feat: rewrite DESIGN.md against D66 and ship the rendered reference (#36)

[thomasluizon]

Ticket 0 of the redesign. Rewrites DESIGN.md against D66 and ships design/reference.html, the page that outranks it under D42. Closes nothing: #36 stays open through both D13 human grants.

Ticket: https://github.com/thomasluizon/orbit-tickets/issues/36 (migrated from ORB-30)

What this PR is

file	change
DESIGN.md	rewritten against D66. One scheme, space-blue accent, Geist trio, near-black canvas
design/reference.html	new. The rendered reference, 305 KB, zero network requests
design/handoff/**	deleted, 40 files
.claude/agents/design-reviewer.md	AI-slop list brought level with the spec, stale notes corrected
.github/workflows/redesign-drift.yml	new. D36 drift control for redesign/main

Base is redesign/main, cut fresh from main for this PR (D36).

The tokens are derived, not picked

Every value was computed in OKLCH against the canvas and measured. The derivation forced four findings that were not in the ticket:

1. White fails on the new accent at 2.41:1. So --fg-on-primary is the canvas ink #020618 (8.35:1), not white. The brand colour is not darkened to rescue white; the thing sitting on it changes instead.
2. --status-frozen collided with the accent. The old #00D3F3 sits 12.0 degrees from #24B4E2, inside the 15 degree band where two hues read as one colour, so streak-freeze and done would have looked identical. Frozen is retired as a hue and renders as a neutral chip plus the snowflake glyph. Thomas decided this in session.
3. --fg-4 failed the 3:1 non-text floor at 2.90:1 and is raised to #5D5D60 (3.03:1).
4. The surface ladder was too tight at the dark end. 0.04 / 0.05 / 0.06 / 0.06 resolved to #151517 against #171719, which is not a step. Widened to 0.04 / 0.06 / 0.08 / 0.12.

The overlay library, per platform

The acceptance criteria required a named library with a rationale, and no candidate existed anywhere in the vault.

• Web: @base-ui/react 1.7.0 (MIT, published 2026-08-04, react ^19 peer, ./dialog subpath). Read live from the registry. Note @base-ui-components/react is the deprecated old name frozen at 1.0.0-rc.0; the registry message is "Package was renamed to @base-ui/react".
• Mobile: @lodev09/react-native-true-sheet 3.11.3, already installed. The library is not the defect, the wrapper is.

The web root cause is visible in the markup: app-overlay.tsx renders <dialog open> rather than calling showModal(), so it gets none of the platform's free behaviour and hand-rolls the focus trap, the body-scroll lock, the portal and Escape, then needs z-[9999] because it never enters the top layer. That z-[9999] violates this repo's own no-arbitrary-zindex ban. Its grabber also has no gesture handler, so web drag-to-dismiss is decorative today and nothing is lost by not replacing it.

Migrating the 83 callers stays R-ticket work, per Out of scope.

Assumptions

Mechanical calls made during the rewrite. None is a Thomas-owned decision; each is sourced.

1. Icon size moves from 22 to Tabler's native 24, and stroke from 1.8/2.2 to the weight-matched 1.5/2/2.5. Tabler's own source is width="24" height="24" stroke-width="2"; an off-grid 22 renders with fractional scaling and looks soft.
2. OKLCH is the derivation space, hex and rgba() the authored notation. @react-native/normalize-colors has no oklch branch, read from the installed package, so an oklch() token is a runtime failure on Android.
3. The overlay panel is authored opaque. A translucent panel over arbitrary content cannot have its text contrast checked.
4. The token vocabulary is not renamed. better-colors prescribes text/border/accent over fg/hairline/primary, but its own first principle says match the project's notation, and the --color-primary versus --color-text-primary collision it guards against does not exist here.
5. prefers-contrast: more is a token-override layer, not a third variant, so D66's "two variants, not twelve" holds.

Manual steps

1. Open design/reference.html from the filesystem and approve the direction. This is human grant 1 (D13). Four controls at the top are the decisions: accent (three candidates), serif on/off, spacing scale A/B, plus light/dark. XP and levels are rendered both ways in section 01.
2. @date-fns/tz ^1.2.0 is a peer of @base-ui/react and is not yet in apps/web. It installs with the R1 overlay ticket, not here.
3. redesign/main now exists on the remote. It has no branch protection; decide whether it should have any.

