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,KAAK,IAAI,CAAA;AAEpH,kHAAkH;AAClH,MAAM,MAAM,yBAAyB,GAAG;IACtC,KAAK,EAAE,QAAQ,GAAG,IAAI,EAAE,CAAA;IACxB,OAAO,CAAC,EAAE,IAAI,CAAA;CACf,CAAA;AAID,MAAM,MAAM,YAAY,GAAG;IACzB,8FAA8F;IAC9F,KAAK,CAAC,EAAE,IAAI,CAAA;IACZ,uEAAuE;IACvE,QAAQ,CAAC,EAAE,MAAM,CAAA;CAClB,CAAA;AAED,MAAM,MAAM,UAAU,GAAG;IACvB,6DAA6D;IAC7D,KAAK,CAAC,EAAE,IAAI,CAAA;IACZ,+DAA+D;IAC/D,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,yFAAyF;IACzF,SAAS,CAAC,EAAE,IAAI,CAAA;CACjB,CAAA;AAED,MAAM,MAAM,aAAa,GAAG;IAC1B,kCAAkC;IAClC,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,gCAAgC;IAChC,MAAM,CAAC,EAAE,IAAI,CAAA;IACb,yCAAyC;IACzC,GAAG,CAAC,EAAE,MAAM,CAAA;CACb,CAAA;AAED,wFAAwF;AACxF,MAAM,MAAM,YAAY,GAAG;IACzB,OAAO,EAAE,OAAO,CAAA;IAChB,SAAS,EAAE,MAAM,CAAA;IACjB,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,MAAM,CAAA;IACd,KAAK,EAAE,IAAI,CAAA;IACX,SAAS,EAAE,MAAM,CAAA;IACjB,KAAK,EAAE,MAAM,CAAA;CACd,CAAA;AAED,eAAO,MAAM,qBAAqB,EAAE,YAQnC,CAAA;AAED,4HAA4H;AAC5H,MAAM,MAAM,oBAAoB,GAAG;IACjC,2DAA2D;IAC3D,QAAQ,EAAE,MAAM,CAAA;IAChB,yDAAyD;IACzD,KAAK,EAAE,MAAM,CAAA;IACb,IAAI,EAAE,SAAS,GAAG,sBAAsB,CAAA;CACzC,CAAA;AAID,eAAO,MAAM,sBAAsB,EAAE,oBAIpC,CAAA;AAED,MAAM,MAAM,aAAa,GAAG;IAC1B,KAAK,CAAC,EAAE,YAAY,CAAA;IACpB,GAAG,CAAC,EAAE,UAAU,CAAA;IAChB,MAAM,CAAC,EAAE,aAAa,CAAA;IACtB,yEAAyE;IACzE,KAAK,CAAC,EAAE,OAAO,CAAC,YAAY,CAAC,CAAA;IAC7B,gFAAgF;IAChF,IAAI,CAAC,EAAE,OAAO,CAAC,oBAAoB,CAAC,CAAA;IACpC,SAAS,CAAC,EAAE,eAAe,CAAA;IAC3B,cAAc,CAAC,EAAE,cAAc,CAAA;CAChC,CAAA;AAED,eAAO,MAAM,sBAAsB;;;;;;;;;;;;;;;;;;;CAMlC,CAAA;AAED,MAAM,WAAW,WAAW;IAC1B,GAAG,EAAE,MAAM,CAAA;IACX,SAAS,EAAE,MAAM,CAAA;CAClB;AA2CD,qBAAa,MAAM;IACjB,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAsB;IAE7C,MAAM,CAAC,WAAW,IAAI,MAAM;IAO5B,OAAO,CAAC,MAAM,CAAmB;IACjC,OAAO,CAAC,MAAM,CAAY;IAC1B,OAAO,CAAC,OAAO,CAAmB;IAClC,OAAO,CAAC,kBAAkB,CAAmB;IAC7C,OAAO,CAAC,MAAM,CAAS;IACvB,OAAO,CAAC,mBAAmB,CAAY;IACvC,OAAO,CAAC,gBAAgB,CAAuB;IAE/C,OAAO,CAAC,KAAK,CAAoC;IACjD,OAAO,CAAC,GAAG,CAAqD;IAChE,OAAO,CAAC,YAAY,CAAkD;IACtE,OAAO,CAAC,kBAAkB,CAAY;IACtC,OAAO,CAAC,SAAS,CAAuB;IACxC,OAAO,CAAC,UAAU,CAAI;IACtB,OAAO,CAAC,cAAc,CAA8B;IACpD,OAAO,CAAC,YAAY,CAAa;IACjC,OAAO,CAAC,aAAa,CAAoB;IACzC,OAAO,CAAC,YAAY,CAAoB;IACxC,OAAO,CAAC,YAAY,CAAoB;IACxC,OAAO,CAAC,mBAAmB,CAAoB;IAC/C,OAAO,CAAC,kBAAkB,CAAoB;IAC9C,OAAO,CAAC,aAAa,CAAoB;IACzC,OAAO,CAAC,YAAY,CAAoB;IACxC,OAAO,CAAC,WAAW,CAAoB;IACvC,OAAO,CAAC,oBAAoB,CAAoB;IAChD,OAAO,CAAC,iBAAiB,CAAoB;IAC7C,OAAO,CAAC,oBAAoB,CAAoB;IAChD,OAAO,CAAC,2BAA2B,CAAqB;IACxD,OAAO,CAAC,eAAe,CAAoB;IAC3C,OAAO,CAAC,2BAA2B,CAAqB;IACxD,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,iCAAiC,CAAqB;IAC9D,OAAO,CAAC,iBAAiB,CAAe;IACxC,OAAO,CAAC,wBAAwB,CAAe;IAC/C,OAAO,CAAC,kBAAkB,CAAa;IACvC,OAAO,CAAC,iBAAiB,CAAa;IACtC,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,iBAAiB,CAAI;IAiB7C,OAAO,CAAC,SAAS,CAAkC;IACnD;0FACsF;IACtF,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,iBAAiB,CAAI;IAC7C;;;;;;;gGAO4F;IAC5F,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,iBAAiB,CAA+B;IACxE,OAAO,CAAC,sBAAsB,CAAa;IAC3C,OAAO,CAAC,kBAAkB,CAAa;IACvC,OAAO,CAAC,eAAe,CAAiB;IACxC,OAAO,CAAC,oBAAoB,CAA0B;IACtD,OAAO,CAAC,uBAAuB,CAA0B;IAIzD,OAAO,CAAC,yBAAyB,CAAoB;IACrD,OAAO,CAAC,sBAAsB,CAAoB;IAClD,OAAO,CAAC,wBAAwB,CAAqB;IACrD,OAAO,CAAC,kBAAkB,CAAe;IACzC,OAAO,CAAC,sBAAsB,CAAY;IAE1C,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAsB;IAQ3D,OAAO,CAAC,YAAY,CAAa;IACjC,OAAO,CAAC,sBAAsB,CAAY;IAC1C,OAAO,CAAC,0BAA0B,CAAY;IAC9C,OAAO,CAAC,QAAQ,CAAC,oBAAoB,CAAsB;IAC3D,OAAO,CAAC,QAAQ,CAAC,wBAAwB,CAAsB;IAC/D,OAAO,CAAC,iBAAiB,CAAoB;IAC7C,OAAO,CAAC,uBAAuB,CAAoB;IACnD,OAAO,CAAC,qBAAqB,CAAoB;IACjD,OAAO,CAAC,wBAAwB,CAAqB;IACrD,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,4BAA4B,CAAqB;IACzD,OAAO,CAAC,gBAAgB,CAAa;IACrC,OAAO,CAAC,cAAc,CAAa;IACnC,OAAO,CAAC,aAAa,CAAI;IACzB,OAAO,CAAC,iBAAiB,CAAuB;IAChD,OAAO,CAAC,eAAe,CAAuB;IAC9C,OAAO,CAAC,kBAAkB,CAAe;IACzC,OAAO,CAAC,yBAAyB,CAAqB;IACtD,OAAO,CAAC,uBAAuB,CAAqB;IACpD,2EAA2E;IAC3E,OAAO,CAAC,mBAAmB,CAA0B;IACrD,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,gBAAgB,CAAI;IAG5C,OAAO,CAAC,kBAAkB,CAAC,CAAW;IACtC,OAAO,CAAC,iBAAiB,CAAC,CAAW;IACrC,OAAO,CAAC,SAAS,CAAQ;IACzB,OAAO,CAAC,gBAAgB,CAAa;IACrC,OAAO,CAAC,kBAAkB,CAAiB;IAC3C,OAAO,CAAC,cAAc,CAAa;IACnC,OAAO,CAAC,WAAW,CAAiB;IACpC,OAAO,CAAC,MAAM,CAAC,QAAQ,CAAC,eAAe,CAAO;IAC9C,OAAO,CAAC,mBAAmB,CAAoB;IAC/C,OAAO,CAAC,mBAAmB,CAAY;IACvC,OAAO,CAAC,mBAAmB,CAAuB;IAClD,OAAO,CAAC,qBAAqB,CAAC,CAAc;IAC5C,OAAO,CAAC,uBAAuB,CAAa;IAC5C,OAAO,CAAC,0BAA0B,CAAC,CAAW;IAC9C,OAAO,CAAC,cAAc,CAAwB;IAE9C,OAAO,CAAC,SAAS,CAAC,CAAiB;IACnC,OAAO,CAAC,cAAc,CAAwD;IAC9E,OAAO,CAAC,aAAa,CAAI;IACzB,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAM;IAEvC,OAAO,CAAC,YAAY,CAAoB;IACxC,OAAO,CAAC,2BAA2B,CAAqB;IACxD,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,8BAA8B,CAAqB;IAC3D,OAAO,CAAC,qBAAqB,CAAe;IAC5C,OAAO,CAAC,WAAW,CAAa;IAChC,OAAO,CAAC,gBAAgB,CAAa;IACrC,OAAO,CAAC,kBAAkB,CAAY;IACtC,OAAO,CAAC,WAAW,CAAwC;IAE3D,OAAO,CAAC,cAAc,CAAmC;IACzD,OAAO,CAAC,eAAe,CAAa;IACpC,OAAO,CAAC,YAAY,CAAgC;IACpD,OAAO,CAAC,eAAe,CAAiC;IACxD,OAAO,CAAC,cAAc,CAA0B;IAChD,OAAO,CAAC,mBAAmB,CAAI;IAG/B,OAAO,CAAC,SAAS,CAAO;IACxB,OAAO,CAAC,cAAc,CAAO;IAG7B,OAAO,CAAC,iBAAiB,CAAqB;IAC9C,OAAO,CAAC,oBAAoB,CAAS;IACrC,OAAO,CAAC,kBAAkB,CAA0B;IAEpD,OAAO,CAAC,aAAa,CAAoB;IACzC,OAAO,CAAC,qBAAqB,CAAI;IACjC,OAAO,CAAC,aAAa,CAAoB;IACzC,OAAO,CAAC,YAAY,CAAI;IACxB,OAAO,CAAC,cAAc,CAAI;IAC1B,OAAO,CAAC,KAAK,CAGZ;IACD,OAAO,CAAC,gBAAgB,CAAsB;IAC9C,OAAO,CAAC,kBAAkB,CAA4B;IACtD,OAAO,CAAC,aAAa,CAAe;IACpC,OAAO,CAAC,aAAa,CAAuB;gBAEhC,MAAM,EAAE,iBAAiB,EAAE,OAAO,CAAC,EAAE,aAAa;IAuB9D,qEAAqE;IACrE,MAAM,CAAC,kBAAkB,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,YAAY;IAcxE,MAAM,CAAC,0BAA0B,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC,oBAAoB,CAAC,GAAG,oBAAoB;IAShG,uEAAuE;IACvE,eAAe,IAAI,YAAY;IAa/B,uBAAuB,IAAI,oBAAoB;IAK/C,uBAAuB,CAAC,KAAK,EAAE,OAAO,CAAC,oBAAoB,CAAC,GAAG,IAAI;IAUnE,OAAO,CAAC,0BAA0B;IAmBlC,wEAAwE;IACxE,eAAe,CAAC,KAAK,EAAE,OAAO,CAAC,YAAY,CAAC,GAAG,IAAI;IAoBnD,OAAO,CAAC,kBAAkB;IAsBpB,IAAI;IAyCV,OAAO,CAAC,WAAW;IA2EnB,OAAO,CAAC,oBAAoB;IAiC5B,OAAO,CAAC,eAAe;IA4qBvB,OAAO,CAAC,WAAW;IAYnB,OAAO,CAAC,YAAY;IAuNpB,OAAO,CAAC,WAAW;IAmBnB,iFAAiF;IACjF,eAAe,CAAC,CAAC,EAAE,IAAI,GAAG,IAAI;