@driftengine/core 3.61.0

package/dist/render/shaders/flat/preamble.js

@@ -0,0 +1,864 @@
+/** Version, inputs, uniforms and the defines the rest of the fragment shader is written against. */
+import { COOKIE_TILE } from '../../lightBudget.js';
+import { CLUSTER_TEXELS, CLUSTER_X, CLUSTER_Y, CLUSTER_Z, LIGHT_REGION_TEXELS, LIGHT_TEXELS, MAX_LIGHTS_PER_CLUSTER, TABLE_WIDTH, } from '../../clusteredLights.js';
+import { MAX_SHADOW_FILTER_TAPS } from '../../renderQuality.js';
+import { FOG_GLSL } from '../fog.js';
+/**
+ * The preamble, built for one light budget.
+ *
+ * **A function rather than the constant it was, and the budget is why.** `MAX_LIGHTS` and
+ * `MAX_AREA_LIGHTS` size twenty and fourteen uniform arrays between them, and an array costs a
+ * whole row of the uniform grid per element whatever its base type — so these two numbers are
+ * most of what the lit shader spends against `MAX_FRAGMENT_UNIFORM_VECTORS`, and on a part that
+ * offers 256 of them they are the difference between a program that links and a black page. See
+ * `uniformVectorBudget.ts` for the counting and for the ladder the renderer walks.
+ *
+ * Rebuilt per call rather than memoised: it is 0.84 ms for the whole fragment source, measured,
+ * and a renderer builds it once at construction.
+ */
+export function preambleGlsl({ maxLights, maxAreaLights }) {
+    return `#version 300 es
+precision highp float;
+
+in vec3 vNormal;
+in vec3 vColor;
+in vec3 vWorldPos;
+in float vEmissive;
+in float vSpecular;
+in vec4 vLightPos;
+in vec2 vUv;
+in vec3 vEmissiveColor;
+in float vRoughness;
+/*
+ * How far a bump may tilt the shading normal, and how much roughness that tilt is worth.
+ *
+ * A surface whose normal wanders cannot hold a highlight narrower than the wander, so the
+ * roughness has to grow with the relief or the picture is sparkle. Stated as a constant pair
+ * because the amplitude is known here rather than measured off the screen.
+ */
+#define RELIEF_TILT 1.6
+#define RELIEF_ROUGHNESS 0.35
+/**
+ * The furthest a texture-derived bump may turn the shading normal, as a tangent.
+ *
+ * 1.0 is 45°. See the cap's own comment at the sampling site for why a limit has to exist at
+ * all: past this the perturbation is the same size as the surface it perturbs, and what comes
+ * out is decided by the last bits of a screen-space derivative rather than by the picture.
+ */
+#define RELIEF_MAX_TILT 1.0
+in float vGrain;
+in float vRelief;
+in vec4 vTangent;
+/*
+ * **Declared after vTangent because the vertex stage declares them there.**
+ *
+ * GLSL ES links varyings by *name*, so on WebGL2 the order is free and this pair sat above
+ * vTangent for a whole build without anything going wrong. The generated path does not: GLSL to
+ * SPIR-V assigns varying locations by *declaration order*, so a stage that lists them in a
+ * different order hands WebGPU a vec4 where it expects a float.
+ *
+ * The failure is total and silent in the picture -- "The component count (4) of the vertex output
+ * at location 11 is different from the component count (1) of the fragment input at location 11",
+ * on a pipeline that is then never created, so the backend draws nothing at all while WebGL2 is
+ * pixel-perfect. Found by this row's own check on its first WebGPU run.
+ *
+ * **The rule this is an instance of: the two stages' varying blocks are one list in two places.**
+ * Adding a varying to one and not the other in the same position is a WebGPU-only defect that no
+ * WebGL2 capture can show.
+ */
+in float vSkyDirect;
+in float vAlpha;
+/*
+ * The refraction thickness lane. **Declared here in the position the vertex stage declares it**,
+ * which is the rule 3.42.0 paid for: GLSL links varyings by name so WebGL2 does not care, and the
+ * generated path assigns locations by declaration order, so a stage that lists them differently
+ * hands WebGPU a mismatched component count and no pipeline is created at all.
+ */
+in float vThickness;
+
+/*
+ * Refraction: the colour the frame had already drawn when the first refracting draw arrived.
+ *
+ * **A uniform and a branch, never a permutation.** One more flatFrag axis is about 247 KB gzipped
+ * paid by every consumer whether or not they refract, against 19.8 KB for a branch -- which is the
+ * price table Track D's three shipped rows measured.
+ *
+ * uRefractStrength at 0 is every draw that shipped before this existed, and it is the default, so
+ * the sampler is bound to an empty texture and read nowhere on a frame that refracts nothing.
+ */
+uniform sampler2D uRefractScene;
+uniform float uRefractStrength;
+uniform vec3 uRefractTint;
+uniform float uRefractThickness;
+flat in int vHasTangents;
+
+/**
+ * Surface colour from an image the consumer supplied, multiplied into the vertex colour.
+ *
+ * The engine still ships and fetches no image assets — see surfaceTexture.ts. What this
+ * samples is whatever the caller handed over, which for a procedurally generated canvas
+ * is zero payload and no pipeline, exactly as vertex colour is.
+ *
+ * Multiplied rather than replacing, so vertex colour keeps working as the per-surface
+ * tint it already is: one greyscale tile image lit by different vertex colours is how a
+ * biome recolours a whole world without a second texture, and a mesh whose texture is
+ * white shades identically to one with no texture at all.
+ */
+uniform sampler2D uAlbedo;
+/** 0 leaves every textured path unevaluated, for the majority of geometry with no image. */
+uniform int uAlbedoEnabled;
+/**
+ * Alpha below which a textured fragment is thrown away entirely, 0 to disable.
+ *
+ * The billboard case: a face, a leaf, a sign is an image with a *shape*, and drawing its
+ * transparent corners as opaque black is the difference between a creature floating in
+ * the dark and a rectangle with a creature on it. A cutout rather than blending because
+ * these are depth-writing world geometry — blending them would need sorting, and a hard
+ * edge is what a cutout is for.
