A palette that claimed to be colour-blind safe and was not

The problem

Every client demo needs charts, tables, a plausible dataset and a brand that looks like theirs. Rebuilding that each time is waste, so this is built once and reskinned.

One accent colour generates every ramp, and a chart theme adapter carries it to all 37 charts. Reskinning is editing one file.

Keeping the system from rotting

A design system with no enforcement decays the first time someone is in a hurry. Application code may only use semantic tokens, and guard:colors fails the build on a raw palette class. Not a lint warning, a failure.

The mistake worth publishing

The categorical chart palette stops at seven colours on purpose. An eighth collapses against the warm members under simulated deuteranopia.

I know that because the palette that shipped first was wrong. Two of its colours were 0.017 apart in perceptual distance under deuteranopia simulation, while a comment directly above them asserted the set was colour-blind safe. It looked fine to me, which is exactly the problem: I am not the person it fails for.

There is now a measurement script and a regression test, so the claim is checked rather than asserted and nobody can re-open it by eye. This is the thing I would want to be judged on out of everything on this site, because the failure was invisible and the fix is permanent.