@photivra/engine 2.21.0 → 2.22.0

package/dist/gpu/wgsl-display-view-quality-kernel.js.map

@@ -1 +1 @@
-{"version":3,"file":"wgsl-display-view-quality-kernel.js","sourceRoot":"","sources":["../../src/gpu/wgsl-display-view-quality-kernel.ts"],"names":[],"mappings":"AAAA,sCAAsC;AAEtC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAiDG;AAEH,OAAO,EAAE,wBAAwB,EAAE,4BAA4B,EAAE,4BAA4B,EAAE,gCAAgC,EAAE,MAAM,qCAAqC,CAAC;AAC7K,OAAO,EAAE,aAAa,EAAE,cAAc,EAAE,eAAe,EAAE,aAAa,EAAE,MAAM,2BAA2B,CAAC;AAC1G,OAAO,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AAErD,MAAM,CAAC,MAAM,mCAAmC,GAAG,EAAE,CAAC;AACtD,qEAAqE;AACrE,MAAM,CAAC,MAAM,4CAA4C,GAAG,CAAC,CAAC;AAC9D,gIAAgI;AAChI,MAAM,CAAC,MAAM,kCAAkC,GAAG,EAAE,CAAC;AACrD,mEAAmE;AACnE,MAAM,CAAC,MAAM,gCAAgC,GAAG,EAAE,CAAC;AACnD,4HAA4H;AAC5H,MAAM,CAAC,MAAM,gCAAgC,GAAG,CAAC,CAAC;AAClD,8DAA8D;AAC9D,MAAM,CAAC,MAAM,8BAA8B,GAAG,CAAC,CAAC;AAChD,iIAAiI;AACjI,MAAM,CAAC,MAAM,gCAAgC,GAAG,CAAC,CAAC;AAClD,gIAAgI;AAChI,MAAM,CAAC,MAAM,kCAAkC,GAAG,EAAE,CAAC;AACrD,iFAAiF;AACjF,MAAM,CAAC,MAAM,oCAAoC,GAAG,EAAE,CAAC;AACvD,0IAA0I;AAC1I,MAAM,CAAC,MAAM,sCAAsC,GAAG,EAAE,CAAC;AACzD,MAAM,CAAC,MAAM,qCAAqC,GAAG,CAAC,CAAC;AACvD,MAAM,sBAAsB,GAAG,CAAC,GAAG,sCAAsC,GAAG,qCAAqC,CAAC;AAClH,+CAA+C;AAC/C,MAAM,CAAC,MAAM,wBAAwB,GAAG,eAAe,CAAC,MAAM,CAAC,MAAM,CAAC;IACpE,kHAAkH;IAClH,QAAQ,EAAE,CAAC;IACX,kFAAkF;IAClF,SAAS,EAAE,GAAG;IACd,0DAA0D;IAC1D,aAAa,EAAE,GAAG;IAClB,2IAA2I;IAC3I,UAAU,EAAE,GAAG;IACf,oBAAoB,EAAE,sBAAsB;IAC5C,IAAI,EAAE,GAAG,GAAG,CAAC,GAAG,sBAAsB;CACvC,CAAC,CAAC;AACH,+CAA+C;AAC/C,MAAM,CAAC,MAAM,yBAAyB,GAAG,eAAe,CAAC,MAAM,CAAC,MAAM,CAAC;IACrE,uFAAuF;IACvF,iBAAiB,EAAE,CAAC;IACpB,oGAAoG;IACpG,iBAAiB,EAAE,CAAC;IACpB,uHAAuH;IACvH,YAAY,EAAE,CAAC;IACf,+FAA+F;IAC/F,YAAY,EAAE,CAAC;IACf,oFAAoF;IACpF,SAAS,EAAE,EAAE;IACb;;;;OAIG;IACH,SAAS,EAAE,EAAE;IACb,uIAAuI;IACvI,kBAAkB,EAAE,EAAE;IACtB,sDAAsD;IACtD,SAAS,EAAE,EAAE;CACd,CAAC,CAAC;AACH,oGAAoG;AACpG,MAAM,CAAC,MAAM,+BAA+B,GAAG,CAAC,CAAC;AACjD;;;;GAIG;AACH,MAAM,CAAC,MAAM,iCAAiC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;AAC1D,iFAAiF;AACjF,MAAM,CAAC,MAAM,iCAAiC,GAAG,IAAI,CAAC,IAAI,CAAC,cAAc,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC;AACrF;;;;GAIG;AACH,MAAM,CAAC,MAAM,sCAAsC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;AAC/D,kIAAkI;AAClI,MAAM,CAAC,MAAM,4BAA4B,GAAG,GAAG,CAAC;AAChD,kEAAkE;AAClE,MAAM,gBAAgB,GAAG,EAAE,CAAC;AAC5B,oFAAoF;AACpF,MAAM,SAAS,GAAG,uBAAuB,CAAC;AAC1C,MAAM,OAAO,GAAG,CAAC,KAAa,EAAU,EAAE;IACxC,MAAM,IAAI,GAAG,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;IACxC,OAAO,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,IAAI,CAAC;AACjD,CAAC,CAAC;AAEF,6EAA6E;AAC7E,MAAM,CAAC,MAAM,gCAAgC,GAAG,UAAU,CAAC;iBAC1C,SAAS;;;2BAGC,OAAO,CAAC,wBAAwB,CAAC;+BAC7B,OAAO,CAAC,4BAA4B,CAAC;4BACxC,OAAO,CAAC,4BAA4B,CAAC;+BAClC,OAAO,CAAC,4BAA4B,CAAC;mCACjC,OAAO,CAAC,gCAAgC,CAAC;gCAC5C,OAAO,CAAC,gBAAgB,CAAC;iCACxB,OAAO,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;sBAC5B,oCAAoC;oBACtC,gCAAgC;qBAC/B,gCAAgC;mBAClC,8BAA8B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAmJhD,CAAC;AAEF,gGAAgG;AAChG,MAAM,aAAa,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;iCAyBA,wBAAwB,CAAC,aAAa;;iCAEtC,wBAAwB,CAAC,aAAa,GAAG,EAAE;;;;;;;;;;;;;;CAc3E,CAAC;AAEF;;;;;GAKG;AACH,MAAM,CAAC,MAAM,iCAAiC,GAAG,UAAU,CAAC,GAAG,gCAAgC;;;2BAGpE,mCAAmC;;mCAE3B,mCAAmC;;;;;;;;;;;;;;;;;CAiBrE,CAAC;AAEF;;;GAGG;AACH,MAAM,CAAC,MAAM,mCAAmC,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;CAqB7D,CAAC;AAEF;;;GAGG;AACH,MAAM,gBAAgB,GAAG,CAAC,UAAmB,EAAU,EAAE,CAAC,UAAU,CAAC;8DACP,UAAU,CAAC,CAAC,CAAC;2DAChB,CAAC,CAAC,CAAC,EAAE;;;;;EAK9D,cAAc;2BACW,kCAAkC;;;;;;;;;;;;;;;CAe5D,CAAC;AAEF;;;;;;;;;;;;;GAaG;AACH,MAAM,CAAC,MAAM,+BAA+B,GAAG,UAAU,CAAC,GAAG,gCAAgC,GAAG,gBAAgB,CAAC,IAAI,CAAC,GAAG,mCAAmC;2BACjI,mCAAmC;;uCAEvB,mCAAmC;;;;;;;;;;;;;;;;;;;;;;;;CAwBzE,CAAC;AAEF;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,CAAC,MAAM,uCAAuC,GAAG,UAAU,CAAC,GAAG,gCAAgC,GAAG,gBAAgB,CAAC,KAAK,CAAC;EAC7H,mCAAmC;2BACV,mCAAmC;;uCAEvB,mCAAmC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA0CzE,CAAC;AAEF,2DAA2D;AAC3D,MAAM,CAAC,MAAM,gCAAgC,GAAG,UAAU,CAAC,GAAG,gCAAgC;0BACpE,cAAc;yBACf,OAAO,CAAC,aAAa,CAAC;2BACpB,OAAO,CAAC,eAAe,CAAC;yBAC1B,OAAO,CAAC,aAAa,CAAC;uBACxB,eAAe,GAAG,aAAa,GAAG,CAAC;yBACjC,wBAAwB,CAAC,UAAU;2BACjC,sBAAsB;8BACnB,yBAAyB,CAAC,iBAAiB;4BAC7C,yBAAyB,CAAC,iBAAiB;0BAC7C,yBAAyB,CAAC,YAAY;wBACxC,yBAAyB,CAAC,YAAY;8BAChC,yBAAyB,CAAC,SAAS;2BACtC,yBAAyB,CAAC,SAAS;8BAChC,yBAAyB,CAAC,kBAAkB;4BAC9C,OAAO,CAAC,iCAAiC,CAAC;;;;;;;;;;;EAWpE,aAAa;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;2BA0EY,mCAAmC;;mCAE3B,mCAAmC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;mBA+CnD,wBAAwB,CAAC,QAAQ;;;;;;;iDAOH,wBAAwB,CAAC,SAAS;;;;;;;mBAOhE,wBAAwB,CAAC,QAAQ;;;;;;;iCAOnB,qCAAqC;;;;;;;;;;;;;;;;;;;;;;;;;;mCA0BnC,qCAAqC;;;sDAGlB,qCAAqC,kCAAkC,qCAAqC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;uCA0C3H,qCAAqC;;;0DAGlB,qCAAqC,kCAAkC,qCAAqC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;mCAsDnI,qCAAqC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;gBA6BxD,kCAAkC;;;;;;;;;;CAUjD,CAAC;AAEF,yIAAyI;AACzI,MAAM,UAAU,qCAAqC,CAAC,kBAA0B;IAC9E,OAAO,iCAAiC,GAAG,IAAI,GAAG,kBAAkB,GAAG,IAAI,CAAC;AAC9E,CAAC","sourcesContent":["// SPDX-License-Identifier: Apache-2.0\n\n/**\n * WGSL kernels of the display view's Quality path on the GPU (#479): the exact area mean with diffraction kernels, the\n * anti-aliasing filter's spots, the AHD candidates and the low-signal correction, in three passes over one band of display\n * rows (docs/WEBGPU.md, \"Display view (Quality) on the GPU\"; the float64 reference is `renderDisplayView` with\n * `createDisplayViewExpectedSignal`, facade/display-view.ts and facade/expected-signal.ts):\n *\n * 1. **Lattice** (`DISPLAY_VIEW_QUALITY_LATTICE_WGSL` plus the material module): one invocation per lattice point of each\n *    lattice group (offsets that share a fractional shift, as the reference groups them): the chart material at\n *    (p + 0.5 + f_x − W/2)·g, −(q + 0.5 + f_y − H/2)·g (the reference's lattice formula, an exact float32 sum times the\n *    scale rounded once), with the module's uncertainty flag in the top bit.\n * 2. **Sites** (`DISPLAY_VIEW_QUALITY_SITES_WGSL`): one invocation per native site of the band's window: the expected\n *    electrons Σ_t V_c(m_t, t) over the taps t of every group (the combined kernel values per material and channel, formed on\n *    the host in float64 as the reference forms them: Σ_members w·Σ_λ A_c(λ)·K_λ·L_m(λ); without diffraction the offset\n *    weights times the material's channel signal), plus the dark electrons; the readout moments (mean k·e, variance k²e + c₀,\n *    third cumulant k³e) and the clip at the white level; and each quantity's running absolute error bound. With the defocus\n *    disc (#492) the variant `DISPLAY_VIEW_QUALITY_SITES_BLOCKED_WGSL` reads a factored table (one neutral row, one row per\n *    spectral material) and sums each group's dense tap rectangle in blocks, so its bound grows with the kernel's width.\n * 3. **Pixels** (`DISPLAY_VIEW_QUALITY_PIXELS_WGSL`): one invocation per display pixel of the band: for every real pixel of\n *    the footprint, the demosaic candidates (one bilinear, or AHD's two directional linear candidates, each with its full\n *    camera-channel covariance), white balance and the highlight clip, the colour matrix, the tone curve and the clip to\n *    [0, 1] with the skewness term, the candidates' average, the Stein covariance with the closed-form B_c, the independent\n *    residual, AHD's choice variance ¼·Var(H − V), the flag for AHD's choice on the noise-free mosaic, which the reference takes and\n *    the kernel does not (#457), and the low-signal correction interpolated trilinearly between the grid\n *    states of a table the host uploaded (the states themselves are computed on the host, #447; the kernel records every\n *    grid key it needs in a bitmap and flags a display pixel whose state the table lacks).\n *\n * **Highlights (#533).** The site pass records whether each site is certainly below white, certainly at or above it, or\n * undecided (`clip_state`, from the site's mean and variance and their running bounds: certain only beyond twice the reference's\n * 8.5 standard deviations after moving the mean by twice its bound, where the reference's clip probability is exactly 0 or 1), in\n * bits 1 and 2 of its flag word. The pixel pass marks a real pixel saturated (P = 1) when a site its demosaic reads is certainly\n * clipped, and the display pixel `highlightUndecided` when a real pixel is neither that nor certainly unclipped. Three flag words\n * per display pixel: the flags, then a bit per real pixel j·nx + i of the footprint with P = 1 (36 bits).\n *\n * **Running bounds.** Every quantity after the site table is carried as a pair (value, absolute bound) in float32: the bound\n * of a sum is the sum of the bounds plus one rounding of the result, of a product |a|·Δb + |b|·Δa plus one rounding, of a\n * square root Δv/(2s) plus 4.5 ULP, of a quotient 2.5 ULP plus the propagated terms; the clipped-normal moments add the\n * Speed kernel's clip constants (CLIP_MEAN_U, CLIP_VARIANCE_U per standard deviation or variance) and propagate their inputs\n * with the moments' first-order derivatives at the pixel's own standardized clip levels (∂E/∂m = P(inside), ∂E/∂s = φa − φb,\n * ∂Var/∂m = −s(ρ_a + ρ_b), ∂Var/∂s = s(2ρ − aρ_a − bρ_b) with ρ_a = 2Φa(a − E[g]), ρ_b = 2Qb(b − E[g]), and the slope's\n * (φa − φb)/s and (aφa − bφb)/s; without noise a slope is exact unless the mean's interval straddles a clip level); the\n * Reinhard quadrature adds REINHARD_MEAN_U and REINHARD_VARIANCE_U. Mixed-sign sums (the AHD\n * kernels, the colour matrix) are therefore bounded by the magnitudes of their terms, not relative to their result. The\n * pixel's expected value and variance leave the kernel with their absolute bounds; the host doubles them (MARGIN) and adds\n * the footprint sum's relative term. Second-order terms (products of bounds) and the bound arithmetic's own float32\n * roundings are what the doubling covers.\n *\n * Precision assumptions are those of the Speed kernel: the W3C WGSL accuracy table (+ − × one rounding in either direction,\n * E = 2^-23; division 2.5 ULP, sqrt 4.5 ULP, exp (3 + 2|x|) ULP, log2 2^-21 absolute on [0.5, 2] and 3 ULP outside; no\n * trigonometric built-in).\n */\n\nimport { DISPLAY_VIEW_CLIP_MEAN_U, DISPLAY_VIEW_CLIP_VARIANCE_U, DISPLAY_VIEW_REINHARD_MEAN_U, DISPLAY_VIEW_REINHARD_VARIANCE_U } from \"./wgsl-display-view-speed-kernel.js\";\nimport { AHD_KEY_FLOOR, AHD_KEY_LEVELS, AHD_KEY_OCTAVES, AHD_KEY_STEPS } from \"../facade/display-view.js\";\nimport { EXACT_SUM_WGSL } from \"./wgsl-exact-sum.js\";\n\nexport const DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE = 64;\n/** Largest footprint per axis (real pixels), as the Speed kernel. */\nexport const DISPLAY_VIEW_QUALITY_MAXIMUM_PIXELS_PER_AXIS = 6;\n/** Most spectral nodes the site passes integrate per site (2.21.0: the per-node partials are private arrays of this length). */\nexport const DISPLAY_VIEW_QUALITY_MAXIMUM_NODES = 96;\n/** Uniform parameter words (see `Params`): 56 words, 224 bytes. */\nexport const DISPLAY_VIEW_QUALITY_PARAM_WORDS = 56;\n/** Words per lattice-group record: lx0, ly0, lw, lh, lbase, first tap (the blocked site pass, #492), cu (f32), cv (f32). */\nexport const DISPLAY_VIEW_QUALITY_GROUP_WORDS = 8;\n/** Words per tap record: group, dp offset, dq offset, pad. */\nexport const DISPLAY_VIEW_QUALITY_TAP_WORDS = 4;\n/** Floats per native site of the site buffer: mean, variance, third cumulant, their three bounds, flag (u32 bits), electrons. */\nexport const DISPLAY_VIEW_QUALITY_SITE_FLOATS = 8;\n/** Floats per display pixel in the output buffer: m0 m1 m2, v00 v11 v22, c01 c02 c12, mean bound, variance bound (absolute). */\nexport const DISPLAY_VIEW_QUALITY_OUTPUT_FLOATS = 11;\n/** Floats per column or row record of the geometry buffer (the Speed layout). */\nexport const DISPLAY_VIEW_QUALITY_GEOMETRY_STRIDE = 20;\n/** Candidate table: entries per phase (the 7 × 7 window of AHD's reach 3) and floats per entry (dx, dy, dh0, dh1, dh2, dv0, dv1, dv2). */\nexport const DISPLAY_VIEW_QUALITY_CANDIDATE_ENTRIES = 49;\nexport const DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS = 8;\nconst CANDIDATE_PHASE_STRIDE = 4 + DISPLAY_VIEW_QUALITY_CANDIDATE_ENTRIES * DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS;\n/** Layout of the aux table (float32 words). */\nexport const DISPLAY_VIEW_QUALITY_AUX = /* @__PURE__ */ Object.freeze({\n  /** Per phase p, channel c, tap t: dx, dy, weight, 0 (the bilinear kernels, 4 taps; unused taps have weight 0). */\n  bilinear: 0,\n  /** Per phase p, channel c: 1 / Σ_t w_t² of the bilinear kernel (host float64). */\n  inverseW2: 192,\n  /** 16 Gauss–Legendre nodes then 16 weights on [−1, 1]. */\n  gaussLegendre: 208,\n  /** Per phase: the entry count (as u32 bits) and 3 pad words, then the entries (dx, dy, dense weights of candidate 0 and 1 per channel). */\n  candidates: 240,\n  candidatePhaseStride: CANDIDATE_PHASE_STRIDE,\n  size: 240 + 4 * CANDIDATE_PHASE_STRIDE\n});\n/** Pixel flag bits of the output flag word. */\nexport const DISPLAY_VIEW_QUALITY_FLAG = /* @__PURE__ */ Object.freeze({\n  /** A site the pixel read had an uncertain lattice material (recompute on the host). */\n  materialUncertain: 1,\n  /** A site the pixel read lies outside the band's site window (a host geometry error; recompute). */\n  siteOutsideWindow: 2,\n  /** The correction needed a grid state the table does not hold (recompute, or compute the state and dispatch again). */\n  stateMissing: 4,\n  /** A sample's pixel variance lies within its bound of the correction threshold (recompute). */\n  thresholdTie: 8,\n  /** Informational: the pixel's expected value includes the low-signal correction. */\n  corrected: 16,\n  /**\n   * #457: the reference takes AHD's choice on the noise-free mosaic (one candidate, or their average on a tie), which the kernel\n   * does not compute; it renders the average. The pixel is flagged (recompute) when Σ_j a_j·½|h(H) − h(V)| of some component\n   * exceeds `DISPLAY_VIEW_QUALITY_CHOICE_SIGMA` times the pixel's noise standard deviation.\n   */\n  ahdChoice: 32,\n  /** #533: a real pixel's saturation is undecided (the host computes the display pixel's count with the reference; its samples stay). */\n  highlightUndecided: 64,\n  /** The bits that require the host's recomputation. */\n  recompute: 47\n});\n/** #533: flag words per display pixel: the flags, then the 36-bit mask of saturated real pixels. */\nexport const DISPLAY_VIEW_QUALITY_FLAG_WORDS = 3;\n/**\n * #457: the largest move, in units of the display pixel's own noise σ, that AHD's choice between its candidates may make to a pixel\n * the kernel keeps (it renders the candidates' average and its bound carries the move); a larger one is recomputed on the host. A\n * design tolerance, not a measurement: 2^-10 σ is a tenth of the 0.01 σ to which the float32 tests hold the GPU to the reference.\n */\nexport const DISPLAY_VIEW_QUALITY_CHOICE_SIGMA = 2 ** -10;\n/** Words of the grid-key bitmap: one bit per grid key (AHD_KEY_LEVELS³ keys). */\nexport const DISPLAY_VIEW_QUALITY_BITMAP_WORDS = Math.ceil(AHD_KEY_LEVELS ** 3 / 32);\n/**\n * A site whose uncertain lattice points (the material module's flag) could move its electrons by more than this fraction of\n * their value is flagged for host recomputation; a smaller uncertainty enters the site's bound instead (the tap weights times\n * the largest value difference between any two materials at that tap).\n */\nexport const DISPLAY_VIEW_QUALITY_UNCERTAINTY_LIMIT = 2 ** -10;\n/** Bound constant of a clipped-normal slope Φ(z) (in units of E): z's 8E through φ·|z| ≤ 0.24 plus the Abramowitz–Stegun 147E. */\nexport const DISPLAY_VIEW_QUALITY_SLOPE_U = 200;\n/** Relative bound of the Reinhard slope E[h′] (in units of E). */\nconst REINHARD_SLOPE_U = 64;\n/** The `E` of the bounds: one WGSL rounding of + − × in either direction, 2^-23. */\nconst E_LITERAL = \"1.1920928955078125e-7\";\nconst literal = (value: number): string => {\n  const text = String(Math.fround(value));\n  return /[.e]/u.test(text) ? text : `${text}.0`;\n};\n\n/** The parameter struct and the bound helpers shared by the three passes. */\nexport const DISPLAY_VIEW_QUALITY_COMMON_WGSL = /* wgsl */ `\nconst E: f32 = ${E_LITERAL};\nconst SLACK: f32 = 1e-14;\nconst CLIP_REACH: f32 = 8.5;\nconst CLIP_MEAN_U: f32 = ${literal(DISPLAY_VIEW_CLIP_MEAN_U)};\nconst CLIP_VARIANCE_U: f32 = ${literal(DISPLAY_VIEW_CLIP_VARIANCE_U)};\nconst CLIP_SLOPE_U: f32 = ${literal(DISPLAY_VIEW_QUALITY_SLOPE_U)};\nconst REINHARD_MEAN_U: f32 = ${literal(DISPLAY_VIEW_REINHARD_MEAN_U)};\nconst REINHARD_VARIANCE_U: f32 = ${literal(DISPLAY_VIEW_REINHARD_VARIANCE_U)};\nconst REINHARD_SLOPE_U: f32 = ${literal(REINHARD_SLOPE_U)};\nconst UNCERTAINTY_LIMIT: f32 = ${literal(2 ** -10)};\nconst STRIDE: u32 = ${DISPLAY_VIEW_QUALITY_GEOMETRY_STRIDE}u;\nconst SITE: u32 = ${DISPLAY_VIEW_QUALITY_SITE_FLOATS}u;\nconst GROUP: u32 = ${DISPLAY_VIEW_QUALITY_GROUP_WORDS}u;\nconst TAP: u32 = ${DISPLAY_VIEW_QUALITY_TAP_WORDS}u;\nstruct Params {\n  count: u32, row0: u32, region_w: u32, region_x: u32,\n  region_y: u32, margin: u32, native_w: u32, native_h: u32,\n  columns: u32, tone: u32, reach: u32, candidates: u32,\n  site_x0: i32, site_y0: i32, site_w: u32, site_h: u32,\n  correction: u32, state_count: u32, taps: u32, groups: u32,\n  lattice_total: u32, gain_at_clip0: u32, gain_at_clip1: u32, gain_at_clip2: u32,\n  k0: f32, k1: f32, k2: f32, reach2: f32,\n  dark0: f32, dark1: f32, dark2: f32, clip_level: f32,\n  c0_0: f32, c0_1: f32, c0_2: f32, threshold: f32,\n  g0: f32, g1: f32, g2: f32, scale_u: f32,\n  m00: f32, m01: f32, m02: f32, m10: f32,\n  m11: f32, m12: f32, m20: f32, m21: f32,\n  m22: f32, scale_v: f32, site_gamma: f32, site_first: u32,\n  site_tiny: f32, nodes: u32, pad_g: f32, pad_h: f32,\n};\n@group(0) @binding(0) var<uniform> params: Params;\n\nfn k_of(c: u32) -> f32 { if (c == 0u) { return params.k0; } if (c == 1u) { return params.k1; } return params.k2; }\nfn dark_of(c: u32) -> f32 { if (c == 0u) { return params.dark0; } if (c == 1u) { return params.dark1; } return params.dark2; }\nfn c0_of(c: u32) -> f32 { if (c == 0u) { return params.c0_0; } if (c == 1u) { return params.c0_1; } return params.c0_2; }\nfn g_of(c: u32) -> f32 { if (c == 0u) { return params.g0; } if (c == 1u) { return params.g1; } return params.g2; }\n// Whether the channel's white-balance gain is the clip level itself (the smallest gain), exactly, in the reference's float64.