@vitreajs/vitrea 0.4.0 → 0.5.0

package/dist/index.d.ts

@@ -2031,6 +2031,111 @@ interface MaterialProfile {
      * 0.0174.
      */
     readonly sizeScatterSpanMax: number;
+    /**
+     * **The body's second scale** (W15 G1, from the measurement of claims §5.69
+     * §1–§2) — the heavy width's gain, the deep value's floor and the deep
+     * value's span top, each read again at dpr 2.
+     *
+     * At 2x the reference's body is a different object from the one the three
+     * constants above describe, and §5.69 measured all of it: the deep interior
+     * is FULLY heavy on the two largest spans and 0.90–0.95 heavy on `rrect-md`
+     * (§2), where the 1x span curve leaves a sharp share of 0.24–0.36 there, and
+     * the heavy component's own width is 8–11 device px against the 1x law's 10
+     * CSS px (§1). One curve fitted at 1x cannot be both, so the deep value's two
+     * constants and the gain get a second reading rather than a scale factor —
+     * the same shape the ramp's anchors already have.
+     *
+     * Each is interpolated by `rampAtScale`: held at the 1x constant at dpr ≤ 1,
+     * at this one from dpr 2 up, linear between. **At dpr 1 every one of them is
+     * inert by construction, whatever it says** — W15's binding rule — so the 1x
+     * material is byte-identical to the W13 bed and the landing's confirmation
+     * measured exactly that: 49 of 49 1x GPU cells identical to the bed with
+     * every capture byte-identical (claims §5.70 §8).
+     *
+     * Consumed at draw time, not folded into a constant: `scatterDeepThickness`
+     * takes the ratio for the floor and the span top, the renderer resolves all
+     * three from the viewport's own ratio and hands them to the optics pass
+     * through the uniforms the shader already reads (`scatter.x`, `lensOval.w`
+     * and `size.x`), and `scatterThickness`, `scatterSharpShare` and the ramp's
+     * area-average projection read the resolved deep value.
+     *
+     * **FITTED** by W15 G1's runtime sweep at
+     * `--profile apple-macos-26.5-2x-light-standard` on the calibration set, the
+     * holdout untouched, and landed at the values below (claims §5.70 §2 and §8;
+     * the four sweeps `results/2026-09-04-w15-body-2x/g1/stage1..stage3` and the
+     * confirmation `confirm-3`, whose 2x holdout was read once for this
+     * configuration). The rule W13 wrote after the paper model over-credited the
+     * mip chain's heavy tap (§5.58 §1) held: no constant here lands on a paper
+     * prediction, and the sweep contradicted the paper estimate.
+     *
+     * **The gain, 4.8.** Stage 2c swept it over 4.0 / 4.4 / 4.8 / 5.2 at the
+     * chosen base and found an INTERIOR MINIMUM at 4.8 — the interior spread
+     * against native reading 0.0335 / 0.0214 / 0.0120 / 0.0217 with the band's
+     * rise at its top too. Through `bodySigmaCssFor` that is a heavy width of
+     * **6 device px** at dpr 2, NARROWER than the 8–11 device px §5.69 §1
+     * bounded as a Gaussian: the estimator carries ±40% on a real capture of a
+     * known law (§5.69 §3) and the mip chain's tap is not a Gaussian, so the
+     * renderer is the fitting instrument and 6 is the width it draws best.
+     *
+     * **The floor, 1.0** — the deep value FULLY heavy at dpr 2, which is what
+     * G0's deep windows read on the two largest spans (§5.69 §2). `kDeep` is
+     * therefore 1 at every span at this ratio, and the depth ramp above it is the
+     * whole body (§5.70 §1).
+     *
+     * **The span top, 256** — the 1x value, unchanged, because a floor of 1
+     * leaves the deep value nothing to rise to and the top has no work at this
+     * ratio. Stage 1 swept it against 128 and the bed did not ask for the change.
+     */
+    readonly sizeScatterGainMax2x: number;
+    readonly sizeScatterFloor2x: number;
+    readonly sizeScatterSpanMax2x: number;
+    /**
+     * **The heavy width's gain at the TOP of the scatter span curve, at dpr 2**
+     * (W15 G1's re-form, from claims §5.70 §4 and §7 and the measurement of
+     * §5.69 §1) — the second-scale term that lets the 2x heavy width GROW with
+     * the span instead of being one number for the whole bed.
+     *
+     * `sizeScatterGainMax2x` above is one gain at dpr 2, and the sweep that fitted
+     * it read the calibration cells, whose spans are 32–128. The reference's heavy
+     * kernel is not one width: §5.69 §1's bounded per-span fit reads 8.0 / 7.5 /
+     * 8.0 / 9.0 / 11.0 device px across the bed's spans, a ratio of 1.375 between
+     * span 160 and span 96. Landing the single gain accordingly left the largest