+ *
+ * Zero by default, so a texture with no alpha to speak of costs one compare and nothing
+ * else changes.
+ */
+uniform float uAlbedoCutout;
+
+/**
+ * A normal map, in the surface's own space.
+ *
+ * highp for the reason the depth samplers give: a lowp sampler may hand back eight bits, and eight
+ * bits of a direction is a surface quantised into visible facets.
+ *
+ * Sampled linear and never sRGB. These bytes are a direction, not a colour; decoding them as
+ * display values bends every normal toward the surface and reads as a green tint over everything.
+ * SurfaceTextureOptions.colorSpace already defaults to linear and says why.
+ */
+uniform highp sampler2D uNormalMap;
+/**
+ * How hard the map turns the shading normal. 0 is no map at all, and gates the whole block.
+ *
+ * A uniform rather than a permutation flag, following uAlbedoEnabled: a fifth boolean would take
+ * the exhaustive permutation sweep from sixteen shaders to thirty-two, and the generated WGSL that
+ * consumers bundle as source from 766 KB to about 1.5 MB. A permutation is the right tool for the
+ * shadow samplers, where a profile with shadows off should declare none of them; it is the wrong
+ * tool for one image on a shader that already declares one.
+ */
+uniform float uNormalStrength;
+
+/**
+ * Occlusion, roughness and metallic in one image. glTF's packing, because that is what an import
+ * carries: R is occlusion, G roughness, B metallic.
+ *
+ * highp for the reason the normal map gives: a lowp sampler may hand back eight bits, and eight
+ * bits of roughness is a highlight that steps between widths as a surface curves.
+ *
+ * Sampled linear and never sRGB. These three numbers *are* the data, not a picture of it, and
+ * decoding them as display values would bend every one of them toward its floor.
+ */
+uniform highp sampler2D uOrmMap;
+/**
+ * 1 when a map is bound. Gates the whole block, following uAlbedoEnabled.
+ *
+ * A uniform rather than a fifth permutation flag: FlatShaderOptions has four booleans and
+ * flatSource.test.ts sweeps all sixteen exhaustively, so a fifth takes that to thirty-two and
+ * roughly doubles the generated WGSL that consumers bundle as source, this engine having no build
+ * step. A permutation is the right tool for the shadow samplers, where a profile with shadows off
+ * should declare none of them. It is the wrong tool for one more image.
+ */
+uniform int uOrmEnabled;
+/**
+ * What each channel is scaled by, component-aligned with the map: r occlusion, g roughness,
+ * b metallic. The alignment is the documentation.
+ *
+ * **r is applied as a mix from 1 and not as a multiply**, because that is what glTF's
+ * occlusionTexture.strength is: 1 + strength * (texel - 1). A multiply at 0 means fully occluded,
+ * which is the opposite of what a caller asking for no occlusion means.
+ */
+uniform vec3 uOrmScale;
+/**
+ * 1 when \`uEnvironment\` holds the GGX-prefiltered cube, 0 when it holds the capture's box chain.
+ *
+ * **Two chains whose levels mean different things, so two pieces of arithmetic downstream.** A box
+ * level is a smaller picture and wants the texel-footprint floor; a prefiltered level is a lobe and
+ * wants a roughness combined in quadrature. Likewise the weight: the split sum is the correct share
+ * of a *prefiltered* environment and asserts an integral that was never performed over a box chain.
+ *
+ * A uniform rather than a permutation flag, following \`uOrmEnabled\`: a fifth boolean doubles the
+ * generated WGSL every consumer bundles as source, and this is a few multiply-adds.
+ */
+uniform float uEnvironmentPrefiltered;
+
+/**
+ * Where a surface glows, and in what colour.
+ *
+ * **sRGB, unlike the two maps above it.** A normal map is a direction and an ORM map is three
+ * numbers; this one is a colour somebody chose, so it is decoded the way the albedo is. Getting
+ * that backwards is not subtle — an emissive map read as linear comes out visibly dark and muddy
+ * where it should be saturated — but it is silent, because a glow that is too dim reads as a
+ * material choice rather than as a decode.
+ *
+ * A uniform gate rather than a permutation flag, for exactly the reasons \`uOrmEnabled\` gives
+ * above, and now with a number behind them: a fifth flag was built and measured at 49% of the
+ * bundle. See \`ARCHITECTURE.md\` §1.
+ */
+uniform highp sampler2D uEmissiveMap;
+/** 1 when a map is bound. Gates the sample, following uOrmEnabled. */
+uniform int uEmissiveMapEnabled;
+/**
+ * What the map is scaled by, per channel. 1 is the identity and the default.
+ *
+ * The counterpart of \`uOrmScale\`: a factor that multiplies a texture is not a value, and this is
+ * where the multiplier lives once an image supplies the shape.
+ */
+uniform vec3 uEmissiveScale;
+
+#if DIRECTIONAL_SHADOWS
+/*
+ * The share marker below is what lets the environment probe exist, and it costs nothing.
+ *
+ * A device's sampler ceiling is per shader stage. This project's adapter offers exactly
+ * sixteen and the widest permutation of this shader declared seventeen — one albedo, these
+ * three maps, twelve point-shadow cubes and the probe's cube — so the probe was the one that
+ * gave way, and reflective surfaces kept a sky-and-ground gradient on hardware that could
+ * have mirrored the room.
+ *
+ * Every one of the fifteen shadow bindings was already one object: the WebGPU renderer hands
+ * them all the same shadow sampler, and flatPass declares them all non-filtering. So
+ * declaring one instead of fifteen changes nothing about how any of them is read, and takes
+ * the count to three. WebGL2 has no such limit and ignores the marker, which is a line
+ * comment to it.
+ *
+ * Marked on each declaration rather than matched on a name in the generator, deliberately.
