CVD-safe default categorical palette

[FarhanAliRaza]

Summary

Replaces the Tableau10 default categorical palette with an eight-slot CVD-safe palette (§20/§36 default theme).

Why: Tableau10's adjacent red/green pair collapsed to ΔE 1.2 under deuteranopia — effectively indistinguishable — and slots 1/5/7/9/10 sat below the chroma floor.

The new palette keeps every step in the OKLCH lightness band shared by light and dark surfaces (L 0.48–0.67) and validates against both reference surfaces (#fcfcfb / #1a1a19):

• worst adjacent-pair CVD ΔE 8.5 (Machado–Oliveira–Fernandes protan/deutan, ≥8 target)
• worst adjacent normal-vision ΔE 19.1 (≥15 floor)
• chroma ≥ 0.10 and ≥3:1 contrast for every slot on both surfaces

The slot order is the safety mechanism — adjacency drives the ΔE gate — so the config comment now warns against re-ordering or extending without re-running the validator.

Also in this PR

• Updated the hardcoded segment colors in the Reflex demo app and the visual-regression chrome shell to the new slots.
• Regenerated examples/reflex/assets/charts/*.html, which also picks up the current embedded render client (colormap stops, colorbar chrome CSS) from recent main commits.

Checks

• pre-commit run --all-files, ruff check, ruff format --check all pass.

[codspeed-hq]

Merging this PR will not alter performance

✅ 94 untouched benchmarks
⏩ 1 skipped benchmark¹

---

_{Comparing cvd-safe-default-palette (5701f77) with main (8ace307)}

Footnotes

1. 1 benchmark was skipped, so the baseline result was used instead. If it was deleted from the codebase, click here and archive it to remove it from the performance reports. ↩

[Alek99]

Reviewed — the palette itself is solid, but the committed demo assets don't match the PR description and actively mislabel colors.

Blocking: the regenerated examples/reflex/assets/charts/*.html files are missing from the commit

The PR body says these assets were regenerated, but the diff only touches 3 files and the committed assets still embed the old Tableau10 colors — custom_chrome.html's spec has "palette":["#4c78a8","#f58518","#54a24b"] and density_scatter.html's density overlay has "color":"#4c78a8". Meanwhile the Reflex demo app's custom legend WAS updated to the new hexes, so the swatch-to-mark mapping is now actively wrong: the legend shows a green #008300 swatch for "Growth" while Growth points render orange #f58518, and magenta #d55181 for "Self serve" while those points render green — worse than before the PR. The visual-regression chrome shell (scripts/visual_regression_smoke.py) wraps the same stale asset, so the mismatch also gets baked into regression screenshots, and the DOM probe only checks element visibility so CI won't catch it.

Fix: run examples/reflex/scripts/build_charts.py and commit the regenerated assets.

Non-blocking:

• The new comment block in python/xy/config.py points at .claude/skills/xy-dataviz/scripts/validate_palette.py, which doesn't exist in the repo — the "re-run the validator before reordering" instruction is unactionable. Either commit the validator or drop the reference.
• tests/test_scatter_matrix.py:107 comment still says the base palette has 10 entries; it's 8 now (assertions are length-agnostic and still pass).

The palette design itself checks out — I independently reproduced the CVD claims (Machado protan/deutan simulation + Lab ΔE76): the old palette does collapse (#54a24b → #e45756 at deutan ΔE76 ≈ 1.7, and 5 slots below 3:1 contrast on light surfaces), while the new one holds worst-case adjacent CVD ΔE76 ≈ 27.6 with contrast 3.05–5.57 on both surface colors. And every palette consumer (Python payload/raster/SVG, JS renderer via the wire palette) indexes modulo length, so shrinking 10→8 is safe; 148 related tests pass.