fix(a11y): pin SPRAWL's chip back to its corner, and guard the sp-* scale

[andraderaul]

Stacked on #290 (the CI fix), so the diff here stays the a11y work alone. GitHub retargets this to main when that merges.

#288 describes this review round in its own body — the sp-* guard, TOUCH_TARGET_ICON, the position find. None of it is in main. The description shipped; the code did not. This lands it.

The one that is a real defect

TOUCH_TARGET_HEIGHT opens with relative, because an ::after overlay needs something to anchor to. SPRAWL's basemap chip named its own absolute before the constant:

'absolute bottom-xs left-xs …',   // ← main
TOUCH_TARGET_HEIGHT,              // ← opens with `relative`, and wins

cn is tailwind-merge, which resolves a position conflict in favour of the last name. So the chip lost absolute and fell out of its corner back into the flow — with every target class present and correct, which is exactly why the className assertion sailed past it. Order swapped, and the new test asserts the position (absolute present, relative absent) rather than the target classes, because that is the axis that broke.

Never shipped: sprawl-touch-targets.md is still unreleased, so the bug lived only between #288 and here. Both constants now carry the ordering rule, and the changeset spells it out for anyone reaching for them next.

Every other caller positions via a parent — I checked all seven — so this was the only exposed site.

The sp-* hole in the scale guard

sp-* is a second scale under the same utilities, and its ends do not line up with the base one:

	runs
base spacing	2xs → 3xl
sp-* macro	xs → 2xl

So gap-3xl is real, p-sp-3xl renders nothing, and sharing the utilities is precisely what makes the second easy to reach for. The deny list now carries the sp- spellings step by step rather than by a prefix rule — the step is the only thing telling the two apart, so a prefix rule would either miss p-sp-3xl or falsely reject p-sp-2xl.

The list is composed from two named halves now (EXTRAPOLATED_STEPS × SECTION_ONLY_GAPS) instead of one hand-written array, and the completeness test holds the new entries to the preset exactly as it holds the old ones — define sp-3xl for real and the list has to give it up. That test also stopped defaulting to {}: a scale whose shape moved now throws by name rather than quietly comparing against nothing, which passes while proving nothing.

Dropped 2sm / 2md / 2lg while here — the multiplier prefix only ever attaches to xs / xl, nobody writes those, and a deny list is read less the more of it is theatre.

TOUCH_TARGET_ICON

Both canvases had independently composed TOUCH_TARGET_HEIGHT + min-w-[44px] inline-flex items-center justify-center. Two real callers is ADR 0014's bar, so it becomes one export. Its sibling TOUCH_TARGET_OVERLAY stays module-local — still one caller.

The asymmetry is the point and is documented on it: height is an overlay, width is real. A centred overlay wide enough for a ~27px icon reaches into its neighbour's across a gap-xs, and two overlapping targets are worse than one small target, because the press that lands stops being the one the user aimed at.

Tests assert the constant, not its expansion

Five target tests compared against string literals like after:h-[44px]. They now match against the exported constant, so respelling a target cannot red them without an actual regression behind it. SPRAWL's is the deliberate exception — it filters relative out, since the chip's own absolute displaces it and either one anchors the overlay.

Two ASCII/GLITCH tests that asserted the same element twice collapsed into one each.

Changeset bumps corrected

deck-kit-a11y-blockers and deck-kit-tablist-and-alerts were marked minor for what are pure fixes — no API moved. Both are now patch. deck-kit-touch-targets stays minor and absorbs TOUCH_TARGET_ICON, so deck-kit's net bump is unchanged; only the changelog reads honestly now.

Recorded rather than silently done

• The byte-identical input[type=range] stylesheets in ASCII and GLITCH. Extracting only the newest rule would split one control's styles across two files — worse than either end. tokens.css is :root-only by its own decision, so element rules need a second stylesheet export and every app's import order re-checked. Its own change.
• CLAUDE.md's scale table split sp-* onto its own row, since the shared-utilities-different-ends trap is the whole reason the guard needed widening. deck-kit's CONTEXT.md gets the matching note.
• TabStrip's roving-focus JSDoc → //. It sat above a const arrow inside a component body, so it was surfacing as that local's hover text.

Verification

• npm run check — clean, 291 files
• npm run typecheck — clean across all 5 workspaces
• npm run test:run — 2120 passed (286 + 363 + 436 + 92 + 943)
• npm run build — all three apps ship min-width:44px and height:44px in their built CSS, which is what proves the new classes survive the purge

🤖 Generated with Claude Code

Why this PR shows no CI

ci.yml gates pull_request into main and feat/** only. Its comment already anticipates stacks — "without it the sub-PRs of a stack land entirely unchecked" — but the allow-list covers feat/**, and this stack is fix/**. So Build and Test will not report here until #290 merges and GitHub retargets this to main.

Not a reason to merge it unchecked — the same commands CI runs are green locally on this exact branch, listed under Verification above. Worth deciding separately whether that allow-list should be [main, 'feat/**', 'fix/**'], or just the base-branch wildcard.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
ascii-art-converter	Ready	Preview	Aug 1, 2026 3:18pm
atlas-sprawl	Ready	Preview	Aug 1, 2026 3:18pm
cyberdeck-golem	Ready	Preview	Aug 1, 2026 3:18pm
glitch-studio	Building	Preview	Aug 1, 2026 3:18pm