refactor(color): emit colors as authored, drop channel companions

[tenphi]

What

Two things existed only to serve the opacity suffix back when it needed numeric channels to write an alpha into:

1. Channel companions — a #name token declared a second variable beside --name-color holding its channels decomposed (--brand-color-oklch: 0.75 0.16 55) plus a matching @property rule, and color emitted --current-color-{space} beside --current-color.
2. Color-space conversion — a token's value was rewritten into the configured colorSpace, so #brand: '#ff8800' declared --brand-color: oklch(0.75 0.16 55).

#268 moved opacity to relative color syntax (oklch(from var(--brand-color) l c h / .5)), which has the browser read the channels off whatever the value resolves to. Neither is load-bearing any more. Both are removed, and a color now passes through untouched: --brand-color: #ff8800.

configure({ colorSpace }) is deprecated — still accepted, inert, warns in development, removed in the next major.

No public contract moves. Nothing exported changes, public-api.md is untouched, and class name hashes are identical (hashing is over the input style object). What changes is the emitted CSS.

Migration

In Color space and in the changeset:

/* before */
color: oklch(var(--brand-color-oklch) / 0.5);
background: oklch(var(--brand-color-oklch));

/* after */
color: oklch(from var(--brand-color) l c h / 0.5);
background: var(--brand-color);

Relative color syntax is strictly more capable: the companion could only carry channels for a color the engine could evaluate at build time, while from <color> works on every <color> — a color-mix(), a light-dark(), a variable Tasty never defined. If you relied on colorSpace for uniform output, author the token in the space you want it emitted in.

Size

Measured against main:

	before	after
main	58.10 kB	56.22 kB	−1.88
core	55.17 kB	53.25 kB	−1.92
static	18.94 kB	16.42 kB	−2.52
zero	33.27 kB	30.81 kB	−2.46
babel-plugin	51.51 kB	48.95 kB	−2.56

Limits tightened to ~0.5 kB headroom rather than left slack. Also: one declaration less per color in every rule, one @property registration less per color token on all four render paths, a pre-pass removed from PropertyTypeResolver.scanDeclarations, and the LRU that memoized the sRGB round-trip is gone. color-space.ts went 706 → 86 lines and is now only about how an alpha is applied; color-math.ts lost 234.

Bugs found along the way

• parseColor rejected most named colors. Its first-character dispatch answered false for any named color starting with a letter it had a shape for, so red, hotpink, lime, orange, violet, coral and teal were rejected and warned about, while blue and green were accepted. The conversion had been masking it. A miss now falls through to the named-color lookup.
• A spurious unable to parse color warning for a token set to a bare keyword, fixed by the same change.
• Silent white on misscaled color-function channels. okhsl(280 80 52) — the % forgotten — clamped to full saturation and rendered white. Now warns once per function in development. See below.

Refactoring

createStyle's color branch went from five outcomes to one expression. Two invariants made that safe, both now pinned by a test rather than assumed: parseColor derives name from the chain it resolved (so name implies color), and a value the converter could handle never resolved to a var() chain. parseSameSpaceFunc/buildSameSpaceString merged, then went entirely with the conversion. processTokens' two branches converged into one resolver. Gone: strToColorSpace, resolveToRgbaValues, colorFuncName, getColorSpaceSuffix, getColorSpaceComponents, convertColorChainToComponentChain, colorInitialValueToComponents, getComponentPropertySyntax, getDefaultComponents, FormatPropertyOptions, rgbToHsl, rgbToOklch, hexToRgbaValues, and the unused isColor on getEffectiveDefinition's result.

Test-only code, enforced

color-math keeps the reverse direction of two conversions the engine never calls — sRGB→OKHSL and lightness→OKHST tone — so the forward ones, which the okhsl()/okhst() plugins use, can be round-tripped in tests rather than pinned to fixtures the implementation itself produced.

Rolldown already drops them (their constants survive only in a source map), but that was true by accident, and knip can't notice if it stops being true — knip.json lists test files as entry points, so anything a test imports reads as used. scripts/check-test-only-code.mjs makes it explicit: it carries the reverse-only closure and fails if any of it reaches an emitted file, and asserts each registered name is still declared in src so it can't silently rot. Validated three ways — passes clean, fails on a forced production re-export (naming the whole closure), fails on a stale entry. Wired into CI.

Checked against tenphi/glaze#94

The producer/writer scale mismatch fixed there does not apply here — createColorFunc already takes channel factors and scales to percentages on output, which is what that PR changed glaze's writers to do. Verified every other boundary too: hslStringToRgb, oklchStringToRgb, strToRgb, and okhstToSrgb's fromTone(t * 100) bridge between the 0-1 lightness scale and the 0-100 tone scale all convert correctly.

One sibling of it was here: the silent clamp described above. Fixed, with mutation-checked tests.

Tests

2067 passing. Coverage moved with the behavior rather than being deleted alongside the functions:

• Cross-space conversion and hue units (deg/turn/rad, negative, >360) → onto strToRgb, still public and still converting.
• Plugin color-function resolution → onto resolveFunctionColor, the generic path that replaced the okhsl branches.
• colorSpace's inertness pinned across all three values; createStyle's full value table and --current-color republishing (including the #current self-reference guard, previously untested) pinned directly.
• Every surviving mention of a companion in the repo is a negative assertion.

Verification

pnpm test 2067 passing · pnpm hygiene + pnpm knip + pnpm check:test-only clean · pnpm build && pnpm size under tightened limits · public-api.md unchanged · snapshot diffs reviewed line by line.

🤖 Generated with Claude Code

[github-actions]

📦 Snapshot release

Published 0.0.0-snapshot.d5c492a.

pnpm add @tenphi/tasty@0.0.0-snapshot.d5c492a

[tenphi]

[P2] Remove the companion promise from public typings

The exported token JSDoc still says color tokens create --name-color-{colorSpace}, although this PR removes that property from every rendering path. The stale contract appears in src/types.ts:113, src/config.ts:414, src/styles/types.ts:742, and src/plugins/types.ts:74, and is emitted into the published declarations. Editor tooltips will therefore direct users toward an undefined CSS property. Please update all four public type comments to say that #name creates only --name-color.

[tenphi]

[P2] Remove the companion promise from public typings

Fixed in 030cffc — all four updated to say #name creates only --name-color.

You were right that it reaches published declarations, so I verified against the built output rather than just the source: no emitted .d.ts mentions the companion any more, and the corrected line is present in dist/config-*.d.ts.

Worth recording why it slipped: my sweep grepped for the resolved suffixes (-color-oklch, -color-rgb, -color-hsl), so the templated --name-color-{colorSpace} matched nothing. Re-swept on colorSpace and color-{ across all of src, which also caught a stale "with no companion variable" aside in overrideColorAlpha's doc comment.