2.2.0: round-trip ΔE by lightness (reference QC scatter) + fault-isolated Analysis tab

Implements the reference report's "B2A1 Roundtrip dEab by quantized L*; Pts vary
from GBD to Neutral" figure, and stops one bad analysis from blanking the page.

New: iccviz::RoundTripByLightness
  32 lightness levels × 64 gamut-boundary points, each repeated eroded toward
  neutral at chroma ×0.8/×0.5/×0.2, last point of each level replaced by neutral.
  8192 points — the reference's exact count. New WASM export
  roundTripByLightness; new plot under the histogram in Profile Statistics.

  An envelope (min/mean/max per band) was built FIRST and discarded after reading
  the actual figure: its entire content is the WITHIN-band shape — error falling
  from the gamut surface to the neutral axis — and any summary erases exactly
  that. Individual points, dotted band separators, 2 px markers.

  Convergence against the published GRACoL2013_CRPC6 figures (0.343 / 0.947 / 2.76):
    first cut                            0.482 / 1.377 / 8.43
    + reference L* range                 0.416 / 1.222 / 8.79
    + L-interpolated GBD                 0.386 / 1.149 / 6.25
    + denser boundary sampling (S=36)    0.411 / 1.248 / 3.22

  Two findings behind those steps:
  - The L* range is INKSET-derived, not gamut-derived: hi = (100 + L_yellow)/2,
    lo = (L_blackpoint + L_blue)/2, reusing HueExtrema + WhiteBlackPoints. That
    reproduces the reference's 18.82–96.85 exactly. The raw gamut extent drags in
    shadow levels the reference excludes, where the gamut is a sliver.
  - A cell-maximum gamut descriptor biases OUTWARD: seeds land just outside the
    gamut, B2A clamps them, and the ΔE reports our sampling rather than the
    profile. The tell was a 13× cliff between the boundary ring and the ×0.8 ring.
    Interpolating in L and sampling denser fixed most of it (max 8.4 → 3.2).

  Residual ~20% high on mean: our GBD comes from a boundary point cloud, the
  reference builds one with a gamut routine we do not have. Ours errs HIGH, the
  conservative direction. Shape and scale match; the absolutes are ours.

Fix: the Analysis tab is fault-isolated
  ErrorBoundary existed in the tree but was wired NOWHERE (its own comment claimed
  otherwise), so any render throw unmounted the whole React tree and left a blank
  browser tab with no message. The Analysis tab — several independent analyses over
  untrusted profile data through a WASM engine — is now boundaried, with a readable
  fallback and resetKey={bytes}.

  Caught this because RtByLightness was mounted inside StatsTable, which does not
  receive `bytes`/`intent`; they resolved as undefined globals and threw. Now
  threaded from ProfileStatsSection. Note the build CANNOT catch that class of bug
  (no scope analysis) and `npx eslint` fails here with "couldn't find
  eslint.config.js" — the repo has ESLint installed but no config, so nothing
  lints. A flat config with no-undef would have caught it; not added here because
  it needs plugin deps and would surface pre-existing findings.

Also
  PlotlyGraph gains `markerSize` so a dense scatter can ask for small dots.
  Removed the superseded per-band aggregation from DeStats / RoundTripResult /
  plot-wrapper rather than leave it unread.
  i18n: dead band_min/mean/max keys removed (verified unreferenced first), new
  analysis_rt_band_pts added across 12 locales; CJK includes the measure word so
  it reads after a numeral. Spreadsheets re-synced.
  MANUAL: Analysis section documents how to read the plot band by band, and why
  its numbers are not comparable with the table above it.

WASM built from a clean iccDEV origin/master worktree.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JkxVtREhRmqbY71R5KnV7e

Author: colourbill-ctrl

MANUAL.md

@@ -215,6 +215,14 @@ The four round-trip types:
 - **Reproducibility (RT2)** — ΔE\*ab between the first and second round trips: how stable a repeated round trip is, independent of the first trip's gamut clipping.
 - **PRMG interoperability** — PCS colours inside the Perceptual Reference Medium Gamut round-tripped once; the ΔE distribution indicates cross-profile interoperability.
 