Verification

Every test scenario the ticket names, run against the committed files:

http://  https://  <link  <script src  @import      0 hits each
em dash / en dash in DESIGN.md and the page         0 / 0
raw hex or rgb outside the single :root block       0 / 0
undeclared var() references                         0 of 120 used
markup errors                                       0
node tools/check-dashes.mjs --files <changed>       exit 0
node tools/check-dashes.mjs --check-baseline        exit 0
node tools/check-copy.mjs --check                   exit 0

Fonts are embedded as data: URIs: 8 latin woff2 faces, 195 KB binary, so the page renders Geist, Space Grotesk, Geist Mono and Instrument Serif with no network.

What is deliberately not here

The 31-rule eslint-rules/ re-derivation, the spacing-scale codification, and the nine warn to error flips. D66 decision 12 re-decides the spacing step values against the rendered reference, and several of those rules encode them, so writing them before grant 1 would bake in values Thomas has not chosen. They land in the next PR on redesign/main.

The 86-rule and per-input verdicts are posted as a comment on the ticket rather than carried in DESIGN.md. DESIGN.md is the guidance an agent reads to build a screen; which external skill was rejected and why does not change how a surface gets built.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
orbit-ui-mobile-web	Ignored		Aug 15, 2026 4:40pm

[thomasluizon]

@pullfrog review

[pullfrog]

Important

The D36 drift workflow cannot observe pushes to main from its current branch, so the synchronization this PR promises is inert until the dispatcher exists on main.

Reviewed changes Reviewed the full design-authority replacement, rendered reference, legacy cleanup, reviewer guidance, and drift automation in this initial review.

• Design authority: Rewrites DESIGN.md around the new near-black, space-blue token system, layout rules, component states, overlays, and accessibility guidance.
• Rendered reference: Adds a self-contained design/reference.html with embedded fonts, four live visual controls, token samples, responsive shell examples, primitives, overlays, and listing states.
• Legacy handoff removal: Deletes the 40-file vendored design/handoff/** source and its assets.
• Reviewer guidance: Aligns .claude/agents/design-reviewer.md with the new design constraints and staged gates.
• Drift automation: Adds .github/workflows/redesign-drift.yml to merge future main changes into redesign/main.

⚠️ The reference clips on phone-width viewports

At a 375px viewport, the document grows to 436px while the visible client is 360px, and the right edge of the mobile-shell sample is clipped unless the reader scrolls horizontally. That makes the rendered authority awkward to inspect on the same class of viewport it is demonstrating.

Technical details

# Keep the rendered reference inside narrow viewports

## Affected sites
- `design/reference.html:280` — `.shell412` has a fixed 412px width and cannot shrink within the page's padded wrapper.
- `design/reference.html:532-543` — the fixed-width sample is rendered directly in the wrapping shell layout.

## Required outcome
- The reference page must avoid document-level horizontal overflow on narrow phones while retaining a true 412px sample whenever enough width is available.

## Suggested approach
- Constrain the sample with `max-width: 100%` and let its inner content adapt below 412px.

ℹ️ Deleted handoff paths remain allowlisted

The PR deletes design/handoff/**, but all 25 allowances for that tree remain in tools/dash-baseline.json. The current checker only rejects growth, so its passing result does not reveal these stale exemptions and the documented shrink-only ratchet remains artificially inflated.

Technical details

# Shrink the dash baseline with the deleted tree

## Affected sites
- `tools/dash-baseline.json:87-111` — 25 entries still allow violations in paths removed by this PR.
- `tools/check-dashes.mjs:149-157` — `--check-baseline` compares only current counts above the baseline and does not reject stale keys.

## Required outcome
- Remove baseline entries for deleted files so the committed baseline represents only extant debt.

## Suggested approach
- Regenerate with `node tools/check-dashes.mjs --write-baseline` after the handoff deletion, then commit the shrink.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[pullfrog]

Important

The integration branch currently bypasses CI, its drift workflow cannot trigger from where it lands, and the approval reference misrenders candidate choices.

Reviewed changes across the D66 design-canon rewrite, rendered decision artifact, legacy handoff removal, reviewer guidance, and redesign-branch automation.

• Design canon — rewrites DESIGN.md around the single space-blue scheme, derived tokens, component rules, accessibility, motion, overlays, and enforcement.
• Rendered reference — adds the offline design/reference.html decision page with mode, accent, serif, and spacing controls plus representative primitives and layouts.
• Legacy retirement — removes the 40-file handoff prototype and updates the design-review agent for the new direction.
• Integration automation — adds a workflow intended to merge each main advance into redesign/main.