+     * span's deep interior about 40% too structured (`rrect-lg` at 2x, spread
+     * 0.1134 against the reference's 0.0810, claims §5.70 §4) while every smaller
+     * cell rose. So the gain takes a second grading, in SPAN:
+     *
+     * ```
+     * gain(span, dpr) = gainAtScale(dpr)
+     *                 + (gainFar(dpr) − gainAtScale(dpr))
+     *                   · smoothstep(sizeSpanMax, sizeScatterSpanMax(dpr), span)
+     * ```
+     *
+     * with `gainAtScale` the existing `scatterGainAtScale` and `gainFar` this
+     * constant interpolated by `rampAtScale` from `sizeScatterGainMax`. The curve
+     * is not a new span statistic: it is the same smoothstep the fourth form's
+     * far-anchor decline already rides (`scatterRampStart`), so the width and the
+     * ramp's start move along one curve above the thickness knee.
+     *
+     * **Inert at dpr 1 by construction.** `gainFar` interpolates from
+     * `sizeScatterGainMax` — the 1x gain, not this constant — so at dpr ≤ 1 the
+     * far gain IS the base gain and the whole term is a flat no-op whatever this
+     * says. That is W15's binding rule (the 1x material does not move) discharged
+     * by the shape of the formula rather than by a capture.
+     *
+     * **LANDED at 9.9** (claims §5.70 §8; W15 Decision Log 3, the user's landing
+     * "the re-form"), and the value is G0's rather than the holdout row's: §5.69
+     * §1's bounded per-span reading is 8 device px at span 96 and 11 at 160, a
+     * ratio of 1.375, and this curve reads 0.352 at span 160, so 4.8 + (9.9 −
+     * 4.8) · 0.352 reproduces that ratio over the base gain of 6 device px. The
+     * heavy width at dpr 2 is therefore 6.0 device px at spans ≤ 96, 6.7 at 128,
+     * 8.2 at 160 and 12.4 at 256 and beyond. The calibration cells neither chose
+     * the value nor rejected it — `stage3` swept 4.8 / 7.5 / 9.9 / 12.5 and only
+     * `rrect-ml` (span 128, a tenth of the grading) moves, its interior spread
+     * 0.1125 → 0.0993 against native 0.1018 at 9.9, which is the sweep's best on
+     * its own objective. The confirmation's holdout `rrect-lg` rose 0.9661 →
+     * 0.9762 with its interior spread 0.0721 against native 0.0810.
+     */
+    readonly sizeScatterGainFar2x: number;
     /**
      * **The body's depth ramp** (W13 G1, from the measurement of claims §5.61 §2
      * and the re-forming its runtime sweep forced) — a near-contour excursion on
@@ -2113,15 +2218,15 @@ interface MaterialProfile {
      *            + (far − thick) · smoothstep(sizeSpanMax, sizeScatterSpanMax, span)
      *
      * Two curves the material already has, one more constant per scale, and no
-     * new span statistic. At 2x `far` sits below the thick anchor, so the null
-     * the third form verified bit-exactly (§5.67 §2) holds by construction: an
-     * excursion that was zero at every span stays zero when its start falls
-     * further.
+     * new span statistic. At 2x the decline is switched off rather than inverted:
+     * `far` and `thick` carry the same number since W15 G1, because G0 read the
+     * reference's 2x start FLAT across the thick spans (claims §5.69 §2, §5.70 §6
+     * (iv)) — the decline is a 1x feature of this form.
      *
-     * PROVISIONAL, all eight. The runtime sweep at this form sets them, because
-     * the paper model over-credited the mip chain's heavy tap once already
-     * (§5.58 §1) and no constant lands on a paper prediction again. What they
-     * carry now:
+     * The 1x three are FITTED (W13 G1's third and fourth sweeps) and the 2x four
+     * are FITTED or read from G0 by W15 G1 (claims §5.70 §2 and §8); no constant
+     * here lands on a paper prediction, because the paper model over-credited the
+     * mip chain's heavy tap once already (§5.58 §1). What they carry:
      *
      * **1x — thin 0.64, thick 0.52, reach 120 device px.** G0's own start
      * readings are 0.637 on `rrect-sm` and 0.642 on the capsule at the thin end
@@ -2132,20 +2237,27 @@ interface MaterialProfile {
      * and is also G0's reading on that cell, so the fit is not traded against the
      * form's own arithmetic.
      *
-     * **2x — thin 0.46, thick 0.17, reach 100 device px, and the excursion clamps
-     * to zero on every cell of the calibration bed.** These are G0's 2x readings
-     * (0.483 / 0.437 on the thin cells, 0.192 / 0.179 / 0.141 on the thick ones),
-     * and every one of them sits BELOW its cell's deep sharp share: the implied
-     * excursions run −0.095 to −0.289 (claims §5.64 §4). `max(0, s₀ − sDeep)` is