IAC9B,gGAAgG;IAChG,eAAe,CAAC,KAAK,EAAE,KAAK,GAAG,IAAI,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,IAAI,GAAG,IAAI;IAoB3E,mIAAmI;IACnI,eAAe,CAAC,KAAK,EAAE,KAAK,GAAG,IAAI,EAAE,QAAQ,CAAC,EAAE,MAAM,EAAE,MAAM,CAAC,EAAE,IAAI,GAAG,IAAI;IAY5E,iBAAiB,IAAI,MAAM;IAG3B,iBAAiB,CAAC,CAAC,EAAE,MAAM,GAAG,IAAI;IAGlC,cAAc,IAAI,MAAM;IAGxB,cAAc,CAAC,CAAC,EAAE,MAAM,GAAG,IAAI;IAG/B,aAAa,IAAI,MAAM;IAGvB,aAAa,CAAC,CAAC,EAAE,MAAM,GAAG,IAAI;IAK9B,OAAO,CAAC,aAAa;IAYrB;;;;OAIG;IACH,OAAO,CAAC,UAAU;IASlB,qFAAqF;IACrF,OAAO,CAAC,QAAQ;IAgBhB,gGAAgG;IAChG,QAAQ,CAAC,OAAO,EAAE,YAAY,GAAG,IAAI;IAMrC,0FAA0F;IAC1F,MAAM,CAAC,OAAO,EAAE,UAAU,GAAG,IAAI;IAUjC,QAAQ,IAAI,QAAQ,CAAC;QAAE,KAAK,EAAE,IAAI,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAC;IAGvD,MAAM,IAAI,QAAQ,CAAC;QAAE,KAAK,EAAE,IAAI,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,IAAI,CAAA;KAAE,CAAC;IAItE,SAAS,CAAC,OAAO,CAAC,EAAE;QAClB,KAAK,CAAC,EAAE,MAAM,CAAA;QACd,MAAM,CAAC,EAAE,MAAM,CAAA;QACf,YAAY,CAAC,EAAE,IAAI,CAAA;QACnB,SAAS,CAAC,EAAE,MAAM,CAAA;QAClB,OAAO,CAAC,EAAE,MAAM,CAAA;QAChB,cAAc,CAAC,EAAE,MAAM,CAAA;QACvB,WAAW,CAAC,EAAE,MAAM,CAAA;QACpB,aAAa,CAAC,EAAE,MAAM,CAAA;QACtB,eAAe,CAAC,EAAE,MAAM,CAAA;QACxB,aAAa,CAAC,EAAE,IAAI,CAAA;QACpB,aAAa,CAAC,EAAE,MAAM,CAAA;KACvB,GAAG,IAAI;IA4BR,OAAO,CAAC,iBAAiB;IAIzB,QAAQ,IAAI,WAAW;IAIvB,aAAa,CAAC,QAAQ,CAAC,EAAE,MAAM,IAAI;IAgBnC,cAAc;IAQd,OAAO;IAkBD,SAAS,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,KAAK,CAAC;IACvC,SAAS,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,KAAK,CAAC;IACrD,SAAS,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,yBAAyB,GAAG,OAAO,CAAC,KAAK,CAAC;IAyB3E,QAAQ,CAAC,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,MAAM,EAAE,WAAW,CAAC,EAAE,WAAW,GAAG,OAAO,CAAC,MAAM,CAAC;IAcxG,WAAW,CAAC,IAAI,EAAE,MAAM,GAAG,IAAI;IAiB/B,aAAa,IAAI,MAAM,EAAE;IAIzB,QAAQ,CAAC,IAAI,EAAE,MAAM,GAAG,KAAK,GAAG,IAAI;IAIpC,qBAAqB,CAAC,gBAAgB,CAAC,EAAE,MAAM,GAAG,KAAK,GAAG,IAAI;IAe9D,kBAAkB,CAAC,SAAS,EAAE,MAAM,EAAE,OAAO,EAAE,iBAAiB,GAAG,IAAI;IASvE,kBAAkB,CAAC,SAAS,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,EAAE,OAAO,EAAE,OAAO,GAAG,IAAI;IAOnF,qBAAqB,CAAC,SAAS,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,GAAG,IAAI;IAOpE,iBAAiB,CAAC,SAAS,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,GAAG,OAAO;IAKnE,YAAY,CAAC,OAAO,EAAE,OAAO,GAAG,IAAI;IAIpC,YAAY,IAAI,OAAO;IAIvB,iBAAiB,CAAC,OAAO,EAAE,OAAO,GAAG,IAAI;IAIzC,iBAAiB,IAAI,OAAO;IAI5B,OAAO,CAAC,eAAe;IAIvB,OAAO,CAAC,eAAe;IAWvB,OAAO,CAAC,kBAAkB;YAOZ,kBAAkB;IAiHhC,OAAO,CAAC,oBAAoB;IAwE5B,OAAO,CAAC,2BAA2B;IA4DnC,OAAO,CAAC,kBAAkB,CAAO;IACjC,OAAO,CAAC,mBAAmB;YAeb,yBAAyB;IA6HvC,OAAO,CAAC,2BAA2B;IAUnC,OAAO,CAAC,mBAAmB;IAU3B,OAAO,CAAC,oBAAoB;YAId,4BAA4B;IAuC1C,OAAO,CAAC,eAAe;IA6CvB,OAAO,CAAC,YAAY;IASpB,OAAO,CAAC,uBAAuB,CAI9B;IAED,OAAO,CAAC,iBAAiB,CA0BxB;IAED,OAAO,CAAC,cAAc;IAStB,OAAO,CAAC,cAAc;YA+CR,iBAAiB;IA6C/B,MAAM;IAiHN,OAAO,CAAC,kBAAkB;IAW1B,OAAO,CAAC,iBAAiB;IAYzB;;;;;OAKG;IACH,OAAO,CAAC,aAAa;IAoBrB;;;;;OAKG;IACH,OAAO,CAAC,YAAY;IAepB;;;;;OAKG;IACH,OAAO,CAAC,cAAc;IAiBtB;;;;;OAKG;IACH,OAAO,CAAC,gBAAgB;IAexB,OAAO,CAAC,oBAAoB;IAY5B,OAAO,CAAC,kBAAkB;IAa1B,OAAO,CAAC,WAAW;CAwBpB"}

