Two token tiers: private ramps, public semantics

Tier one is OKLCH ramps (--jolly-neutral-*, --jolly-accent-*, danger/warning/success), never read by a component. Tier two is roughly thirty semantic aliases, the only names components use and the only names consumers override; light and dark select ramp indices through var() rather than inlining hand-picked pairs. Ramp privacy is conventional, not enforced.

The accent is seeded from #4488ff (used 24 times across voxel-map and voxel-model) and is split in two, because one value cannot do both jobs: #3a6fc2 measures 3.5:1 as text on #131b24 but 5.0:1 beneath white. Peer colours rotate hue on the golden angle at fixed lightness and chroma, so any number of peers stay mutually distinguishable and equally legible.

Considered Options

• A single flat tier. Thirteen names do not stretch to 45 components with eight states, and light and dark drift apart.
• A third tier of per-component tokens. About 180 public names to document and keep stable, for a library maintained by a small team.
• Hand-authored hex ramps. Perceptual spacing drifts between steps, and every peer colour needs hand checking.
• One accent token for fill and text. Measured above; it cannot satisfy both.
• Seeding the accent from pixel-art's blue. It would shift a blue used four times more often.