\nfn gain_at_clip(c: u32) -> bool { if (c == 0u) { return params.gain_at_clip0 == 1u; } if (c == 1u) { return params.gain_at_clip1 == 1u; } return params.gain_at_clip2 == 1u; }\nfn m_of(r: u32, c: u32) -> f32 {\n  let i = r * 3u + c;\n  switch (i) {\n    case 0u: { return params.m00; } case 1u: { return params.m01; } case 2u: { return params.m02; }\n    case 3u: { return params.m10; } case 4u: { return params.m11; } case 5u: { return params.m12; }\n    case 6u: { return params.m20; } case 7u: { return params.m21; } default: { return params.m22; }\n  }\n}\nfn channel_of(x: i32, y: i32) -> u32 {\n  if ((y & 1) == 0) { return select(1u, 0u, (x & 1) == 0); }\n  return select(2u, 1u, (x & 1) == 0);\n}\n// A value with its absolute error bound (x: value, y: bound). First-order rules; the host doubles the final bound.\nfn b_add(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> { let v = a.x + b.x; return vec2<f32>(v, a.y + b.y + E * abs(v)); }\nfn b_sub(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> { let v = a.x - b.x; return vec2<f32>(v, a.y + b.y + E * abs(v)); }\nfn b_mul(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> { let v = a.x * b.x; return vec2<f32>(v, abs(a.x) * b.y + abs(b.x) * a.y + E * abs(v)); }\n// Times an exact constant (a kernel weight, an integer, a power of two).\nfn b_scale(a: vec2<f32>, k: f32) -> vec2<f32> { let v = a.x * k; return vec2<f32>(v, a.y * abs(k) + E * abs(v)); }\n// #533: 0 when a site of mean m (bound dm) and variance v (bound dv) is certainly below white, 1 certainly at or above it, 2\n// undecided (the Speed kernel's rule: twice the reference's 8.5 standard deviations, the mean moved by twice its bound).\nfn clip_state(m: f32, dm: f32, v: f32, dv: f32) -> u32 {\n  let gap = 1.0 - m;\n  let wide = 4.0 * params.reach2 * (v + 2.0 * dv);\n  let below = gap - 2.0 * dm;\n  let above = gap + 2.0 * dm;\n  if (below > 0.0 && below * below > wide) { return 0u; }\n  if (above < 0.0 && above * above > wide) { return 1u; }\n  return 2u;\n}\n// A host constant rounded once to float32 (a gain, a matrix entry).\nfn b_rounded(k: f32) -> vec2<f32> { return vec2<f32>(k, E * abs(k)); }\nfn b_div(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> { let v = a.x / b.x; return vec2<f32>(v, (a.y + abs(v) * b.y) / abs(b.x) + 2.5 * E * abs(v)); }\n// |sqrt(v + d) - sqrt(v)| <= min(d / (2 sqrt v), sqrt d), plus the 4.5 ULP of the built-in.\nfn b_sqrt(a: vec2<f32>) -> vec2<f32> {\n  let s = sqrt(max(a.x, 0.0));\n  if (s > 0.0) { return vec2<f32>(s, min(a.y / (2.0 * s), sqrt(a.y)) + 4.5 * E * s); }\n  return vec2<f32>(0.0, sqrt(a.y));\n}\n// Standard normal density, as the reference forms it, and Abramowitz & Stegun 26.2.17 as (Φ, 1 − Φ).\nfn phi(z: f32) -> f32 { return 0.3989422804014327 * exp(-0.5 * z * z); }\nfn pq(z: f32) -> vec2<f32> {\n  let x = abs(z);\n  let t = 1.0 / (1.0 + 0.2316419 * x);\n  let q = phi(x) * t * (0.319381530 + t * (-0.356563782 + t * (1.781477937 + t * (-1.821255978 + t * 1.330274429))));\n  if (z >= 0.0) { return vec2<f32>(1.0 - q, q); }\n  return vec2<f32>(q, 1.0 - q);\n}\nstruct Moments { mean: vec2<f32>, variance: vec2<f32>, slope: vec2<f32> }\n// Moments of min(Y, top) for Y ~ N(m, v) with their bounds: Var = v·(Φ − zφ + z²Q − δ²), δ = φ − zQ (the Speed forms).\nfn upper_clip(m: vec2<f32>, v: vec2<f32>, top: f32) -> Moments {\n  var out: Moments;\n  if (!(v.x > 0.0)) {\n    // No noise on either side (a zero bound): the reference's slope is 1 or 0 by m < top, decided when m's interval lies on one side.\n    let decided = v.y == 0.0 && (m.x - m.y >= top || m.x + m.y < top);\n    out.mean = vec2<f32>(min(m.x, top), m.y); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(select(0.0, 1.0, m.x < top), select(1.0, 0.0, decided));\n    return out;\n  }\n  let s = b_sqrt(v);\n  let z = (top - m.x) / s.x;\n  if (z > CLIP_REACH) { out.mean = m; out.variance = v; out.slope = vec2<f32>(1.0, 0.0); return out; }\n  if (z < -CLIP_REACH) { out.mean = vec2<f32>(top, 0.0); out.variance = vec2<f32>(0.0, 0.0); out.slope = vec2<f32>(0.0, 0.0); return out; }\n  let p = pq(z);\n  let d = phi(z);\n  let delta = d - z * p.y;\n  let ratio = p.x - z * d + z * z * p.y - delta * delta;\n  // First-order propagation of the inputs' bounds: with g = min(Z, z), E[g] = −δ and ρ = Var(g) = ratio, ∂E/∂m = Φ, ∂E/∂s = φ,\n  // ∂Var/∂m = −s·ρ_z, ∂Var/∂s = s·(2ρ − z·ρ_z) with ρ_z = 2Q·(z + δ), ∂Φ/∂m = −φ/s and ∂Φ/∂s = −zφ/s.\n  let rho_z = 2.0 * p.y * (z + delta);\n  out.mean = vec2<f32>(m.x - s.x * delta, p.x * m.y + d * s.y + CLIP_MEAN_U * E * s.x);\n  out.variance = vec2<f32>(v.x * max(0.0, ratio), s.x * abs(rho_z) * m.y + s.x * abs(2.0 * ratio - z * rho_z) * s.y + CLIP_VARIANCE_U * E * v.x);\n  out.slope = vec2<f32>(p.x, min(1.0, (d * m.y + abs(z) * d * s.y) / s.x + CLIP_SLOPE_U * E));\n  return out;\n}\n// Moments of clip(Y, 0, 1) for Y with mean m, variance v and standardized skewness γ (|γ| ≤ 1), with their bounds (the\n// Speed forms: W = m + s·g(Z), E[g] = aΦa + φa − φb + bQb, E[g²] = a²Φa + (Φb − Φa) − (bφb − aφa) + b²Qb, the Edgeworth\n// shift (s·γ/6)·(aφa − bφb) on the mean and −shift·(2·E[Y] + shift) on the variance).\nfn unit_clip(m: vec2<f32>, v: vec2<f32>, skew: vec2<f32>) -> Moments {\n  var out: Moments;\n  if (!(v.x > 0.0)) {\n    let decided = v.y == 0.0 && ((m.x - m.y > 0.0 && m.x + m.y < 1.0) || m.x + m.y <= 0.0 || m.x - m.y >= 1.0);\n    out.mean = vec2<f32>(clamp(m.x, 0.0, 1.0), m.y); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(select(0.0, 1.0, m.x > 0.0 && m.x < 1.0), select(1.0, 0.0, decided));\n    return out;\n  }\n  let s = b_sqrt(v);\n  let a = -m.x / s.x;\n  let b = (1.0 - m.x) / s.x;\n  if (a < -CLIP_REACH && b > CLIP_REACH) { out.mean = m; out.variance = v; out.slope = vec2<f32>(1.0, 0.0); return out; }\n  var pa = 0.0; var da = 0.0;\n  if (!(a < -CLIP_REACH)) { pa = pq(a).x; da = phi(a); }\n  var pb = 1.0; var qb = 0.0; var db = 0.0;\n  if (!(b > CLIP_REACH)) { let p = pq(b); pb = p.x; qb = p.y; db = phi(b); }\n  let inside = pb - pa;\n  let gamma = clamp(skew.x, -1.0, 1.0);\n  let aa = a * da;\n  let bb = b * db;\n  let eg = a * pa + (da - db) + b * qb;\n  let eg2 = a * a * pa + inside - (bb - aa) + b * b * qb;\n  let normal_mean = m.x + s.x * eg;\n  var mean = normal_mean;\n  var shift = 0.0;\n  if (gamma != 0.0) { shift = s.x * gamma / 6.0 * (aa - bb); mean = normal_mean + shift; }\n  // |aφa − bφb| ≤ 0.48, so |shift| ≤ 0.08·s·|γ| and a wrong γ moves it by at most 0.16·s.\n  let d_shift = 0.1 * s.x * min(2.0, skew.y) + 0.1 * abs(gamma) * s.y + E * abs(shift);\n  // First-order propagation of the inputs' bounds: with ρ = E[g²] − E[g]², ∂E[g]/∂a = Φa, ∂E[g]/∂b = Qb, ∂E[g²]/∂a = 2aΦa,\n  // ∂E[g²]/∂b = 2bQb, so ρ_a = 2Φa(a − E[g]), ρ_b = 2Qb(b − E[g]); ∂E/∂m = inside, ∂E/∂s = φa − φb, ∂Var/∂m = −s(ρ_a + ρ_b),\n  // ∂Var/∂s = s(2ρ − aρ_a − bρ_b); the slope's derivatives are (φa − φb)/s and (aφa − bφb)/s.\n  let rho = eg2 - eg * eg;\n  let rho_a = 2.0 * pa * (a - eg);\n  let rho_b = 2.0 * qb * (b - eg);\n  let d_mean = m.y * inside + abs(da - db) * s.y + 2.0 * CLIP_MEAN_U * E * s.x + d_shift;\n  // The normal part X = v·ρ and its bound; the reference's variance is max(0, X − shift·(2·E[Y] + shift)), so it lies in\n  // [0, X_cpu] when the shift term is non-negative on both sides (shift > Δshift), which caps the bound by max(Var, X + ΔX).\n  let normal_variance = v.x * rho;\n  let d_normal = s.x * abs(rho_a + rho_b) * m.y + s.x * abs(2.0 * rho - a * rho_a - b * rho_b) * s.y + 2.0 * CLIP_VARIANCE_U * E * v.x;\n  let variance = max(0.0, normal_variance - shift * (2.0 * normal_mean + shift));\n  var d_var = d_normal + 2.0 * (d_shift * (abs(normal_mean) + abs(shift)) + abs(shift) * d_mean);\n  if (shift > d_shift) { d_var = min(d_var, max(variance, max(0.0, normal_variance) + d_normal)); }\n  out.mean = vec2<f32>(mean, d_mean);\n  out.variance = vec2<f32>(variance, d_var);\n  out.slope = vec2<f32>(inside, min(1.0, (abs(da - db) * m.y + abs(aa - bb) * s.y) / s.x + 2.0 * CLIP_SLOPE_U * E));\n  return out;\n}\n`;\n\n/** The Gauss–Legendre Reinhard moments read the aux table, so they live in the pixel kernel. */\nconst REINHARD_WGSL = /* wgsl */ `\n// Positive Reinhard per channel then the clip to [0, 1]: h(x) = x/(1 + x) for x > 0, else 0, by the reference's composite\n// 16-point Gauss–Legendre rule in standardized units (the Speed forms), with the bounds of the quadrature (REINHARD_*_U)\n// and of its inputs (|h′| ≤ 1, |h″| ≤ 2).\nfn reinhard_moments(m: vec2<f32>, v: vec2<f32>) -> Moments {\n  var out: Moments;\n  if (!(v.x > 0.0)) {\n    if (m.x > 0.0) {\n      let h = m.x / (1.0 + m.x);\n      out.mean = vec2<f32>(h, m.y + 4.0 * E * h); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(1.0 / ((1.0 + m.x) * (1.0 + m.x)), min(1.0, 2.0 * m.y + 6.0 * E));\n      return out;\n    }\n    out.mean = vec2<f32>(0.0, m.y); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(0.0, 1.0);\n    return out;\n  }\n  let s = b_sqrt(v);\n  let zlo = max(-m.x / s.x, -CLIP_REACH);\n  if (!(CLIP_REACH > zlo)) { out.mean = vec2<f32>(0.0, m.y + s.y); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(0.0, 1.0); return out; }\n  let hm = select(0.0, m.x / (1.0 + m.x), m.x > 0.0);\n  var e1 = 0.0; var e2 = 0.0; var ed = 0.0;\n  let zwidth = (CLIP_REACH - zlo) / 4.0;\n  for (var p = 0u; p < 4u; p = p + 1u) {\n    let zc = zlo + (f32(p) + 0.5) * zwidth;\n    let zhalf = zwidth / 2.0;\n    for (var k = 0u; k < 16u; k = k + 1u) {\n      let z = zc + zhalf * aux[${DISPLAY_VIEW_QUALITY_AUX.gaussLegendre}u + k];\n      let x = m.x + s.x * z;\n      let weight = zhalf * aux[${DISPLAY_VIEW_QUALITY_AUX.gaussLegendre + 16}u + k] * phi(z);\n      let h = x / (1.0 + x);\n      e1 = e1 + weight * h;\n      e2 = e2 + weight * (h - hm) * (h - hm);\n      ed = ed + weight / ((1.0 + x) * (1.0 + x));\n    }\n  }\n  let shifted = e1 - hm;\n  let variance = max(0.0, e2 - shifted * shifted);\n  out.mean = vec2<f32>(e1, m.y + s.y + REINHARD_MEAN_U * E * abs(e1));\n  out.variance = vec2<f32>(variance, 2.0 * s.x * m.y + 2.0 * v.y + REINHARD_VARIANCE_U * E * variance);\n  out.slope = vec2<f32>(ed, min(1.0, 2.0 * m.y + 2.0 * s.y + REINHARD_SLOPE_U * E));\n  return out;\n}\n`;\n\n/**\n * Pass 1: the material of every lattice point of every group (the module's `material_at`, its uncertainty in bit 31; the output\n * is `lattice_out`, since a material module may declare its own `lattice`, as #491's composite module does). The group\n * records hold each group's lattice window (lx0, ly0, lw, lh), its first index in the lattice buffer and the point constants\n * c_u = 0.5 + f_x − W/2 and c_v = 0.5 + f_y − H/2 (exact in float32, checked by the host).\n */\nexport const DISPLAY_VIEW_QUALITY_LATTICE_WGSL = /* wgsl */ `${DISPLAY_VIEW_QUALITY_COMMON_WGSL}\n@group(0) @binding(1) var<storage, read> groups: array<u32>;\n@group(0) @binding(2) var<storage, read_write> lattice_out: array<u32>;\n@compute @workgroup_size(${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE})\nfn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) wg: vec3<u32>) {\n  let i = gid.x + gid.y * wg.x * ${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE}u;\n  if (i >= params.lattice_total) { return; }\n  var g = 0u;\n  for (var k = 1u; k < params.groups; k = k + 1u) { if (i >= groups[k * GROUP + 4u]) { g = k; } }\n  let base = g * GROUP;\n  let lx0 = bitcast<i32>(groups[base]);\n  let ly0 = bitcast<i32>(groups[base + 1u]);\n  let lw = groups[base + 2u];\n  let local = i - groups[base + 4u];\n  let q = ly0 + i32(local / lw);\n  let p = lx0 + i32(local % lw);\n  let u = (f32(p) + bitcast<f32>(groups[base + 6u])) * params.scale_u;\n  let v = (f32(q) + bitcast<f32>(groups[base + 7u])) * params.scale_v;\n  material_uncertain = false;\n  let m = material_at(vec2<f32>(u, v));\n  lattice_out[i] = m | select(0u, 0x80000000u, material_uncertain);\n}\n`;\n\n/**\n * The end of pass 2, shared by both site passes: from the site's electrons e and their absolute bound de, the readout moments\n * (mean k·e, variance k²e + c₀, third cumulant k³e), the clip at the white level and their bounds, written to site i.\n */\nexport const DISPLAY_VIEW_QUALITY_SITE_TAIL_WGSL = /* wgsl */ `\nfn write_site(i: u32, c: u32, e: f32, de: f32, flag: u32) {\n  let kc = k_of(c);\n  let mean = kc * e;\n  let variance = kc * kc * e + c0_of(c);\n  let third = kc * kc * kc * e;\n  var m = vec2<f32>(mean, kc * de + 2.0 * E * mean);\n  var v = vec2<f32>(variance, kc * kc * de + 4.0 * E * variance);\n  let gap = 1.0 - mean;\n  // #533: bits 1 (certainly clipped) and 2 (undecided) of the flag word.\n  let state = clip_state(m.x, m.y, v.x, v.y);\n  let bits = flag | select(0u, 2u, state == 1u) | select(0u, 4u, state == 2u);\n  if (!(gap > 0.0 && gap * gap > params.reach2 * variance)) {\n    let clipped = upper_clip(m, v, 1.0);\n    m = clipped.mean; v = clipped.variance;\n  }\n  let o = i * SITE;\n  sites[o] = m.x; sites[o + 1u] = v.x; sites[o + 2u] = third;\n  sites[o + 3u] = m.y; sites[o + 4u] = v.y; sites[o + 5u] = kc * kc * kc * de + 5.0 * E * third;\n  sites[o + 6u] = bitcast<f32>(bits); sites[o + 7u] = e;\n}\n`;\n\n/**\n * The bindings of the site passes (pass 2); the blocked pass reads no tap records, so it declares no binding 2 (an automatic\n * layout leaves out a binding the entry point does not use, and the host binds it only for the tap-list pass).\n */\nconst siteBindingsWgsl = (tapRecords: boolean): string => /* wgsl */ `\n@group(0) @binding(1) var<storage, read> groups: array<u32>;${tapRecords ? `\n@group(0) @binding(2) var<storage, read> taps: array<u32>;` : \"\"}\n@group(0) @binding(3) var<storage, read> table: array<f32>;\n@group(0) @binding(4) var<storage, read> lattice: array<u32>;\n@group(0) @binding(5) var<storage, read_write> sites: array<f32>;\n@group(0) @binding(6) var<storage, read> spread: array<f32>;\n${EXACT_SUM_WGSL}\nalias Nodes = array<f32, ${DISPLAY_VIEW_QUALITY_MAXIMUM_NODES}>;\nfn add_tap(p: ptr<function, Nodes>, c: u32, t: u32, m: u32) {\n  let N = params.nodes;\n  for (var n = 0u; n < N; n = n + 1u) { (*p)[n] = (*p)[n] + table[(c * params.taps + t) * N + n] * table[(3u * params.taps + m) * N + n]; }\n}\nfn add_nodes(p: ptr<function, Nodes>, q: ptr<function, Nodes>) { for (var n = 0u; n < params.nodes; n = n + 1u) { (*p)[n] = (*p)[n] + (*q)[n]; (*q)[n] = 0.0; } }\nfn node_total(p: ptr<function, Nodes>, dark: f32) -> f32 {\n  var anchor = exact_exponent(dark);\n  for (var k = 0u; k < params.nodes; k = k + 1u) { if ((*p)[k] > 0.0) { anchor = max(anchor, exact_exponent((*p)[k])); } }\n  var sum: ExactSum;\n  sum.anchor = anchor;\n  for (var k = 0u; k < params.nodes; k = k + 1u) { exact_add(&sum, (*p)[k]); }\n  exact_add(&sum, dark);\n  return exact_round(&sum);\n}\n`;\n\n/**\n * Pass 2: the expected electrons, readout moments and white clip of every native site of the band's window, with their bounds.\n * 2.21.0 (signals integrated from the material spectrum per site): the site's electrons are Σ_t Σ_n C_c(t, n)·L_{m_t}(λ_n) +\n * dark_c, from the channel's per-node kernel values and the spectrum of the material each tap reads (`table`: C at\n * (c·T + t)·N + n, the spectra from 3·T·N), accumulated per node over the taps, then the node partials and the dark electrons\n * summed in integer arithmetic and rounded once (EXACT_SUM_WGSL). Every term is non-negative and carries two host roundings (E/2\n * each), its product (E) and at most T − 1 additions, then the final rounding (E), and the dark electrons their host rounding\n * (E/2), so the sum carries at most (T + 2.5)·E, inside (T + 3)·E\n * (`params.site_gamma`),\n * plus `params.site_tiny`, the host's absolute bound on what a device that flushes below the float32 normal range can drop. An uncertain lattice point adds the tap's value spread\n * over the materials (`spread`, host float64) to the site's uncertainty, which enters its bounds, or flags the site when it\n * exceeds UNCERTAINTY_LIMIT of the electrons. One dispatch covers the `params.count` sites from `params.site_first` (#492: the\n * host splits a band's sites over several dispatches so that no dispatch gathers more than its tap budget).