+Below the histogram, **Round-trip ΔE by lightness** answers the question the summary figures cannot: *where* in the tone scale the inversion struggles, and *how far into the gamut* the trouble reaches.
+
+It samples 32 lightness levels. At each level it takes 64 points **on the gamut boundary**, then repeats them eroded toward the neutral axis at 80 %, 50 % and 20 % of their chroma — 8,192 points in all, each plotted individually, with dotted separators between levels.
+
+Read it band by band. Within one band the points run from the gamut surface on the left to neutral on the right, so the error should **fall away** across the band: colours near the boundary are the hardest to invert, colours near neutral the easiest. A band whose error stays high all the way to neutral is a profile in trouble at that lightness. The tallest spikes mark the lightness levels where the gamut boundary itself is least well inverted.
+
+Every seed is inside the gamut by construction, so what you see is genuine `B2A`/`A2B` disagreement rather than colours being clipped for being unreachable. Note that this plot has its **own** sampling — deliberately weighted toward the gamut boundary — so its mean and maximum are higher than, and not comparable with, the table above. It follows the rendering intent but not the round-trip type.
+
 A profile lacking the device↔PCS transforms a metric needs shows a *not applicable* note; a device space too wide to sample is reported as skipped rather than as an error.
 
 #### Extrema Colorimetry

frontend/package-lock.json

@@ -1,6 +1,6 @@
 {
   "name": "profiletool-frontend",
-  "version": "2.1.0",
+  "version": "2.2.0",
   "description": "ICC profile viewer and validation UI",
   "type": "module",
   "scripts": {

frontend/package.json

@@ -329,6 +329,10 @@ <h4>Profile Statistics</h4>
 <li><strong>Reproducibility (RT2)</strong> — ΔE*ab between the first and second round trips: how stable a repeated round trip is, independent of the first trip's gamut clipping.</li>
 <li><strong>PRMG interoperability</strong> — PCS colours inside the Perceptual Reference Medium Gamut round-tripped once; the ΔE distribution indicates cross-profile interoperability.</li>
 </ul>
+<p>Below the histogram, <strong>Round-trip ΔE by lightness</strong> answers the question the summary figures cannot: <em>where</em> in the tone scale the inversion struggles, and <em>how far into the gamut</em> the trouble reaches.</p>
+<p>It samples 32 lightness levels. At each level it takes 64 points <strong>on the gamut boundary</strong>, then repeats them eroded toward the neutral axis at 80 %, 50 % and 20 % of their chroma — 8,192 points in all, each plotted individually, with dotted separators between levels.</p>
+<p>Read it band by band. Within one band the points run from the gamut surface on the left to neutral on the right, so the error should <strong>fall away</strong> across the band: colours near the boundary are the hardest to invert, colours near neutral the easiest. A band whose error stays high all the way to neutral is a profile in trouble at that lightness. The tallest spikes mark the lightness levels where the gamut boundary itself is least well inverted.</p>
+<p>Every seed is inside the gamut by construction, so what you see is genuine <code>B2A</code>/<code>A2B</code> disagreement rather than colours being clipped for being unreachable. Note that this plot has its <strong>own</strong> sampling — deliberately weighted toward the gamut boundary — so its mean and maximum are higher than, and not comparable with, the table above. It follows the rendering intent but not the round-trip type.</p>
 <p>A profile lacking the device↔PCS transforms a metric needs shows a <em>not applicable</em> note; a device space too wide to sample is reported as skipped rather than as an error.</p>
 <h4>Extrema Colorimetry</h4>
 <p>The ends of the profile's reproduction range — the numbers a print operator usually reads first.</p>

frontend/public/help.html

@@ -6,8 +6,8 @@ e416013b730ef48882d781115ffd41772238c6bad1dd2326fc080ed5110bf653  iccjson.mjs
 5efa97180dd0235152cd167c847a6d7c370f132e86faec9e451300ea3715313f  iccjson.wasm
 3f6fa5c27187c9d99756bac44d3b2aad943847ea0786413af9803af4018a6fae  iccconstruct.mjs
 55880e6e58fc6a84bf03d15bba16233eff5b0db42fc996440548847a67df5e6e  iccconstruct.wasm
-63ce2f51dca5f2d141c4c317e95dff40ed8e1d8e45afe7bbce2ba7dfd3771b55  iccplot.mjs
-b57a8f8d174054fcdd516082f8e1a62b60b0c74a5f354966e04947ef3615aef5  iccplot.wasm
+054553512d5c57f30ae64a4f5532e1f0d9c45596a18ee9a980e473d9e3a719e2  iccplot.mjs
+d5eb8baeebcad4b936bfd12c1b3877626df9d317d91e07bfdf118eff22c608c0  iccplot.wasm
 dabe0d142bf4c7590eaec75a921933a65e7fb565e653a4865ec71651b6b128ab  iccpawg.mjs
 89d76edecb7070ee94c3b2018a69ec69fc88698a2e7e4eac971583a4fd22c3b1  iccpawg.wasm
 dae04d36df7afa645c70e856cb9719fd285d66574ebe7495b8edbb0c39105c08  iccimage.mjs

