Fix defects found in a full simulation review

[veillette]

Reviewed the simulation end to end — physics model, state handling, views, i18n, accessibility, tests, build — and fixed everything found. Every change was verified against a real production build in Chromium (WebGL via SWiftShader) at 1024×700, 1280×800 and 1600×900, not only in unit tests.

The physics core came out clean and was verified independently rather than taken on trust: phase is continuous at both slab interfaces for all n, w and both travel directions; the quarter-wave preset genuinely converts 45° linear light to circular (|E| after the plate is constant at 5.000 across a period, versus 0 → 7.07 before it); the half-wave plate mirrors the polarization plane (135° → 45°); B = k̂ × E is perpendicular for every polarization and flips with direction. No changes were needed there.

Scene framing behind the control panels

The three.js canvas spans the whole window and the Scenery panels are painted over it, so a screen's scene only clears its controls if it is shifted left and zoomed out enough to fit beside them. On the three two-column screens it was not: roughly 30% of the play area sat behind the panels, including the right transverse frame and the right-end E-vector arrow — exactly the transmitted light those screens exist to show. This also contradicted viewOffset's own comment ("clearing the right-hand panel stack").

viewOffset alone could not fix it — clearing the panels at the old zoom would have needed a half-width of ~200 px against the ~420 px actually drawn. Screens now declare a panelLayout ("singleColumn" | "twoColumn") and WaveScreenView picks the matching projection offset and starting camera range from one SCENE_FRAMING table; WaveSceneCamera takes the starting range as a constructor argument so Reset All returns to it. The widest window is the tightest fit, because ScreenView scaling grows the panels faster than the extra width it buys, so the two-column range is set from that case. Intro is unchanged.

Wave Plates benefits most: the whole plate, both end frames, the incoming linear light and the outgoing circular light are now visible together.

Lab permalinks lost the preset selection

queryStringFromState() never emits preset=, but stateFromQueryParameters() only compared the state against the preset named in the URL. Every link produced by "Copy link" therefore reopened with the physics correct and the "Show me" menu reading Custom — 19 of the 20 presets affected.

The selection is now recovered by matching the clamped state against every preset. The 20 preset states are mutually distinct, and a new test asserts that so the match stays unambiguous. The round-trip test previously covered 3 presets and only asserted state, which is why this passed; it now covers all 20 and asserts the selection too.

Accessibility

The material insert checkbox announced a fixed "Insert material" while its visible label is the panel title ("Polarizer (dichroic filter)", "Wave plate (birefringent slab)"), so a speech-input user saying "Polarizer" could not activate it — WCAG 2.5.3 Label in Name. The accessible name now embeds the visible title via a new controls.material.insertPattern string in all three locales. Confirmed in the DOM on each screen, e.g. "Insert: Polarizer (dichroic filter)".

Smaller fixes

• Locale key-parity guard skipped Spanish. StringManager asserted en ↔ fr only, so a Spanish file missing a key would have compiled and thrown at runtime. The locales are in parity today (124 keys each); the added pair keeps them that way.
• Shared scratch vector. WaveDisplayNode used one module-level Vector3 as both the camera position and the arrow direction — correct only because of the order render() happened to call them in. Separate scratches now.
• "Copy link" confirmation timer did not cancel its pending timeout, so a second click had its label cleared early by the first click's timer. Each copy restarts the two seconds.
• "Custom" was offered in the preset menu even though selecting it does nothing — it reports a state the model sets after a manual edit. It is now listed only while it is the selection, in the visible list and in the parallel DOM alike (the option element is hidden / display: none when a preset is selected).
• Signed retardation readout showed "−90°" whenever the horizontal index was the larger one. A plate is specified by how much retardation it introduces, and the "Fast axis" readout already states the direction in words, so the readouts quote the magnitude; Δφ itself keeps its sign. The degrees rounding moved onto the model so the panel and the accessible summary cannot disagree.
• Intro's panel stack was the only one without maxHeight, so a longer locale could push it past the screen edge.
• Absorption preference is global but only does anything where a material absorbs, so its description now names the screens it affects, in all three locales.
• Fuzz spec lint warnings — the three useExplicitType warnings were the only noise left in npm run lint; annotating the constants makes the whole gate silent.

Verification

npm run check, npm run lint and npm run build are clean, with lint now completely warning-free.

69 tests pass, up from 61. New coverage: all 20 presets round-tripping through the permalink mapping with their selection asserted, preset-state distinctness, and a new tests/WavePlatesModel.test.ts covering retardation, its displayed magnitude (including the negative-Δφ case that is awkward to reach in a browser) and the exact QWP/HWP presets.

Browser checks across all four screens plus the permalink, clipboard and accessible-name paths reported 0 console errors.

CLAUDE.md and doc/implementation-notes.md are updated for the new test suite and the per-screen framing decision.

Reviewer notes

• The Lab's insert checkbox now announces "Insert: Material", which is slightly stiff where the title is already the generic "Material". I kept the shared pattern for consistency with the other two screens rather than special-casing Lab — easy to change if you would rather it read "Insert material" there.
• CAMERA_RANGE_TWO_COLUMN = 29 was tuned empirically against screenshots at the three window sizes above. It is a visual-design number; if you want the wave larger and are willing to accept less clearance, that is the one value to move.

---

Generated by Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None