\n */\nexport const DISPLAY_VIEW_QUALITY_SITES_WGSL = /* wgsl */ `${DISPLAY_VIEW_QUALITY_COMMON_WGSL}${siteBindingsWgsl(true)}${DISPLAY_VIEW_QUALITY_SITE_TAIL_WGSL}\n@compute @workgroup_size(${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE})\nfn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) wg: vec3<u32>) {\n  let local = gid.x + gid.y * wg.x * ${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE}u;\n  if (local >= params.count) { return; }\n  let i = params.site_first + local;\n  let ii = i % params.site_w;\n  let jj = i / params.site_w;\n  let x = params.site_x0 + i32(ii);\n  let y = params.site_y0 + i32(jj);\n  let c = channel_of(x, y);\n  var uncertainty = 0.0;\n  let T = params.taps;\n  var partial: Nodes;\n  for (var t = 0u; t < T; t = t + 1u) {\n    let tb = t * TAP;\n    let g = taps[tb] * GROUP;\n    let index = groups[g + 4u] + (jj + taps[tb + 2u]) * groups[g + 2u] + ii + taps[tb + 1u];\n    let raw = lattice[index];\n    if ((raw >> 31u) != 0u) { uncertainty = uncertainty + spread[c * T + t]; }\n    add_tap(&partial, c, t, raw & 0x7fffffffu);\n  }\n  let e = node_total(&partial, dark_of(c));\n  let flag = select(0u, 1u, uncertainty > UNCERTAINTY_LIMIT * e);\n  let de = params.site_gamma * e + uncertainty + params.site_tiny;\n  write_site(i, c, e, de, flag);\n}\n`;\n\n/**\n * Pass 2 with the defocus disc (#492): the same site electrons, summed in blocks.\n *\n * - **Spectral table.** The pass reads the tap-list pass's table (2.21.0): the channel's per-node kernel values C_c(t, n) and each\n *   material's spectrum, so the signal of each tap is integrated from its material's spectrum when the site is read.\n * - **Blocked sum.** Every group's taps form a dense cw × ch rectangle (the combined kernel's window, row-major from the group's\n *   first tap, `groups[g + 5]`), so a site sums each row of a group into a row partial per node, the rows into a group partial\n *   and the groups into the total, then the nodes in integer arithmetic. No tap records are read: the tap's lattice offset\n *   within its group is its (dx, dy).\n *\n * Every term is non-negative and carries two host roundings (u each, u = E/2: the kernel value and the spectrum), one product\n * (E) and at most cw − 1 + ch − 1 + G − 1 inexact additions, then the integer sum of the node partials and the dark electrons,\n * rounded once (E), and the dark electrons' host rounding (E/2), so the electrons carry at most (cw_max + ch_max + G + 0.5)·E, inside (cw_max + ch_max + G + 3)·E (`params.site_gamma`;\n * recursive summation of the up to 4·(2·96 + 2)² = 150 544 taps of a combined radius of 96 would carry (T + 3)·E ≈ 1.8 %\n * instead). The uncertainty, summed the same way, is scaled by 1 + site_gamma, and `params.site_tiny` (host float64, rounded up)\n * bounds what a device that flushes results below the float32 normal range can drop (the tap-list pass adds it too).\n */\nexport const DISPLAY_VIEW_QUALITY_SITES_BLOCKED_WGSL = /* wgsl */ `${DISPLAY_VIEW_QUALITY_COMMON_WGSL}${siteBindingsWgsl(false)}\n${DISPLAY_VIEW_QUALITY_SITE_TAIL_WGSL}\n@compute @workgroup_size(${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE})\nfn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) wg: vec3<u32>) {\n  let local = gid.x + gid.y * wg.x * ${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE}u;\n  if (local >= params.count) { return; }\n  let i = params.site_first + local;\n  let ii = i % params.site_w;\n  let jj = i / params.site_w;\n  let x = params.site_x0 + i32(ii);\n  let y = params.site_y0 + i32(jj);\n  let c = channel_of(x, y);\n  let T = params.taps;\n  var total: Nodes;\n  var group_sum: Nodes;\n  var row_sum: Nodes;\n  var uncertainty = 0.0;\n  for (var k = 0u; k < params.groups; k = k + 1u) {\n    let g = k * GROUP;\n    let lw = groups[g + 2u];\n    let cw = lw - params.site_w + 1u;\n    let ch = groups[g + 3u] - params.site_h + 1u;\n    let lattice0 = groups[g + 4u] + jj * lw + ii;\n    let tap0 = groups[g + 5u];\n    var group_uncertainty = 0.0;\n    for (var dy = 0u; dy < ch; dy = dy + 1u) {\n      let row = lattice0 + dy * lw;\n      let row_tap = tap0 + dy * cw;\n      var row_uncertainty = 0.0;\n      for (var dx = 0u; dx < cw; dx = dx + 1u) {\n        let raw = lattice[row + dx];\n        let t = row_tap + dx;\n        if ((raw >> 31u) != 0u) { row_uncertainty = row_uncertainty + spread[c * T + t]; }\n        add_tap(&row_sum, c, t, raw & 0x7fffffffu);\n      }\n      add_nodes(&group_sum, &row_sum);\n      group_uncertainty = group_uncertainty + row_uncertainty;\n    }\n    add_nodes(&total, &group_sum);\n    uncertainty = uncertainty + group_uncertainty;\n  }\n  let e = node_total(&total, dark_of(c));\n  let flag = select(0u, 1u, uncertainty > UNCERTAINTY_LIMIT * e);\n  let de = params.site_gamma * e + uncertainty * (1.0 + params.site_gamma) + params.site_tiny;\n  write_site(i, c, e, de, flag);\n}\n`;\n\n/** Pass 3: the display pixels (see the module comment). */\nexport const DISPLAY_VIEW_QUALITY_PIXELS_WGSL = /* wgsl */ `${DISPLAY_VIEW_QUALITY_COMMON_WGSL}\nconst KEY_LEVELS: u32 = ${AHD_KEY_LEVELS}u;\nconst KEY_STEPS: f32 = ${literal(AHD_KEY_STEPS)};\nconst KEY_OCTAVES: f32 = ${literal(AHD_KEY_OCTAVES)};\nconst KEY_FLOOR: f32 = ${literal(AHD_KEY_FLOOR)};\nconst KEY_TOP: u32 = ${AHD_KEY_OCTAVES * AHD_KEY_STEPS - 1}u;\nconst CAND_BASE: u32 = ${DISPLAY_VIEW_QUALITY_AUX.candidates}u;\nconst CAND_STRIDE: u32 = ${CANDIDATE_PHASE_STRIDE}u;\nconst FLAG_UNCERTAIN: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.materialUncertain}u;\nconst FLAG_OUTSIDE: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.siteOutsideWindow}u;\nconst FLAG_STATE: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.stateMissing}u;\nconst FLAG_TIE: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.thresholdTie}u;\nconst FLAG_CORRECTED: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.corrected}u;\nconst FLAG_CHOICE: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.ahdChoice}u;\nconst FLAG_HIGHLIGHT: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.highlightUndecided}u;\nconst CHOICE_SIGMA: f32 = ${literal(DISPLAY_VIEW_QUALITY_CHOICE_SIGMA)};\n@group(0) @binding(1) var<storage, read> aux: array<f32>;\n@group(0) @binding(2) var<storage, read> geometry: array<f32>;\n@group(0) @binding(3) var<storage, read> sites: array<f32>;\n@group(0) @binding(4) var<storage, read> state_keys: array<i32>;\n@group(0) @binding(5) var<storage, read> state_values: array<f32>;\n@group(0) @binding(6) var<storage, read_write> bitmap: array<atomic<u32>>;\n@group(0) @binding(7) var<storage, read_write> out: array<f32>;\n@group(0) @binding(8) var<storage, read_write> flags: array<u32>;\n\nvar<private> pixel_flags: u32;\n${REINHARD_WGSL}\nfn reflect_x(x: i32) -> i32 { if (x < 0) { return -x; } if (x >= i32(params.native_w)) { return 2 * i32(params.native_w) - 2 - x; } return x; }\nfn reflect_y(y: i32) -> i32 { if (y < 0) { return -y; } if (y >= i32(params.native_h)) { return 2 * i32(params.native_h) - 2 - y; } return y; }\n// The site buffer offset of native site (x, y) (reflected at the raster's borders), or a flag when it lies outside the window.\nfn site_at(x: i32, y: i32) -> u32 {\n  let rx = reflect_x(x) - params.site_x0;\n  let ry = reflect_y(y) - params.site_y0;\n  if (rx < 0 || ry < 0 || rx >= i32(params.site_w) || ry >= i32(params.site_h)) { pixel_flags = pixel_flags | FLAG_OUTSIDE; return 0u; }\n  let o = (u32(ry) * params.site_w + u32(rx)) * SITE;\n  pixel_flags = pixel_flags | (bitcast<u32>(sites[o + 6u]) & 1u);\n  return o;\n}\n// #533: the clip state bits (1 certainly clipped, 2 undecided) of the site at offset o.\nfn clip_bits(o: u32) -> u32 { return (bitcast<u32>(sites[o + 6u]) >> 1u) & 3u; }\n// The offset of a grid key in the sorted state table, or state_count when absent.\nfn find_state(key: i32) -> u32 {\n  var lo = 0u; var hi = params.state_count;\n  while (lo < hi) {\n    let mid = (lo + hi) / 2u;\n    if (state_keys[mid] < key) { lo = mid + 1u; } else { hi = mid; }\n  }\n  if (lo < params.state_count && state_keys[lo] == key) { return lo; }\n  return params.state_count;\n}\n// The flat-neighbourhood correction at channel means (m0, m1, m2) with its bound: trilinear between the eight grid states\n// around (log2(clamp(m, 2^-12, 1)) + 12)·8; a missing state flags the pixel. The bound of each axis's coordinate times the\n// largest state difference along that axis bounds the interpolation's dependence on the coordinates' error.\nfn flat_correction(m0: vec2<f32>, m1: vec2<f32>, m2: vec2<f32>, corr: ptr<function, array<vec2<f32>, 3>>) {\n  var base: array<u32, 3>; var frac: array<f32, 3>; var dg: array<f32, 3>;\n  for (var i = 0u; i < 3u; i = i + 1u) {\n    var mi = m0; if (i == 1u) { mi = m1; } if (i == 2u) { mi = m2; }\n    let clamped = clamp(mi.x, KEY_FLOOR, 1.0);\n    let g = (log2(clamped) + KEY_OCTAVES) * KEY_STEPS;\n    let b = min(u32(max(floor(g), 0.0)), KEY_TOP);\n    base[i] = b; frac[i] = g - f32(b);\n    // log2: 1/ln 2 per relative input error, 2^-21 absolute or 3 ULP (≤ 36E on [2^-12, 1]), the sum and the product.\n    var slope = 0.0; if (mi.x > KEY_FLOOR && mi.x < 1.0) { slope = 1.4426950408889634 * mi.y / clamped; }\n    dg[i] = KEY_STEPS * (slope + 4.76837158203125e-7 + 36.0 * E) + 2.0 * E * (g + 1.0);\n  }\n  var values: array<f32, 24>;\n  var missing = false;\n  for (var corner = 0u; corner < 8u; corner = corner + 1u) {\n    let i0 = base[0] + (corner & 1u); let i1 = base[1] + ((corner >> 1u) & 1u); let i2 = base[2] + ((corner >> 2u) & 1u);\n    let key = i32((i0 * KEY_LEVELS + i1) * KEY_LEVELS + i2);\n    var weight = 1.0;\n    for (var i = 0u; i < 3u; i = i + 1u) { weight = weight * select(1.0 - frac[i], frac[i], ((corner >> i) & 1u) == 1u); }\n    if (weight == 0.0) { for (var c = 0u; c < 3u; c = c + 1u) { values[corner * 3u + c] = 0.0; } continue; }\n    atomicOr(&bitmap[u32(key) / 32u], 1u << (u32(key) % 32u));\n    let at = find_state(key);\n    if (at >= params.state_count) { missing = true; continue; }\n    for (var c = 0u; c < 3u; c = c + 1u) { values[corner * 3u + c] = state_values[at * 3u + c]; }\n  }\n  if (missing) { pixel_flags = pixel_flags | FLAG_STATE; for (var c = 0u; c < 3u; c = c + 1u) { (*corr)[c] = vec2<f32>(0.0, 0.0); } return; }\n  for (var c = 0u; c < 3u; c = c + 1u) {\n    var sum = 0.0; var mag = 0.0;\n    for (var corner = 0u; corner < 8u; corner = corner + 1u) {\n      var weight = 1.0;\n      for (var i = 0u; i < 3u; i = i + 1u) { weight = weight * select(1.0 - frac[i], frac[i], ((corner >> i) & 1u) == 1u); }\n      sum = sum + weight * values[corner * 3u + c];\n      mag = mag + abs(weight * values[corner * 3u + c]);\n    }\n    // The largest state difference along each axis (over the four edges), times that axis's coordinate bound.\n    var slope_term = 0.0;\n    for (var i = 0u; i < 3u; i = i + 1u) {\n      var largest = 0.0;\n      for (var corner = 0u; corner < 8u; corner = corner + 1u) {\n        if (((corner >> i) & 1u) == 1u) { continue; }\n        largest = max(largest, abs(values[(corner | (1u << i)) * 3u + c] - values[corner * 3u + c]));\n      }\n      slope_term = slope_term + dg[i] * largest;\n    }\n    (*corr)[c] = vec2<f32>(sum, slope_term + 12.0 * E * mag);\n  }\n}\n@compute @workgroup_size(${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE})\nfn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) wg: vec3<u32>) {\n  let i = gid.x + gid.y * wg.x * ${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE}u;\n  if (i >= params.count) { return; }\n  pixel_flags = 0u;\n  let rx = i % params.region_w;\n  let ry = params.row0 + i / params.region_w;\n  let col = rx * STRIDE;\n  let row = (params.columns + ry) * STRIDE;\n  let col_first = bitcast<i32>(geometry[col]);\n  let nx = bitcast<u32>(geometry[col + 1u]);\n  let row_first = bitcast<i32>(geometry[row]);\n  let ny = bitcast<u32>(geometry[row + 1u]);\n  let margin = i32(params.margin);\n  let reinhard = params.tone == 1u;\n  let count = params.candidates;\n  let share = select(1.0, 0.5, count == 2u);\n  // B_c of the footprint from the host's axis sums (each over its scale squared): red, green, blue.\n  let xb_even = geometry[col + 8u]; let xb_odd = geometry[col + 9u];\n  let xpp_even = geometry[col + 10u]; let xpp_odd = geometry[col + 11u];\n  let xqq_even = geometry[col + 12u]; let xqq_odd = geometry[col + 13u];\n  let xpq_even = geometry[col + 14u]; let xpq_odd = geometry[col + 15u];\n  let yb_even = geometry[row + 8u]; let yb_odd = geometry[row + 9u];\n  let ypp_even = geometry[row + 10u]; let ypp_odd = geometry[row + 11u];\n  let yqq_even = geometry[row + 12u]; let yqq_odd = geometry[row + 13u];\n  let ypq_even = geometry[row + 14u]; let ypq_odd = geometry[row + 15u];\n  let b0 = xb_even * yb_even;\n  let b1 = (xpp_odd * ypp_even + xqq_odd * ypp_even + xpp_odd * yqq_even + 2.0 * xpq_odd * ypp_even + 2.0 * xpp_odd * ypq_even + 2.0 * xpq_odd * ypq_even) +\n    (xpp_even * ypp_odd + xqq_even * ypp_odd + xpp_even * yqq_odd + 2.0 * xpq_even * ypp_odd + 2.0 * xpp_even * ypq_odd + 2.0 * xpq_even * ypq_odd);\n  let b2 = xb_odd * yb_odd;\n  var s = array<vec2<f32>, 3>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n  var cov = array<vec2<f32>, 6>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n  var res = array<vec2<f32>, 3>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n  // #457: Σ_j a_j·½|h(H) − h(V)| per component, how far the reference's choice can move the value from the candidates' average.\n  var choice = array<f32, 3>(0.0, 0.0, 0.0);\n  // #533: the saturated real pixels (bit j·nx + i) and whether a real pixel is undecided.\n  var mask0 = 0u; var mask1 = 0u; var undecided = false;\n  for (var j = 0u; j < ny; j = j + 1u) {\n    let ay = geometry[row + 2u + j];\n    for (var ii = 0u; ii < nx; ii = ii + 1u) {\n      // The overlap weight a_j: two host roundings (the overlaps over the scale) and the product.\n      let aj = vec2<f32>(geometry[col + 2u + ii] * ay, 3.0 * E * abs(geometry[col + 2u + ii] * ay));\n      let x = col_first + i32(ii) + margin;\n      let y = row_first + i32(j) + margin;\n      let phase = u32(y & 1) * 2u + u32(x & 1);\n      // Site variance of each channel through the bilinear kernels, over Σw² (for B_c).\n      var csv: array<vec2<f32>, 3>;\n      for (var c = 0u; c < 3u; c = c + 1u) {\n        var variance = vec2<f32>(0.0, 0.0);\n        let tb = ${DISPLAY_VIEW_QUALITY_AUX.bilinear}u + (phase * 3u + c) * 16u;\n        for (var t = 0u; t < 4u; t = t + 1u) {\n          let w = aux[tb + 4u * t + 2u];\n          if (w == 0.0) { continue; }\n          let o = site_at(x + i32(aux[tb + 4u * t]), y + i32(aux[tb + 4u * t + 1u]));\n          variance = b_add(variance, b_scale(vec2<f32>(sites[o + 1u], sites[o + 4u]), w * w));\n        }\n        csv[c] = b_mul(variance, b_rounded(aux[${DISPLAY_VIEW_QUALITY_AUX.inverseW2}u + phase * 3u + c]));\n      }\n      let cb = CAND_BASE + phase * CAND_STRIDE;\n      let entries = bitcast<u32>(aux[cb]);\n      // #533: the sites the demosaic reads (the bilinear kernels and every candidate's entries), as the reference's sample does.\n      var clip_any = 0u;\n      for (var c = 0u; c < 3u; c = c + 1u) {\n        let tb = ${DISPLAY_VIEW_QUALITY_AUX.bilinear}u + (phase * 3u + c) * 16u;\n        for (var t = 0u; t < 4u; t = t + 1u) {\n          if (aux[tb + 4u * t + 2u] == 0.0) { continue; }\n          clip_any = clip_any | clip_bits(site_at(x + i32(aux[tb + 4u * t]), y + i32(aux[tb + 4u * t + 1u])));\n        }\n      }\n      for (var n = 0u; n < entries; n = n + 1u) {\n        let eb = cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u;\n        var used = false;\n        for (var k = 0u; k < 3u * count; k = k + 1u) { if (aux[eb + 2u + k] != 0.0) { used = true; } }\n        if (used) { clip_any = clip_any | clip_bits(site_at(x + i32(aux[eb]), y + i32(aux[eb + 1u]))); }\n      }\n      if ((clip_any & 1u) != 0u) {\n        let bit = j * nx + ii;\n        if (bit < 32u) { mask0 = mask0 | (1u << bit); } else { mask1 = mask1 | (1u << (bit - 32u)); }\n      } else if ((clip_any & 2u) != 0u) { undecided = true; }\n      var e_mean = array<vec2<f32>, 3>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n      var cm_average = array<vec2<f32>, 3>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n      var pixel_variance = vec2<f32>(0.0, 0.0);\n      var bs: array<vec2<f32>, 3>; var hs: array<vec2<f32>, 3>;\n      var hv: array<vec2<f32>, 3>;\n      // #457: each candidate's rendered mean, for the reference's choice between them.\n      var h_cand = array<vec2<f32>, 6>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n      for (var d = 0u; d < count; d = d + 1u) {\n        // Camera-channel mean and covariance of candidate d from its dense weights over the window.\n        var cm: array<vec2<f32>, 3>;\n        var C: array<vec2<f32>, 6>;\n        // Whether every site a channel reads is exactly at white (a zero bound: the reference's site is exactly 1 too); the\n        // dyadic kernel weights then sum exactly on both sides, so the channel's mean carries no error (the host checks the weights).