+ * flatPass records what a name rule cost: a point-shadow cube's name does not end the way the
+ * three directional maps' names do, and a pattern wide enough to catch it also caught the
+ * probe's colour cube — whose mip chain is its roughness, so it needs a filtering sampler and
+ * cannot share a non-filtering one.
+ *
+ * Two things this comment may not contain, both learned here. No backticks: this is inside a
+ * template literal and one of them ends the shader. And no uniform names from the probe's
+ * half of the file, because a test asserts that a profile without a probe mentions them
+ * nowhere — a comment is source too, and it failed that test first.
+ */
+uniform highp sampler2D uStaticShadowMap;  // wgsl:share shadow
+uniform highp sampler2D uPeeledShadowMap;  // wgsl:share shadow
+uniform highp sampler2D uDynamicShadowMap;  // wgsl:share shadow
+/** 0 disables directional shadows when their source is not emitting. */
+uniform float uShadowStrength;
+uniform float uShadowMapSize;
+/** Orthographic depth span, used to keep the residual bias in world metres. */
+uniform float uShadowDepthSpan;
+/** Horizontal world-space reach over which directional occlusion dissolves. */
+uniform float uShadowMaxDistance;
+/** 0 for a cheaper single-static-depth quality profile. */
+uniform int uPeeledShadowEnabled;
+/** Maximum horizontal projection per vertical metre for directional shadows. */
+uniform float uShadowMaxSlope;
+#endif
+/** One sampling budget drives both directional and point-light filtering. */
+uniform int uShadowFilterTaps;
+uniform vec3 uDirectionalDir;      // surface → dominant source, normalized
+uniform vec3 uDirectionalColor;
+uniform vec3 uAmbient;
+/**
+ * Ambient arriving from below, for a hemispheric fill.
+ *
+ * Ambient from a single colour lights a floor and a ceiling identically, which is the
+ * one thing real ambient never does: skylight comes from above and bounce comes off
+ * whatever you are standing on. A room whose floor is carpet and whose ceiling is white
+ * tile reads as neither if both get the same fill — the whole scene takes one flat tint.
+ *
+ * Defaults to uAmbient, so a consumer that sets only the one colour gets exactly the
+ * uniform fill it always did.
+ */
+uniform vec3 uAmbientGround;
+${FOG_GLSL}
+uniform vec3 uCameraPos;
+/** Optional world plane; reflection passes retain only the viewer's side. */
+uniform vec4 uClipPlane;
+uniform int uClipEnabled;
+uniform float uEmissiveGain;
+uniform float uNightFactor;
+#if NIGHT_EMISSIVE
+/**
+ * Emission that only shows where the dominant light does not reach. 0 is off and is the default.
+ *
+ * The case it exists for is a lamp field on something that turns. Where the lights are is a
+ * property of the surface, so it is authored per vertex; whether they are visible is a property
+ * of where the sun is, and until this term nothing multiplied the two. Measured on the rotating
+ * planet it was reported from, a patch bright enough to read on the night side was sixteen times
+ * too bright on the day side, and the ratio is fixed by the lighting rather than by the value.
+ *
+ * The cut at the terminator is hard on purpose. Softening it is what puts the lights back on the
+ * day side, which is the whole of the defect being fixed.
+ */
+uniform float uNightEmissive;
+#endif
+/**
+ * How solid this draw is, 0 to 1.
+ *
+ * One uniform rather than a second material, for the same reason uTint is one:
+ * everything else about a translucent piece of the world — its lighting, its fog,
+ * its emissive — is what the flat shader already does, and duplicating the whole
+ * of that to change one channel is how a renderer ends up with two shaders that
+ * disagree about the sun. The caller owns the blend state; this owns the number.
+ */
+uniform float uOpacity;
+
+/**
+ * Whether this draw is accumulating into the order-independent transparency buffers, and
+ * therefore weights its own output. 0 is every draw that is not, which is every draw on every
+ * published scene.
+ *
+ * **A uniform rather than a permutation, and the cost is why.** Writing the accumulation and the
+ * revealage from one pass needs a second fragment output, which is another \`flatFrag\` axis —
+ * measured in \`scripts/wgsl.ts\` at about 247 KB gzipped, doubling the sixteen fragment variants,
+ * and paid by every consumer whether or not they ever switch this on. A branch on a uniform costs
+ * a compare, and the geometry is submitted twice instead. See \`orderIndependent.ts\`.
+ *
+ * **The revealage pass does not read this.** It draws with this at 0 and a blend of
+ * \`(ZERO, ONE_MINUS_SRC_ALPHA)\`, so the colour it computes is multiplied away and only its alpha
+ * reaches the target — which is exactly the point: both passes run the same code to decide what
+ * alpha this fragment has, so they cannot disagree about it.
+ */
+uniform float uOitWeighted;
+
+/**
+ * Whether the fragment stage runs any lighting: ambient, the sun, point lights, specular,
+ * reflectivity, procedural grain, emissive and the arrival highlight. 1 keeps every one of
+ * them, which is what every draw did before this existed and what an opaque draw always
+ * does. 0 leaves the surface exactly its own vertex colour times its own texture — three.js
+ * calls that \`meshBasicMaterial\`, and it is what a flat glow shell or a flat backdrop plate
+ * needs: colour and opacity, and nothing else touching it.
+ *
+ * **A uniform branch rather than a fifth permutation, and that is a cost decision, not a
+ * style one.** \`POINT_SHADOWS\`, \`ENVIRONMENT_PROBE\` and the like are \`#if\`'d out because
+ * leaving them in adds a sampler or a uniform-block member a profile without the feature
+ * would still declare and still pay a bind group layout for — sixteen shader permutations
+ * already, doubled again by the translucent pipeline on WebGPU. This adds no binding and no
+ * block member the shader does not already carry the size of; it is one \`int\` beside
+ * \`uOpacity\`, read once. And the branch it drives is on a value that is the same for every
+ * fragment in one draw call, which is exactly the guarantee \`uReflectivity\` and
+ * \`uAlbedoEnabled\` already lean on above: a compiler asked to predicate a whole wavefront on
+ * one uniform skips the untaken side rather than computing both and discarding one, which is
+ * only true because nothing here reads it through a *varying*-dependent branch.