⚠️ The integration branch bypasses the repository gates

Every existing PR workflow still filters its base to main, so PRs targeting the new shared redesign/main branch do not run Build, Lint, Type Check, Unit Tests, Guards, architecture drift, or the other normal verification. PR 735's actual rollup confirms the gap: only GitGuardian and Vercel registered, which lets each R-ticket enter the long-lived integration branch without the checks that normally stop regressions.

Technical details

# Run the normal gates on redesign PRs

## Affected sites
- [`test.yml:3-6`](https://github.com/thomasluizon/orbit-ui-mobile/blob/b9ef378d05c9ce784cadddf7b58d54bed284e199/.github/workflows/test.yml#L3-L6) — the build, lint, type-check, and unit-test workflow accepts only PRs targeting `main`.
- [`guards.yml:14-23`](https://github.com/thomasluizon/orbit-ui-mobile/blob/b9ef378d05c9ce784cadddf7b58d54bed284e199/.github/workflows/guards.yml#L14-L23) — the deterministic repository gates also accept only PRs targeting `main`.
- `.github/workflows/{arch-map,dependency-review,mutation,perf,react-doctor,sonarcloud,visual}.yml` — each has the same `pull_request.branches: [main]` filter.
- PR 735 check rollup at `b9ef378` — no repository build, test, or guard jobs were created.

## Required outcome
- Every R-ticket PR targeting `redesign/main` receives the same applicable verification before it enters the shared integration branch.
- Treat branch protection as a separate decision; the workflows must first create checks for this base branch at all.

⚠️ The approval page does not faithfully render two decisions

The page that outranks DESIGN.md lets dark accent rules override the light palette and keeps the tint pinned to candidate A, while spacing candidate B omits its documented 48px and 64px steps. Human grant 1 could therefore approve an option from a rendering that is neither contrast-correct nor complete.

Technical details

# Make every approval control render its declared candidate

## Affected sites
- [`design/reference.html:129-165`](https://github.com/thomasluizon/orbit-ui-mobile/blob/b9ef378d05c9ce784cadddf7b58d54bed284e199/design/reference.html#L129-L165) — the later `[data-accent]` blocks override `[data-mode="light"]`. In a real browser, light + calm computed `--primary-soft: #43B3DB` at 2.31:1 on `#FAFAFA`; light + B computed `#0080A4` at 4.35:1. Those miss the page's stated 4.5:1 text floor, and calm also misses the 3:1 graphic floor.
- [`design/reference.html:93`](https://github.com/thomasluizon/orbit-ui-mobile/blob/b9ef378d05c9ce784cadddf7b58d54bed284e199/design/reference.html#L93) — `--primary-dim` remains `rgba(36,180,226,0.14)` for every mode and accent even though each control updates `--primary-rgb`.
- [`design/reference.html:166-170`](https://github.com/thomasluizon/orbit-ui-mobile/blob/b9ef378d05c9ce784cadddf7b58d54bed284e199/design/reference.html#L166-L170) and [`426-437`](https://github.com/thomasluizon/orbit-ui-mobile/blob/b9ef378d05c9ce784cadddf7b58d54bed284e199/design/reference.html#L426-L437) — candidate B renders `4 8 12 16 20 24 32 40 56`, omitting 48 and 64 from the twelve-step candidate specified in [`DESIGN.md:69-72`](https://github.com/thomasluizon/orbit-ui-mobile/blob/b9ef378d05c9ce784cadddf7b58d54bed284e199/DESIGN.md#L69-L72).

## Required outcome
- Every selectable mode/accent combination uses the intended mode-specific candidate, derives its tint from that candidate, and clears the advertised contrast floors.
- Spacing B renders the complete candidate set exactly as `DESIGN.md` defines it.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[pullfrog]

Important

The new grant-state commits leave the design authorities and rendered decision controls disagreeing, so both automated reviews and the human approval pass can act on obsolete or contradictory rules.

Reviewed changes Reviewed the two commits after b9ef378 that incorporated the first grant decisions and expanded the rendered reference.

• Finalized grant decisions: Removed the provisional serif, selected spacing A, and kept XP as a row rather than a ring.
• Reworked the accent choice: Replaced the space-blue candidates with ten measured hues, two fill treatments, and an emerald/rose/orange shortlist.
• Added the hover canon: Introduced hover tokens, timing variants, transition rules, and rendered surface/control examples.
• Refined the reference layouts: Adjusted card/row padding, overlay sizing, shells, status examples, and control organization.