package/dist/engine.js

@@ -65,6 +65,23 @@ export class Engine {
         this.lightData = new Float32Array(64);
         this.lightCount = 0;
         this.resizeObserver = null;
+        // 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).
+        this.hdrFormat = "rgba16float";
         // [exposure, invGamma, _, _,  bloomTint.x, bloomTint.y, bloomTint.z, bloomIntensity]
         this.compositeUniformData = new Float32Array(8);
         this.bloomBlitUniformData = new Float32Array(4);
@@ -276,11 +293,19 @@ 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 = "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) {
             throw new Error("Failed to get WebGPU context.");
@@ -438,7 +463,7 @@ export class Engine {
         // composite pass writes the swapchain. Tonemap moved to composite so bloom
         // (added next) can run on linear HDR.
         const standardBlend = {
-            format: Engine.HDR_FORMAT,
+            format: this.hdrFormat,
             blend: {
                 color: {
                     srcFactor: "src-alpha",
@@ -452,10 +477,21 @@ 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 = { 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 = {
+            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 = [standardBlend, maskBlend];
         const shaderModule = this.device.createShaderModule({
             label: "default model shader",
@@ -889,21 +925,21 @@ 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" },
         });
         // ─── Composite: HDR + bloom → Filmic → swapchain (premultiplied) ───
@@ -921,6 +957,9 @@ 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: {} },
             ],
         });
         const compositeShader = this.device.createShaderModule({
@@ -1029,13 +1068,13 @@ 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,
             });
             // Bloom-mask MRT attachments — same dims + MSAA as HDR so they share the render pass.
@@ -1064,14 +1103,14 @@ 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 = [];
@@ -1186,6 +1225,7 @@ export class Engine {
                         { binding: 1, resource: compositeBloomView },
                         { binding: 2, resource: this.bloomSampler },
                         { binding: 3, resource: { buffer: this.compositeUniformBuffer } },
+                        { binding: 4, resource: this.maskResolveView },
                     ],
                 });
             }
@@ -2320,13 +2360,17 @@ export class Engine {
 }
 Engine.instance = null;
 Engine.MULTISAMPLE_COUNT = 4;
-Engine.HDR_FORMAT = "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). */
 Engine.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. */
-Engine.BLOOM_MASK_FORMAT = "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. */
+Engine.BLOOM_MASK_FORMAT = "rg8unorm";
 Engine.BLOOM_MAX_LEVELS = 7;
 Engine.SHADOW_MAP_SIZE = 2048;