frontend/public/wasm/SHA256SUMS

@@ -2,7 +2,7 @@
 import { useEffect, useState } from 'react'
 import {
   enumerateVisualizations, renderGraph, tagEvalInfo, gamutVolume, roundTripStats,
-  whiteBlackPoints, hueExtrema, shadowInkPaths,
+  whiteBlackPoints, hueExtrema, shadowInkPaths, roundTripByLightness,
 } from '../lib/vizPlot.js'
 import { channelColor } from './viz/colors.js'
 import { labToRgb } from '../lib/rasterDecode.js'
@@ -219,7 +219,7 @@ function ProfileStatsSection({ bytes, t }) {
       {/* Dynamic, code-grounded description of the selected round-trip type. */}
       <p className={styles.typeDesc}>{t(typeMeta.descKey) || typeMeta.descFallback}</p>
 
-      <StatsTable gamut={gamut} rts={rts} type={type} t={t} />
+      <StatsTable gamut={gamut} rts={rts} type={type} bytes={bytes} intent={intent} t={t} />
     </Collapsible>
   )
 }
@@ -229,7 +229,7 @@ function ProfileStatsSection({ bytes, t }) {
 // crosses the WASM→JS boundary, so each field is shape-guarded before formatting —
 // a partial or malformed result degrades to '—' or a note, never a render-time
 // TypeError that would blank the whole Analysis tab.
-function StatsTable({ gamut, rts, type, t }) {
+function StatsTable({ gamut, rts, type, bytes, intent, t }) {
   if (gamut.loading || rts.loading) {
     return <div className={styles.loading}><span className={styles.spinner} /> {t('analysis_loading') || 'Analysing…'}</div>
   }
@@ -340,6 +340,8 @@ function StatsTable({ gamut, rts, type, t }) {
         </div>
       )}
 
+      <RtByLightness bytes={bytes} intent={intent} t={t} />
+
       {/* Worst-error colour: the device/PCS colour where this round trip fails hardest. */}
       {worst && (
         <p className={styles.rtGamut}>
@@ -350,6 +352,73 @@ function StatsTable({ gamut, rts, type, t }) {
   )
 }
 
+// ── Round-trip ΔE by quantized lightness ──────────────────────────────────────
+// The reference QC report's scatter: every seed drawn individually, grouped into
+// lightness bands, with the points inside each band marching from the gamut boundary
+// inward to neutral. The shape WITHIN a band is the content — error is largest on the
+// gamut surface and falls away toward the neutral axis — so this is one place where a
+// summary (an envelope, a mean line) would erase exactly what the plot is for.
+//
+// It carries its own seeding, so it follows the rendering intent but NOT the
+// round-trip type selector, and its numbers are not comparable with the table above:
+// it deliberately over-weights the gamut boundary, where inversion is hardest.
+function RtByLightness({ bytes, intent, t }) {
+  const [state, setState] = useState({ loading: true })
+
+  useEffect(() => {
+    let cancelled = false
+    setState({ loading: true })
+    roundTripByLightness(bytes, intent).then(
+      (r) => { if (!cancelled) setState({ loading: false, data: r }) },
+      (e) => { if (!cancelled) setState({ loading: false, error: e.message }) },
+    )
+    return () => { cancelled = true }
+  }, [bytes, intent])
+
+  if (state.loading) {
+    return <div className={styles.loading}><span className={styles.spinner} /> {t('analysis_loading') || 'Analysing…'}</div>
+  }
+  // A profile without the AToB/BToA pair (or one whose lightness range cannot be
+  // derived) simply doesn't get this plot; the section above still stands.
+  if (state.error || !state.data) return null
+
+  const d = state.data
+  const yHi = d.max > 0 ? d.max * 1.05 : 1
+  const bandW = (d.hiL - d.loL) / (d.levels - 1)
+
+  // Band separators as ONE series: vertical segments joined by null-y breaks, rather
+  // than 32 separate series cluttering the legend.
+  const sep = []
+  for (const L of d.levelL) { sep.push(L, 0, L, yHi, L, null) }
+
+  const graph = {
+    xAxis: { label: t('analysis_rt_band_xaxis') || 'L* (lightness of the requested colour)', min: d.loL, max: d.hiL + bandW },
+    yAxis: { label: t('analysis_rt_band_yaxis') || 'ΔE*ab', min: 0, max: yHi },
+    // Neither series is named: this plot shows no legend (one data series plus
+    // reference geometry), so a name would only be dead weight in the payload.
+    series: [
+      { id: 'sep', role: 'hint', shape: 'polyline', points: sep, color: '#b9bec7' },
+      {
+        id: 'de', role: 'primary', shape: 'scatter', color: '#2f6fb3', markerSize: 2,
+        points: d.x.flatMap((x, i) => [x, d.de[i]]),
+      },
+    ],
+  }
+
+  return (
+    <>
+      <p className={styles.rtGamut}>
+        <strong>{t('analysis_rt_band_heading') || 'Round-trip ΔE by lightness'}</strong>{' '}
+        <span className={styles.histCount}>
+          ({d.n.toLocaleString()} {t('analysis_rt_band_pts') || 'points, gamut boundary → neutral'};{' '}
+          {t('analysis_rt_mean') || 'Mean'} {d.mean.toFixed(3)} · P90 {d.p90.toFixed(3)} · {t('analysis_rt_max') || 'Max'} {d.max.toFixed(2)})
+        </span>
+      </p>
+      <PlotlyGraph graph={graph} storageKey="profiletool.rtBandHeight" defaultH={280} />
+    </>
+  )
+}
+
 function NeutralSection({ bytes, profileClass, tables, t }) {
   // Output (printer) profiles only — others get a localized error.
   const isOutput = String(profileClass || '').toLowerCase().includes('output')

