[lawrencecchen]

Summary

The feat-cmux-next tip failed four webviews checks (seen by merge-batch at a8615fd, still red at 25c1584). The pre-push hook refused every squash that touches webviews. This PR fixes each finding in the code. No checker, baseline or allowlist changes.

Check	Finding at 25c1584	Culprit	Fix
check:browser-lanes	test/composer-draft-overflow.test.ts:434 samples in a raw rAF loop	dc0fd32 (Austin Wang, #28672 work, direct push)	The sampler reads through window.cmuxAfterPaint, which open already installs
check:design-tokens	ContextSection.css padding 2px 6px 0	fbea666 (#28969)	4px 8px 0; the link's own 2px side padding is taken back (margin-right: -2px) so its text stays on the rows' 8px trailing edge
check:type-scale	pages/apps/styles.css: 8 raw sizes, baseline 7 (new .apps-instance 12px)	d13d17a (direct push)	var(--text-detail) (12px) with its letter spacing
vp lint	SectionList.tsx:73 jsx-a11y(prefer-tag-over-role) on role="group"	6b7c28f (direct push)	Reasoned disable: the suggested tags (fieldset, details, optgroup) are for form controls or disclosure, not a labelled run of nav links (same as DrillList)

vp fmt --check is already green at this tip (#28927).

Changelog

none

Observed / Not run

• Observed: bun run check in webviews/ at base 25c1584: 4 red: check:browser-lanes, check:design-tokens, check:type-scale, vp lint. At head 95efc8b (rebased on a0d51f8): exit 0.
• Observed: scripts/web-lint.sh passed; prepush-gate.py: no new violations.
• Observed: CMUX_BROWSER_TESTS=1 bun test test/composer-draft-overflow.test.ts -t "never under a button|inset comes back": 5 pass, 0 fail (Chromium).
• Not run: the WebKit lane of that test (WebKit is not installed on this Linux VM). No visual capture of the 2px change in the Context section's Compact row (the text position is the same; the row is 2px taller at the top).

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

---

Summary by cubic

Fixes four red webviews checks on the feat-cmux-next tip so the pre-push hook no longer blocks webviews squashes.

• check:browser-lanes: the draft-overflow sampler now reads through window.cmuxAfterPaint instead of a hand-rolled rAF + MessageChannel loop.
• check:design-tokens: the Context section's Compact row padding moves from 2px 6px 0 to 4px 8px 0; the link's own 2px side padding is offset so its text keeps the rows' 8px trailing edge.
• check:type-scale: .apps-instance uses var(--text-detail) with its letter spacing instead of a raw 12px.
• vp lint: SectionList gets a reasoned disable for role="group" since the suggested tags (fieldset, details, optgroup) target form controls or disclosure, not navigation links.

^{Written for commit 95efc8b. Summary will update on new commits.}

[lawrencecchen]

Review (claude-review-subagent): SHIP at 95efc8b

Correctness review of the 4-file diff (HEAD~1..HEAD). The diff changes no checker, baseline or allowlist; only the 4 source files listed in the PR body changed.

Verification (on this head, Linux VM)

• bun run check in webviews/: exit 0. check:browser-lanes ("Browser test lane imports checked"), check:design-tokens ("ratchet passed: 2136 recorded violations"), check:type-scale ("ok, 314 raw font sizes left in the baseline"), vp lint (clean), vp fmt --check and the type check pass. The check:system-blue and check:ui-review lines are warnings on files this PR does not touch.
• CMUX_BROWSER_TESTS=1 bun test test/composer-draft-overflow.test.ts -t "never under a button|inset comes back": 5 pass, 0 fail (Chromium). WebKit is not installed here, so its lane was skipped. That is the same gap the PR body reports.

Per change

• composer-draft-overflow.test.ts: the old loop posted a MessageChannel message from each rAF callback and ran check in that task. cmuxAfterPaint (test/support/paintedFrames.ts) does the same per frame: it asks for the next rAF in the callback, then runs the read as a task posted after the frame. It also runs a pending read at the first input event, before the page's handlers, so it never reads a layout that no frame painted. That is the same or stricter. check returns undefined, never false, so sampling keeps going for the whole test, as before. open calls installPaintedFrames(page) (line 192) before any watchButtonEdge call, and the global Window declaration comes in through the existing import. Not a hack: the test now uses the shared reader that the lane check requires.
• ContextSection.css: the old horizontal geometry was container padding 6px plus the link's own 2px padding (inspector.css .acpmux-chat-inspector-link { padding: 0 2px }), so the text ended 8px from the edge. The new geometry is 8px padding with a -2px margin on the link plus the link's 2px padding, so the text still ends at 8px. That matches the rows (.acpmux-chat-inspector-row padding 0 8px). The focus ring and hover area are unchanged (the link box is in the same place). The only visual change is 2px more space above the Compact link, which the PR body states. The design-token scanner does not check margin, so the -2px could look like a way around it. I judge it a legitimate optical-alignment offset: it cancels a specific component padding, has a comment, and lets the container sit on the 4/8 scale. The other choice, changing the shared link padding, would move Environment's Save, Connectors and the subagent Back link.
• apps/styles.css: --text-detail is calc(12px * var(--cmux-interface-scale, 1)) (pages/shared/desktop.css), so the size is the same 12px at scale 1. It now also follows the interface scale and gets the token's letter spacing, the same as the other --text-detail uses in this file (lines 147, 730, 801). This is the intended token use.
• SectionList.tsx: the markup and ARIA are unchanged (role="group" with aria-labelledby stays), so accessibility does not change. A labelled group of nav links inside <nav> is correct ARIA. fieldset/details/optgroup would be wrong here. The disable has a reason and follows the same pattern as src/ui/DrillList.tsx:187.

Findings

• None blocking.
• Nit (no action needed): the PR body says the Context row is "2px taller at the top" and gives no capture. The change is small and the PR states it.

No hacks found: each fix changes the code that was flagged, and no checker inputs changed.

[lawrencecchen]

Superseded: direct pushes 2d93485, a19e400 and e7bab27 fixed the same four findings on feat-cmux-next while this PR was in review, and this PR now conflicts with them. bun run check in webviews/ exits 0 at a3f620e. No ready-to-land: nothing is left to fix. Differences that are not fixed (follow-up only if wanted): the Context Compact padding is now calc(var(--ds-space-1) / 2) calc(var(--ds-space-3) / 2) 0, still 2px/6px (off the 4/8 scale); .apps-instance has no --text-detail--letter-spacing; the composer test keeps its own rAF loop with a raf-read: comment instead of cmuxAfterPaint.