package/dist/shaders/body.d.ts

@@ -1,2 +1,2 @@
-export declare const BODY_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_B: f32 = 3.141592653589793;\nconst BODY_ROUGHNESS: f32 = 0.3;\n// Dump: \u5C42\u6743\u91CD.002 Blend; \u8FD0\u7B97.007 POWER exponent Value_001; \u80CC\u666F Color; \u8FD0\u7B97.004 after invert\nconst BODY_RIM2_LAYER_BLEND: f32 = 0.20000000298023224;\nconst BODY_RIM2_POW: f32 = 1.4300000667572021;\nconst BODY_RIM2_BG: vec3f = vec3f(1.0, 0.4303792119026184, 0.3315804898738861);\nconst BODY_WARM_AO_MUL: f32 = 0.30000001192092896;\nconst BODY_SPECULAR: f32 = 0.5;\nconst BODY_MIX_NPR: f32 = 0.5;\n// EEVEE Light Clamp equivalent \u2014 caps firefly specular from noise-bumped NDF aliasing.\nconst BODY_SPEC_CLAMP: f32 = 10.0;\n\n// smoothstep-based ramp: t*t*(3-2*t) between two color stops\nfn ramp_ease(f: f32, p0: f32, c0: vec4f, p1: f32, c1: vec4f) -> vec4f {\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@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) mask: f32,\n};\n\n@fragment fn fs(input: VertexOutput) -> FSOut {\n  let alpha = material.alpha;\n  if (alpha < 0.001) { discard; }\n\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\n  let tex_color = textureSample(diffuseTexture, diffuseSampler, input.uv).rgb;\n  let shadow = sampleShadow(input.worldPos, n);\n\n  // \u2550\u2550\u2550 TOON MASK: ShaderToRGB \u2192 ramp.008 CONSTANT [0\u2192black, 0.2966\u2192white] \u2550\u2550\u2550\n  let ndotl_raw = shader_to_rgb_diffuse(n, l, sun, light.ambientColor.xyz, shadow);\n  let toon = ramp_constant(ndotl_raw, 0.0, vec4f(0,0,0,1), 0.2966, vec4f(1,1,1,1)).r;\n\n  // \u2550\u2550\u2550 TOON COLOR: Mix.004 A=HueSat, B=HueSat.001, Fac=ramp.008 (R) \u2550\u2550\u2550\n  let shadow_tint = hue_sat_id(2.0, 0.3499999940395355, 1.0, tex_color);\n  let lit_tint = hue_sat_id(1.5, 1.0, 1.0, tex_color);\n  let toon_color = mix_blend(toon, shadow_tint, lit_tint);\n  let bc = bright_contrast(toon_color, 0.1, 0.2);\n\n  // \u2550\u2550\u2550 AO CHAIN: AO \u2192 ramp CONSTANT [0\u2192white, 0.5995\u2192black] \u2192 Mix.003 \u2550\u2550\u2550\n  let ao = 1.0; // ao_fake(n, v) \u2014 no SSAO yet; inline 1.0 so the ramp/mix chain folds at compile time.\n  let ao_ramp = ramp_constant(ao, 0.0, vec4f(1,1,1,1), 0.5995, vec4f(0,0,0,1)).r;\n  let ao_mixed = mix_blend(ao_ramp, bc, vec3f(0.8301780223846436, 0.3345769941806793, 0.27946099638938904));\n\n  // \u2550\u2550\u2550 EMISSION.003 (strength=4.0) \u2550\u2550\u2550\n  let emission3 = ao_mixed * 4.0;\n\n  // \u2550\u2550\u2550 WARM: \u989C\u8272\u6E10\u53D8.008 \u2192 \u8FD0\u7B97.006 ADD +0.5 (m_graphs) \u2192 clamp \u2192 \u989C\u8272\u6E10\u53D8.003 \u2550\u2550\u2550\n  let ao_inv = invert_f(1.0, ao_ramp);\n  let warm_str = ao_inv * BODY_WARM_AO_MUL;\n  let warm_input = clamp(toon + 0.5, 0.0, 1.0);\n  let warm_color = ramp_cardinal(warm_input, 0.2409,\n    vec4f(0.2426, 0.068, 0.0588, 1.0), 0.4663,\n    vec4f(0.6677, 0.5024, 0.5126, 1.0)).rgb;\n  let warm_emission = warm_color * warm_str;\n\n  // \u2550\u2550\u2550 RIM 1: \u83F2\u6D85\u5C14 \u00D7 \u5C42\u6743\u91CD.001 Facing Blend=0.24 \u2192 \u81EA\u53D1\u5149 Strength \u2550\u2550\u2550\n  let rim1_str = fresnel(2.0, n, v) * layer_weight_facing(0.24000005424022675, n, v);\n  let rim1 = vec3f(0.984157919883728, 0.6110184788703918, 0.5736401677131653) * rim1_str;\n\n  // \u2550\u2550\u2550 RIM 2: \u5C42\u6743\u91CD.002 Facing \u2192 \u8FD0\u7B97.007 POWER \u2192 \u989C\u8272\u6E10\u53D8.010 EASE \u2192 MixShader.002 Fac \u2550\u2550\u2550\n  let facing_raw = layer_weight_facing(BODY_RIM2_LAYER_BLEND, n, v);\n  let facing_pow = math_power(facing_raw, BODY_RIM2_POW);\n  let rim2_fac = ramp_ease(facing_pow, 0.0, vec4f(0,0,0,1), 0.5052, vec4f(1,1,1,1)).r;\n  let rim2_mixed = mix(emission3, BODY_RIM2_BG, rim2_fac);\n\n  // \u2550\u2550\u2550 NPR STACK: AddShader chain (no bright gate in body) \u2550\u2550\u2550\n  let add0 = rim1 + rim2_mixed;\n  let npr_stack = add0 + warm_emission;\n\n  // \u2550\u2550\u2550 PRINCIPLED BSDF: noise bump, GGX specular, SSS from AO \u2550\u2550\u2550\n  // Mapping loc=rot=0 \u2192 plain scale multiply, inline.\n  let noise_val = tex_noise_d2(input.worldPos * vec3f(1.0, 1.0, 1.5), 1.0);\n  let noise_ramp = ramp_linear(noise_val, 0.0, vec4f(0,0,0,1), 1.0, vec4f(1,1,1,1)).r;\n  let bumped_n = bump_lh(0.324644535779953, noise_ramp, n, input.worldPos);\n\n  let principled_base = mix_blend(noise_ramp, bc, vec3f(0.6831911206245422, 0.19474034011363983, 0.13732507824897766));\n  let p_emission = bc * 0.2;\n\n  // Reuse 'ao' (ao_fake(n, v) above) \u2014 identical inputs, avoid a second procedural AO pass.\n  let sss = ramp_linear(ao, 0.003, vec4f(0,0,0,1), 1.0, vec4f(0.0786, 0.0786, 0.0786, 1.0)).r;\n\n  // \u539F\u7406\u5316BSDF (EEVEE port): metallic=0, specular=0.5, roughness=0.3, specular_tint=0.\n  let NL = max(dot(bumped_n, l), 0.0);\n  let NV = max(dot(bumped_n, v), 1e-4);\n\n  // f0/f90 per gpu_shader_material_principled.glsl \u2014 specular_tint=0 \u2192 dielectric_f0_color=white.\n  let f0 = vec3f(0.08 * BODY_SPECULAR);\n  let f90 = mix(f0, vec3f(1.0), sqrt(BODY_SPECULAR));\n  let brdf_lut = brdf_lut_sample(NV, BODY_ROUGHNESS);\n  let reflection_color = F_brdf_multi_scatter(f0, f90, brdf_lut.xy);\n\n  // Direct glossy \u2014 bsdf_ggx already includes NL; no F applied here (tinted after accum).\n  // ltc_brdf_scale: EEVEE direct path uses LTC; split-sum LUT path is rescaled to match.