+ */
+uniform int uLightingEnabled;
+/**
+ * Whether this draw recedes into the medium: the atmospheric haze and the underwater tint
+ * both, folded together because both are the *camera's* medium rather than a property of the
+ * surface — see \`mediumFog\`/\`mediumColor\` in \`fog.ts\`. 1 keeps both, matching every draw
+ * before this existed.
+ *
+ * **Independent of \`uLightingEnabled\`, and that is the point of it being a second uniform
+ * rather than the same one.** "Unlit but still fading into the haze" is a real thing to
+ * want — a marker at the edge of view distance should still say it is far away even if
+ * nothing about it is shaded — so coupling the two would take away a look this shader can
+ * otherwise express for free. The caller chooses the combination; this only offers both.
+ *
+ * **The trade-off worth stating: with this at 0, the underwater tint is skipped too, which
+ * is the one place this draw can now disagree with the rest of a submerged world.** Every
+ * other pass applies \`uUnderwaterFactor\` unconditionally so that crossing the surface changes
+ * the whole scene together — see the comment on \`waterTransmission\` below. A translucent draw
+ * that asked to be unfogged is asking for its colour untouched by the medium full stop, which
+ * is what "exactly as its colour says" has to mean for it; the cost is that such a draw will
+ * not tint if the camera goes under water, unlike everything else in the frame.
+ */
+uniform int uFogEnabled;
+
+/**
+ * A pass-level scale over the grain the geometry declared, 0 to 1.
+ *
+ * Which surfaces are mineral is aGrain, per vertex, because it is a property of a
+ * material and nothing else in the frame can know it. This is the override on top: an
+ * imported model that declares nothing is drawn with the pass's own answer, and a scene
+ * can dial a whole draw down without rebuilding its geometry.
+ */
+uniform float uGrain;
+/*
+ * The material's own relief: how strong, and how many bumps to a metre.
+ *
+ * Two numbers rather than one because they are what separates the surfaces this exists for.
+ * Asphalt is coarse and deep, orange peel on paint is fine and shallow, cast concrete sits
+ * between them, and the geometry carrying them is identical in all three cases. See
+ * Renderer.setSurfaceRelief.
+ */
+uniform float uRelief;
+uniform float uReliefCycles;
+/**
+ * How hard the bound surface texture's own luminance turns the shading normal, 0 for not at all.
+ *
+ * **The other half of relief, and the half a photograph can give.** \`uRelief\` invents structure
+ * from noise, which is right for asphalt and concrete because their structure has no particular
+ * arrangement. A rock, a bark, a beaten-metal plate does: the bumps are *where the picture says
+ * they are*, and noise cannot know. This reads the height off the image already bound as colour,
+ * so one photograph lights as the surface it is a photograph of.
+ *
+ * **The image serves both roles rather than a second sampler being added**, which is a decision
+ * and not a shortcut. This shader's widest permutation already declares one sampler more than
+ * this project's adapter offers, and is legal only because of the \`wgsl:share\` markers on the
+ * shadow cubemaps; a second image sampler here is not the free addition it would be in a smaller
+ * shader. It is also what the consumer this was built for actually does — three.js's own
+ * \`bumpMap\` is a height field, and a scene handing one image to \`map\` and \`bumpMap\` together
+ * is the ordinary case rather than an unusual one.
+ *
+ * Scaled exactly as three.js's \`bumpScale\` is, so a scene porting one carries the number across
+ * rather than refitting it, and negative inverts the relief there and here alike. Not clamped to
+ * 0..1 like its neighbours for the same reason: this is an amplitude against a luminance
+ * gradient, not a fraction of anything.
+ */
+uniform float uTextureRelief;
+
+/**
+ * How much of the environment this material mirrors, 0 to 1.
+ *
+ * Specular here is a highlight lobe and nothing else: a surface reflected the four lamps
+ * above it and not the room around it. That is why polished paint read as opaque no matter
+ * how tight the lobe was made, because what makes a car body look wet is mostly the world
+ * in it rather than the lights on it.
+ *
+ * Off by default, so every scene written before it is unchanged.
+ */
+uniform float uReflectivity;
+/**
+ * How bright the environment these surfaces reflect is, as a multiplier on it.
+ *
+ * **This exists because a metal has no other way to be lit.** Its Fresnel base is its own albedo,
+ * so \`f0\` is 1 and the environment blend reaches full weight at every angle: what a metal shows
+ * is \`environment * albedo\` and nothing else — no ambient, no direct diffuse. When the only
+ * environment available is a low-resolution cube of a room, a metal is exactly as bright as that
+ * room's own surfaces happen to be, and in an interior lit by a handful of lamps that is dark. A
+ * consumer reported it as an imported model reading dark in every world at once, which is the tell
+ * that it is not any one world's lighting.
+ *
+ * It is a **stand-in for image-based lighting** and it says so. A real prefiltered environment
+ * carries the sky and the sources a probe of nearby geometry never sees, and it would make this
+ * unnecessary; until there is one, the honest thing is one named number rather than thirty worlds
+ * each re-lit around the same missing term. See the \`image-based-lighting\` row in
+ * docs/CAPABILITIES.md.
+ *
+ * Applies to whatever the surface reflects, probe or hemispheric approximation alike, and to the
+ * probe's own irradiance where a caller asked for that. **1 is the default and the identity**, so
+ * every scene written before this is unchanged.
+ */
+uniform float uEnvironmentGain;
+/**
+ * The room itself, as a cubemap baked from one point, and whether one exists.
+ *
+ * highp because a lowp sampler may hand back eight bits and this one carries an image the eye
+ * looks *at* rather than a term that modulates something else. See the depth samplers.
