a11y(presets): GUI3 preset accessibility (#2338 #2339 #2345 #2350 #2352)

[kpoineal]

Summary

Addresses all 5 NVDA screen-reader issues against the Presets UI 3 beta. No behaviour or visual regressions. One reviewer (Lovell) required.

Changes per issue

Closes #2338 — Preset parameter controls programmatic labels

All interactive controls now have programmatic accessible names via htmlFor/id (temperature, top_p, context size, top_k, repeat penalty, steps, CFG scale, engine hint, AutoOpt result, llamacpp_args, sdcpp_args). Image width/height use aria-label since they share one visual label.

Closes #2339 — Advanced backend/device fields discoverable

llamacpp_backend and llamacpp_device converted from plain text inputs to input[list] + datalist, exposing known selectable values (9 backends, 7 devices) while keeping free-text for custom values.

Closes #2345 — Preset card richer accessible description

Card overlay button gains aria-describedby pointing to a sr-only span with: starter/manual-args status, applies_to capabilities, parameter summary, prompt name, tools state.

Closes #2350 — Capability chip radio semantics

Capability chip container gains role=radiogroup. Each chip button gains role=radio + aria-checked. Radio chosen over aria-pressed because toggleCap is single-select.

Closes #2352 — AutoOpt run selection state

Each AutoOpt run button gains aria-pressed={selectedAutoRunId === run.id}. NVDA announces pressed/not-pressed state.

Test results

61 passed / 7 skipped / 0 failed (was 50/7/0). 11 new tests A35-A45.

Review notes

Needs human reviewer (not author). Arrow key navigation for radiogroup is a follow-up enhancement.

[kpoineal]

Internal review: CHANGES SUGGESTED (advisory — human approval required)

Authored by Mattingly; reviewed by Lovell. Reviewer lockout in effect — Mattingly may NOT self-revise; a different agent must address the blocking issue.

[BLOCKER] #2350 — role="radiogroup" declared without arrow-key navigation (PresetManager.tsx ~lines 937–943)
The APG Radio Group pattern requires ArrowDown/Up to cycle focus through role="radio" children. Without an onKeyDown handler on the radiogroup, NVDA enters forms mode, expects arrow keys to work, and gets silence. This leaves capability chips less operable than the prior aria-pressed approach (which Tab/Space handled correctly). Deferring arrow-key nav as "a follow-up enhancement" while declaring role="radiogroup" is a WCAG 2.1.1 violation, not a gap. Fix: (a) add onKeyDown to the radiogroup handling ArrowUp/Down/Home/End; or (b) revert chips to aria-pressed toggle buttons — fully operable today with no keyboard model required.

Test gap — no ArrowKey coverage (tests/a11y.spec.ts A41–A43)
A41–A43 verify the role/aria-checked exist but no test asserts keyboard.press('ArrowDown') moves focus to the next chip. The suite passes today even with broken navigation.

All other items — no blocking issues

• a11y: Preset parameter controls are not programmatically labelled #2338: htmlFor/id labeling + aria-label for shared width×height — correct.
• a11y: Advanced backend/device fields in Presets are not discoverable enough #2339: input[list]+datalist for backend/device — correct discoverable pattern; free-text preserved.
• a11y: Preset cards hide important metadata from screen readers #2345: aria-describedby sr-only metadata on card overlay button — correct.
• a11y: AutoOpt run selection state is visual-only #2352: aria-pressed on AutoOpt run buttons — right choice given structural complexity.
• 11 new tests (A35–A45) cover all five issues; no existing tests weakened.

Action required: Assign a different agent (not Mattingly) to add arrow-key navigation to the capability chip radiogroup and a corresponding Playwright ArrowDown focus test.

[kpoineal]

Review finding #2350 item 13 is addressed in commit 5beb8a3.

Switched capability chips from role="radiogroup" / role="radio" / aria-checked to role="group" / plain buttons / aria-pressed. Tests A41-A43 updated to assert toggle-button semantics. Full suite: 61 passed, 7 skipped, 0 failed.

[kpoineal]

Internal review (re-review): LGTM (advisory — human approval required)

Reviewer: Lovell | Re-review date: 2026-06-22

The fix correctly resolves the WCAG 2.1.1 keyboard gap I flagged in the first pass. Findings:

• role="radio" / aria-checked fully removed. grep on the branch shows zero occurrences of role="radio", aria-checked, or role="radiogroup" in PresetManager.tsx. Clean sweep.
• Toggle-button semantics are correct. Container carries role="group" aria-label="Applies to capabilities" (PresetManager.tsx ~line 937); each chip button carries aria-pressed={appliesTo.includes(cap)} with no role override — plain <button> semantics, keyboard-accessible via Tab + Enter/Space, no arrow-key contract. WCAG 2.1.1 gap is closed.
• Tests A41–A43 are non-trivial and precise. A41 asserts role === 'group' and aria-label is truthy. A42 loops every chip, explicitly asserts role !== 'radio', and validates aria-pressed is strictly "true" or "false". A43 uses evaluateAll to count pressed states and asserts trueCount === 1 and falseCount === total - 1. No existing tests weakened.
• Click/select behavior unchanged. onClick={() => toggleCap(cap)} is untouched; only the ARIA attributes were added. No visual or functional regression risk.

One minor advisory note (non-blocking): A41 checks aria-label is truthy rather than matching the specific string — acceptable coverage, but worth tightening if the label text ever needs enforcement.

No blocking issues. Cleared for human approval.

[fl0rianr]

LGTM