Semantic button colors

[aksOps]

Question

Dogfood finding (user, 2026-08-20): the buttons in the card-detail popup are visually identical - all render on the same Raised surface. The web UI colored buttons by meaning; that is the bar.

Scope:

• widget.Button gains a semantic variant (Primary / Success / Danger / Neutral, or equivalent) backed by theme tokens with focused/hovered/blurred states per variant, honest at 256 colors (audit the new slots like the palette's existing ones).
• Apply consistently across every button surface: card-detail action row (Edit=primary, Ship/Restore=success, Delete=danger, Close=neutral), editor and settings (Save=primary, Cancel=neutral, Remove/Delete=danger, Test=neutral), ship-guard and kill dialogs (Ship anyway=danger or warn, Tick everything=success, Cancel=neutral - pick per spec judgment and record), huh Confirm restyled to match.
• Amend the spec (design/tui-spec branch) with the variant tokens and the assignment table.
• Regenerate affected goldens. Interaction FROZEN; coverage gates hold.

Part of #136.

[aksOps]

Shipped in #161 (c5073eb).

widget.Button carries a semantic Variant — Neutral, Primary, Success, Danger — backed by theme tokens with per-variant blurred, hovered and focused states, all built once by the style factory. The hue carries the meaning and the state carries the elevation: a blurred button wears its variant as a tint on the resting surface, a hovered one wears the tint as a fill, and a focused one wears the saturated hue. Neutral is the zero value, so a caller that states no meaning gets the calmest surface rather than an accidental accent.

Tokens. Four new palette slots: StatusAlarm #b31f14 (124), TintPrimary #a8b6ff (147), TintSuccess #7fe0b0 (115), TintDanger #ffa7a0 (217). All four indices were previously free, so the real-collision set of the 256 audit is still empty and the alias set is unchanged.

Armed is FgBase bold on StatusAlarm for every variant. It needed its own fill: a purge button arms from the Danger variant, so an armed button wearing the focused-Danger fill would be the one state a user must not misread.

Contrast. A contrast audit joins the 256 audit in theme/audit_test.go — every variant, every state, in truecolor and again after quantization, against the WCAG AA floor of 4.5:1.

Variant	Blurred	Hovered	Focused	Armed
Neutral	8.61 / 8.39	5.82 / 5.50	7.76 / 7.88	5.51 / 6.41
Primary	5.41 / 4.82	9.93 / 9.28	6.02 / 5.71	5.51 / 6.41
Success	6.60 / 5.74	12.13 / 11.03	8.25 / 7.11	5.51 / 6.41
Danger	5.65 / 5.57	10.38 / 10.71	6.26 / 6.29	5.51 / 6.41

(truecolor / x256; worst cell 4.82.) A second test keeps the four variants separable at 256 colors, so the finding this closes cannot come back on the terminals least able to afford it.

Assignment. Card detail: Edit primary; Check / Restore success; Kill / Del / Unlink / Purge danger; Drift / Comment / Link / Close neutral; Save comment, Add link, Check selected, Update baseline primary; Delete / Confirm delete danger. Editor: Save card primary, the rest neutral. Settings: Save* primary, Test* and + Add integration neutral, Remove danger. Ship guard: Tick everything success, Ship anyway danger, Cancel neutral. Kill prompt: both kill choices danger. Purge: danger. ADR split: Propose stories and Add selected (N) primary. Issue import: Import primary.

Recorded judgment calls: Ship anyway is danger rather than warn — it overrides a guard the board raised, and the set spends no fifth token family on one button; Unlink is danger (a delete with a softer name) while Link stays neutral (it opens a picker); the AI actions are neutral because a pane has one primary affirmative and it is always Save.

huh's Confirm carries the Neutral pair. That is a limitation, not a preference: huh exposes one button pair per Confirm while the pair spans two meanings (Cancel and Ship anyway), so painting it with either would lie about the other. The hued treatment lives in the ButtonGroup form of the same guard.

Spec amended on design/tui-spec (12c8b9b): new §1.9 for the variant tokens, state matrix and contrast audit, new §5.4 for the assignment table and the judgment calls, plus §1.5, §1.7, §5.1, §5.2 and §6.1.

Interaction frozen; only the four affected color goldens changed. Coverage 96.9% total, every package at or above 95%, internal/tui at 96.6%.