+ *
+ * Sampled at a mip level chosen by roughness, which is the cheapest honest gloss blur there
+ * is: a polished surface takes level 0 and a satin one takes a blurred level, for the price of
+ * a fetch rather than a pass. This is the documented exception to the textureLod rule in
+ * AGENTS.md 2026-08-07 — the level is *wanted* here — and it is still an explicit level rather
+ * than an implicit derivative, so the rule holds either way.
+ *
+ * When no probe has been baked this stays at zero and the gradient below is used instead. It is
+ * gated on a flag rather than on the texture being bound, because an unbaked cubemap contains
+ * whatever the driver left in it, and a car mirroring uninitialised memory is worse than a car
+ * mirroring an approximation.
+ */
+#if ENVIRONMENT_PROBE
+uniform highp sampler2DArray uEnvironment;
+uniform float uEnvironmentEnabled;
+uniform float uEnvironmentMaxLod;
+/**
+ * Texels across one probe's map at level 0, which the footprint arithmetic needs and
+ * \`uEnvironmentMaxLod\` no longer carries.
+ *
+ * They were one number while the environment was a cube: the chain ran to a single texel a face,
+ * so the top level *was* the log of the face size. The chain now stops one level below the cosine
+ * convolution, so the top level is 4 while the map is 256 across, and a footprint expressed as
+ * \`exp2(uEnvironmentMaxLod)\` would say sixteen. That reads as every metal in the scene being
+ * under-sampled and forced to a roughness it does not have.
+ */
+uniform float uEnvironmentEdge;
+/**
+ * The level holding the cosine convolution, which is the scene's diffuse ambient.
+ *
+ * **One binding carries both integrals**, and this is the line between them. Levels up to
+ * \`uEnvironmentMaxLod\` are the reflection at the roughness each stands for; this one is the
+ * room's diffuse light, convolved against the cosine at bake time. Folding them into one array is
+ * what leaves WebGL2's last guaranteed texture unit free.
+ *
+ * **It replaces nine spherical-harmonic coefficients**, which were projected on the CPU, read back
+ * off the GPU and uploaded again — a synchronous \`readPixels\` on one backend and a \`mapAsync\`
+ * answering a frame or two later on the other, with a second gate uniform to cover the gap. A
+ * second-order fit also could not carry a sharp source; an eight-texel octahedral image carries
+ * strictly more than nine coefficients did, and it costs a fetch rather than a readback.
+ */
+uniform float uEnvironmentIrradianceLevel;
+/**
+ * Where the probes stand, as a lattice: the first probe, the reciprocal of the step, and the
+ * counts.
+ *
+ * **Four uniforms whatever the probe count, and that is not an economy but a requirement.** This
+ * shader already declares 35 array uniforms totalling 385 uniform vectors against WebGL2's
+ * guaranteed \`MAX_FRAGMENT_UNIFORM_VECTORS\` of 224, so it is 1.7x over the guarantee before a
+ * probe is placed. Sixteen probes of position and extent would add about 48 vectors to a budget
+ * nobody can raise. A lattice needs none: the layer index is arithmetic.
+ *
+ * Counts of \`(1, 1, 1)\` is a single probe, which is what every scene that bakes one environment
+ * becomes, and it is not a special case anywhere below.
+ */
+uniform vec3 uProbeGridOrigin;
+uniform vec3 uProbeGridInvSpacing;
+uniform vec3 uProbeGridCounts;
+/**
+ * Whether the grid supplies the scene's diffuse ambient, or only its reflections.
+ *
+ * \`ProbeBakeOptions.irradiance\`, at grid scope. A bake used to do two things inseparably: fill
+ * a cube that reflective surfaces sample, and replace every diffuse surface's ambient with the
+ * room's own light. A consumer who wanted a car to mirror the village also got the whole scene
+ * relit at whatever moment the bake landed, and filed it four times over two days as four
+ * different bugs.
+ *
+ * Zero here declines the second half and the hemispheric gradient below is what lights the scene,
+ * exactly as it did before any probe existed.
+ */
+uniform float uProbeGridAmbient;
+#endif
+/** A world-space box that lights up regardless of time of day (run finished). */
+uniform vec3 uHighlightMin;
+uniform vec3 uHighlightMax;
+uniform float uHighlightGain;
+
+/*
+ * The froxel table, and the numbers needed to find a fragment's place in it.
+ *
+ * **Every layout constant is interpolated from \`clusteredLights.ts\` rather than written here.**
+ * The binner, this shader and the WGSL kernel are three readers of one layout, and two of them
+ * agreeing because both contain the number 8 is the shape of defect the conformance script exists
+ * to catch. There is one definition and all three read it.
+ *
+ * **Declared unconditionally rather than behind a permutation, and that is a measured reversal.**
+ * \`ARCHITECTURE.md\` §1 makes a feature inside the lit pass a compile-on-demand permutation, and a
+ * fifth flag doubled the generated WGSL from sixteen variants to thirty-two — **914 KB to 1,906 KB
+ * raw and 400,728 to 597,638 bytes gzipped, a 49% rise on every consumer including those who never
+ * turn this on.** Gzip cannot dedupe near-identical copies across that span; its window is 32 KB
+ * and a permutation is larger than that. So this is a runtime branch on a uniform, and what it
+ * costs is stated below rather than hidden. \`ARCHITECTURE.md\` carries the finding.
+ *
+ * **\`texelFetch\` rather than \`texture\`, and it is not only a filtering choice.** An integer
+ * texture cannot be filtered at all, so the sampler this declaration would otherwise bring is dead
+ * weight against a per-stage ceiling; the \`wgsl:share\` marker gives it the shadow group's one,
+ * which it never uses. It also means the 2026-08-07 rule does not apply here: \`texelFetch\` takes
+ * no derivative, so reading it inside the light loop needs no explicit level and cannot be the
+ * flattened-branch cost that rule is about.