\n        var exact = array<bool, 3>(true, true, true);\n        for (var c = 0u; c < 3u; c = c + 1u) { cm[c] = vec2<f32>(0.0, 0.0); }\n        for (var k = 0u; k < 6u; k = k + 1u) { C[k] = vec2<f32>(0.0, 0.0); }\n        for (var n = 0u; n < entries; n = n + 1u) {\n          let eb = cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u + 2u + d * 3u;\n          let w0 = aux[eb]; let w1 = aux[eb + 1u]; let w2 = aux[eb + 2u];\n          if (w0 == 0.0 && w1 == 0.0 && w2 == 0.0) { continue; }\n          let o = site_at(x + i32(aux[cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u]), y + i32(aux[cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u + 1u]));\n          let sm = vec2<f32>(sites[o], sites[o + 3u]);\n          let sv = vec2<f32>(sites[o + 1u], sites[o + 4u]);\n          let white = sm.y == 0.0 && sm.x == 1.0;\n          exact[0] = exact[0] && (white || w0 == 0.0); exact[1] = exact[1] && (white || w1 == 0.0); exact[2] = exact[2] && (white || w2 == 0.0);\n          cm[0] = b_add(cm[0], b_scale(sm, w0)); cm[1] = b_add(cm[1], b_scale(sm, w1)); cm[2] = b_add(cm[2], b_scale(sm, w2));\n          C[0] = b_add(C[0], b_scale(sv, w0 * w0)); C[1] = b_add(C[1], b_scale(sv, w0 * w1)); C[2] = b_add(C[2], b_scale(sv, w0 * w2));\n          C[3] = b_add(C[3], b_scale(sv, w1 * w1)); C[4] = b_add(C[4], b_scale(sv, w1 * w2)); C[5] = b_add(C[5], b_scale(sv, w2 * w2));\n        }\n        for (var c = 0u; c < 3u; c = c + 1u) { if (exact[c]) { cm[c].y = 0.0; } cm_average[c] = b_add(cm_average[c], b_scale(cm[c], share)); }\n        // White balance and the highlight clip at the smallest gain. A channel exactly at white whose gain is the clip level is\n        // exactly at the clip level on both sides.\n        var bm: array<vec2<f32>, 3>; var bv: array<vec2<f32>, 3>;\n        for (var c = 0u; c < 3u; c = c + 1u) {\n          let gc = b_rounded(g_of(c));\n          var um = b_mul(gc, cm[c]);\n          if (exact[c] && cm[c].x == 1.0 && gain_at_clip(c)) { um = vec2<f32>(params.clip_level, 0.0); }\n          let uv = b_mul(b_mul(gc, gc), C[select(select(0u, 3u, c == 1u), 5u, c == 2u)]);\n          let gap = params.clip_level - um.x;\n          if (gap > 0.0 && gap * gap > params.reach2 * uv.x) { bm[c] = um; bv[c] = uv; bs[c] = gc; }\n          else { let clipped = upper_clip(um, uv, params.clip_level); bm[c] = clipped.mean; bv[c] = clipped.variance; bs[c] = b_mul(gc, clipped.slope); }\n        }\n        // Colour matrix with the full channel covariance, then the rendering of each component.\n        var hm: array<vec2<f32>, 3>;\n        for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n          let r0 = b_rounded(m_of(comp, 0u)); let r1 = b_rounded(m_of(comp, 1u)); let r2 = b_rounded(m_of(comp, 2u));\n          let mean = b_add(b_add(b_mul(r0, bm[0]), b_mul(r1, bm[1])), b_mul(r2, bm[2]));\n          var variance = b_add(b_add(b_mul(b_mul(r0, r0), bv[0]), b_mul(b_mul(r1, r1), bv[1])), b_mul(b_mul(r2, r2), bv[2]));\n          let x01 = b_mul(b_mul(b_mul(r0, r1), b_mul(bs[0], bs[1])), C[1]);\n          let x02 = b_mul(b_mul(b_mul(r0, r2), b_mul(bs[0], bs[2])), C[2]);\n          let x12 = b_mul(b_mul(b_mul(r1, r2), b_mul(bs[1], bs[2])), C[4]);\n          variance = b_add(variance, b_scale(b_add(b_add(x01, x02), x12), 2.0));\n          var moments: Moments;\n          if (!reinhard && mean.x > 0.0 && mean.x < 1.0 && mean.x * mean.x > params.reach2 * variance.x && (1.0 - mean.x) * (1.0 - mean.x) > params.reach2 * variance.x) {\n            moments.mean = mean; moments.variance = variance; moments.slope = vec2<f32>(1.0, 0.0);\n          } else if (reinhard) {\n            moments = reinhard_moments(mean, variance);\n          } else {\n            // κ₃ of the component: Σ_sites (Σ_c M_comp,c·Γ_c·w_c,site)³·κ₃(site), over the candidate's dense weights.\n            var third = vec2<f32>(0.0, 0.0);\n            let k0 = b_mul(r0, bs[0]); let k1 = b_mul(r1, bs[1]); let k2 = b_mul(r2, bs[2]);\n            for (var n = 0u; n < entries; n = n + 1u) {\n              let eb = cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u + 2u + d * 3u;\n              let w0 = aux[eb]; let w1 = aux[eb + 1u]; let w2 = aux[eb + 2u];\n              if (w0 == 0.0 && w1 == 0.0 && w2 == 0.0) { continue; }\n              let o = site_at(x + i32(aux[cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u]), y + i32(aux[cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u + 1u]));\n              let coefficient = b_add(b_add(b_scale(k0, w0), b_scale(k1, w1)), b_scale(k2, w2));\n              let cube = b_mul(b_mul(coefficient, coefficient), coefficient);\n              third = b_add(third, b_mul(cube, vec2<f32>(sites[o + 2u], sites[o + 5u])));\n            }\n            var skew = vec2<f32>(0.0, 0.0);\n            if (variance.x > 0.0) { skew = b_div(third, b_mul(variance, b_sqrt(variance))); }\n            moments = unit_clip(mean, variance, skew);\n          }\n          hm[comp] = moments.mean; hv[comp] = moments.variance; hs[comp] = moments.slope;\n          h_cand[d * 3u + comp] = moments.mean;\n          e_mean[comp] = b_add(e_mean[comp], b_scale(moments.mean, share));\n          pixel_variance = vec2<f32>(max(pixel_variance.x, moments.variance.x), max(pixel_variance.y, moments.variance.y));\n        }\n        // Linear (Stein) part with B_c·σ_c², and the pixel's own linear variance aᵀCa for the residual.\n        var an: array<vec2<f32>, 9>;\n        for (var c = 0u; c < 3u; c = c + 1u) {\n          an[c] = b_mul(b_mul(hs[0], b_rounded(m_of(0u, c))), bs[c]);\n          an[3u + c] = b_mul(b_mul(hs[1], b_rounded(m_of(1u, c))), bs[c]);\n          an[6u + c] = b_mul(b_mul(hs[2], b_rounded(m_of(2u, c))), bs[c]);\n          var bc = b0; if (c == 1u) { bc = b1; } if (c == 2u) { bc = b2; }\n          let wgt = b_mul(b_mul(b_scale(aj, share), b_rounded(bc)), csv[c]);\n          cov[0] = b_add(cov[0], b_mul(b_mul(wgt, an[c]), an[c])); cov[1] = b_add(cov[1], b_mul(b_mul(wgt, an[c]), an[3u + c])); cov[2] = b_add(cov[2], b_mul(b_mul(wgt, an[c]), an[6u + c]));\n          cov[3] = b_add(cov[3], b_mul(b_mul(wgt, an[3u + c]), an[3u + c])); cov[4] = b_add(cov[4], b_mul(b_mul(wgt, an[3u + c]), an[6u + c])); cov[5] = b_add(cov[5], b_mul(b_mul(wgt, an[6u + c]), an[6u + c]));\n        }\n        for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n          var l = vec2<f32>(0.0, 0.0);\n          for (var p = 0u; p < 3u; p = p + 1u) {\n            for (var q = 0u; q < 3u; q = q + 1u) {\n              var cij = C[0]; if (p + q == 1u) { cij = C[1]; } if ((p == 0u && q == 2u) || (p == 2u && q == 0u)) { cij = C[2]; } if (p == 1u && q == 1u) { cij = C[3]; }\n              if ((p == 1u && q == 2u) || (p == 2u && q == 1u)) { cij = C[4]; } if (p == 2u && q == 2u) { cij = C[5]; }\n              l = b_add(l, b_mul(b_mul(an[comp * 3u + p], an[comp * 3u + q]), cij));\n            }\n          }\n          let t = vec2<f32>(share * max(0.0, hv[comp].x - l.x), share * (hv[comp].y + l.y));\n          res[comp] = b_add(res[comp], b_mul(b_mul(aj, aj), t));\n        }\n      }\n      if (count == 2u) {\n        // #457: the reference takes AHD's choice on the noise-free mosaic, one candidate or (a tie) their average, so its value\n        // lies within half the candidates' difference plus their bounds of the average: the bound carries that, and the move\n        // accumulates in choice, which flags the pixel below when it exceeds CHOICE_SIGMA of the pixel's noise σ.\n        for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n          let diff = abs(h_cand[comp].x - h_cand[3u + comp].x);\n          let both = h_cand[comp].y + h_cand[3u + comp].y;\n          choice[comp] = choice[comp] + aj.x * 0.5 * diff;\n          e_mean[comp].y = e_mean[comp].y + 0.5 * (diff + both);\n        }\n      }\n      for (var comp = 0u; comp < 3u; comp = comp + 1u) { s[comp] = b_add(s[comp], b_mul(aj, e_mean[comp])); }\n      if (count == 2u) {\n        // AHD's per-pixel choice: ¼·Var(H − V) at the last candidate's slopes, linear in the sites.\n        var x0 = vec2<f32>(0.0, 0.0); var x1 = vec2<f32>(0.0, 0.0); var x2 = vec2<f32>(0.0, 0.0);\n        for (var n = 0u; n < entries; n = n + 1u) {\n          let eb = cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u;\n          let o = site_at(x + i32(aux[eb]), y + i32(aux[eb + 1u]));\n          let d0 = b_scale(bs[0], aux[eb + 2u] - aux[eb + 5u]);\n          let d1 = b_scale(bs[1], aux[eb + 3u] - aux[eb + 6u]);\n          let d2 = b_scale(bs[2], aux[eb + 4u] - aux[eb + 7u]);\n          let sv = vec2<f32>(sites[o + 1u], sites[o + 4u]);\n          let a = b_mul(hs[0], b_add(b_add(b_mul(b_rounded(params.m00), d0), b_mul(b_rounded(params.m01), d1)), b_mul(b_rounded(params.m02), d2)));\n          let b = b_mul(hs[1], b_add(b_add(b_mul(b_rounded(params.m10), d0), b_mul(b_rounded(params.m11), d1)), b_mul(b_rounded(params.m12), d2)));\n          let c = b_mul(hs[2], b_add(b_add(b_mul(b_rounded(params.m20), d0), b_mul(b_rounded(params.m21), d1)), b_mul(b_rounded(params.m22), d2)));\n          x0 = b_add(x0, b_mul(b_mul(a, a), sv)); x1 = b_add(x1, b_mul(b_mul(b, b), sv)); x2 = b_add(x2, b_mul(b_mul(c, c), sv));\n        }\n        let ajaj = b_mul(aj, aj);\n        res[0] = b_add(res[0], b_mul(ajaj, b_scale(x0, 0.25))); res[1] = b_add(res[1], b_mul(ajaj, b_scale(x1, 0.25))); res[2] = b_add(res[2], b_mul(ajaj, b_scale(x2, 0.25)));\n      }\n      if (params.correction == 1u) {\n        // The low-signal correction where the pixel noise exceeds the threshold; a decision within its bound is flagged.\n        if (abs(pixel_variance.x - params.threshold) <= pixel_variance.y) { pixel_flags = pixel_flags | FLAG_TIE; }\n        if (pixel_variance.x > params.threshold) {\n          var corr: array<vec2<f32>, 3>;\n          flat_correction(cm_average[0], cm_average[1], cm_average[2], &corr);\n          for (var comp = 0u; comp < 3u; comp = comp + 1u) { s[comp] = b_add(s[comp], b_mul(aj, corr[comp])); }\n          pixel_flags = pixel_flags | FLAG_CORRECTED;\n        }\n      }\n    }\n  }\n  // #457: a pixel whose candidates' choice could move it by more than CHOICE_SIGMA of its own noise σ is recomputed on the host.\n  if (choice[0] > CHOICE_SIGMA * sqrt(max(cov[0].x + res[0].x, 0.0)) || choice[1] > CHOICE_SIGMA * sqrt(max(cov[3].x + res[1].x, 0.0)) ||\n      choice[2] > CHOICE_SIGMA * sqrt(max(cov[5].x + res[2].x, 0.0))) { pixel_flags = pixel_flags | FLAG_CHOICE; }\n  let o = i * ${DISPLAY_VIEW_QUALITY_OUTPUT_FLOATS}u;\n  out[o] = s[0].x; out[o + 1u] = s[1].x; out[o + 2u] = s[2].x;\n  out[o + 3u] = cov[0].x + res[0].x; out[o + 4u] = cov[3].x + res[1].x; out[o + 5u] = cov[5].x + res[2].x;\n  out[o + 6u] = cov[1].x; out[o + 7u] = cov[2].x; out[o + 8u] = cov[4].x;\n  out[o + 9u] = max(s[0].y, max(s[1].y, s[2].y));\n  out[o + 10u] = max(max(cov[0].y + res[0].y, cov[3].y + res[1].y), max(cov[5].y + res[2].y, max(cov[1].y, max(cov[2].y, cov[4].y))));\n  flags[3u * i] = pixel_flags | select(0u, FLAG_HIGHLIGHT, undecided);\n  flags[3u * i + 1u] = mask0;\n  flags[3u * i + 2u] = mask1;\n}\n`;\n\n/** The lattice pass with a material module appended (`fn material_at(p: vec2<f32>) -> u32`, `var<private> material_uncertain: bool`). */\nexport function assembleDisplayViewQualityLatticeWgsl(materialModuleWgsl: string): string {\n  return DISPLAY_VIEW_QUALITY_LATTICE_WGSL + \"\\n\" + materialModuleWgsl + \"\\n\";\n}\n"]}
+{"version":3,"file":"wgsl-display-view-quality-kernel.js","sourceRoot":"","sources":["../../src/gpu/wgsl-display-view-quality-kernel.ts"],"names":[],"mappings":"AAAA,sCAAsC;AAEtC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+CG;AAEH,OAAO,EAAE,wBAAwB,EAAE,4BAA4B,EAAE,4BAA4B,EAAE,gCAAgC,EAAE,MAAM,qCAAqC,CAAC;AAC7K,OAAO,EAAE,cAAc,EAAE,MAAM,qBAAqB,CAAC;AACrD,OAAO,EAAE,8BAA8B,EAAE,MAAM,qCAAqC,CAAC;AAErF,MAAM,CAAC,MAAM,mCAAmC,GAAG,EAAE,CAAC;AACtD,qEAAqE;AACrE,MAAM,CAAC,MAAM,4CAA4C,GAAG,CAAC,CAAC;AAC9D,gIAAgI;AAChI,MAAM,CAAC,MAAM,kCAAkC,GAAG,EAAE,CAAC;AACrD,mEAAmE;AACnE,MAAM,CAAC,MAAM,gCAAgC,GAAG,EAAE,CAAC;AACnD,4HAA4H;AAC5H,MAAM,CAAC,MAAM,gCAAgC,GAAG,CAAC,CAAC;AAClD,8DAA8D;AAC9D,MAAM,CAAC,MAAM,8BAA8B,GAAG,CAAC,CAAC;AAChD,iIAAiI;AACjI,MAAM,CAAC,MAAM,gCAAgC,GAAG,CAAC,CAAC;AAClD,gIAAgI;AAChI,MAAM,CAAC,MAAM,kCAAkC,GAAG,EAAE,CAAC;AACrD,iFAAiF;AACjF,MAAM,CAAC,MAAM,oCAAoC,GAAG,EAAE,CAAC;AACvD,0IAA0I;AAC1I,MAAM,CAAC,MAAM,sCAAsC,GAAG,EAAE,CAAC;AACzD,MAAM,CAAC,MAAM,qCAAqC,GAAG,CAAC,CAAC;AACvD,MAAM,sBAAsB,GAAG,CAAC,GAAG,sCAAsC,GAAG,qCAAqC,CAAC;AAClH,+CAA+C;AAC/C,MAAM,CAAC,MAAM,wBAAwB,GAAG,eAAe,CAAC,MAAM,CAAC,MAAM,CAAC;IACpE,kHAAkH;IAClH,QAAQ,EAAE,CAAC;IACX,kFAAkF;IAClF,SAAS,EAAE,GAAG;IACd,0DAA0D;IAC1D,aAAa,EAAE,GAAG;IAClB,2IAA2I;IAC3I,UAAU,EAAE,GAAG;IACf,oBAAoB,EAAE,sBAAsB;IAC5C,IAAI,EAAE,GAAG,GAAG,CAAC,GAAG,sBAAsB;CACvC,CAAC,CAAC;AACH,+CAA+C;AAC/C,MAAM,CAAC,MAAM,yBAAyB,GAAG,eAAe,CAAC,MAAM,CAAC,MAAM,CAAC;IACrE,uFAAuF;IACvF,iBAAiB,EAAE,CAAC;IACpB,oGAAoG;IACpG,iBAAiB,EAAE,CAAC;IACpB,6GAA6G;IAC7G;;;;OAIG;IACH,SAAS,EAAE,EAAE;IACb,uIAAuI;IACvI,kBAAkB,EAAE,EAAE;IACtB,sDAAsD;IACtD,SAAS,EAAE,EAAE;CACd,CAAC,CAAC;AACH,oGAAoG;AACpG,MAAM,CAAC,MAAM,+BAA+B,GAAG,CAAC,CAAC;AACjD;;;;GAIG;AACH,MAAM,CAAC,MAAM,iCAAiC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;AAC1D;;;;GAIG;AACH,MAAM,CAAC,MAAM,sCAAsC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;AAC/D,kIAAkI;AAClI,MAAM,CAAC,MAAM,4BAA4B,GAAG,GAAG,CAAC;AAChD,kEAAkE;AAClE,MAAM,gBAAgB,GAAG,EAAE,CAAC;AAC5B,oFAAoF;AACpF,MAAM,SAAS,GAAG,uBAAuB,CAAC;AAC1C,MAAM,OAAO,GAAG,CAAC,KAAa,EAAU,EAAE;IACxC,MAAM,IAAI,GAAG,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;IACxC,OAAO,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,IAAI,CAAC;AACjD,CAAC,CAAC;AAEF,6EAA6E;AAC7E,MAAM,CAAC,MAAM,gCAAgC,GAAG,UAAU,CAAC;iBAC1C,SAAS;;;2BAGC,OAAO,CAAC,wBAAwB,CAAC;+BAC7B,OAAO,CAAC,4BAA4B,CAAC;4BACxC,OAAO,CAAC,4BAA4B,CAAC;+BAClC,OAAO,CAAC,4BAA4B,CAAC;mCACjC,OAAO,CAAC,gCAAgC,CAAC;gCAC5C,OAAO,CAAC,gBAAgB,CAAC;iCACxB,OAAO,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC;sBAC5B,oCAAoC;oBACtC,gCAAgC;qBAC/B,gCAAgC;mBAClC,8BAA8B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAmJhD,CAAC;AAEF,gGAAgG;AAChG,MAAM,aAAa,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;iCAyBA,wBAAwB,CAAC,aAAa;;iCAEtC,wBAAwB,CAAC,aAAa,GAAG,EAAE;;;;;;;;;;;;;;CAc3E,CAAC;AAEF;;;;;GAKG;AACH,MAAM,CAAC,MAAM,iCAAiC,GAAG,UAAU,CAAC,GAAG,gCAAgC;;;2BAGpE,mCAAmC;;mCAE3B,mCAAmC;;;;;;;;;;;;;;;;;CAiBrE,CAAC;AAEF;;;GAGG;AACH,MAAM,CAAC,MAAM,mCAAmC,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;CAqB7D,CAAC;AAEF;;;GAGG;AACH,MAAM,gBAAgB,GAAG,CAAC,UAAmB,EAAU,EAAE,CAAC,UAAU,CAAC;8DACP,UAAU,CAAC,CAAC,CAAC;2DAChB,CAAC,CAAC,CAAC,EAAE;;;;;EAK9D,cAAc;2BACW,kCAAkC;;;;;;;;;;;;;;;CAe5D,CAAC;AAEF;;;;;;;;;;;;;GAaG;AACH,MAAM,CAAC,MAAM,+BAA+B,GAAG,UAAU,CAAC,GAAG,gCAAgC,GAAG,gBAAgB,CAAC,IAAI,CAAC,GAAG,mCAAmC;2BACjI,mCAAmC;;uCAEvB,mCAAmC;;;;;;;;;;;;;;;;;;;;;;;;CAwBzE,CAAC;AAEF;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,CAAC,MAAM,uCAAuC,GAAG,UAAU,CAAC,GAAG,gCAAgC,GAAG,gBAAgB,CAAC,KAAK,CAAC;EAC7H,mCAAmC;2BACV,mCAAmC;;uCAEvB,mCAAmC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA0CzE,CAAC;AAEF,2DAA2D;AAC3D;;;;GAIG;AACH,MAAM,CAAC,MAAM,qCAAqC,GAAG,UAAU,CAAC,GAAG,gCAAgC;yBAC1E,wBAAwB,CAAC,UAAU;2BACjC,sBAAsB;8BACnB,yBAAyB,CAAC,iBAAiB;4BAC7C,yBAAyB,CAAC,iBAAiB;2BAC5C,yBAAyB,CAAC,SAAS;8BAChC,yBAAyB,CAAC,kBAAkB;4BAC9C,OAAO,CAAC,iCAAiC,CAAC;;;;;;;sCAOhC,wBAAwB,CAAC,IAAI;;;EAGjE,aAAa;;;;;;;;;;;;;;2BAcY,mCAAmC;;mCAE3B,mCAAmC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;mBAkDnD,wBAAwB,CAAC,QAAQ;;;;;;;iDAOH,wBAAwB,CAAC,SAAS;;;;;;;mBAOhE,wBAAwB,CAAC,QAAQ;;;;;;;iCAOnB,qCAAqC;;;;;;;;;;;;;;;;;;;;;;;;mCAwBnC,qCAAqC;;;sDAGlB,qCAAqC,kCAAkC,qCAAqC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;uCAmD3H,qCAAqC;;;0DAGlB,qCAAqC,kCAAkC,qCAAqC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;mCAiHnI,qCAAqC;;;;;;;;;;;;;;;;;;;;;;;;;;gBA0BxD,kCAAkC;;;;;;;;;;CAUjD,CAAC;AAEF,sJAAsJ;AACtJ,MAAM,UAAU,oCAAoC,CAAC,cAAsB,8BAA8B;IACvG,OAAO,qCAAqC,GAAG,IAAI,GAAG,WAAW,CAAC;AACpE,CAAC;AACD,MAAM,CAAC,MAAM,gCAAgC,GAAG,eAAe,CAAC,oCAAoC,EAAE,CAAC;AAEvG,yIAAyI;AACzI,MAAM,UAAU,qCAAqC,CAAC,kBAA0B;IAC9E,OAAO,iCAAiC,GAAG,IAAI,GAAG,kBAAkB,GAAG,IAAI,CAAC;AAC9E,CAAC","sourcesContent":["// SPDX-License-Identifier: Apache-2.0\n\n/**\n * WGSL kernels of the display view's Quality path on the GPU (#479): the exact area mean with diffraction kernels, the\n * anti-aliasing filter's spots and the AHD candidates, in three passes over one band of display\n * rows (docs/WEBGPU.md, \"Display view (Quality) on the GPU\"; the float64 reference is `renderDisplayView` with\n * `createDisplayViewExpectedSignal`, facade/display-view.ts and facade/expected-signal.ts):\n *\n * 1. **Lattice** (`DISPLAY_VIEW_QUALITY_LATTICE_WGSL` plus the material module): one invocation per lattice point of each\n *    lattice group (offsets that share a fractional shift, as the reference groups them): the chart material at\n *    (p + 0.5 + f_x − W/2)·g, −(q + 0.5 + f_y − H/2)·g (the reference's lattice formula, an exact float32 sum times the\n *    scale rounded once), with the module's uncertainty flag in the top bit.\n * 2. **Sites** (`DISPLAY_VIEW_QUALITY_SITES_WGSL`): one invocation per native site of the band's window: the expected\n *    electrons Σ_t V_c(m_t, t) over the taps t of every group (the combined kernel values per material and channel, formed on\n *    the host in float64 as the reference forms them: Σ_members w·Σ_λ A_c(λ)·K_λ·L_m(λ); without diffraction the offset\n *    weights times the material's channel signal), plus the dark electrons; the readout moments (mean k·e, variance k²e + c₀,\n *    third cumulant k³e) and the clip at the white level; and each quantity's running absolute error bound. With the defocus\n *    disc (#492) the variant `DISPLAY_VIEW_QUALITY_SITES_BLOCKED_WGSL` reads a factored table (one neutral row, one row per\n *    spectral material) and sums each group's dense tap rectangle in blocks, so its bound grows with the kernel's width.