\n  let spec_direct_raw = bsdf_ggx(bumped_n, l, v, NL, NV, BODY_ROUGHNESS) * sun * shadow * ltc_brdf_scale_from_lut(brdf_lut);\n  let spec_direct = min(spec_direct_raw, vec3f(BODY_SPEC_CLAMP));\n  // Indirect glossy \u2014 flat world probe (solid color). Phase 2 adds cubemap.\n  let spec_indirect = light.ambientColor.xyz;\n  let spec_radiance = (spec_direct + spec_indirect) * reflection_color;\n\n  // Indirect diffuse = base_color \u00D7 L_w per Blender closure_eval_surface_lib.glsl line 302;\n  // probe_evaluate_world_diff returns radiance (SH-projected, not cosine-convolved).\n  // No (1-F) factor per EEVEE \u2014 it doesn't energy-conserve spec<->diffuse.\n  let diffuse_radiance = principled_base * (sun * NL * shadow / PI_B + light.ambientColor.xyz);\n  let principled = diffuse_radiance + spec_radiance + p_emission + vec3f(sss);\n\n  // \u6DF7\u5408\u7740\u8272\u5668.001: Shader=\u76F8\u52A0\u7740\u8272\u5668.001, Shader_001=\u539F\u7406\u5316BSDF\n  let final_color = mix(npr_stack, principled, BODY_MIX_NPR);\n\n  var out: FSOut;\n  out.color = vec4f(final_color, alpha);\n  out.mask = 1.0;\n  return out;\n}\n\n";
+export declare const BODY_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_B: f32 = 3.141592653589793;\nconst BODY_ROUGHNESS: f32 = 0.3;\nconst BODY_RIM2_LAYER_BLEND: f32 = 0.20000000298023224;\nconst BODY_RIM2_POW: f32 = 1.4300000667572021;\nconst BODY_RIM2_BG: vec3f = vec3f(1.0, 0.4303792119026184, 0.3315804898738861);\nconst BODY_WARM_STR: f32 = 0.30000001192092896;\nconst BODY_SPECULAR: f32 = 0.5;\nconst BODY_MIX_NPR: f32 = 0.5;\n// EEVEE Light Clamp equivalent \u2014 caps firefly specular from noise-bumped NDF aliasing.\nconst BODY_SPEC_CLAMP: f32 = 10.0;\n\n// smoothstep-based ramp: t*t*(3-2*t) between two color stops\nfn ramp_ease(f: f32, p0: f32, c0: vec4f, p1: f32, c1: vec4f) -> vec4f {\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@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) mask: f32,\n};\n\n@fragment fn fs(input: VertexOutput) -> FSOut {\n  let alpha = material.alpha;\n  if (alpha < 0.001) { discard; }\n\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\n  let tex_color = textureSample(diffuseTexture, diffuseSampler, input.uv).rgb;\n  let shadow = sampleShadow(input.worldPos, n);\n\n  let ndotl_raw = shader_to_rgb_diffuse(n, l, sun, light.ambientColor.xyz, shadow);\n  let toon = ramp_constant(ndotl_raw, 0.0, vec4f(0,0,0,1), 0.2966, vec4f(1,1,1,1)).r;\n\n  let shadow_tint = hue_sat_id(2.0, 0.3499999940395355, 1.0, tex_color);\n  let lit_tint = hue_sat_id(1.5, 1.0, 1.0, tex_color);\n  let toon_color = mix_blend(toon, shadow_tint, lit_tint);\n  let bc = bright_contrast(toon_color, 0.1, 0.2);\n\n  let emission3 = bc * 4.0;\n\n  let warm_input = clamp(toon + 0.5, 0.0, 1.0);\n  let warm_color = ramp_cardinal(warm_input, 0.2409,\n    vec4f(0.2426, 0.068, 0.0588, 1.0), 0.4663,\n    vec4f(0.6677, 0.5024, 0.5126, 1.0)).rgb;\n  let warm_emission = warm_color * BODY_WARM_STR;\n\n  let rim1_str = fresnel(2.0, n, v) * layer_weight_facing(0.24000005424022675, n, v);\n  let rim1 = vec3f(0.984157919883728, 0.6110184788703918, 0.5736401677131653) * rim1_str;\n\n  let facing_raw = layer_weight_facing(BODY_RIM2_LAYER_BLEND, n, v);\n  let facing_pow = math_power(facing_raw, BODY_RIM2_POW);\n  let rim2_fac = ramp_ease(facing_pow, 0.0, vec4f(0,0,0,1), 0.5052, vec4f(1,1,1,1)).r;\n  let rim2_mixed = mix(emission3, BODY_RIM2_BG, rim2_fac);\n\n  let npr_stack = rim1 + rim2_mixed + warm_emission;\n\n  // Noise bump \u2014 Mapping loc=rot=0 folds to a plain scale multiply.\n  let noise_val = tex_noise_d2(input.worldPos * vec3f(1.0, 1.0, 1.5), 1.0);\n  let noise_ramp = ramp_linear(noise_val, 0.0, vec4f(0,0,0,1), 1.0, vec4f(1,1,1,1)).r;\n  let bumped_n = bump_lh(0.324644535779953, noise_ramp, n, input.worldPos);\n\n  let principled_base = mix_blend(noise_ramp, bc, vec3f(0.6831911206245422, 0.19474034011363983, 0.13732507824897766));\n  let p_emission = bc * 0.2;\n\n  // Principled BSDF (EEVEE port): metallic=0, specular=0.5, roughness=0.3, specular_tint=0.\n  let NL = max(dot(bumped_n, l), 0.0);\n  let NV = max(dot(bumped_n, v), 1e-4);\n\n  // f0/f90 per gpu_shader_material_principled.glsl \u2014 specular_tint=0 \u2192 dielectric_f0_color=white.\n  let f0 = vec3f(0.08 * BODY_SPECULAR);\n  let f90 = mix(f0, vec3f(1.0), sqrt(BODY_SPECULAR));\n  let brdf_lut = brdf_lut_sample(NV, BODY_ROUGHNESS);\n  let reflection_color = F_brdf_multi_scatter(f0, f90, brdf_lut.xy);\n\n  // Direct glossy \u2014 bsdf_ggx already includes NL; no F applied here (tinted after accum).\n  // ltc_brdf_scale: EEVEE direct path uses LTC; split-sum LUT path is rescaled to match.\n  let spec_direct_raw = bsdf_ggx(bumped_n, l, v, NL, NV, BODY_ROUGHNESS) * sun * shadow * ltc_brdf_scale_from_lut(brdf_lut);\n  let spec_direct = min(spec_direct_raw, vec3f(BODY_SPEC_CLAMP));\n  // Indirect glossy \u2014 flat world probe (solid color). Phase 2 adds cubemap.\n  let spec_indirect = light.ambientColor.xyz;\n  let spec_radiance = (spec_direct + spec_indirect) * reflection_color;\n\n  // Indirect diffuse = base_color \u00D7 L_w per Blender closure_eval_surface_lib.glsl line 302;\n  // probe_evaluate_world_diff returns radiance (SH-projected, not cosine-convolved).\n  // No (1-F) factor per EEVEE \u2014 it doesn't energy-conserve spec<->diffuse.\n  let diffuse_radiance = principled_base * (sun * NL * shadow / PI_B + light.ambientColor.xyz);\n  let principled = diffuse_radiance + spec_radiance + p_emission;\n\n  let final_color = mix(npr_stack, principled, BODY_MIX_NPR);\n\n  var out: FSOut;\n  out.color = vec4f(final_color, alpha);\n  out.mask = 1.0;\n  return out;\n}\n\n";
 //# sourceMappingURL=body.d.ts.map