+ */
+#define CLUSTER_X ${CLUSTER_X}
+#define CLUSTER_Y ${CLUSTER_Y}
+#define CLUSTER_Z ${CLUSTER_Z}
+#define CLUSTER_TEXELS ${CLUSTER_TEXELS}
+#define MAX_LIGHTS_PER_CLUSTER ${MAX_LIGHTS_PER_CLUSTER}
+#define LIGHT_TEXELS ${LIGHT_TEXELS}
+#define LIGHT_REGION_TEXELS ${LIGHT_REGION_TEXELS}
+#define CLUSTER_TABLE_WIDTH ${TABLE_WIDTH}
+uniform highp usampler2D uClusterTable;  // wgsl:share shadow
+/**
+ * The view matrix, so a fragment can find its own froxel.
+ *
+ * **From view space rather than from \`gl_FragCoord\`, and that is deliberate.** A WebGL2
+ * framebuffer counts y upward and a WebGPU one counts it downward, so a tile row derived from the
+ * fragment's window coordinate would be mirrored on one backend — the same trap the 2026-08-17
+ * rule about \`dFdy\` is written for, arriving through a different door. View space has no such
+ * disagreement, and it is also the space the binner works in, so the shader and the binner answer
+ * the same question with the same arithmetic.
+ */
+uniform mat4 uView;
+/** near, far, tan(fovY/2) and aspect: the four numbers the froxel grid is defined by. */
+uniform vec4 uClusterFrustum;
+/**
+ * 1 when the light loop reads the froxel table, 0 when it reads the uniform slots.
+ *
+ * A uniform, so the branch is uniform control flow and every fragment in a draw takes the same
+ * arm — which is what keeps it cheap and what keeps the 2026-08-07 rule satisfied. The table is
+ * only ever read with \`texelFetch\`, which takes no derivative, so nothing here needs an explicit
+ * level either way.
+ *
+ * **What this costs when it is 0**: one texture unit, one matrix and one vec4 in the block, and a
+ * compare per light. What it buys is that a consumer who never enables it does not download a
+ * second copy of the whole shader — see the note above the table's declaration.
+ */
+uniform int uClustered;
+
+/**
+ * One texel of the table, addressed by its flat index.
+ *
+ * The width divides both regions exactly, which is why this is a mask and a shift rather than a
+ * divide — see \`TABLE_WIDTH\` for the arithmetic that guarantees it.
+ */
+uvec4 clusterTexel(int texel) {
+  return texelFetch(
+    uClusterTable,
+    ivec2(texel % CLUSTER_TABLE_WIDTH, texel / CLUSTER_TABLE_WIDTH),
+    0);
+}
+
+#define MAX_LIGHTS ${maxLights}
+/*
+ * How many times the light loop may go round.
+ *
+ * The clustered arm iterates a froxel's own list and the fixed arm iterates the uniform slots, so
+ * the bound differs; GLSL needs a constant either way. Defined here rather than at the loop so
+ * both numbers sit beside the budgets they come from.
+ */
+/*
+ * **Both arms have the same bound, and that equality is deliberate.** There is one loop, so a
+ * larger clustered cap would raise the bound for the fixed path too, and a scene that never asks
+ * for froxels would carry a loop of 28 where it carries 16 today. \`MAX_LIGHTS_PER_CLUSTER\` is
+ * pinned equal to \`MAX_LIGHTS\` for that reason; a test in \`textureUnitBudget.test.ts\` asserts it.
+ */
+#define LIGHT_LOOP_MAX MAX_LIGHTS
+uniform int uLightCount;
+/** 0 keeps the radius-shaped falloff; 1 is physical inverse-square with a soft cutoff. */
+uniform int uLightFalloff;
+uniform vec3 uLightPos[MAX_LIGHTS];
+uniform vec3 uLightColor[MAX_LIGHTS];
+uniform float uLightRadius[MAX_LIGHTS];
+/*
+ * Each light's emitter radius in metres. Zero is a mathematical point, which is what every
+ * source was until a highlight narrower than a fragment turned out to be the cause of speckle
+ * on glossy paint. See sphereLobe.
+ */
+uniform float uLightSourceRadius[MAX_LIGHTS];
+/*
+ * How present each light is, 0 to 1: the ramp it crosses the edge of view range on,
+ * and the one it trades the last slot on. A world carries more lights than
+ * MAX_LIGHTS, so membership changes as the camera moves, and without this it changes
+ * by switching. See pointLightSelection.ts.
+ */
+uniform float uLightWeight[MAX_LIGHTS];
+/**
+ * Where each spot points, normalised, in world space. Unused by a light with no cone.
+ *
+ * Declared for every light rather than as a separate spot array, because a spot is a point light
+ * with a direction and giving it arrays of its own would put a second loop in the largest shader
+ * here — see \`ARCHITECTURE.md\` section 1 on what a fifth permutation cost.
+ */
+uniform vec3 uLightDir[MAX_LIGHTS];
+/**
+ * The cosine of each light's inner cone angle, then of its outer.
+ *
+ * A light with no cone carries -1 and -2, which makes \`smoothstep(outer, inner, dot)\` exactly 1
+ * for every direction on the sphere. See \`POINT_LIGHT_COS_OUTER\`: the pair is chosen so a point
+ * light collapses to the arithmetic it had rather than to something very close to it.
+ */
+uniform vec2 uLightCone[MAX_LIGHTS];
+/**
+ * Which row of the photometric atlas each light uses, or a negative index for none.
+ *
+ * A float rather than an int because it travels through the froxel record as one, and because the
+ * shader turns it into a texture coordinate immediately anyway.
+ */
+uniform float uLightIesProfile[MAX_LIGHTS];
+/**
+ * Where each light's photometric azimuth zero points, in world space.
+ *
+ * **Read only by an asymmetric profile**, which is a street light or a wall washer and is almost
+ * nothing else — every axially symmetric fixture ignores it, and a zero vector or one parallel to
+ * the aim leaves the profile on its first plane. Orthogonalised against the light's direction in
+ * the shader, so a caller hands the fixture's own forward and the aim decides the rest.