\n * 3. **Pixels** (`DISPLAY_VIEW_QUALITY_PIXELS_WGSL`): one invocation per display pixel of the band: for every real pixel of\n *    the footprint, the demosaic candidates (one bilinear, or AHD's two directional linear candidates, each with its full\n *    camera-channel covariance), white balance and the highlight clip, the colour matrix, the tone curve and the clip to\n *    [0, 1] with the skewness term, the candidates' average, the Stein covariance with the closed-form B_c, the independent\n *    residual, AHD's choice variance ¼·Var(H − V), the flag for AHD's choice on the noise-free mosaic, which the reference takes and\n *    the kernel does not (#457). 2.22.0: there is no low-signal correction (docs/DISPLAY_VIEW.md, \"AHD at low signal\").\n *\n * **Highlights (#533).** The site pass records whether each site is certainly below white, certainly at or above it, or\n * undecided (`clip_state`, from the site's mean and variance and their running bounds: certain only beyond twice the reference's\n * 8.5 standard deviations after moving the mean by twice its bound, where the reference's clip probability is exactly 0 or 1), in\n * bits 1 and 2 of its flag word. The pixel pass marks a real pixel saturated (P = 1) when a site its demosaic reads is certainly\n * clipped, and the display pixel `highlightUndecided` when a real pixel is neither that nor certainly unclipped. Three flag words\n * per display pixel: the flags, then a bit per real pixel j·nx + i of the footprint with P = 1 (36 bits).\n *\n * **Running bounds.** Every quantity after the site table is carried as a pair (value, absolute bound) in float32: the bound\n * of a sum is the sum of the bounds plus one rounding of the result, of a product |a|·Δb + |b|·Δa plus one rounding, of a\n * square root Δv/(2s) plus 4.5 ULP, of a quotient 2.5 ULP plus the propagated terms; the clipped-normal moments add the\n * Speed kernel's clip constants (CLIP_MEAN_U, CLIP_VARIANCE_U per standard deviation or variance) and propagate their inputs\n * with the moments' first-order derivatives at the pixel's own standardized clip levels (∂E/∂m = P(inside), ∂E/∂s = φa − φb,\n * ∂Var/∂m = −s(ρ_a + ρ_b), ∂Var/∂s = s(2ρ − aρ_a − bρ_b) with ρ_a = 2Φa(a − E[g]), ρ_b = 2Qb(b − E[g]), and the slope's\n * (φa − φb)/s and (aφa − bφb)/s; without noise a slope is exact unless the mean's interval straddles a clip level); the\n * Reinhard quadrature adds REINHARD_MEAN_U and REINHARD_VARIANCE_U. Mixed-sign sums (the AHD\n * kernels, the colour matrix) are therefore bounded by the magnitudes of their terms, not relative to their result. The\n * pixel's expected value and variance leave the kernel with their absolute bounds; the host doubles them (MARGIN) and adds\n * the footprint sum's relative term. Second-order terms (products of bounds) and the bound arithmetic's own float32\n * roundings are what the doubling covers.\n *\n * Precision assumptions are those of the Speed kernel: the W3C WGSL accuracy table (+ − × one rounding in either direction,\n * E = 2^-23; division 2.5 ULP, sqrt 4.5 ULP, exp (3 + 2|x|) ULP, log2 2^-21 absolute on [0.5, 2] and 3 ULP outside; no\n * trigonometric built-in).\n */\n\nimport { DISPLAY_VIEW_CLIP_MEAN_U, DISPLAY_VIEW_CLIP_VARIANCE_U, DISPLAY_VIEW_REINHARD_MEAN_U, DISPLAY_VIEW_REINHARD_VARIANCE_U } from \"./wgsl-display-view-speed-kernel.js\";\nimport { EXACT_SUM_WGSL } from \"./wgsl-exact-sum.js\";\nimport { DISPLAY_VIEW_COUPLED_STUB_WGSL } from \"./wgsl-display-view-coupled-stub.js\";\n\nexport const DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE = 64;\n/** Largest footprint per axis (real pixels), as the Speed kernel. */\nexport const DISPLAY_VIEW_QUALITY_MAXIMUM_PIXELS_PER_AXIS = 6;\n/** Most spectral nodes the site passes integrate per site (2.21.0: the per-node partials are private arrays of this length). */\nexport const DISPLAY_VIEW_QUALITY_MAXIMUM_NODES = 96;\n/** Uniform parameter words (see `Params`): 56 words, 224 bytes. */\nexport const DISPLAY_VIEW_QUALITY_PARAM_WORDS = 56;\n/** Words per lattice-group record: lx0, ly0, lw, lh, lbase, first tap (the blocked site pass, #492), cu (f32), cv (f32). */\nexport const DISPLAY_VIEW_QUALITY_GROUP_WORDS = 8;\n/** Words per tap record: group, dp offset, dq offset, pad. */\nexport const DISPLAY_VIEW_QUALITY_TAP_WORDS = 4;\n/** Floats per native site of the site buffer: mean, variance, third cumulant, their three bounds, flag (u32 bits), electrons. */\nexport const DISPLAY_VIEW_QUALITY_SITE_FLOATS = 8;\n/** Floats per display pixel in the output buffer: m0 m1 m2, v00 v11 v22, c01 c02 c12, mean bound, variance bound (absolute). */\nexport const DISPLAY_VIEW_QUALITY_OUTPUT_FLOATS = 11;\n/** Floats per column or row record of the geometry buffer (the Speed layout). */\nexport const DISPLAY_VIEW_QUALITY_GEOMETRY_STRIDE = 20;\n/** Candidate table: entries per phase (the 7 × 7 window of AHD's reach 3) and floats per entry (dx, dy, dh0, dh1, dh2, dv0, dv1, dv2). */\nexport const DISPLAY_VIEW_QUALITY_CANDIDATE_ENTRIES = 49;\nexport const DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS = 8;\nconst CANDIDATE_PHASE_STRIDE = 4 + DISPLAY_VIEW_QUALITY_CANDIDATE_ENTRIES * DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS;\n/** Layout of the aux table (float32 words). */\nexport const DISPLAY_VIEW_QUALITY_AUX = /* @__PURE__ */ Object.freeze({\n  /** Per phase p, channel c, tap t: dx, dy, weight, 0 (the bilinear kernels, 4 taps; unused taps have weight 0). */\n  bilinear: 0,\n  /** Per phase p, channel c: 1 / Σ_t w_t² of the bilinear kernel (host float64). */\n  inverseW2: 192,\n  /** 16 Gauss–Legendre nodes then 16 weights on [−1, 1]. */\n  gaussLegendre: 208,\n  /** Per phase: the entry count (as u32 bits) and 3 pad words, then the entries (dx, dy, dense weights of candidate 0 and 1 per channel). */\n  candidates: 240,\n  candidatePhaseStride: CANDIDATE_PHASE_STRIDE,\n  size: 240 + 4 * CANDIDATE_PHASE_STRIDE\n});\n/** Pixel flag bits of the output flag word. */\nexport const DISPLAY_VIEW_QUALITY_FLAG = /* @__PURE__ */ Object.freeze({\n  /** A site the pixel read had an uncertain lattice material (recompute on the host). */\n  materialUncertain: 1,\n  /** A site the pixel read lies outside the band's site window (a host geometry error; recompute). */\n  siteOutsideWindow: 2,\n  // Bits 4, 8 and 16 (the low-signal correction's state, threshold and applied flags) are unused since 2.22.0.\n  /**\n   * #457: the reference takes AHD's choice on the noise-free mosaic (one candidate, or their average on a tie), which the kernel\n   * does not compute; it renders the average. The pixel is flagged (recompute) when Σ_j a_j·½|h(H) − h(V)| of some component\n   * exceeds `DISPLAY_VIEW_QUALITY_CHOICE_SIGMA` times the pixel's noise standard deviation.\n   */\n  ahdChoice: 32,\n  /** #533: a real pixel's saturation is undecided (the host computes the display pixel's count with the reference; its samples stay). */\n  highlightUndecided: 64,\n  /** The bits that require the host's recomputation. */\n  recompute: 35\n});\n/** #533: flag words per display pixel: the flags, then the 36-bit mask of saturated real pixels. */\nexport const DISPLAY_VIEW_QUALITY_FLAG_WORDS = 3;\n/**\n * #457: the largest move, in units of the display pixel's own noise σ, that AHD's choice between its candidates may make to a pixel\n * the kernel keeps (it renders the candidates' average and its bound carries the move); a larger one is recomputed on the host. A\n * design tolerance, not a measurement: 2^-10 σ is a tenth of the 0.01 σ to which the float32 tests hold the GPU to the reference.\n */\nexport const DISPLAY_VIEW_QUALITY_CHOICE_SIGMA = 2 ** -10;\n/**\n * A site whose uncertain lattice points (the material module's flag) could move its electrons by more than this fraction of\n * their value is flagged for host recomputation; a smaller uncertainty enters the site's bound instead (the tap weights times\n * the largest value difference between any two materials at that tap).\n */\nexport const DISPLAY_VIEW_QUALITY_UNCERTAINTY_LIMIT = 2 ** -10;\n/** Bound constant of a clipped-normal slope Φ(z) (in units of E): z's 8E through φ·|z| ≤ 0.24 plus the Abramowitz–Stegun 147E. */\nexport const DISPLAY_VIEW_QUALITY_SLOPE_U = 200;\n/** Relative bound of the Reinhard slope E[h′] (in units of E). */\nconst REINHARD_SLOPE_U = 64;\n/** The `E` of the bounds: one WGSL rounding of + − × in either direction, 2^-23. */\nconst E_LITERAL = \"1.1920928955078125e-7\";\nconst literal = (value: number): string => {\n  const text = String(Math.fround(value));\n  return /[.e]/u.test(text) ? text : `${text}.0`;\n};\n\n/** The parameter struct and the bound helpers shared by the three passes. */\nexport const DISPLAY_VIEW_QUALITY_COMMON_WGSL = /* wgsl */ `\nconst E: f32 = ${E_LITERAL};\nconst SLACK: f32 = 1e-14;\nconst CLIP_REACH: f32 = 8.5;\nconst CLIP_MEAN_U: f32 = ${literal(DISPLAY_VIEW_CLIP_MEAN_U)};\nconst CLIP_VARIANCE_U: f32 = ${literal(DISPLAY_VIEW_CLIP_VARIANCE_U)};\nconst CLIP_SLOPE_U: f32 = ${literal(DISPLAY_VIEW_QUALITY_SLOPE_U)};\nconst REINHARD_MEAN_U: f32 = ${literal(DISPLAY_VIEW_REINHARD_MEAN_U)};\nconst REINHARD_VARIANCE_U: f32 = ${literal(DISPLAY_VIEW_REINHARD_VARIANCE_U)};\nconst REINHARD_SLOPE_U: f32 = ${literal(REINHARD_SLOPE_U)};\nconst UNCERTAINTY_LIMIT: f32 = ${literal(2 ** -10)};\nconst STRIDE: u32 = ${DISPLAY_VIEW_QUALITY_GEOMETRY_STRIDE}u;\nconst SITE: u32 = ${DISPLAY_VIEW_QUALITY_SITE_FLOATS}u;\nconst GROUP: u32 = ${DISPLAY_VIEW_QUALITY_GROUP_WORDS}u;\nconst TAP: u32 = ${DISPLAY_VIEW_QUALITY_TAP_WORDS}u;\nstruct Params {\n  count: u32, row0: u32, region_w: u32, region_x: u32,\n  region_y: u32, margin: u32, native_w: u32, native_h: u32,\n  columns: u32, tone: u32, reach: u32, candidates: u32,\n  site_x0: i32, site_y0: i32, site_w: u32, site_h: u32,\n  unused_16: u32, unused_17: u32, taps: u32, groups: u32,\n  lattice_total: u32, gain_at_clip0: u32, gain_at_clip1: u32, gain_at_clip2: u32,\n  k0: f32, k1: f32, k2: f32, reach2: f32,\n  dark0: f32, dark1: f32, dark2: f32, clip_level: f32,\n  c0_0: f32, c0_1: f32, c0_2: f32, unused_35: f32,\n  g0: f32, g1: f32, g2: f32, scale_u: f32,\n  m00: f32, m01: f32, m02: f32, m10: f32,\n  m11: f32, m12: f32, m20: f32, m21: f32,\n  m22: f32, scale_v: f32, site_gamma: f32, site_first: u32,\n  site_tiny: f32, nodes: u32, pad_g: f32, pad_h: f32,\n};\n@group(0) @binding(0) var<uniform> params: Params;\n\nfn k_of(c: u32) -> f32 { if (c == 0u) { return params.k0; } if (c == 1u) { return params.k1; } return params.k2; }\nfn dark_of(c: u32) -> f32 { if (c == 0u) { return params.dark0; } if (c == 1u) { return params.dark1; } return params.dark2; }\nfn c0_of(c: u32) -> f32 { if (c == 0u) { return params.c0_0; } if (c == 1u) { return params.c0_1; } return params.c0_2; }\nfn g_of(c: u32) -> f32 { if (c == 0u) { return params.g0; } if (c == 1u) { return params.g1; } return params.g2; }\n// Whether the channel's white-balance gain is the clip level itself (the smallest gain), exactly, in the reference's float64.\nfn gain_at_clip(c: u32) -> bool { if (c == 0u) { return params.gain_at_clip0 == 1u; } if (c == 1u) { return params.gain_at_clip1 == 1u; } return params.gain_at_clip2 == 1u; }\nfn m_of(r: u32, c: u32) -> f32 {\n  let i = r * 3u + c;\n  switch (i) {\n    case 0u: { return params.m00; } case 1u: { return params.m01; } case 2u: { return params.m02; }\n    case 3u: { return params.m10; } case 4u: { return params.m11; } case 5u: { return params.m12; }\n    case 6u: { return params.m20; } case 7u: { return params.m21; } default: { return params.m22; }\n  }\n}\nfn channel_of(x: i32, y: i32) -> u32 {\n  if ((y & 1) == 0) { return select(1u, 0u, (x & 1) == 0); }\n  return select(2u, 1u, (x & 1) == 0);\n}\n// A value with its absolute error bound (x: value, y: bound). First-order rules; the host doubles the final bound.\nfn b_add(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> { let v = a.x + b.x; return vec2<f32>(v, a.y + b.y + E * abs(v)); }\nfn b_sub(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> { let v = a.x - b.x; return vec2<f32>(v, a.y + b.y + E * abs(v)); }\nfn b_mul(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> { let v = a.x * b.x; return vec2<f32>(v, abs(a.x) * b.y + abs(b.x) * a.y + E * abs(v)); }\n// Times an exact constant (a kernel weight, an integer, a power of two).\nfn b_scale(a: vec2<f32>, k: f32) -> vec2<f32> { let v = a.x * k; return vec2<f32>(v, a.y * abs(k) + E * abs(v)); }\n// #533: 0 when a site of mean m (bound dm) and variance v (bound dv) is certainly below white, 1 certainly at or above it, 2\n// undecided (the Speed kernel's rule: twice the reference's 8.5 standard deviations, the mean moved by twice its bound).\nfn clip_state(m: f32, dm: f32, v: f32, dv: f32) -> u32 {\n  let gap = 1.0 - m;\n  let wide = 4.0 * params.reach2 * (v + 2.0 * dv);\n  let below = gap - 2.0 * dm;\n  let above = gap + 2.0 * dm;\n  if (below > 0.0 && below * below > wide) { return 0u; }\n  if (above < 0.0 && above * above > wide) { return 1u; }\n  return 2u;\n}\n// A host constant rounded once to float32 (a gain, a matrix entry).\nfn b_rounded(k: f32) -> vec2<f32> { return vec2<f32>(k, E * abs(k)); }\nfn b_div(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> { let v = a.x / b.x; return vec2<f32>(v, (a.y + abs(v) * b.y) / abs(b.x) + 2.5 * E * abs(v)); }\n// |sqrt(v + d) - sqrt(v)| <= min(d / (2 sqrt v), sqrt d), plus the 4.5 ULP of the built-in.\nfn b_sqrt(a: vec2<f32>) -> vec2<f32> {\n  let s = sqrt(max(a.x, 0.0));\n  if (s > 0.0) { return vec2<f32>(s, min(a.y / (2.0 * s), sqrt(a.y)) + 4.5 * E * s); }\n  return vec2<f32>(0.0, sqrt(a.y));\n}\n// Standard normal density, as the reference forms it, and Abramowitz & Stegun 26.2.17 as (Φ, 1 − Φ).\nfn phi(z: f32) -> f32 { return 0.3989422804014327 * exp(-0.5 * z * z); }\nfn pq(z: f32) -> vec2<f32> {\n  let x = abs(z);\n  let t = 1.0 / (1.0 + 0.2316419 * x);\n  let q = phi(x) * t * (0.319381530 + t * (-0.356563782 + t * (1.781477937 + t * (-1.821255978 + t * 1.330274429))));\n  if (z >= 0.0) { return vec2<f32>(1.0 - q, q); }\n  return vec2<f32>(q, 1.0 - q);\n}\nstruct Moments { mean: vec2<f32>, variance: vec2<f32>, slope: vec2<f32> }\n// Moments of min(Y, top) for Y ~ N(m, v) with their bounds: Var = v·(Φ − zφ + z²Q − δ²), δ = φ − zQ (the Speed forms).\nfn upper_clip(m: vec2<f32>, v: vec2<f32>, top: f32) -> Moments {\n  var out: Moments;\n  if (!(v.x > 0.0)) {\n    // No noise on either side (a zero bound): the reference's slope is 1 or 0 by m < top, decided when m's interval lies on one side.\n    let decided = v.y == 0.0 && (m.x - m.y >= top || m.x + m.y < top);\n    out.mean = vec2<f32>(min(m.x, top), m.y); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(select(0.0, 1.0, m.x < top), select(1.0, 0.0, decided));\n    return out;\n  }\n  let s = b_sqrt(v);\n  let z = (top - m.x) / s.x;\n  if (z > CLIP_REACH) { out.mean = m; out.variance = v; out.slope = vec2<f32>(1.0, 0.0); return out; }\n  if (z < -CLIP_REACH) { out.mean = vec2<f32>(top, 0.0); out.variance = vec2<f32>(0.0, 0.0); out.slope = vec2<f32>(0.0, 0.0); return out; }\n  let p = pq(z);\n  let d = phi(z);\n  let delta = d - z * p.y;\n  let ratio = p.x - z * d + z * z * p.y - delta * delta;\n  // First-order propagation of the inputs' bounds: with g = min(Z, z), E[g] = −δ and ρ = Var(g) = ratio, ∂E/∂m = Φ, ∂E/∂s = φ,\n  // ∂Var/∂m = −s·ρ_z, ∂Var/∂s = s·(2ρ − z·ρ_z) with ρ_z = 2Q·(z + δ), ∂Φ/∂m = −φ/s and ∂Φ/∂s = −zφ/s.\n  let rho_z = 2.0 * p.y * (z + delta);\n  out.mean = vec2<f32>(m.x - s.x * delta, p.x * m.y + d * s.y + CLIP_MEAN_U * E * s.x);\n  out.variance = vec2<f32>(v.x * max(0.0, ratio), s.x * abs(rho_z) * m.y + s.x * abs(2.0 * ratio - z * rho_z) * s.y + CLIP_VARIANCE_U * E * v.x);\n  out.slope = vec2<f32>(p.x, min(1.0, (d * m.y + abs(z) * d * s.y) / s.x + CLIP_SLOPE_U * E));\n  return out;\n}\n// Moments of clip(Y, 0, 1) for Y with mean m, variance v and standardized skewness γ (|γ| ≤ 1), with their bounds (the\n// Speed forms: W = m + s·g(Z), E[g] = aΦa + φa − φb + bQb, E[g²] = a²Φa + (Φb − Φa) − (bφb − aφa) + b²Qb, the Edgeworth\n// shift (s·γ/6)·(aφa − bφb) on the mean and −shift·(2·E[Y] + shift) on the variance).\nfn unit_clip(m: vec2<f32>, v: vec2<f32>, skew: vec2<f32>) -> Moments {\n  var out: Moments;\n  if (!(v.x > 0.0)) {\n    let decided = v.y == 0.0 && ((m.x - m.y > 0.0 && m.x + m.y < 1.0) || m.x + m.y <= 0.0 || m.x - m.y >= 1.0);\n    out.mean = vec2<f32>(clamp(m.x, 0.0, 1.0), m.y); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(select(0.0, 1.0, m.x > 0.0 && m.x < 1.0), select(1.0, 0.0, decided));\n    return out;\n  }\n  let s = b_sqrt(v);\n  let a = -m.x / s.x;\n  let b = (1.0 - m.x) / s.x;\n  if (a < -CLIP_REACH && b > CLIP_REACH) { out.mean = m; out.variance = v; out.slope = vec2<f32>(1.0, 0.0); return out; }\n  var pa = 0.0; var da = 0.0;\n  if (!(a < -CLIP_REACH)) { pa = pq(a).x; da = phi(a); }\n  var pb = 1.0; var qb = 0.0; var db = 0.0;\n  if (!(b > CLIP_REACH)) { let p = pq(b); pb = p.x; qb = p.y; db = phi(b); }\n  let inside = pb - pa;\n  let gamma = clamp(skew.x, -1.0, 1.0);\n  let aa = a * da;\n  let bb = b * db;\n  let eg = a * pa + (da - db) + b * qb;\n  let eg2 = a * a * pa + inside - (bb - aa) + b * b * qb;\n  let normal_mean = m.x + s.x * eg;\n  var mean = normal_mean;\n  var shift = 0.0;\n  if (gamma != 0.0) { shift = s.x * gamma / 6.0 * (aa - bb); mean = normal_mean + shift; }\n  // |aφa − bφb| ≤ 0.48, so |shift| ≤ 0.08·s·|γ| and a wrong γ moves it by at most 0.16·s.