-     * therefore zero everywhere on this bed at dpr 2. That is not a special case
-     * and not a disabled facet — it is the law evaluating to no excursion where
-     * the reference's own contour share is heavier than vitrea's deep value. It
-     * follows that **the 2x gap is a deep-value gap and not a ramp gap**: the
-     * span law's floor, knee and top are what miss at that scale, and no
-     * one-signed excursion above them can express it. That fit is W13's Deferred
-     * charter and not this constant's business.
-     *
-     * **far — 0.20 at 1x (FITTED on the W14 bed), 0.15 at 2x.** The fourth
+     * **2x — thin 0.46, thick 0.21, far 0.21, reach 100 device px, and the ramp
+     * is now the WHOLE body above the deep value** (W15 G1, claims §5.70 §1–§2).
+     * W13 left these inert for the deep value's sake: at `sizeScatterFloor2x` 0.4
+     * every anchor sat below its cell's deep sharp share and `max(0, s₀ − sDeep)`
+     * was bit-exactly zero on the whole bed. W15 G1 fitted that deep value fully
+     * heavy (`sizeScatterFloor2x` 1.0), so the deep sharp share at dpr 2 is 0 and
+     * the anchors act on the numbers they were carrying — which is why the wave's
+     * three gaps turned out to be one gap (§5.70 §6 (i)).
+     *
+     * The values are G0's readings where the sweep was flat and the sweep's where
+     * it was not. **Thin 0.46** is G0's u 6 reading exactly (0.494 / 0.468 on the
+     * thin cells) and stage 2b's best, over a top nine points within 0.002 of
+     * each other. **Thick and far 0.21**: G0's 2x starts on the thick cells are
+     * 0.163 / 0.199 / 0.210 at their own widths and 0.192 / 0.199 / 0.195 at one
+     * common width — flat with span, not declining — so the two anchors take one
+     * number and the fourth form's decline is off at this scale; stage 2a
+     * measured the objective flat over 0.17–0.25 and worse at 0.33. **Reach 100
+     * device px** = 50 CSS px against G0's measured 49 / 53 / 56, and stage 2b
+     * found 100 and 130 indifferent.
+     *
+     * **far — 0.20 at 1x (FITTED on the W14 bed), 0.21 at 2x.** The fourth
      * sweep (`results/2026-09-03-w13-ramp/g1/sweep-4/`) swept far over eight
      * points from 0.52 (the third form) down to 0.15: S1 and S4 pass at every
      * one, only `rrect-ml`'s rows move and they improve monotonically as far
@@ -2155,8 +2267,14 @@ interface MaterialProfile {
      * is far = 0.207 through this form — and 0.20 is the measured grid point that
      * carries it. The confirmation read `rrect-lg` at `ssimMean` +0.0056 against
      * the W14 bed (the third form: −0.0026) with its interior 12% over the
-     * reference (from 33% over). At 2x far is below the thick anchor so that the
-     * null holds — re-verified bit-exact under this form.
+     * reference (from 33% over). At 2x far equals the thick anchor, so the start
+     * is flat past the knee at that scale — G0's own reading (§5.69 §2), and the
+     * holdout's band supports it (`rrect-lg` +0.0106, claims §5.70 §8).
+     *
+     * At dpr 1 none of the 2x four is read at all: `rampAtScale` holds every one
+     * of them at its 1x anchor below dpr 1, so the 1x material is byte-identical
+     * to the W13 bed and the landing's confirmation measured that on 49 of 49 1x
+     * GPU cells (claims §5.70 §8).
      */
     readonly sizeScatterRampStartThin1x: number;
     readonly sizeScatterRampStartThick1x: number;
@@ -2438,6 +2556,10 @@ interface MaterialProfilePatch {
     readonly sizeScatterGainMax?: number;
     readonly sizeScatterFloor?: number;
     readonly sizeScatterSpanMax?: number;
+    readonly sizeScatterGainMax2x?: number;
+    readonly sizeScatterFloor2x?: number;
+    readonly sizeScatterSpanMax2x?: number;
+    readonly sizeScatterGainFar2x?: number;
     readonly sizeScatterRampStartThin1x?: number;
     readonly sizeScatterRampStartThick1x?: number;
     readonly sizeScatterRampStartFar1x?: number;

package/dist/index.js

@@ -1092,7 +1092,7 @@ function isHealthy(state) {
 
 // src/renderer-seam.ts
 async function loadWebGPURendererModule() {
-  return import('./dist-RGWZAZ6J.js');
+  return import('./dist-OHNXYEZP.js');
 }
 async function loadWebGPURenderer() {
   const { createWebGPURenderer } = await loadWebGPURendererModule();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vitreajs/vitrea",
-  "version": "0.4.0",
+  "version": "0.5.0",
   "description": "Framework-agnostic Liquid Glass material runtime: scene model, capability resolution, material policy, accessibility policy.",
   "license": "Apache-2.0",
   "homepage": "https://github.com/SSFSKIM/designer",