+ *
+ * **It cannot be derived from the direction.** There is no continuous field of unit vectors
+ * tangent to a sphere, so any reference built from the aim alone flips somewhere — and the obvious
+ * constructions put that flip exactly where these fixtures point, \`cross(worldUp, dir)\` being
+ * singular for a light aimed straight down.
+ */
+uniform vec3 uLightIesAxis[MAX_LIGHTS];
+/**
+ * Which tile of the cookie atlas each light projects, or a negative index for none.
+ *
+ * A cookie is a mask the fixture throws — a window frame, a grille, foliage — and it is oriented
+ * by \`uLightIesAxis\`, the same reference a photometric profile's azimuth uses, because both are
+ * answering the same question about a fixture: which way is up.
+ */
+uniform float uLightCookie[MAX_LIGHTS];
+/**
+ * Photometric profiles: one row per fixture, a fixture's intensity by vertical angle across it.
+ *
+ * **Every row spans the same 0 to 180 degree arc**, whatever the file measured, which is what lets
+ * a light carry only a row index — see \`iesProfile.ts\`. With no profile loaded this holds a
+ * single row of ones, the multiplicative identity, because a declared sampler needs a complete
+ * texture whether or not the branch reads it and a row of zeros would switch off every light.
+ */
+uniform sampler2D uIesAtlas;
+/** Rows in the atlas, so a row index becomes a coordinate. At least 1. */
+uniform float uIesAtlasRows;
+/**
+ * Horizontal planes each profile occupies, so a row is \`profile * planes + plane\`.
+ *
+ * **1 for an atlas with nothing asymmetric in it**, which is nearly every atlas, and the whole
+ * plane path is branched away on that — see \`photometric\` in \`main.ts\`. One count for the atlas
+ * rather than one per profile, because a per-profile count would need a second number in a light's
+ * record to divide by, and a light's record is where space is expensive.
+ */
+uniform float uIesPlaneCount;
+/**
+ * Cookies: one row of square tiles, one tile a fixture's mask.
+ *
+ * **0 when no consumer has loaded one, and the whole projection is branched away on it** — a
+ * \`uniform float\` is provably uniform control flow, so a scene with no cookie performs no
+ * arithmetic and no fetch. With none loaded this holds a single white texel, the multiplicative
+ * identity, because a declared sampler needs a complete texture whether or not the branch reads it
+ * and a black stand-in would switch off every light that reached it.
+ */
+uniform sampler2D uCookieAtlas;
+uniform float uCookieTiles;
+/**
+ * Half a texel of a tile, so a cookie's edge cannot reach its neighbour in the atlas.
+ *
+ * A row of tiles filtered bilinearly bleeds across the seam between two of them, and the seam is
+ * exactly where a cone's edge lands — so the neighbour's colour would appear as a rim on every
+ * cookie. Derived from \`COOKIE_TILE\` rather than written twice.
+ */
+#define COOKIE_INSET (0.5 / ${COOKIE_TILE}.0)
+/**
+ * Rectangular area lights, and how many of them are live.
+ *
+ * **Four, and a branch on this count rather than a shader permutation.** \`ARCHITECTURE.md\` section
+ * 1 says a lit-pass feature earns a permutation only when carrying it compiled-in costs more than
+ * doubling the shader corpus, and this session measured both sides of that: thirty lines of shared
+ * code cost 6,215 gzipped bytes across the sixteen permutations, and fifty cost 10,517 — while the
+ * fifth flag clustered lighting nearly took cost 196,910. So the whole area-light term compiled in
+ * unconditionally is roughly a twentieth of what permuting it would cost, and the count being zero
+ * is what makes it free at runtime.
+ *
+ * Four because a rectangle is a *fixture* — a window, a softbox, a panel — and a scene with more
+ * than a handful is a scene that wants them baked. Raising it is a uniform-array size and an
+ * iteration of the loop, exactly as \`MAX_LIGHTS\` is.
+ */
+#define MAX_AREA_LIGHTS ${maxAreaLights}
+uniform int uAreaLightCount;
+/** Centre of each rectangle, in world space. */
+uniform vec3 uAreaLightPos[MAX_AREA_LIGHTS];
+/** Colour times intensity, as the point lights carry it. */
+uniform vec3 uAreaLightColor[MAX_AREA_LIGHTS];
+/** The rectangle's in-plane axes, unit length, and its half extents along each. */
+uniform vec3 uAreaLightRight[MAX_AREA_LIGHTS];
+uniform vec3 uAreaLightUp[MAX_AREA_LIGHTS];
+uniform vec2 uAreaLightSize[MAX_AREA_LIGHTS];
+/**
+ * Whether the rectangle emits from both faces, 1 or 0.
+ *
+ * A window is one-sided and a hanging panel is two-sided, and the difference is visible: a
+ * one-sided light behind a surface must contribute nothing rather than lighting it from behind,
+ * which is what an unsigned form factor would do.
+ */
+uniform float uAreaLightTwoSided[MAX_AREA_LIGHTS];
+/** Pi, for turning a dot product into the angle a photometric row is indexed by. */
+#define PI_IES 3.14159265359
+/** A whole turn, for mapping an azimuth onto the plane grid. */
+#define TAU_IES 6.28318530718
+/** Index into the light arrays that each shadow cubemap belongs to, -1 if unused. */
+/**
+ * Sampled shadow maps — one per shaded light, deliberately.
+ *
+ * Any number below MAX_LIGHTS means some lights illuminate without occluding,
+ * and *which* ones changes as the camera moves. That reads as shadows switching
+ * on and off as you walk up to a lamp, which is worse than either extreme: a
+ * missing shadow is merely flat, but one that appears and vanishes draws the
+ * eye straight to it. Matching the two budgets is what makes the effect
+ * impossible rather than rare.
+ *
+ * **One array texture, a layer per light, where there were twelve cubemaps.** That is
+ * where eleven of WebGL2's sixteen guaranteed texture units came back from, and it is
+ * also why the arrays below are indexed by the light rather than by a slot: GLSL ES
+ * cannot index a sampler array with a non-constant expression, so selecting a cubemap
+ * meant ten unrolled comparison arms. A layer is an ordinary integer, so it is a lookup.