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

@@ -1 +1 @@
-{"version":3,"file":"body.d.ts","sourceRoot":"","sources":["../../src/shaders/body.ts"],"names":[],"mappings":"AAIA,eAAO,MAAM,gBAAgB,wrgCAyN5B,CAAA"}
+{"version":3,"file":"body.d.ts","sourceRoot":"","sources":["../../src/shaders/body.ts"],"names":[],"mappings":"AAIA,eAAO,MAAM,gBAAgB,gu7BAoM5B,CAAA"}

package/dist/shaders/body.js

@@ -69,11 +69,10 @@ fn sampleShadow(worldPos: vec3f, n: vec3f) -> f32 {
 
 const PI_B: f32 = 3.141592653589793;
 const BODY_ROUGHNESS: f32 = 0.3;
-// Dump: 层权重.002 Blend; 运算.007 POWER exponent Value_001; 背景 Color; 运算.004 after invert
 const BODY_RIM2_LAYER_BLEND: f32 = 0.20000000298023224;
 const BODY_RIM2_POW: f32 = 1.4300000667572021;
 const BODY_RIM2_BG: vec3f = vec3f(1.0, 0.4303792119026184, 0.3315804898738861);
-const BODY_WARM_AO_MUL: f32 = 0.30000001192092896;
+const BODY_WARM_STR: f32 = 0.30000001192092896;
 const BODY_SPECULAR: f32 = 0.5;
 const BODY_MIX_NPR: f32 = 0.5;
 // EEVEE Light Clamp equivalent — caps firefly specular from noise-bumped NDF aliasing.
@@ -131,49 +130,33 @@ struct FSOut {
   let tex_color = textureSample(diffuseTexture, diffuseSampler, input.uv).rgb;
   let shadow = sampleShadow(input.worldPos, n);
 
-  // ═══ TOON MASK: ShaderToRGB → ramp.008 CONSTANT [0→black, 0.2966→white] ═══
   let ndotl_raw = shader_to_rgb_diffuse(n, l, sun, light.ambientColor.xyz, shadow);
   let toon = ramp_constant(ndotl_raw, 0.0, vec4f(0,0,0,1), 0.2966, vec4f(1,1,1,1)).r;
 