\n  let d_shift = 0.1 * s.x * min(2.0, skew.y) + 0.1 * abs(gamma) * s.y + E * abs(shift);\n  // First-order propagation of the inputs' bounds: with ρ = E[g²] − E[g]², ∂E[g]/∂a = Φa, ∂E[g]/∂b = Qb, ∂E[g²]/∂a = 2aΦa,\n  // ∂E[g²]/∂b = 2bQb, so ρ_a = 2Φa(a − E[g]), ρ_b = 2Qb(b − E[g]); ∂E/∂m = inside, ∂E/∂s = φa − φb, ∂Var/∂m = −s(ρ_a + ρ_b),\n  // ∂Var/∂s = s(2ρ − aρ_a − bρ_b); the slope's derivatives are (φa − φb)/s and (aφa − bφb)/s.\n  let rho = eg2 - eg * eg;\n  let rho_a = 2.0 * pa * (a - eg);\n  let rho_b = 2.0 * qb * (b - eg);\n  let d_mean = m.y * inside + abs(da - db) * s.y + 2.0 * CLIP_MEAN_U * E * s.x + d_shift;\n  // The normal part X = v·ρ and its bound; the reference's variance is max(0, X − shift·(2·E[Y] + shift)), so it lies in\n  // [0, X_cpu] when the shift term is non-negative on both sides (shift > Δshift), which caps the bound by max(Var, X + ΔX).\n  let normal_variance = v.x * rho;\n  let d_normal = s.x * abs(rho_a + rho_b) * m.y + s.x * abs(2.0 * rho - a * rho_a - b * rho_b) * s.y + 2.0 * CLIP_VARIANCE_U * E * v.x;\n  let variance = max(0.0, normal_variance - shift * (2.0 * normal_mean + shift));\n  var d_var = d_normal + 2.0 * (d_shift * (abs(normal_mean) + abs(shift)) + abs(shift) * d_mean);\n  if (shift > d_shift) { d_var = min(d_var, max(variance, max(0.0, normal_variance) + d_normal)); }\n  out.mean = vec2<f32>(mean, d_mean);\n  out.variance = vec2<f32>(variance, d_var);\n  out.slope = vec2<f32>(inside, min(1.0, (abs(da - db) * m.y + abs(aa - bb) * s.y) / s.x + 2.0 * CLIP_SLOPE_U * E));\n  return out;\n}\n`;\n\n/** The Gauss–Legendre Reinhard moments read the aux table, so they live in the pixel kernel. */\nconst REINHARD_WGSL = /* wgsl */ `\n// Positive Reinhard per channel then the clip to [0, 1]: h(x) = x/(1 + x) for x > 0, else 0, by the reference's composite\n// 16-point Gauss–Legendre rule in standardized units (the Speed forms), with the bounds of the quadrature (REINHARD_*_U)\n// and of its inputs (|h′| ≤ 1, |h″| ≤ 2).\nfn reinhard_moments(m: vec2<f32>, v: vec2<f32>) -> Moments {\n  var out: Moments;\n  if (!(v.x > 0.0)) {\n    if (m.x > 0.0) {\n      let h = m.x / (1.0 + m.x);\n      out.mean = vec2<f32>(h, m.y + 4.0 * E * h); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(1.0 / ((1.0 + m.x) * (1.0 + m.x)), min(1.0, 2.0 * m.y + 6.0 * E));\n      return out;\n    }\n    out.mean = vec2<f32>(0.0, m.y); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(0.0, 1.0);\n    return out;\n  }\n  let s = b_sqrt(v);\n  let zlo = max(-m.x / s.x, -CLIP_REACH);\n  if (!(CLIP_REACH > zlo)) { out.mean = vec2<f32>(0.0, m.y + s.y); out.variance = vec2<f32>(0.0, v.y); out.slope = vec2<f32>(0.0, 1.0); return out; }\n  let hm = select(0.0, m.x / (1.0 + m.x), m.x > 0.0);\n  var e1 = 0.0; var e2 = 0.0; var ed = 0.0;\n  let zwidth = (CLIP_REACH - zlo) / 4.0;\n  for (var p = 0u; p < 4u; p = p + 1u) {\n    let zc = zlo + (f32(p) + 0.5) * zwidth;\n    let zhalf = zwidth / 2.0;\n    for (var k = 0u; k < 16u; k = k + 1u) {\n      let z = zc + zhalf * aux[${DISPLAY_VIEW_QUALITY_AUX.gaussLegendre}u + k];\n      let x = m.x + s.x * z;\n      let weight = zhalf * aux[${DISPLAY_VIEW_QUALITY_AUX.gaussLegendre + 16}u + k] * phi(z);\n      let h = x / (1.0 + x);\n      e1 = e1 + weight * h;\n      e2 = e2 + weight * (h - hm) * (h - hm);\n      ed = ed + weight / ((1.0 + x) * (1.0 + x));\n    }\n  }\n  let shifted = e1 - hm;\n  let variance = max(0.0, e2 - shifted * shifted);\n  out.mean = vec2<f32>(e1, m.y + s.y + REINHARD_MEAN_U * E * abs(e1));\n  out.variance = vec2<f32>(variance, 2.0 * s.x * m.y + 2.0 * v.y + REINHARD_VARIANCE_U * E * variance);\n  out.slope = vec2<f32>(ed, min(1.0, 2.0 * m.y + 2.0 * s.y + REINHARD_SLOPE_U * E));\n  return out;\n}\n`;\n\n/**\n * Pass 1: the material of every lattice point of every group (the module's `material_at`, its uncertainty in bit 31; the output\n * is `lattice_out`, since a material module may declare its own `lattice`, as #491's composite module does). The group\n * records hold each group's lattice window (lx0, ly0, lw, lh), its first index in the lattice buffer and the point constants\n * c_u = 0.5 + f_x − W/2 and c_v = 0.5 + f_y − H/2 (exact in float32, checked by the host).\n */\nexport const DISPLAY_VIEW_QUALITY_LATTICE_WGSL = /* wgsl */ `${DISPLAY_VIEW_QUALITY_COMMON_WGSL}\n@group(0) @binding(1) var<storage, read> groups: array<u32>;\n@group(0) @binding(2) var<storage, read_write> lattice_out: array<u32>;\n@compute @workgroup_size(${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE})\nfn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) wg: vec3<u32>) {\n  let i = gid.x + gid.y * wg.x * ${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE}u;\n  if (i >= params.lattice_total) { return; }\n  var g = 0u;\n  for (var k = 1u; k < params.groups; k = k + 1u) { if (i >= groups[k * GROUP + 4u]) { g = k; } }\n  let base = g * GROUP;\n  let lx0 = bitcast<i32>(groups[base]);\n  let ly0 = bitcast<i32>(groups[base + 1u]);\n  let lw = groups[base + 2u];\n  let local = i - groups[base + 4u];\n  let q = ly0 + i32(local / lw);\n  let p = lx0 + i32(local % lw);\n  let u = (f32(p) + bitcast<f32>(groups[base + 6u])) * params.scale_u;\n  let v = (f32(q) + bitcast<f32>(groups[base + 7u])) * params.scale_v;\n  material_uncertain = false;\n  let m = material_at(vec2<f32>(u, v));\n  lattice_out[i] = m | select(0u, 0x80000000u, material_uncertain);\n}\n`;\n\n/**\n * The end of pass 2, shared by both site passes: from the site's electrons e and their absolute bound de, the readout moments\n * (mean k·e, variance k²e + c₀, third cumulant k³e), the clip at the white level and their bounds, written to site i.\n */\nexport const DISPLAY_VIEW_QUALITY_SITE_TAIL_WGSL = /* wgsl */ `\nfn write_site(i: u32, c: u32, e: f32, de: f32, flag: u32) {\n  let kc = k_of(c);\n  let mean = kc * e;\n  let variance = kc * kc * e + c0_of(c);\n  let third = kc * kc * kc * e;\n  var m = vec2<f32>(mean, kc * de + 2.0 * E * mean);\n  var v = vec2<f32>(variance, kc * kc * de + 4.0 * E * variance);\n  let gap = 1.0 - mean;\n  // #533: bits 1 (certainly clipped) and 2 (undecided) of the flag word.\n  let state = clip_state(m.x, m.y, v.x, v.y);\n  let bits = flag | select(0u, 2u, state == 1u) | select(0u, 4u, state == 2u);\n  if (!(gap > 0.0 && gap * gap > params.reach2 * variance)) {\n    let clipped = upper_clip(m, v, 1.0);\n    m = clipped.mean; v = clipped.variance;\n  }\n  let o = i * SITE;\n  sites[o] = m.x; sites[o + 1u] = v.x; sites[o + 2u] = third;\n  sites[o + 3u] = m.y; sites[o + 4u] = v.y; sites[o + 5u] = kc * kc * kc * de + 5.0 * E * third;\n  sites[o + 6u] = bitcast<f32>(bits); sites[o + 7u] = e;\n}\n`;\n\n/**\n * The bindings of the site passes (pass 2); the blocked pass reads no tap records, so it declares no binding 2 (an automatic\n * layout leaves out a binding the entry point does not use, and the host binds it only for the tap-list pass).\n */\nconst siteBindingsWgsl = (tapRecords: boolean): string => /* wgsl */ `\n@group(0) @binding(1) var<storage, read> groups: array<u32>;${tapRecords ? `\n@group(0) @binding(2) var<storage, read> taps: array<u32>;` : \"\"}\n@group(0) @binding(3) var<storage, read> table: array<f32>;\n@group(0) @binding(4) var<storage, read> lattice: array<u32>;\n@group(0) @binding(5) var<storage, read_write> sites: array<f32>;\n@group(0) @binding(6) var<storage, read> spread: array<f32>;\n${EXACT_SUM_WGSL}\nalias Nodes = array<f32, ${DISPLAY_VIEW_QUALITY_MAXIMUM_NODES}>;\nfn add_tap(p: ptr<function, Nodes>, c: u32, t: u32, m: u32) {\n  let N = params.nodes;\n  for (var n = 0u; n < N; n = n + 1u) { (*p)[n] = (*p)[n] + table[(c * params.taps + t) * N + n] * table[(3u * params.taps + m) * N + n]; }\n}\nfn add_nodes(p: ptr<function, Nodes>, q: ptr<function, Nodes>) { for (var n = 0u; n < params.nodes; n = n + 1u) { (*p)[n] = (*p)[n] + (*q)[n]; (*q)[n] = 0.0; } }\nfn node_total(p: ptr<function, Nodes>, dark: f32) -> f32 {\n  var anchor = exact_exponent(dark);\n  for (var k = 0u; k < params.nodes; k = k + 1u) { if ((*p)[k] > 0.0) { anchor = max(anchor, exact_exponent((*p)[k])); } }\n  var sum: ExactSum;\n  sum.anchor = anchor;\n  for (var k = 0u; k < params.nodes; k = k + 1u) { exact_add(&sum, (*p)[k]); }\n  exact_add(&sum, dark);\n  return exact_round(&sum);\n}\n`;\n\n/**\n * Pass 2: the expected electrons, readout moments and white clip of every native site of the band's window, with their bounds.\n * 2.21.0 (signals integrated from the material spectrum per site): the site's electrons are Σ_t Σ_n C_c(t, n)·L_{m_t}(λ_n) +\n * dark_c, from the channel's per-node kernel values and the spectrum of the material each tap reads (`table`: C at\n * (c·T + t)·N + n, the spectra from 3·T·N), accumulated per node over the taps, then the node partials and the dark electrons\n * summed in integer arithmetic and rounded once (EXACT_SUM_WGSL). Every term is non-negative and carries two host roundings (E/2\n * each), its product (E) and at most T − 1 additions, then the final rounding (E), and the dark electrons their host rounding\n * (E/2), so the sum carries at most (T + 2.5)·E, inside (T + 3)·E\n * (`params.site_gamma`),\n * plus `params.site_tiny`, the host's absolute bound on what a device that flushes below the float32 normal range can drop. An uncertain lattice point adds the tap's value spread\n * over the materials (`spread`, host float64) to the site's uncertainty, which enters its bounds, or flags the site when it\n * exceeds UNCERTAINTY_LIMIT of the electrons. One dispatch covers the `params.count` sites from `params.site_first` (#492: the\n * host splits a band's sites over several dispatches so that no dispatch gathers more than its tap budget).\n */\nexport const DISPLAY_VIEW_QUALITY_SITES_WGSL = /* wgsl */ `${DISPLAY_VIEW_QUALITY_COMMON_WGSL}${siteBindingsWgsl(true)}${DISPLAY_VIEW_QUALITY_SITE_TAIL_WGSL}\n@compute @workgroup_size(${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE})\nfn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) wg: vec3<u32>) {\n  let local = gid.x + gid.y * wg.x * ${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE}u;\n  if (local >= params.count) { return; }\n  let i = params.site_first + local;\n  let ii = i % params.site_w;\n  let jj = i / params.site_w;\n  let x = params.site_x0 + i32(ii);\n  let y = params.site_y0 + i32(jj);\n  let c = channel_of(x, y);\n  var uncertainty = 0.0;\n  let T = params.taps;\n  var partial: Nodes;\n  for (var t = 0u; t < T; t = t + 1u) {\n    let tb = t * TAP;\n    let g = taps[tb] * GROUP;\n    let index = groups[g + 4u] + (jj + taps[tb + 2u]) * groups[g + 2u] + ii + taps[tb + 1u];\n    let raw = lattice[index];\n    if ((raw >> 31u) != 0u) { uncertainty = uncertainty + spread[c * T + t]; }\n    add_tap(&partial, c, t, raw & 0x7fffffffu);\n  }\n  let e = node_total(&partial, dark_of(c));\n  let flag = select(0u, 1u, uncertainty > UNCERTAINTY_LIMIT * e);\n  let de = params.site_gamma * e + uncertainty + params.site_tiny;\n  write_site(i, c, e, de, flag);\n}\n`;\n\n/**\n * Pass 2 with the defocus disc (#492): the same site electrons, summed in blocks.\n *\n * - **Spectral table.** The pass reads the tap-list pass's table (2.21.0): the channel's per-node kernel values C_c(t, n) and each\n *   material's spectrum, so the signal of each tap is integrated from its material's spectrum when the site is read.\n * - **Blocked sum.** Every group's taps form a dense cw × ch rectangle (the combined kernel's window, row-major from the group's\n *   first tap, `groups[g + 5]`), so a site sums each row of a group into a row partial per node, the rows into a group partial\n *   and the groups into the total, then the nodes in integer arithmetic. No tap records are read: the tap's lattice offset\n *   within its group is its (dx, dy).\n *\n * Every term is non-negative and carries two host roundings (u each, u = E/2: the kernel value and the spectrum), one product\n * (E) and at most cw − 1 + ch − 1 + G − 1 inexact additions, then the integer sum of the node partials and the dark electrons,\n * rounded once (E), and the dark electrons' host rounding (E/2), so the electrons carry at most (cw_max + ch_max + G + 0.5)·E, inside (cw_max + ch_max + G + 3)·E (`params.site_gamma`;\n * recursive summation of the up to 4·(2·96 + 2)² = 150 544 taps of a combined radius of 96 would carry (T + 3)·E ≈ 1.8 %\n * instead). The uncertainty, summed the same way, is scaled by 1 + site_gamma, and `params.site_tiny` (host float64, rounded up)\n * bounds what a device that flushes results below the float32 normal range can drop (the tap-list pass adds it too).\n */\nexport const DISPLAY_VIEW_QUALITY_SITES_BLOCKED_WGSL = /* wgsl */ `${DISPLAY_VIEW_QUALITY_COMMON_WGSL}${siteBindingsWgsl(false)}\n${DISPLAY_VIEW_QUALITY_SITE_TAIL_WGSL}\n@compute @workgroup_size(${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE})\nfn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) wg: vec3<u32>) {\n  let local = gid.x + gid.y * wg.x * ${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE}u;\n  if (local >= params.count) { return; }\n  let i = params.site_first + local;\n  let ii = i % params.site_w;\n  let jj = i / params.site_w;\n  let x = params.site_x0 + i32(ii);\n  let y = params.site_y0 + i32(jj);\n  let c = channel_of(x, y);\n  let T = params.taps;\n  var total: Nodes;\n  var group_sum: Nodes;\n  var row_sum: Nodes;\n  var uncertainty = 0.0;\n  for (var k = 0u; k < params.groups; k = k + 1u) {\n    let g = k * GROUP;\n    let lw = groups[g + 2u];\n    let cw = lw - params.site_w + 1u;\n    let ch = groups[g + 3u] - params.site_h + 1u;\n    let lattice0 = groups[g + 4u] + jj * lw + ii;\n    let tap0 = groups[g + 5u];\n    var group_uncertainty = 0.0;\n    for (var dy = 0u; dy < ch; dy = dy + 1u) {\n      let row = lattice0 + dy * lw;\n      let row_tap = tap0 + dy * cw;\n      var row_uncertainty = 0.0;\n      for (var dx = 0u; dx < cw; dx = dx + 1u) {\n        let raw = lattice[row + dx];\n        let t = row_tap + dx;\n        if ((raw >> 31u) != 0u) { row_uncertainty = row_uncertainty + spread[c * T + t]; }\n        add_tap(&row_sum, c, t, raw & 0x7fffffffu);\n      }\n      add_nodes(&group_sum, &row_sum);\n      group_uncertainty = group_uncertainty + row_uncertainty;\n    }\n    add_nodes(&total, &group_sum);\n    uncertainty = uncertainty + group_uncertainty;\n  }\n  let e = node_total(&total, dark_of(c));\n  let flag = select(0u, 1u, uncertainty > UNCERTAINTY_LIMIT * e);\n  let de = params.site_gamma * e + uncertainty * (1.0 + params.site_gamma) + params.site_tiny;\n  write_site(i, c, e, de, flag);\n}\n`;\n\n/** Pass 3: the display pixels (see the module comment). */\n/*\n * 2.22.0: in the luminance branch the nested loops read the colour matrix rows from one local array (`mm`) instead of calling\n * `m_of` (a `switch`): SwiftShader's compiler ended the GPU process on the `switch` inside those loops (after about 44 s). The\n * values are the same (docs/WEBGPU.md, \"Coupled rendering steps\").\n */\nexport const DISPLAY_VIEW_QUALITY_PIXELS_BODY_WGSL = /* wgsl */ `${DISPLAY_VIEW_QUALITY_COMMON_WGSL}\nconst CAND_BASE: u32 = ${DISPLAY_VIEW_QUALITY_AUX.candidates}u;\nconst CAND_STRIDE: u32 = ${CANDIDATE_PHASE_STRIDE}u;\nconst FLAG_UNCERTAIN: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.materialUncertain}u;\nconst FLAG_OUTSIDE: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.siteOutsideWindow}u;\nconst FLAG_CHOICE: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.ahdChoice}u;\nconst FLAG_HIGHLIGHT: u32 = ${DISPLAY_VIEW_QUALITY_FLAG.highlightUndecided}u;\nconst CHOICE_SIGMA: f32 = ${literal(DISPLAY_VIEW_QUALITY_CHOICE_SIGMA)};\n@group(0) @binding(1) var<storage, read> aux: array<f32>;\n@group(0) @binding(2) var<storage, read> geometry: array<f32>;\n@group(0) @binding(3) var<storage, read> sites: array<f32>;\n@group(0) @binding(4) var<storage, read_write> out: array<f32>;\n@group(0) @binding(5) var<storage, read_write> flags: array<u32>;\n// #593, #601: the coupled rendering steps' constants, after the aux table's own entries (one zero word without them).\nfn cp_k(i: u32) -> f32 { return aux[${DISPLAY_VIEW_QUALITY_AUX.size}u + i]; }\n\nvar<private> pixel_flags: u32;\n${REINHARD_WGSL}\nfn reflect_x(x: i32) -> i32 { if (x < 0) { return -x; } if (x >= i32(params.native_w)) { return 2 * i32(params.native_w) - 2 - x; } return x; }\nfn reflect_y(y: i32) -> i32 { if (y < 0) { return -y; } if (y >= i32(params.native_h)) { return 2 * i32(params.native_h) - 2 - y; } return y; }\n// The site buffer offset of native site (x, y) (reflected at the raster's borders), or a flag when it lies outside the window.\nfn site_at(x: i32, y: i32) -> u32 {\n  let rx = reflect_x(x) - params.site_x0;\n  let ry = reflect_y(y) - params.site_y0;\n  if (rx < 0 || ry < 0 || rx >= i32(params.site_w) || ry >= i32(params.site_h)) { pixel_flags = pixel_flags | FLAG_OUTSIDE; return 0u; }\n  let o = (u32(ry) * params.site_w + u32(rx)) * SITE;\n  pixel_flags = pixel_flags | (bitcast<u32>(sites[o + 6u]) & 1u);\n  return o;\n}\n// #533: the clip state bits (1 certainly clipped, 2 undecided) of the site at offset o.