Nine Keyboard Widgets That Were Never Implemented

Why Care?

Seven members declared role="menu", role="listbox" or role="radiogroup" — each of which is a promise to a screen-reader user: this is a widget you navigate with arrows, and Tab takes you out of it. Not one implemented it. Two declared role="menu" with no menuitem children; one declared a radio group containing no radios.

Seven members got it wrong independently. That is the tell: the roving-tabindex contract is the organ.

This is the first arc with tests, and it had to be

Button, Chip and CardRow are appearance with a little behaviour — a probe screenshot catches their defects. Selector is the inverse. A broken keyboard renders perfectly.

So 99 tests, most written before the component existed. And they immediately earned their keep by finding that my own components were broken:

• SelectWrapper--ClickBody was keyboard-dead. display: contents generates no layout box, so the button was not focusable — while reporting tabIndex="0", carrying a correct accessible name, and working with a mouse. Three engineers found it independently; one measured tab stops going 37 → 29, exactly the eight instances.

• Selector--Menu broke its own headline promise three ways, each revealed only by fixing the one before. Fixing a focus bug is how you find the next focus bug.

• A fixture simpler than every call site. Its Escape test passed while both adopters were broken, because the fixture's trigger was a local const no member could clear — and closing a menu is exactly when the anchor goes away. The test and the realistic call site disagreed, and the test won.

The two best moves of the arc were made by agents, not by me

A refusal that produced a fix. person-enrichment stopped mid-adoption because SearchBox--Autocomplete could not take an initial value — and its affiliation name arrives pre-filled from the person's email domain. Adopting would have silently dropped a value the operator was looking at. It marked its eight red assertions it.fails — which passes while the defect is present and fails the moment it is fixed. A live tripwire rather than a silenced test. All eight flipped on the first run after the fix; not one defect had to be re-discovered by hand.

A refusal that was right in the strongest sense. Two icon-only links could not adopt ExternalLink, because an aria-label suppresses the visually-hidden new-tab notice — the one thing the component exists to add. One of those sites already cleared the target floor with its own rules, so adopting would have turned a measured pass into a measured failure.

The gate was punishing the practice

An engineer deleted a rule carrying z-index: 5, wrote a comment explaining the deletion, and F4 re-reported the literal from the prose. The member sat at its old count until the sentence was reworded.

Documenting a removed defect re-created it in the gate — a direct incentive never to explain a deletion, in a codebase whose entire practice is explaining things in place. F8 had the same blind spot and had been reported twice as a finding without anyone fixing the cause.

The script's own header warns about a checker that reports success because it failed to look. This was the inverse.

And my own census had the same bug. DevelopersMenu was on the nine-surface list because a grep hit a code comment. The ninth surface never existed.

What a browser drive caught that 99 tests could not

Button declares white-space: nowrap and a fixed height. A Selector option declared neither. So every member that swapped Button rows for options silently converted a clip into a wrap — one real row measured 48px against its neighbours' 29px, its label squeezed to the min-content of its first word.

jsdom has no layout. It passed every test in every adopting member, and was found on the one member where somebody looked at pixels.

A component that replaces another must match its layout defaults, not just its role and its keyboard. The parts nobody declares are the parts that change silently.

Defects the sweeps found in members

• shell's workspace switcher rendered a dead socket and a healthy one identically — four of six connection states fell through to the same label, with the only trace of failure in a title nobody hovers.

• Six members were still shipping one identical broken status recipe after nine migrations: five states, three colour treatments.

• A tag popup was thirteen tab stops — Tab took the caret out of the input mid-word.

• Six of the federation's seven missing-noopener links were in two files, each opened page holding window.opener access back into ours.

• All nine hand-rolled checkboxes measured 13×13 — 54% of the target floor, every one. Two had no accessible name at all.

• A keyboard hint that had started lying. A ↵ badge advertised "Enter picks the first match" — true when written, false after. Removing it was the point: a hint that lies is the defect class this rollout exists to remove.

Two hand-rolled guards disagreed, and the weaker one was shipping a bug

Both async search surfaces guarded against stale responses. One guarded only the success path — its catch was unguarded, so a rejection for a term the operator had already moved past emptied the suggestions and painted an error over correct, on-screen results.

That disagreement is the argument for promoting a guard into a component rather than documenting it.

What is not done

• Snippet props cannot be typed with a member's own row type. Three adopters have now hand-rolled the identical id-lookup workaround. That is the threshold; it wants generics.

• Every adopter keeps a document-level Escape in addition to the Selector's — it covers Escape while focus is still on the trigger. Third copy; smells like a trigger primitive.

• InteractiveBadge has one sighting. Not built — the same rule that keeps Selector--Radio unbuilt.

• F7 still has no mechanical check, and 11 of 22 units still have no tests.

Related

• [[2026-09-13_07_CardRow-Twenty-Members-In-One-Evening-And-A-Primitive-That-Failed-Silently]]

• [[../context-v/plans/Build-Selector-And-Close-The-Keyboard-Contract]]

• [[../context-v/specs/SearchBox-LiveFilter-And-Autocomplete]]

• [[../context-v/loops/Adopt-Selector-In-One-Member]]

• [[../context-v/issues/A-Component-Test-Whose-Fixture-Is-Simpler-Than-Its-Call-Sites]]