+ */
+#if POINT_SHADOWS
+/** Which array layer holds this light's shadow, or -1 for none. */
+uniform int uPointShadowLayer[MAX_LIGHTS];
+uniform float uPointShadowFar[MAX_LIGHTS];
+uniform float uPointShadowNear[MAX_LIGHTS];
+/** Emitter radius per light — drives how soft its shadows are. */
+uniform float uPointShadowSize[MAX_LIGHTS];
+// How present each bound map is, 0 to 1. See the ramp at the sampling site.
+uniform float uPointShadowWeight[MAX_LIGHTS];
+/*
+ * **highp, and it is load-bearing.** GLSL ES gives a fragment shader's samplers a default
+ * precision of lowp, whatever the file declares for its floats: that line is about float,
+ * not about a sampler. A lowp sampler is entitled to return eight bits.
+ *
+ * This holds DEPTH_COMPONENT24 — see pointShadowArray.ts — carrying a radius as a fraction
+ * of the light's far plane. Eight bits of that is a shadow quantised to eight steps of the
+ * light's range, which reads as concentric rings around every lamp. The conversion is only a
+ * multiply now rather than the perspective divide a cubemap needed, so the error is no longer
+ * *magnified*; it is still eight bits where twenty-four were stored.
+ *
+ * The 2D shadow maps a hundred lines up already carry the qualifier, and the composite pass
+ * was found reading depth at eight bits in 0.13.0. This is the same fault in the third place
+ * it exists.
+ */
+uniform highp sampler2DArray uPointShadows;  // wgsl:share shadow
+
+/**
+ * Moving casters are separate from persistent world maps. Two live maps exist
+ * only so ownership can crossfade when another light becomes nearer.
+ *
+ * Layers of the same array, and addressed by the light rather than by the pair, because a
+ * light owns at most one of the two at a time.
+ */
+uniform int uLivePointShadowLayer[MAX_LIGHTS];
+uniform float uLivePointShadowFar[MAX_LIGHTS];
+uniform float uLivePointShadowNear[MAX_LIGHTS];
+uniform float uLivePointShadowSize[MAX_LIGHTS];
+uniform float uLivePointShadowWeight[MAX_LIGHTS];
+
+/**
+ * A rectangular area light's occlusion, in the same array a point light's lives in.
+ *
+ * **The same array because there is no second texture unit to put one in.** WebGL2 guarantees
+ * sixteen and the lit pass binds sixteen — \`FlatShaderOptions.environmentProbe\` documents the
+ * same wall from the other side, a seventeenth declared sampler silently sharing unit 0 with a
+ * \`sampler2D\` shadow map. So an area shadow is a layer above the point pool rather than a
+ * texture of its own, which also means **it rides \`POINT_SHADOWS\`**: a consumer whose scene is
+ * lit by rectangles alone still needs that flag on, and \`RenderQuality.pointShadows\` is where it
+ * is turned on. The renderer says so by name rather than drawing an unoccluded rectangle.
+ *
+ * Indexed by area-light slot, which is the same index the arrays above it use, because
+ * \`selectAreaLights\` fills both from the caller's list in order. \`-1\` in a layer says this
+ * rectangle has no image — a slot that does not cast, or one whose first bake has not landed.
+ */
+uniform int uAreaShadowLayer[MAX_AREA_LIGHTS];
+uniform float uAreaShadowFar[MAX_AREA_LIGHTS];
+uniform float uAreaShadowNear[MAX_AREA_LIGHTS];
+/** How present the image is, so a first bake arrives rather than appears. */
+uniform float uAreaShadowWeight[MAX_AREA_LIGHTS];
+/**
+ * The layer holding movers only, composed with the one above by multiplication.
+ *
+ * One per casting rectangle rather than a shared pair, which is where this differs from the point
+ * lights above: those borrow two live maps and crossfade ownership because fifty lamps compete for
+ * them, and at most \`MAX_AREA_LIGHTS\` rectangles exist, so each simply owns one. Nothing changes
+ * hands, so nothing has to fade across a handover.
+ */
+uniform int uLiveAreaShadowLayer[MAX_AREA_LIGHTS];
+uniform float uLiveAreaShadowFar[MAX_AREA_LIGHTS];
+uniform float uLiveAreaShadowNear[MAX_AREA_LIGHTS];
+uniform float uLiveAreaShadowWeight[MAX_AREA_LIGHTS];
+#endif
+
+/**
+ * Omnidirectional shadow test against a depth cubemap.
+ *
+ * The comparison happens in *linear* distance, not in stored depth. Perspective
+ * depth is wildly non-linear: near the far plane a bias of a few thousandths
+ * spans metres of world space, which silently makes everything unshadowed. So
+ * the sampled depth is converted back to a distance and compared in metres,
+ * where a bias means what it looks like it means.
+ */
+/**
+ * Sample offsets forming a rough sphere, for filtering a cubemap shadow.
+ * Perturbing the *direction* is how PCF works on a cube — there is no 2D UV
+ * neighbourhood to step across the way there is for the sun's cascade.
+ */
+#if POINT_SHADOWS
+const vec3 PCF_OFFSETS[${MAX_SHADOW_FILTER_TAPS}] = vec3[${MAX_SHADOW_FILTER_TAPS}](
+  vec3( 1.0,  1.0,  1.0), vec3( 1.0, -1.0, -1.0), vec3(-1.0,  1.0, -1.0), vec3(-1.0, -1.0,  1.0),
+  vec3( 1.0, -1.0,  1.0), vec3(-1.0,  1.0,  1.0), vec3( 1.0,  1.0, -1.0), vec3(-1.0, -1.0, -1.0),
+  vec3( 1.0,  0.0,  0.0), vec3(-1.0,  0.0,  0.0), vec3( 0.0,  1.0,  0.0), vec3( 0.0, -1.0,  0.0)
+);
+#endif
+
+/** A finite point-light shadow loses strength with the light that casts it. */`;
+}