\nfn clip_bits(o: u32) -> u32 { return (bitcast<u32>(sites[o + 6u]) >> 1u) & 3u; }\n@compute @workgroup_size(${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE})\nfn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) wg: vec3<u32>) {\n  let i = gid.x + gid.y * wg.x * ${DISPLAY_VIEW_QUALITY_WORKGROUP_SIZE}u;\n  if (i >= params.count) { return; }\n  pixel_flags = 0u;\n  let rx = i % params.region_w;\n  let ry = params.row0 + i / params.region_w;\n  let col = rx * STRIDE;\n  let row = (params.columns + ry) * STRIDE;\n  let col_first = bitcast<i32>(geometry[col]);\n  let nx = bitcast<u32>(geometry[col + 1u]);\n  let row_first = bitcast<i32>(geometry[row]);\n  let ny = bitcast<u32>(geometry[row + 1u]);\n  let margin = i32(params.margin);\n  let reinhard = params.tone == 1u;\n  let count = params.candidates;\n  // #593: whether the last candidate took the coupled slopes, and those slopes (for AHD's choice variance).\n  var cp_used = false;\n  var cp_a: array<vec2<f32>, 9>;\n  let share = select(1.0, 0.5, count == 2u);\n  // B_c of the footprint from the host's axis sums (each over its scale squared): red, green, blue.\n  let xb_even = geometry[col + 8u]; let xb_odd = geometry[col + 9u];\n  let xpp_even = geometry[col + 10u]; let xpp_odd = geometry[col + 11u];\n  let xqq_even = geometry[col + 12u]; let xqq_odd = geometry[col + 13u];\n  let xpq_even = geometry[col + 14u]; let xpq_odd = geometry[col + 15u];\n  let yb_even = geometry[row + 8u]; let yb_odd = geometry[row + 9u];\n  let ypp_even = geometry[row + 10u]; let ypp_odd = geometry[row + 11u];\n  let yqq_even = geometry[row + 12u]; let yqq_odd = geometry[row + 13u];\n  let ypq_even = geometry[row + 14u]; let ypq_odd = geometry[row + 15u];\n  let b0 = xb_even * yb_even;\n  let b1 = (xpp_odd * ypp_even + xqq_odd * ypp_even + xpp_odd * yqq_even + 2.0 * xpq_odd * ypp_even + 2.0 * xpp_odd * ypq_even + 2.0 * xpq_odd * ypq_even) +\n    (xpp_even * ypp_odd + xqq_even * ypp_odd + xpp_even * yqq_odd + 2.0 * xpq_even * ypp_odd + 2.0 * xpp_even * ypq_odd + 2.0 * xpq_even * ypq_odd);\n  let b2 = xb_odd * yb_odd;\n  var s = array<vec2<f32>, 3>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n  var cov = array<vec2<f32>, 6>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n  var res = array<vec2<f32>, 3>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n  // #457: Σ_j a_j·½|h(H) − h(V)| per component, how far the reference's choice can move the value from the candidates' average.\n  var choice = array<f32, 3>(0.0, 0.0, 0.0);\n  // #533: the saturated real pixels (bit j·nx + i) and whether a real pixel is undecided.\n  var mask0 = 0u; var mask1 = 0u; var undecided = false;\n  for (var j = 0u; j < ny; j = j + 1u) {\n    let ay = geometry[row + 2u + j];\n    for (var ii = 0u; ii < nx; ii = ii + 1u) {\n      // The overlap weight a_j: two host roundings (the overlaps over the scale) and the product.\n      let aj = vec2<f32>(geometry[col + 2u + ii] * ay, 3.0 * E * abs(geometry[col + 2u + ii] * ay));\n      let x = col_first + i32(ii) + margin;\n      let y = row_first + i32(j) + margin;\n      let phase = u32(y & 1) * 2u + u32(x & 1);\n      // Site variance of each channel through the bilinear kernels, over Σw² (for B_c).\n      var csv: array<vec2<f32>, 3>;\n      for (var c = 0u; c < 3u; c = c + 1u) {\n        var variance = vec2<f32>(0.0, 0.0);\n        let tb = ${DISPLAY_VIEW_QUALITY_AUX.bilinear}u + (phase * 3u + c) * 16u;\n        for (var t = 0u; t < 4u; t = t + 1u) {\n          let w = aux[tb + 4u * t + 2u];\n          if (w == 0.0) { continue; }\n          let o = site_at(x + i32(aux[tb + 4u * t]), y + i32(aux[tb + 4u * t + 1u]));\n          variance = b_add(variance, b_scale(vec2<f32>(sites[o + 1u], sites[o + 4u]), w * w));\n        }\n        csv[c] = b_mul(variance, b_rounded(aux[${DISPLAY_VIEW_QUALITY_AUX.inverseW2}u + phase * 3u + c]));\n      }\n      let cb = CAND_BASE + phase * CAND_STRIDE;\n      let entries = bitcast<u32>(aux[cb]);\n      // #533: the sites the demosaic reads (the bilinear kernels and every candidate's entries), as the reference's sample does.\n      var clip_any = 0u;\n      for (var c = 0u; c < 3u; c = c + 1u) {\n        let tb = ${DISPLAY_VIEW_QUALITY_AUX.bilinear}u + (phase * 3u + c) * 16u;\n        for (var t = 0u; t < 4u; t = t + 1u) {\n          if (aux[tb + 4u * t + 2u] == 0.0) { continue; }\n          clip_any = clip_any | clip_bits(site_at(x + i32(aux[tb + 4u * t]), y + i32(aux[tb + 4u * t + 1u])));\n        }\n      }\n      for (var n = 0u; n < entries; n = n + 1u) {\n        let eb = cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u;\n        var used = false;\n        for (var k = 0u; k < 3u * count; k = k + 1u) { if (aux[eb + 2u + k] != 0.0) { used = true; } }\n        if (used) { clip_any = clip_any | clip_bits(site_at(x + i32(aux[eb]), y + i32(aux[eb + 1u]))); }\n      }\n      if ((clip_any & 1u) != 0u) {\n        let bit = j * nx + ii;\n        if (bit < 32u) { mask0 = mask0 | (1u << bit); } else { mask1 = mask1 | (1u << (bit - 32u)); }\n      } else if ((clip_any & 2u) != 0u) { undecided = true; }\n      var e_mean = array<vec2<f32>, 3>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n      var bs: array<vec2<f32>, 3>; var hs: array<vec2<f32>, 3>;\n      var hv: array<vec2<f32>, 3>;\n      // #457: each candidate's rendered mean, for the reference's choice between them.\n      var h_cand = array<vec2<f32>, 6>(vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0), vec2<f32>(0.0));\n      for (var d = 0u; d < count; d = d + 1u) {\n        // Camera-channel mean and covariance of candidate d from its dense weights over the window.\n        var cm: array<vec2<f32>, 3>;\n        var C: array<vec2<f32>, 6>;\n        // Whether every site a channel reads is exactly at white (a zero bound: the reference's site is exactly 1 too); the\n        // dyadic kernel weights then sum exactly on both sides, so the channel's mean carries no error (the host checks the weights).\n        var exact = array<bool, 3>(true, true, true);\n        for (var c = 0u; c < 3u; c = c + 1u) { cm[c] = vec2<f32>(0.0, 0.0); }\n        for (var k = 0u; k < 6u; k = k + 1u) { C[k] = vec2<f32>(0.0, 0.0); }\n        for (var n = 0u; n < entries; n = n + 1u) {\n          let eb = cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u + 2u + d * 3u;\n          let w0 = aux[eb]; let w1 = aux[eb + 1u]; let w2 = aux[eb + 2u];\n          if (w0 == 0.0 && w1 == 0.0 && w2 == 0.0) { continue; }\n          let o = site_at(x + i32(aux[cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u]), y + i32(aux[cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u + 1u]));\n          let sm = vec2<f32>(sites[o], sites[o + 3u]);\n          let sv = vec2<f32>(sites[o + 1u], sites[o + 4u]);\n          let white = sm.y == 0.0 && sm.x == 1.0;\n          exact[0] = exact[0] && (white || w0 == 0.0); exact[1] = exact[1] && (white || w1 == 0.0); exact[2] = exact[2] && (white || w2 == 0.0);\n          cm[0] = b_add(cm[0], b_scale(sm, w0)); cm[1] = b_add(cm[1], b_scale(sm, w1)); cm[2] = b_add(cm[2], b_scale(sm, w2));\n          C[0] = b_add(C[0], b_scale(sv, w0 * w0)); C[1] = b_add(C[1], b_scale(sv, w0 * w1)); C[2] = b_add(C[2], b_scale(sv, w0 * w2));\n          C[3] = b_add(C[3], b_scale(sv, w1 * w1)); C[4] = b_add(C[4], b_scale(sv, w1 * w2)); C[5] = b_add(C[5], b_scale(sv, w2 * w2));\n        }\n        for (var c = 0u; c < 3u; c = c + 1u) { if (exact[c]) { cm[c].y = 0.0; } }\n        // #593: reconstruct within the blend's reach integrates the whole chain below; out of its reach it is the unclipped white\n        // balance. An undecided reach is recomputed on the host.\n        var cp_chain_run = false;\n        if (cp_on() && cp_reconstruct()) {\n          let reach = cp_reconstruct_reach(vec3<f32>(cm[0].x, cm[1].x, cm[2].x), vec3<f32>(C[0].x, C[3].x, C[5].x));\n          if (reach == 2u) { pixel_flags = pixel_flags | FLAG_UNCERTAIN; }\n          cp_chain_run = reach >= 1u;\n        }\n        let cp_unclipped = cp_on() && cp_reconstruct();\n        // White balance and the highlight clip at the smallest gain. A channel exactly at white whose gain is the clip level is\n        // exactly at the clip level on both sides.\n        var bm: array<vec2<f32>, 3>; var bv: array<vec2<f32>, 3>;\n        for (var c = 0u; c < 3u; c = c + 1u) {\n          let gc = b_rounded(g_of(c));\n          var um = b_mul(gc, cm[c]);\n          if (exact[c] && cm[c].x == 1.0 && gain_at_clip(c) && !cp_unclipped) { um = vec2<f32>(params.clip_level, 0.0); }\n          let uv = b_mul(b_mul(gc, gc), C[select(select(0u, 3u, c == 1u), 5u, c == 2u)]);\n          let gap = params.clip_level - um.x;\n          if (cp_unclipped || (gap > 0.0 && gap * gap > params.reach2 * uv.x)) { bm[c] = um; bv[c] = uv; bs[c] = gc; }\n          else { let clipped = upper_clip(um, uv, params.clip_level); bm[c] = clipped.mean; bv[c] = clipped.variance; bs[c] = b_mul(gc, clipped.slope); }\n        }\n        // Colour matrix with the full channel covariance, then the rendering of each component.\n        var hm: array<vec2<f32>, 3>;\n        for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n          let r0 = b_rounded(m_of(comp, 0u)); let r1 = b_rounded(m_of(comp, 1u)); let r2 = b_rounded(m_of(comp, 2u));\n          let mean = b_add(b_add(b_mul(r0, bm[0]), b_mul(r1, bm[1])), b_mul(r2, bm[2]));\n          var variance = b_add(b_add(b_mul(b_mul(r0, r0), bv[0]), b_mul(b_mul(r1, r1), bv[1])), b_mul(b_mul(r2, r2), bv[2]));\n          let x01 = b_mul(b_mul(b_mul(r0, r1), b_mul(bs[0], bs[1])), C[1]);\n          let x02 = b_mul(b_mul(b_mul(r0, r2), b_mul(bs[0], bs[2])), C[2]);\n          let x12 = b_mul(b_mul(b_mul(r1, r2), b_mul(bs[1], bs[2])), C[4]);\n          variance = b_add(variance, b_scale(b_add(b_add(x01, x02), x12), 2.0));\n          var moments: Moments;\n          if (!reinhard && mean.x > 0.0 && mean.x < 1.0 && mean.x * mean.x > params.reach2 * variance.x && (1.0 - mean.x) * (1.0 - mean.x) > params.reach2 * variance.x) {\n            moments.mean = mean; moments.variance = variance; moments.slope = vec2<f32>(1.0, 0.0);\n          } else if (reinhard) {\n            moments = reinhard_moments(mean, variance);\n          } else {\n            // κ₃ of the component: Σ_sites (Σ_c M_comp,c·Γ_c·w_c,site)³·κ₃(site), over the candidate's dense weights.\n            var third = vec2<f32>(0.0, 0.0);\n            let k0 = b_mul(r0, bs[0]); let k1 = b_mul(r1, bs[1]); let k2 = b_mul(r2, bs[2]);\n            for (var n = 0u; n < entries; n = n + 1u) {\n              let eb = cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u + 2u + d * 3u;\n              let w0 = aux[eb]; let w1 = aux[eb + 1u]; let w2 = aux[eb + 2u];\n              if (w0 == 0.0 && w1 == 0.0 && w2 == 0.0) { continue; }\n              let o = site_at(x + i32(aux[cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u]), y + i32(aux[cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u + 1u]));\n              let coefficient = b_add(b_add(b_scale(k0, w0), b_scale(k1, w1)), b_scale(k2, w2));\n              let cube = b_mul(b_mul(coefficient, coefficient), coefficient);\n              third = b_add(third, b_mul(cube, vec2<f32>(sites[o + 2u], sites[o + 5u])));\n            }\n            var skew = vec2<f32>(0.0, 0.0);\n            if (variance.x > 0.0) { skew = b_div(third, b_mul(variance, b_sqrt(variance))); }\n            moments = unit_clip(mean, variance, skew);\n          }\n          hm[comp] = moments.mean; hv[comp] = moments.variance; hs[comp] = moments.slope;\n        }\n        // #593, #601: the coupled steps (wgsl-display-view-coupled.ts) replace the moments and give the slopes ∂out/∂c.\n        var an: array<vec2<f32>, 9>;\n        cp_used = false;\n        if (cp_on()) {\n          if (!cp_chain_run && cp_luminance()) {\n            var mu = vec3<f32>(0.0); var dmu = vec3<f32>(0.0); var dsv = vec3<f32>(0.0);\n            var srow = array<vec3<f32>, 3>(vec3<f32>(0.0), vec3<f32>(0.0), vec3<f32>(0.0));\n            var cc = array<f32, 9>(C[0].x, C[1].x, C[2].x, C[1].x, C[3].x, C[4].x, C[2].x, C[4].x, C[5].x);\n            var mm = array<vec3<f32>, 3>(vec3<f32>(params.m00, params.m01, params.m02), vec3<f32>(params.m10, params.m11, params.m12), vec3<f32>(params.m20, params.m21, params.m22));\n            for (var r = 0u; r < 3u; r = r + 1u) {\n              mu[r] = mm[r][0] * bm[0].x + mm[r][1] * bm[1].x + mm[r][2] * bm[2].x;\n              dmu[r] = abs(mm[r][0]) * bm[0].y + abs(mm[r][1]) * bm[1].y + abs(mm[r][2]) * bm[2].y;\n              dsv[r] = mm[r][0] * mm[r][0] * bv[0].y + mm[r][1] * mm[r][1] * bv[1].y + mm[r][2] * mm[r][2] * bv[2].y;\n              for (var q = 0u; q < 3u; q = q + 1u) {\n                var v = 0.0;\n                for (var a = 0u; a < 3u; a = a + 1u) { for (var b = 0u; b < 3u; b = b + 1u) {\n                  var bab = bv[a].x; if (a != b) { bab = bs[a].x * bs[b].x * cc[a * 3u + b]; }\n                  v = v + mm[r][a] * mm[q][b] * bab;\n                } }\n                srow[r][q] = v;\n              }\n            }\n            let t = cp_tone(mu, srow[0], srow[1], srow[2], dmu, dsv);\n            if (t.flag) { pixel_flags = pixel_flags | FLAG_UNCERTAIN; }\n            if (cp_desaturate() && t.top) { cp_chain_run = true; }\n            else {\n              let rows = array<vec3<f32>, 3>(t.a0, t.a1, t.a2);\n              for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n                hm[comp] = vec2<f32>(t.hm[comp], t.bm[comp]); hv[comp] = vec2<f32>(t.hv[comp], t.bv[comp]);\n                for (var c = 0u; c < 3u; c = c + 1u) {\n                  let jmc = rows[comp].x * m_of(0u, c) + rows[comp].y * m_of(1u, c) + rows[comp].z * m_of(2u, c);\n                  let ajm = abs(rows[comp].x * m_of(0u, c)) + abs(rows[comp].y * m_of(1u, c)) + abs(rows[comp].z * m_of(2u, c));\n                  let value = jmc * bs[c].x;\n                  an[comp * 3u + c] = vec2<f32>(value, ajm * bs[c].y + abs(value) * (t.ju + 4.0 * E));\n                }\n              }\n              cp_used = true;\n            }\n          } else if (!cp_chain_run && cp_desaturate() && !reinhard) {\n            for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n              let top = hm[comp].x + CLIP_REACH * sqrt(hv[comp].x);\n              if (hm[comp].x > 0.0 && top >= 0.999) { cp_chain_run = true; }\n              if (hm[comp].x > 0.0 && hm[comp].x - 2.0 * hm[comp].y < 0.999 && abs(top - 0.999) <= 2.0 * (hm[comp].y + CLIP_REACH * min(sqrt(hv[comp].y), hv[comp].y / (2.0 * sqrt(hv[comp].x)))) + 64.0 * E) { pixel_flags = pixel_flags | FLAG_UNCERTAIN; }\n            }\n          }\n          if (cp_chain_run) {\n            let t = cp_chain(vec3<f32>(cm[0].x, cm[1].x, cm[2].x), C[0].x, C[1].x, C[2].x, C[3].x, C[4].x, C[5].x, vec3<f32>(cm[0].y, cm[1].y, cm[2].y), vec3<f32>(C[0].y, C[3].y, C[5].y));\n            if (t.flag) { pixel_flags = pixel_flags | FLAG_UNCERTAIN; }\n            let rows = array<vec3<f32>, 3>(t.a0, t.a1, t.a2);\n            for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n              hm[comp] = vec2<f32>(t.hm[comp], t.bm[comp]); hv[comp] = vec2<f32>(t.hv[comp], t.bv[comp]);\n              for (var c = 0u; c < 3u; c = c + 1u) { an[comp * 3u + c] = vec2<f32>(rows[comp][c], abs(rows[comp][c]) * t.ju); }\n            }\n            cp_used = true;\n          }\n          if (cp_used) { for (var k = 0u; k < 9u; k = k + 1u) { cp_a[k] = an[k]; } }\n        }\n        for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n          h_cand[d * 3u + comp] = hm[comp];\n          e_mean[comp] = b_add(e_mean[comp], b_scale(hm[comp], share));\n        }\n        // Linear (Stein) part with B_c·σ_c², and the pixel's own linear variance aᵀCa for the residual.\n        for (var c = 0u; c < 3u; c = c + 1u) {\n          if (!cp_used) {\n            an[c] = b_mul(b_mul(hs[0], b_rounded(m_of(0u, c))), bs[c]);\n            an[3u + c] = b_mul(b_mul(hs[1], b_rounded(m_of(1u, c))), bs[c]);\n            an[6u + c] = b_mul(b_mul(hs[2], b_rounded(m_of(2u, c))), bs[c]);\n          }\n          var bc = b0; if (c == 1u) { bc = b1; } if (c == 2u) { bc = b2; }\n          let wgt = b_mul(b_mul(b_scale(aj, share), b_rounded(bc)), csv[c]);\n          cov[0] = b_add(cov[0], b_mul(b_mul(wgt, an[c]), an[c])); cov[1] = b_add(cov[1], b_mul(b_mul(wgt, an[c]), an[3u + c])); cov[2] = b_add(cov[2], b_mul(b_mul(wgt, an[c]), an[6u + c]));\n          cov[3] = b_add(cov[3], b_mul(b_mul(wgt, an[3u + c]), an[3u + c])); cov[4] = b_add(cov[4], b_mul(b_mul(wgt, an[3u + c]), an[6u + c])); cov[5] = b_add(cov[5], b_mul(b_mul(wgt, an[6u + c]), an[6u + c]));\n        }\n        for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n          var l = vec2<f32>(0.0, 0.0);\n          for (var p = 0u; p < 3u; p = p + 1u) {\n            for (var q = 0u; q < 3u; q = q + 1u) {\n              var cij = C[0]; if (p + q == 1u) { cij = C[1]; } if ((p == 0u && q == 2u) || (p == 2u && q == 0u)) { cij = C[2]; } if (p == 1u && q == 1u) { cij = C[3]; }\n              if ((p == 1u && q == 2u) || (p == 2u && q == 1u)) { cij = C[4]; } if (p == 2u && q == 2u) { cij = C[5]; }\n              l = b_add(l, b_mul(b_mul(an[comp * 3u + p], an[comp * 3u + q]), cij));\n            }\n          }\n          let t = vec2<f32>(share * max(0.0, hv[comp].x - l.x), share * (hv[comp].y + l.y));\n          res[comp] = b_add(res[comp], b_mul(b_mul(aj, aj), t));\n        }\n      }\n      if (count == 2u) {\n        // #457: the reference takes AHD's choice on the noise-free mosaic, one candidate or (a tie) their average, so its value\n        // lies within half the candidates' difference plus their bounds of the average: the bound carries that, and the move\n        // accumulates in choice, which flags the pixel below when it exceeds CHOICE_SIGMA of the pixel's noise σ.