fix: refresh generated tooltips when the accessible name changes

[FutuRiata]

The effect in SketchForgeEditor.tsx that copies a button's accessible name into its title returns early on every button that already has one:

if (button.title) {
  return;
}

The MutationObserver right below it watches aria-label on the whole subtree, so the intent to keep the tooltip in step with the name is clearly there. But the guard makes the observer a no-op: the first pass gives each unlabelled button a title, and every later pass hits the guard and leaves. A generated tooltip is written once and never again.

Where it shows. Toggle buttons, whose accessible name is the thing that changes. ShapeInspector.tsx:480,483:

aria-label={locked ? "Unlock shape" : "Lock shape"}
aria-label={shape.hidden ? "Show shape" : "Hide shape"}

After one click the tooltip states the opposite of what the button will do: hovering the lock on an already-locked shape reads "Lock shape", and clicking it unlocks. The sketch image lock in SketchWorkspace.tsx:1241 behaves the same way. Found in a production build, by hovering, not by reading code.

Why the guard cannot simply go. I walked the JSX with the TypeScript parser rather than trusting grep. Of 152 <button> elements in apps/web/src, 53 set title in the markup, and on 23 of those the title deliberately differs from the accessible name:

Site	title	accessible name
WorkplaneViewport.tsx:5048-5053 (view cube, 6 buttons)	Top view (5)	Top view
WorkplaneViewport.tsx:5077	Place workplane (W)	Place workplane
WorkplaneViewport.tsx:5059,5064	Show controls	Show camera controls
SketchForgeEditor.tsx:9983	Add box at the sketch origin, or drag it onto the sketch	Box (from the span)
SketchForgeEditor.tsx:10032,10035	Undo	Sketch undo
ShapeInspector.tsx:528	#D41721	Set color #d41721
SketchWorkspace.tsx:1238	Lock image (L)	Lock sketch image
SketchWorkspace.tsx:1247	Unlock image before deleting	Delete sketch image
TransformOverlay.tsx:215,251	Rotate	none at all
page.tsx:1771,1873	Project options	Project options for <name>

Overwriting those would lose every keyboard shortcut hint in the viewport, so the fix has to tell "this tooltip was generated" from "this tooltip was authored".

What this does. The helper remembers what it wrote in a WeakMap keyed by the element and refreshes only a title that still matches its own last value. A title that came from the markup, or one that changed behind its back, is left alone. Buttons with a markup title take the same cheap early exit as before, so the observer's hot path is unchanged, and nothing is added to the DOM.

The logic moves into src/lib/buttonTooltips.ts so it can be covered by a unit test in the existing style; the effect in SketchForgeEditor.tsx is three lines now. If you would rather keep it inline, say so and I will fold it back — the WeakMap is the only part that matters.

Transcribing the shipped applyTitles() body verbatim into a scratch test and asserting the refresh makes it fail on Lock shape where Unlock shape is expected, which is the regression the new test pins down.

269 tests pass, the 264 existing ones plus five new; tsc --noEmit is clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TAoEuBEJyJtkEGapsp6G22