-  // ═══ TOON COLOR: Mix.004 A=HueSat, B=HueSat.001, Fac=ramp.008 (R) ═══
   let shadow_tint = hue_sat_id(2.0, 0.3499999940395355, 1.0, tex_color);
   let lit_tint = hue_sat_id(1.5, 1.0, 1.0, tex_color);
   let toon_color = mix_blend(toon, shadow_tint, lit_tint);
   let bc = bright_contrast(toon_color, 0.1, 0.2);
 
-  // ═══ AO CHAIN: AO → ramp CONSTANT [0→white, 0.5995→black] → Mix.003 ═══
-  let ao = 1.0; // ao_fake(n, v) — no SSAO yet; inline 1.0 so the ramp/mix chain folds at compile time.
-  let ao_ramp = ramp_constant(ao, 0.0, vec4f(1,1,1,1), 0.5995, vec4f(0,0,0,1)).r;
-  let ao_mixed = mix_blend(ao_ramp, bc, vec3f(0.8301780223846436, 0.3345769941806793, 0.27946099638938904));
+  let emission3 = bc * 4.0;
 
-  // ═══ EMISSION.003 (strength=4.0) ═══
-  let emission3 = ao_mixed * 4.0;
-
-  // ═══ WARM: 颜色渐变.008 → 运算.006 ADD +0.5 (m_graphs) → clamp → 颜色渐变.003 ═══
-  let ao_inv = invert_f(1.0, ao_ramp);
-  let warm_str = ao_inv * BODY_WARM_AO_MUL;
   let warm_input = clamp(toon + 0.5, 0.0, 1.0);
   let warm_color = ramp_cardinal(warm_input, 0.2409,
     vec4f(0.2426, 0.068, 0.0588, 1.0), 0.4663,
     vec4f(0.6677, 0.5024, 0.5126, 1.0)).rgb;
-  let warm_emission = warm_color * warm_str;
+  let warm_emission = warm_color * BODY_WARM_STR;
 
-  // ═══ RIM 1: 菲涅尔 × 层权重.001 Facing Blend=0.24 → 自发光 Strength ═══
   let rim1_str = fresnel(2.0, n, v) * layer_weight_facing(0.24000005424022675, n, v);
   let rim1 = vec3f(0.984157919883728, 0.6110184788703918, 0.5736401677131653) * rim1_str;
 
-  // ═══ RIM 2: 层权重.002 Facing → 运算.007 POWER → 颜色渐变.010 EASE → MixShader.002 Fac ═══
   let facing_raw = layer_weight_facing(BODY_RIM2_LAYER_BLEND, n, v);
   let facing_pow = math_power(facing_raw, BODY_RIM2_POW);
   let rim2_fac = ramp_ease(facing_pow, 0.0, vec4f(0,0,0,1), 0.5052, vec4f(1,1,1,1)).r;
   let rim2_mixed = mix(emission3, BODY_RIM2_BG, rim2_fac);
 
-  // ═══ NPR STACK: AddShader chain (no bright gate in body) ═══
-  let add0 = rim1 + rim2_mixed;
-  let npr_stack = add0 + warm_emission;
+  let npr_stack = rim1 + rim2_mixed + warm_emission;
 
-  // ═══ PRINCIPLED BSDF: noise bump, GGX specular, SSS from AO ═══
-  // Mapping loc=rot=0 → plain scale multiply, inline.
+  // Noise bump — Mapping loc=rot=0 folds to a plain scale multiply.
   let noise_val = tex_noise_d2(input.worldPos * vec3f(1.0, 1.0, 1.5), 1.0);
   let noise_ramp = ramp_linear(noise_val, 0.0, vec4f(0,0,0,1), 1.0, vec4f(1,1,1,1)).r;
   let bumped_n = bump_lh(0.324644535779953, noise_ramp, n, input.worldPos);
@@ -181,10 +164,7 @@ struct FSOut {
   let principled_base = mix_blend(noise_ramp, bc, vec3f(0.6831911206245422, 0.19474034011363983, 0.13732507824897766));
   let p_emission = bc * 0.2;
 
-  // Reuse 'ao' (ao_fake(n, v) above) — identical inputs, avoid a second procedural AO pass.
-  let sss = ramp_linear(ao, 0.003, vec4f(0,0,0,1), 1.0, vec4f(0.0786, 0.0786, 0.0786, 1.0)).r;
-
-  // 原理化BSDF (EEVEE port): metallic=0, specular=0.5, roughness=0.3, specular_tint=0.
+  // Principled BSDF (EEVEE port): metallic=0, specular=0.5, roughness=0.3, specular_tint=0.
   let NL = max(dot(bumped_n, l), 0.0);
   let NV = max(dot(bumped_n, v), 1e-4);
 
@@ -206,9 +186,8 @@ struct FSOut {
   // probe_evaluate_world_diff returns radiance (SH-projected, not cosine-convolved).
   // No (1-F) factor per EEVEE — it doesn't energy-conserve spec<->diffuse.
   let diffuse_radiance = principled_base * (sun * NL * shadow / PI_B + light.ambientColor.xyz);
-  let principled = diffuse_radiance + spec_radiance + p_emission + vec3f(sss);
+  let principled = diffuse_radiance + spec_radiance + p_emission;
 
-  // 混合着色器.001: Shader=相加着色器.001, Shader_001=原理化BSDF
   let final_color = mix(npr_stack, principled, BODY_MIX_NPR);
 
   var out: FSOut;