⚠️ The grant state is contradictory across authorities

DESIGN.md now settles the spacing scale and leaves hue/fill open, but its locked summary and the design-review agent still mandate space blue, canvas ink on every primary, and provisional spacing. An R-ticket review would therefore ignore a real off-scale value while rejecting valid shortlisted dark fills that use white text.

Technical details

# Align every authority with the current grant state

## Affected sites
- `DESIGN.md:2,22` — the locked summary still names a deep space-blue accent.
- `DESIGN.md:67-70` — spacing is now explicitly chosen and enforceable.
- `DESIGN.md:177-202` — hue and fill remain open, and dark fills intentionally use white foregrounds.
- `.claude/agents/design-reviewer.md:37` — still instructs reviewers not to flag off-scale spacing because the values are provisional.
- `.claude/agents/design-reviewer.md:39` — still requires space blue and canvas ink on primary, contradicting every dark-fill shortlist option.
- PR description — still describes the removed space-blue/serif/provisional-spacing state and its old manual controls.

## Required outcome
- `DESIGN.md`, the rendered reference, the PR description, and the design-review checklist must agree on which decisions are settled and which remain open.
- The design reviewer must enforce the chosen spacing scale without rejecting a valid grant-1 accent/fill candidate.

⚠️ The rendered authority can represent impossible active states

The page can render rose with a light fill while still marking Emerald selected, and it lets the rejected spacing B override the settled scale. Its default medium timing computes 280ms/180ms while labels say 180ms/120ms, the row demo never applies the promised strong hairline, and --primary-rgb remains the light-fill byte even when --primary is dark; because this page outranks the prose, these are conflicting authority states rather than demo polish.

Technical details

# Keep the rendered canon synchronized with its active controls

## Affected sites
- `design/reference.html:95-106,158-173` — `--primary` switches between light and dark fills, but `--primary-rgb` always resolves to the light-fill RGB.
- `design/reference.html:137-178` — medium is `180ms` for controls and `280ms` for surfaces, while spacing B still activates the removed `20/40/56` values.
- `design/reference.html:198` — row hover changes only the background despite `DESIGN.md:598` and the rendered explanation requiring `--hairline-strong` too.
- `design/reference.html:339-377,746-766` — direct hue/fill changes do not clear or update the three preset buttons; a browser run left Emerald pressed after selecting rose plus light fill.
- `design/reference.html:408,466,474,479,489` — visible timing labels describe the fast preset while medium is selected and active.

## Required outcome
- Every rendered selection must have one synchronized pressed state and one corresponding semantic-token state.
- Settled alternatives must not remain able to override the canon that the page declares authoritative.
- Visible timing and hover descriptions must match the selected tokens and the behavior they demonstrate.
- `--primary-rgb` must describe the active `--primary` value for every hue, fill, and mode combination.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[pullfrog]

Important

The newly settled dark-fill hover state fails the contrast floor for every remaining accent, and the hover-speed decision also slows press feedback beyond its fixed contract.

Reviewed changes Reviewed 2fd8b22, which settled the fill and hover-speed decisions and reduced the rendered choice to three accent hues.

• Settled the fill: Removed the light-fill treatment and standardized all candidates on dark fills with white foregrounds.
• Settled hover timing: Removed the speed selector and fixed controls at 240ms and surfaces at 380ms.
• Reduced the accent choice: Removed the comparison grid and retained emerald, rose, and orange as the only interactive candidates.
• Synchronized token state: Updated candidate RGB values and removed the redundant preset/spacing control paths.

⚠️ Every primary button fails AA on hover

Each chosen fill starts only slightly above the 4.5:1 white-text floor, then --primary-hover mixes it 12% toward white and removes that margin. A real-browser measurement gives 3.65-3.70:1 on hover for all three candidates, so the rendered canon makes its primary CTA inaccessible in the state this commit just settled.

Technical details

# Preserve text contrast through the primary hover state

## Affected sites
- `design/reference.html:45-55` — the three dark fills are explicitly the lightest values at which white clears 4.5:1.
- `design/reference.html:94-95` — `--primary-hover` moves each threshold fill 12% toward its white foreground.
- `design/reference.html:174` — every primary pill uses that lighter value on hover.
- `DESIGN.md:189-197` — the settled treatment says every candidate clears all three floors and always carries white text.

## Browser evidence
- Emerald hover resolves to `rgb(58,150,103)`: white contrast `3.66:1`.
- Rose hover resolves to `rgb(201,100,151)`: white contrast `3.65:1`.
- Orange hover resolves to `rgb(205,105,57)`: white contrast `3.70:1`.