frontend/public/wasm/iccplot.mjs

@@ -2,6 +2,7 @@
 import { lazy, Suspense, useEffect, useState } from 'react'
 import HeaderTable from './HeaderTable.jsx'
 import TagTable from './TagTable.jsx'
+import ErrorBoundary from './ErrorBoundary.jsx'
 import { preloadPlotly } from './viz/plotly.js'
 import { useT } from '../i18n.jsx'
 import { TAB_DEFS as TABS } from '../lib/tabs.js'
@@ -122,9 +123,22 @@ export default function ProfileViewer({
           </Suspense>
         )}
         {active === 'Analysis'   && (
-          <Suspense fallback={<div className={styles.loading}>{t('analysis_loading') || 'Analysing…'}</div>}>
-            <AnalysisPanel bytes={bytes} profileClass={data.header?.['Profile Class']} />
-          </Suspense>
+          /* Boundaried: the Analysis tab hosts several independent analyses, each
+             fed by untrusted profile data through a WASM engine. Without this, one
+             bad render takes the WHOLE React tree down and the browser tab goes
+             blank — no message, nothing to act on. Reset when the profile changes. */
+          <ErrorBoundary
+            resetKey={bytes}
+            fallback={(err) => (
+              <div className={styles.panelError}>
+                <strong>{t('error_label')}</strong> {err?.message || String(err)}
+              </div>
+            )}
+          >
+            <Suspense fallback={<div className={styles.loading}>{t('analysis_loading') || 'Analysing…'}</div>}>
+              <AnalysisPanel bytes={bytes} profileClass={data.header?.['Profile Class']} />
+            </Suspense>
+          </ErrorBoundary>
         )}
         {active === 'XML'        && (
           <Suspense fallback={<div className={styles.loading}>{t('loading_xml_editor')}</div>}>

frontend/public/wasm/iccplot.wasm

@@ -45,6 +45,20 @@
   line-height: 1.4;
 }
 .partialIcon { font-size: 16px; flex-shrink: 0; }
+
+/* Fallback shown when a tab's panel throws during render (see the ErrorBoundary
+   around the Analysis tab). Inset rather than full-bleed like .partialBanner,
+   because it replaces the panel's content rather than heading the whole viewer. */
+.panelError {
+  margin: 16px;
+  padding: 12px 16px;
+  border-radius: var(--radius);
+  background: var(--error-tint);
+  border: 1px solid #f4cccc;
+  color: var(--error);
+  font-size: 12.5px;
+  line-height: 1.5;
+}
 :global(body.dark) .partialBanner { border-bottom-color: #5a2424; }
 
 .validBadge {