[gustmrg]

Stacked on #329.

An untouched window never went idle: every blinking Wisp kept Chromium recalculating styles on every frame, even though the eyes only move for 4% of the time. Nothing runs between blinks now, and the blink looks the same. Menus and panels also become readable sooner, and moving between tooltips no longer replays their fade.

Blink

• useWispBlink starts a 240ms animation with element.animate when a blink is due, and replaces the endless wisp-blink CSS animation.
• It keeps the six-second period, the phase each name gets from hashSeed, and the same Wisps: the xl picture and the conversation list.
• CSS reduced-motion rules don't reach animations started from script, so the hook checks prefers-reduced-motion itself. It also skips hidden windows.

Measured over 10 seconds in an idle window with seven Wisps:

	Before	After
Animations running between blinks	7	0
Style recalculations	601	176
Main thread time	0.50s	0.27s
Layouts	165	176

Layouts stay about the same because they come from the blink itself, which still moves the eyes. These are single runs under xvfb, so read them as an order of magnitude.

Other changes

• --ease-out is now cubic-bezier(0.23, 1, 0.32, 1), and the --animate-* entrance tokens use it. This also changes every ease-out utility, such as the splash fade and the connection progress bar.
• Dropdown, Popover, Select, Context menu, Tooltip, and the update toast ask for ease-out. They were falling back to the built-in ease.
• TooltipContent skips its fade with data-instant:animate-none. Base UI sets data-instant when one tooltip follows another, when focus opens one, and when Escape dismisses one.
• ToggleSwitch moves its thumb with translate instead of left, so toggling no longer triggers layout.
• chat-subline-pulse scales an ::after disc from 1 to 3 and fades it, instead of repainting box-shadow on every frame. The size and fade match the old pulse.

Testing

• npm run typecheck, npm run lint, npm run format:check, and npm run build pass. Lint reports the same 2 warnings as main.
• npm test on Node 22: 2448 tests pass.
• New tests in use-wisp-blink.test.tsx cover the timing per name, which Wisps blink, reduced motion, and stopping on unmount.
• Built app with the fake agent, measured over DevTools:
  • The table above, plus 13 blinks in 10 seconds with scale keyframes over 240ms.
  • The first tooltip fades with the new curve. The next one has data-instant="delay" and animation-name: none.
  • The context menu and details panel report cubic-bezier(0.23, 1, 0.32, 1).
  • The switch thumb moves from 3px to 17px with translate.
  • The pulse animates only opacity and scale, on ::after, with no box-shadow.
• The pulse was checked on an element mounted by hand, because the fake agent's working state is too brief to measure.

To check by hand before merging:

• The blink and the working pulse look the same as before on a real display.
• The stronger curve feels right on menus, the details panel, and the splash fade.

Not in this PR

• Each blink still costs layout, because it scales SVG eyes with the scale property. fix(ui): cut the main thread cost of working Wisps and panel drags #332 moves it to transform, which costs less.
• transition-[height] on the recording meter's bars and background-position on the connection shimmer still avoid the compositor. fix(ui): cut the main thread cost of working Wisps and panel drags #332 fixes both.
• The global prefers-reduced-motion rule removes all motion, including fades that would be safe to keep. Changing it is an accessibility decision of its own.
• SWE-268 may remove the working pulse. This PR only makes it cheaper while it exists.

Refs SWE-268