## Required outcome
- White text must remain at or above 4.5:1 in the primary button's default, hover, active, and disabled-visible states for every candidate and mode.
- The hover treatment must remain visibly distinct without invalidating the chosen foreground contract.

⚠️ The hover-speed choice also replaces the press timing

The new 240ms control-hover token also drives the pill's scale transition, so the rendered press now takes the same 240ms as hover. That contradicts both the fixed 150ms press contract and the stated reason for choosing a slower hover than press.

Technical details

# Keep press response independent from hover timing

## Affected sites
- `design/reference.html:126` — `--dur-fast` is now 240ms for the selected control-hover speed.
- `design/reference.html:183` — the same token controls both background hover and press scale.
- `design/reference.html:421-423` — the authority now explicitly presents press as 240ms.
- `DESIGN.md:394` — press feedback is fixed at 150ms.
- `DESIGN.md:586-592` — hover is deliberately slower than press, with controls at 240ms.

## Required outcome
- Press scale must retain its 150ms response independently of the selected 240ms control-hover transition.
- The rendered labels, tokens, and transition declarations must describe the same timing model.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[pullfrog]

Important

The new brand comparison breaks its light-mode text floor, and the brand contract now differs between the rendered authority, prose, and design reviewer.

Reviewed changes Reviewed 5f3ae37, which unbound completion from the accent and added the rendered brand-mark comparison.

• Neutralized completion: Moved --status-done onto the foreground ramp and updated the habit-row, derivation, and accent-rationing rules so finished work never takes the accent.
• Defined the identity mark: Added the Orbit/Astra shape distinction, mark asset sizes, 16px redraw, wordmark, lockup, and logo-body accent exception.
• Rendered the shortlist: Added side-by-side emerald, rose, and orange brand specimens across store, lockup, navigation, Astra, and in-context habit examples.

⚠️ The brand comparison breaks light-mode text contrast

The newly added brand columns hard-override --primary-soft to dark-canvas values, so switching the page to Light leaves every candidate's mini-tab accent text below 4.5:1. Since this page is the grant artifact and explicitly claims all candidates clear every floor, the light rendering is not a faithful basis for approval.

Technical details

# Preserve the light-mode text token in every brand column

## Affected sites
- `design/reference.html:143-150` — light mode maps `--primary-soft` to the candidate's `--hue-lmtext` token.
- `design/reference.html:300-302` — each new `.brandcol` overrides that mapping with the dark-canvas `*-text` token regardless of mode.
- `design/reference.html:485,539,593` — the 11px active mini-tab text consumes the overridden token directly on the light surface.

## Browser evidence
- Emerald rendered `#008B56` at 4.17:1 on `#FAFAFA`; its `#008451` light token measures 4.55:1.
- Rose rendered `#C2508C` at 4.17:1; its `#BB4986` light token measures 4.56:1.
- Orange rendered `#C85716` at 4.16:1; its `#C15109` light token measures 4.52:1.

## Required outcome
- The side-by-side candidate comparison must remain independent of the selected hue while resolving text roles through each candidate's mode-correct token.
- Every rendered text role must meet the documented 4.5:1 floor in both modes.

⚠️ The new identity contract has no single authority

DESIGN.md fixes exact lockup geometry and a new accent allowlist, but the rendered page uses different dimensions and the design reviewer still requires accented done dots while omitting the logo-body exception. A future implementation can therefore conform to one authority and still be rejected by another.

Technical details

# Align the rendered and reviewable identity contract

## Affected sites
- `DESIGN.md:30,163,216` — the mark's body is the only non-state accent use, and completion is explicitly neutral.
- `.claude/agents/design-reviewer.md:19` — the review checklist still includes done dots in the accent allowlist and does not allow the logo body.
- `DESIGN.md:36` — the horizontal lockup is fixed at a 28px mark, 12px gap, and 22px wordmark.
- `design/reference.html:313-314` — the authority page renders that lockup with a 10px gap and 26px wordmark.

## Required outcome
- The rendered reference and prose must specify one byte-exact lockup geometry.
- The design-review checklist must enforce the new completion and logo accent semantics rather than the superseded list.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[pullfrog]

Important

The expanded reference still contradicts the accent-rationing contract and publishes nonfunctional switch semantics, so it cannot yet serve as the authoritative implementation reference.

Reviewed changes Reviewed the three commits after the prior Pullfrog review, covering the accent-restraint fixes, expanded primitive reference, and narrowed accent shortlist.