\n        for (var comp = 0u; comp < 3u; comp = comp + 1u) {\n          let diff = abs(h_cand[comp].x - h_cand[3u + comp].x);\n          let both = h_cand[comp].y + h_cand[3u + comp].y;\n          choice[comp] = choice[comp] + aj.x * 0.5 * diff;\n          e_mean[comp].y = e_mean[comp].y + 0.5 * (diff + both);\n        }\n      }\n      for (var comp = 0u; comp < 3u; comp = comp + 1u) { s[comp] = b_add(s[comp], b_mul(aj, e_mean[comp])); }\n      if (count == 2u) {\n        // AHD's per-pixel choice: ¼·Var(H − V) at the last candidate's slopes, linear in the sites.\n        var x0 = vec2<f32>(0.0, 0.0); var x1 = vec2<f32>(0.0, 0.0); var x2 = vec2<f32>(0.0, 0.0);\n        for (var n = 0u; n < entries; n = n + 1u) {\n          let eb = cb + 4u + n * ${DISPLAY_VIEW_QUALITY_CANDIDATE_FLOATS}u;\n          let o = site_at(x + i32(aux[eb]), y + i32(aux[eb + 1u]));\n          let d0 = b_scale(bs[0], aux[eb + 2u] - aux[eb + 5u]);\n          let d1 = b_scale(bs[1], aux[eb + 3u] - aux[eb + 6u]);\n          let d2 = b_scale(bs[2], aux[eb + 4u] - aux[eb + 7u]);\n          let sv = vec2<f32>(sites[o + 1u], sites[o + 4u]);\n          var a = b_mul(hs[0], b_add(b_add(b_mul(b_rounded(params.m00), d0), b_mul(b_rounded(params.m01), d1)), b_mul(b_rounded(params.m02), d2)));\n          var b = b_mul(hs[1], b_add(b_add(b_mul(b_rounded(params.m10), d0), b_mul(b_rounded(params.m11), d1)), b_mul(b_rounded(params.m12), d2)));\n          var c = b_mul(hs[2], b_add(b_add(b_mul(b_rounded(params.m20), d0), b_mul(b_rounded(params.m21), d1)), b_mul(b_rounded(params.m22), d2)));\n          if (cp_used) {\n            // #593: the coupled slopes ∂out/∂c of the last candidate.\n            let e0 = aux[eb + 2u] - aux[eb + 5u]; let e1 = aux[eb + 3u] - aux[eb + 6u]; let e2 = aux[eb + 4u] - aux[eb + 7u];\n            a = b_add(b_add(b_scale(cp_a[0], e0), b_scale(cp_a[1], e1)), b_scale(cp_a[2], e2));\n            b = b_add(b_add(b_scale(cp_a[3], e0), b_scale(cp_a[4], e1)), b_scale(cp_a[5], e2));\n            c = b_add(b_add(b_scale(cp_a[6], e0), b_scale(cp_a[7], e1)), b_scale(cp_a[8], e2));\n          }\n          x0 = b_add(x0, b_mul(b_mul(a, a), sv)); x1 = b_add(x1, b_mul(b_mul(b, b), sv)); x2 = b_add(x2, b_mul(b_mul(c, c), sv));\n        }\n        let ajaj = b_mul(aj, aj);\n        res[0] = b_add(res[0], b_mul(ajaj, b_scale(x0, 0.25))); res[1] = b_add(res[1], b_mul(ajaj, b_scale(x1, 0.25))); res[2] = b_add(res[2], b_mul(ajaj, b_scale(x2, 0.25)));\n      }\n    }\n  }\n  // #457: a pixel whose candidates' choice could move it by more than CHOICE_SIGMA of its own noise σ is recomputed on the host.\n  if (choice[0] > CHOICE_SIGMA * sqrt(max(cov[0].x + res[0].x, 0.0)) || choice[1] > CHOICE_SIGMA * sqrt(max(cov[3].x + res[1].x, 0.0)) ||\n      choice[2] > CHOICE_SIGMA * sqrt(max(cov[5].x + res[2].x, 0.0))) { pixel_flags = pixel_flags | FLAG_CHOICE; }\n  let o = i * ${DISPLAY_VIEW_QUALITY_OUTPUT_FLOATS}u;\n  out[o] = s[0].x; out[o + 1u] = s[1].x; out[o + 2u] = s[2].x;\n  out[o + 3u] = cov[0].x + res[0].x; out[o + 4u] = cov[3].x + res[1].x; out[o + 5u] = cov[5].x + res[2].x;\n  out[o + 6u] = cov[1].x; out[o + 7u] = cov[2].x; out[o + 8u] = cov[4].x;\n  out[o + 9u] = max(s[0].y, max(s[1].y, s[2].y));\n  out[o + 10u] = max(max(cov[0].y + res[0].y, cov[3].y + res[1].y), max(cov[5].y + res[2].y, max(cov[1].y, max(cov[2].y, cov[4].y))));\n  flags[3u * i] = pixel_flags | select(0u, FLAG_HIGHLIGHT, undecided);\n  flags[3u * i + 1u] = mask0;\n  flags[3u * i + 2u] = mask1;\n}\n`;\n\n/** The pixel pass (`DISPLAY_VIEW_QUALITY_PIXELS_BODY_WGSL`) with the coupled steps' text: the stub by default, the coupled steps on demand (#593). */\nexport function assembleDisplayViewQualityPixelsWgsl(coupledWgsl: string = DISPLAY_VIEW_COUPLED_STUB_WGSL): string {\n  return DISPLAY_VIEW_QUALITY_PIXELS_BODY_WGSL + \"\\n\" + coupledWgsl;\n}\nexport const DISPLAY_VIEW_QUALITY_PIXELS_WGSL = /* @__PURE__ */ assembleDisplayViewQualityPixelsWgsl();\n\n/** The lattice pass with a material module appended (`fn material_at(p: vec2<f32>) -> u32`, `var<private> material_uncertain: bool`). */\nexport function assembleDisplayViewQualityLatticeWgsl(materialModuleWgsl: string): string {\n  return DISPLAY_VIEW_QUALITY_LATTICE_WGSL + \"\\n\" + materialModuleWgsl + \"\\n\";\n}\n"]}

package/dist/gpu/wgsl-display-view-speed-kernel.d.ts

@@ -1,43 +1,3 @@
-/**
- * WGSL kernel of the display view's Speed path on the GPU (#478): one invocation per display pixel of the exact
- * area mean (docs/DISPLAY_VIEW.md, "Exact area mean") without diffraction, with the bilinear demosaic and chart
- * materials. It evaluates `renderDisplayView`'s per-sample loop (facade/display-view.ts) for every real pixel of the
- * display pixel's footprint, in the reference's operation order where the order matters, in float32:
- *
- * 1. the real pixels the footprint covers and their overlap weights a_j = w_j / A (the host computes the 1-D overlaps
- *    divided by the scale in float64; the kernel multiplies the two axes);
- * 2. for every native site the real pixels' demosaic reads (the footprint plus one site of reach), the expected
- *    electrons: the scene material at the pixel aperture's query offsets (the material module, `material_at`), the
- *    material's channel signal integrated in the kernel from its spectrum (Σ_λ A_c(λ)·L(λ), `material_signal`), or (#581) the image-plane
- *    module's channel signal at the point (`scene_value`), summed over the offsets in the reference's order, plus the dark
- *    electrons;
- * 3. the readout moments of each site (mean k·e, variance k²·e + c₀, third cumulant k³·e) and the clip at the white
- *    level (the clipped-normal moments, in a cancellation-free standardized form);
- * 4. per real pixel the bilinear kernels, white balance and highlight clip, the colour matrix, the tone curve
- *    (identity, or positive Reinhard by the reference's 4 × 16-point Gauss–Legendre rule) and the clip to [0, 1] with
- *    the first-order skewness term;
- * 5. the display pixel's expected value Σ_j a_j·E[L_j], the linear (Stein) covariance Σ_c B_c·σ_c²·a_c·a_cᵀ with the
- *    closed-form B_c of the footprint, and the independent residual Σ_j a_j²·r_j.
- *
- * It outputs, per display pixel, the three means, the three variances (linear plus residual), the three covariances
- * and two running error terms: the absolute error bound of the mean and of the variance that the clip steps
- * contribute (a fixed formula evaluated from the pixel's own float32 values; docs/WEBGPU.md, "Display view (Speed)").
- * #533: it also decides each real pixel's saturation probability where the float32 values leave no doubt: a site is
- * certainly unclipped or certainly clipped when its standardized distance from white, with the site's mean moved by twice
- * `DISPLAY_VIEW_SITE_RELATIVE_U`·E of itself, still exceeds twice the reference's 8.5 (where the reference's probability is
- * exactly 0 or 1), and a real pixel is certainly saturated (P = 1) when a site its demosaic reads is certainly clipped and
- * certainly not (P = 0) when all are certainly unclipped. Three flag words per display pixel: bit 0 of the first the
- * material flag, bit 1 a real pixel left undecided (the host computes the pixel's count with the reference); the other two
- * a bit per real pixel j·nx + i of the footprint with P = 1.
- * The host adds the output quantization term, draws the noise, encodes and rounds (`display-view-speed-gpu.ts`), so
- * the noise draw is the CPU's bit for bit.
- *
- * Precision: float32 with an a-priori bound per step from the WGSL accuracy table (+ − × within one float32 rounding in
- * either direction, E = 2^-23; division 2.5 ULP, sqrt 4.5 ULP, exp 3 + 2|x| ULP; no trigonometric built-in is used). Every quotient that is a constant of
- * the request (the overlaps over the scale, the axis sums over the squared scale, 1/Σw² of the demosaic kernels) is
- * computed on the host in float64 and uploaded. Data-dependent quotients (z = (t − m)/s, the Abramowitz–Stegun t,
- * the skewness) carry the 2.5 ULP bound.
- */
 export declare const DISPLAY_VIEW_WORKGROUP_SIZE = 64;
 /** Largest footprint per axis (real pixels): the site window is (n + 2)² ≤ 64 entries of private storage. */
 export declare const DISPLAY_VIEW_GPU_MAXIMUM_PIXELS_PER_AXIS = 6;
@@ -102,4 +62,4 @@ export declare const DISPLAY_VIEW_SPEED_KERNEL_WGSL: string;
  */
 export declare const DISPLAY_VIEW_MATERIAL_SCENE_VALUE_WGSL = "\nvar<private> scene_m: u32 = 0xffffffffu;\nvar<private> scene_c: u32 = 0u;\nvar<private> scene_s: f32 = 0.0;\nfn scene_value(p: vec2<f32>, c: u32) -> f32 {\n  let m = material_at(p);\n  if (m != scene_m || c != scene_c) { scene_s = material_signal(m, c); scene_m = m; scene_c = c; }\n  return scene_s;\n}\n";
 /** The full shader: the kernel plus a material module (see `DISPLAY_VIEW_SPEED_KERNEL_WGSL`), and the table adapter unless the module defines `scene_value`. */
-export declare function assembleDisplayViewSpeedWgsl(materialModuleWgsl: string): string;
+export declare function assembleDisplayViewSpeedWgsl(materialModuleWgsl: string, coupledWgsl?: string): string;

package/dist/gpu/wgsl-display-view-speed-kernel.js

@@ -39,6 +39,7 @@
  * computed on the host in float64 and uploaded. Data-dependent quotients (z = (t − m)/s, the Abramowitz–Stegun t,
  * the skewness) carry the 2.5 ULP bound.
  */
+import { DISPLAY_VIEW_COUPLED_STUB_WGSL } from "./wgsl-display-view-coupled-stub.js";
 export const DISPLAY_VIEW_WORKGROUP_SIZE = 64;
 /** Largest footprint per axis (real pixels): the site window is (n + 2)² ≤ 64 entries of private storage. */
 export const DISPLAY_VIEW_GPU_MAXIMUM_PIXELS_PER_AXIS = 6;
@@ -102,6 +103,11 @@ const literal = (value) => {
  * declares how many roundings it adds (`DisplayViewGpuMaterialModule.extraSiteRoundings`). The module may declare bindings
  * in @group(1) and use the constants E and SLACK declared here.
  */
+/*
+ * 2.22.0: in the luminance branch the nested loops read the colour matrix rows from one local array (`mm`) instead of calling
+ * `m_of` (a `switch`): SwiftShader's compiler ended the GPU process on the `switch` inside those loops (after about 44 s). The
+ * values are the same (docs/WEBGPU.md, "Coupled rendering steps").
+ */
 export const DISPLAY_VIEW_SPEED_KERNEL_WGSL = /* wgsl */ `
 const E: f32 = ${E_LITERAL};
 const SLACK: f32 = 1e-14;
@@ -130,6 +136,8 @@ m22: f32, scale_u: f32, scale_v: f32, pad3: f32,
 @group(0) @binding(4) var<storage, read_write> out: array<f32>;
 @group(0) @binding(5) var<storage, read_write> flags: array<u32>;
 
+fn cp_k(i: u32) -> f32 { return aux[${DISPLAY_VIEW_AUX.size}u + i]; }
+
 var<private> site_mean: array<f32, 64>;
 var<private> site_var: array<f32, 64>;
 var<private> site_third: array<f32, 64>;
@@ -243,6 +251,7 @@ ed = ed + weight / ((1.0 + x) * (1.0 + x));
 let shifted = e1 - hm;
 return vec3<f32>(e1, max(0.0, e2 - shifted * shifted), ed);
 }
+
 fn reflect_x(x: i32) -> i32 { if (x < 0) { return -x; } if (x >= i32(params.native_w)) { return 2 * i32(params.native_w) - 2 - x; } return x; }
 fn reflect_y(y: i32) -> i32 { if (y < 0) { return -y; } if (y >= i32(params.native_h)) { return 2 * i32(params.native_h) - 2 - y; } return y; }
 fn channel_of(x: i32, y: i32) -> u32 {
@@ -362,6 +371,14 @@ let bit = j * nx + ii;
 if (bit < 32u) { mask0 = mask0 | (1u << bit); } else { mask1 = mask1 | (1u << (bit - 32u)); }
 } else if (near_white) { undecided = true; }
 
+var cp_chain_run = false;
+if (cp_on() && cp_reconstruct()) {
+let reach = cp_reconstruct_reach(vec3<f32>(cm[0], cm[1], cm[2]), vec3<f32>(cvv[0], cvv[1], cvv[2]));
+if (reach == 2u) { material_uncertain = true; }
+cp_chain_run = reach >= 1u;
+}
+let cp_unclipped = cp_on() && cp_reconstruct();
+
 var bm: array<f32, 3>; var bv: array<f32, 3>; var bs: array<f32, 3>; var bbm: array<f32, 3>; var bbv: array<f32, 3>;
 for (var c = 0u; c < 3u; c = c + 1u) {
 let gc = g_of(c);
@@ -370,7 +387,7 @@ let uv = gc * gc * cvv[c];
 let gap = params.clip_level - um;
 bbm[c] = gc * CLIP_MEAN_U * cb[c];
 bbv[c] = gc * gc * CLIP_VARIANCE_U * cb[c] * cb[c];
-if (gap > 0.0 && gap * gap > params.reach2 * uv) { bm[c] = um; bv[c] = uv; bs[c] = gc; }
+if (cp_unclipped || (gap > 0.0 && gap * gap > params.reach2 * uv)) { bm[c] = um; bv[c] = uv; bs[c] = gc; }
 else {
 let clipped = upper_clip(um, uv, params.clip_level);
 bm[c] = clipped.x; bv[c] = clipped.y; bs[c] = gc * clipped.z;
@@ -405,14 +422,60 @@ bv_c = bv_c + 2.0 * CLIP_VARIANCE_U * variance;
 }
 hbm[comp] = bm_c; hbv[comp] = bv_c;
 }
+
+
+var jm = array<vec3<f32>, 3>(vec3<f32>(0.0), vec3<f32>(0.0), vec3<f32>(0.0));
+var ju = 0.0;
+var cp_used = false;
+if (cp_on()) {
+if (!cp_chain_run && cp_luminance()) {
+var mu = vec3<f32>(0.0); var dmu = vec3<f32>(0.0); var dsv = vec3<f32>(0.0);
+var srow = array<vec3<f32>, 3>(vec3<f32>(0.0), vec3<f32>(0.0), vec3<f32>(0.0));
+var mm = array<vec3<f32>, 3>(vec3<f32>(params.m00, params.m01, params.m02), vec3<f32>(params.m10, params.m11, params.m12), vec3<f32>(params.m20, params.m21, params.m22));
+for (var r = 0u; r < 3u; r = r + 1u) {
+mu[r] = mm[r][0] * bm[0] + mm[r][1] * bm[1] + mm[r][2] * bm[2];
+dmu[r] = (abs(mm[r][0]) * bbm[0] + abs(mm[r][1]) * bbm[1] + abs(mm[r][2]) * bbm[2]) * E;
+dsv[r] = (mm[r][0] * mm[r][0] * bbv[0] + mm[r][1] * mm[r][1] * bbv[1] + mm[r][2] * mm[r][2] * bbv[2]) * E;
+for (var q = 0u; q < 3u; q = q + 1u) { srow[r][q] = mm[r][0] * mm[q][0] * bv[0] + mm[r][1] * mm[q][1] * bv[1] + mm[r][2] * mm[q][2] * bv[2]; }
+}
+let t = cp_tone(mu, srow[0], srow[1], srow[2], dmu, dsv);
+if (t.flag) { material_uncertain = true; }
+if (cp_desaturate() && t.top) { cp_chain_run = true; }
+else {
+let rows = array<vec3<f32>, 3>(t.a0, t.a1, t.a2);
+for (var comp = 0u; comp < 3u; comp = comp + 1u) {
+hm[comp] = t.hm[comp]; hv[comp] = t.hv[comp]; hbm[comp] = t.bm[comp] / E; hbv[comp] = t.bv[comp] / E;
+for (var c = 0u; c < 3u; c = c + 1u) { jm[comp][c] = (rows[comp].x * m_of(0u, c) + rows[comp].y * m_of(1u, c) + rows[comp].z * m_of(2u, c)) * bs[c]; }
+}
+ju = t.ju / E; cp_used = true;
+}
+} else if (!cp_chain_run && cp_desaturate() && params.tone != 1u) {
+
+for (var comp = 0u; comp < 3u; comp = comp + 1u) {
+let top = hm[comp] + CLIP_REACH * sqrt(hv[comp]);
+if (hm[comp] > 0.0 && top >= 0.999) { cp_chain_run = true; }
+if (hm[comp] > 0.0 && hm[comp] - 2.0 * hbm[comp] * E < 0.999 && abs(top - 0.999) <= 2.0 * (hbm[comp] * E + CLIP_REACH * min(sqrt(hbv[comp] * E), hbv[comp] * E / (2.0 * sqrt(hv[comp])))) + 64.0 * E) { material_uncertain = true; }
+}
+}
+if (cp_chain_run) {
+let dcm = vec3<f32>(CLIP_MEAN_U * cb[0], CLIP_MEAN_U * cb[1], CLIP_MEAN_U * cb[2]) * E;
+let dcv = vec3<f32>(CLIP_VARIANCE_U * cb[0] * cb[0], CLIP_VARIANCE_U * cb[1] * cb[1], CLIP_VARIANCE_U * cb[2] * cb[2]) * E;
+let t = cp_chain(vec3<f32>(cm[0], cm[1], cm[2]), cvv[0], 0.0, 0.0, cvv[1], 0.0, cvv[2], dcm, dcv);
+if (t.flag) { material_uncertain = true; }
+jm[0] = t.a0; jm[1] = t.a1; jm[2] = t.a2;
+for (var comp = 0u; comp < 3u; comp = comp + 1u) { hm[comp] = t.hm[comp]; hv[comp] = t.hv[comp]; hbm[comp] = t.bm[comp] / E; hbv[comp] = t.bv[comp] / E; }
+ju = t.ju / E; cp_used = true;
+}
+}
 s0 = s0 + aj * hm[0]; s1 = s1 + aj * hm[1]; s2 = s2 + aj * hm[2];
 
 var l0 = 0.0; var l1 = 0.0; var l2 = 0.0;
 var lb0 = 0.0; var lb1 = 0.0; var lb2 = 0.0;
 for (var c = 0u; c < 3u; c = c + 1u) {
-let a0 = hs[0] * m_of(0u, c) * bs[c];
-let a1 = hs[1] * m_of(1u, c) * bs[c];
-let a2 = hs[2] * m_of(2u, c) * bs[c];
+var a0 = hs[0] * m_of(0u, c) * bs[c];
+var a1 = hs[1] * m_of(1u, c) * bs[c];
+var a2 = hs[2] * m_of(2u, c) * bs[c];
+if (cp_used) { a0 = jm[0][c]; a1 = jm[1][c]; a2 = jm[2][c]; }
 var bc = b0; if (c == 1u) { bc = b1; } if (c == 2u) { bc = b2; }
 let wgt = aj * bc * csv[c];
 c00 = c00 + wgt * a0 * a0; c01 = c01 + wgt * a0 * a1; c02 = c02 + wgt * a0 * a2;
@@ -420,8 +483,8 @@ c11 = c11 + wgt * a1 * a1; c12 = c12 + wgt * a1 * a2; c22 = c22 + wgt * a2 * a2;
 l0 = l0 + a0 * a0 * cvv[c]; l1 = l1 + a1 * a1 * cvv[c]; l2 = l2 + a2 * a2 * cvv[c];
 
 
-lb0 = lb0 + (a0 * a0) * bbv[c]; lb1 = lb1 + (a1 * a1) * bbv[c]; lb2 = lb2 + (a2 * a2) * bbv[c];
-        bound_var = bound_var + aj * bc * (bbv[c] * aux[${DISPLAY_VIEW_AUX.inverseW2}u + phase * 3u + c]) * max(a0 * a0, max(a1 * a1, a2 * a2));
+lb0 = lb0 + (a0 * a0) * (bbv[c] + 2.0 * ju * cvv[c]); lb1 = lb1 + (a1 * a1) * (bbv[c] + 2.0 * ju * cvv[c]); lb2 = lb2 + (a2 * a2) * (bbv[c] + 2.0 * ju * cvv[c]);
+        bound_var = bound_var + aj * bc * ((bbv[c] + 2.0 * ju * cvv[c]) * aux[${DISPLAY_VIEW_AUX.inverseW2}u + phase * 3u + c]) * max(a0 * a0, max(a1 * a1, a2 * a2));
 }
 r0 = r0 + aj * aj * max(0.0, hv[0] - l0);
 r1 = r1 + aj * aj * max(0.0, hv[1] - l1);
@@ -455,8 +518,8 @@ return scene_s;
 }
 `;
 /** The full shader: the kernel plus a material module (see `DISPLAY_VIEW_SPEED_KERNEL_WGSL`), and the table adapter unless the module defines `scene_value`. */
-export function assembleDisplayViewSpeedWgsl(materialModuleWgsl) {
+export function assembleDisplayViewSpeedWgsl(materialModuleWgsl, coupledWgsl = DISPLAY_VIEW_COUPLED_STUB_WGSL) {
     const adapter = /\bfn scene_value\(/u.test(materialModuleWgsl) ? "" : DISPLAY_VIEW_MATERIAL_SCENE_VALUE_WGSL;
-    return DISPLAY_VIEW_SPEED_KERNEL_WGSL + "\n" + materialModuleWgsl + "\n" + adapter;
+    return DISPLAY_VIEW_SPEED_KERNEL_WGSL + "\n" + materialModuleWgsl + "\n" + adapter + "\n" + coupledWgsl;
 }
 //# sourceMappingURL=wgsl-display-view-speed-kernel.js.map