feat: admit Select

[inureyes]

Admits Select, proposed in #16.

The proposal asked two questions. Both are answered here rather than left open, and both are reversible if a maintainer disagrees: this is additive, and no existing component changes.

Is it wanted? Yes, and rule 2 is the reason. backend.ai-go and continuum-hub both ship this component, with the same SelectOption shape and the same core props. That is what the second-consumer test asks for, and it is the part that usually fails: the same check disqualified MultiSelect, which exists in one product, and TextInput, where the two products share a directory name and nothing else.

Superset or intersection? The superset: the label, the search filter, invalid, and aria-describedby come too. An intersection-only admission would leave the hub wrapping this package to re-add them, which is the fork the admission exists to end. The extra props are generic view models rather than one product's categories, and backend.ai-go reads them as additive, so it can adopt without touching a call site.

What changed on the way in

The locale key. The empty-search line called t("common.select.noOptions"). It is noOptionsLabel now, with an English default, the shape closeLabel and moreTabsLabel already use. That was the one thing the proposal named as blocking rule 1.

The focus border. .select__trigger:focus and .select__search-input:focus-visible painted their border from the bare --token-colorPrimary. That accent measures 2.61:1 against white in orange-light, under the 3:1 that WCAG 2.2 SC 1.4.11 sets for a focus indicator. They mix the accent with the text colour now, which is the form CONTRIBUTING documents for a border-drawn ring. This package's own guard caught it on the first test run, which is a fair advertisement for the guard: the component had shipped in two products with that border.

Nothing else in the CSS differs from the hub's copy. The diff is those two declarations.

Three tokens. --token-colorTextDisabled, --token-colorTextPlaceholder and --token-colorPrimaryBgHover were read but undeclared, which the token contract guard refuses. The first two are #767676 rather than the customary #bfbfbf, which measures 2.2:1 on white: a placeholder naming a field and a disabled value explaining why it cannot be changed are both text a reader is expected to read.

Tests

Fifteen, where both products had none. That is rule 3, and it is not paperwork: the point of sharing a component is being able to change it afterwards.

They cover the selected value and the open state, aria-selected on the right option, selection by click and by Enter, arrow-key movement through aria-activedescendant, Escape closing without choosing, a disabled option staying visible and unselectable, the search filter over both label and description, the empty-search line in English and in a caller's own words, the label naming the trigger, invalid, aria-describedby, and onBlur.

Three perturbations, each verified: hardcoding the empty label fails one, dropping aria-invalid fails one, ignoring option.disabled fails one.

One shim in the test setup: jsdom implements no layout and therefore no scrollIntoView, which the component calls to keep the active option visible.

pnpm run verify passes: 535 tests, 140 files packed.

Not in this pull request

The hub still has its copy; retiring it is lablup/continuum-hub#1001's half, and it needs this released first. --token-controlLabelOffset, held back in #15 for want of a labelled control, is now worth revisiting, and still depends on --token-lineHeightSM, which this package does not declare.

Refs #16