• Tightened accent restraint: Neutralized badges, Astra, completion, and selection tint while adding field, OTP, switch, error, disabled, and listing specimens.
• Expanded primitive coverage: Added TabBar, FAB, ProgressRing, NavHeader, SettingsGroup, StatTile, InfoCard, PlanCard, and Satellite examples.
• Narrowed the shortlist: Removed emerald and kept warm orange and rose synchronized across the selector, tokens, and brand comparison.

⚠️ The expanded kit breaks the closed accent allowlist

The new kit section says it renders the sanctioned accent surfaces, but it also spends the accent on a completed 100% ring, InfoCard and PlanCard surfaces, and the Satellite arc. None is in the closed “whole list,” and selected state is limited to a glyph and label, so the rendered authority now directs future implementations to violate the prose.

Technical details

# Establish one accent-rationing contract

## Affected sites
- `DESIGN.md:214,217` — selected accent is limited to a glyph and label, and the six-use list declares itself exhaustive.
- `DESIGN.md:334-340` — the primitive table separately assigns accent tint or strokes to PlanCard, InfoCard, and Satellite.
- `design/reference.html:880` — the newly rendered 100% ProgressRing remains fully primary despite the ban on accent marking completion.
- `design/reference.html:914,919,923` — the newly rendered InfoCard, selected PlanCard, and Satellite use `--primary*` outside the closed list.

## Required outcome
- `DESIGN.md`, the rendered page, and the design-review checklist must define one exhaustive set of accent uses.
- Every rendered specimen must follow that set, including completion and selected-card treatment.

## Open questions for the human
- Are PlanCard, InfoCard, and Satellite intended additions to the allowlist, or should these specimens be neutralized?

⚠️ The switch specimens have semantics but no operable control

The two new role="switch" specimens are keyboard-focusable, but the accessibility tree exposes both without a name and pressing Space leaves their state unchanged. This publishes a nonfunctional keyboard and screen-reader contract in the page that implementations are told to follow.

Technical details

# Make the rendered switch contract operable

## Affected sites
- `design/reference.html:779-780` — both switches use `role="switch"`, `tabindex="0"`, and fixed `aria-checked` values without an associated accessible name or activation behavior.
- `design/reference.html:992-1001` — the only script handles the mode and hue segmented controls; no switch click or keyboard path exists.

## Required outcome
- Each switch must have an accessible name and update its visual state plus `aria-checked` from click, Enter, and Space.
- If these are intentionally static illustrations, remove the interactive role, focusability, and pointer affordance instead of announcing controls that do nothing.

## Evidence
- A real `file://` browser accessibility snapshot reported two unnamed switches.
- Focusing the checked switch and pressing Space left both classes and `aria-checked` values unchanged.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[pullfrog]

Important

The revised accent-role contract still assigns selection semantics to a static informational card, so the rendered authority does not yet define one enforceable rationing model.

Reviewed changes Reviewed 1a644a9, which removed the inert drift workflow and revised the fixes for accent rationing and switch operability.

• Removed inert drift automation: Deleted the workflow that could not observe pushes to main, leaving it to land separately against the correct base.
• Reframed accent rationing: Replaced the component allowlist with four semantic roles, neutralized the completed ProgressRing, and clarified the PlanCard and Satellite specimens.
• Made switches operable: Replaced the role-bearing div elements with labelled native buttons and synchronized their visual and ARIA state on activation.
• Expanded design-review checks: Added the newer glass, sparkle, pill, default-theme, and redundant-subtitle slop indicators.

ℹ️ The PR description advertises removed automation

1a644a9 deletes .github/workflows/redesign-drift.yml so the automation can land separately on main, but the PR description still lists that workflow as a delivered file. The review and merge record would therefore say D36 drift control shipped in this PR when it did not.

Technical details

# Keep the PR scope description current

## Affected sites
- PR description, `What this PR is` table — still lists `.github/workflows/redesign-drift.yml` as new.
- `1a644a9` — removes the workflow because a definition living only on `redesign/main` cannot observe pushes to `main`.

## Required outcome
- The PR description must no longer claim that this branch delivers drift automation.
- If a follow-up PR or ticket now owns the workflow, link it as deferred work rather than as a file shipped here.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[pullfrog]

ℹ️ No new issues found in 69b5fc9.

The commit consistently removes the invalid static InfoCard accent variant from DESIGN.md and the rendered authority. Browser verification found two .infocard.quiet specimens, zero .infocard.accent specimens, and no page or console errors.

Previously reported PR-wide requested changes remain outstanding, so this incremental pass does not approve the PR.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}