Cut a render pass and keep the divider lines off the paint path

[Guiw5]

Two changes, both measured. The interesting part of this PR is what it does not do: profiling killed two planned optimisations.

Changes

Derive the ready flag. useReady kept a state map that every measurement wrote to twice (register false, then true), each write costing a render pass, and the callbacks were invoked during render through useMemo. Readiness is just "the viewport and the content have a size", so useSnapPoints derives it now and useReady is gone.

Composite the divider lines. The header and footer lines lived in box-shadow with calc(var(--rsbs-content-opacity) * 0.125) as the alpha. Animating a colour inside a shadow repaints both elements every frame. They are pseudo elements now, with a fixed colour and an animated opacity, which stays on the compositor. Pixel output is unchanged.

Measured

React commits from click to data-rsbs-state="open", React Profiler, 4x CPU throttle:

	commits	of which cascading
before	6	2
after	4	2

The two remaining cascading commits come from setMounted in index.tsx, not from this change.

What the profiling said, and what I dropped because of it

I traced a sustained drag and an open on the scrollable fixture at 6x CPU throttle and aggregated the trace by rendering phase.

Drag window, 4575 ms:

phase	share
script	~84%
style recalc	9.6%
paint	2.0%
layout	1.0%

Open window, 1331 ms:

phase	share
script	~82%
style recalc	5.7%
paint	2.0%
layout	1.7%

Dropped: the heightMode="transform" idea. The premise was that animating height forces layout every frame and that moving to transform would be the big win. Layout is 1-2% of the time. The change would have cost the sticky footer its position and broken the scroll viewport at small snap points, in exchange for almost nothing.

Dropped: removing XState. It is 27% of the bundle, so it looked like the obvious cut. During a drag it is 0.4% of sampled CPU, and the drag is not CPU-bound at all — a third of the trace window is idle, p50 frame time is 16.7 ms at 6x throttle. Removing it is a rewrite of the interruption semantics for a bundle saving on a chunk that consuming apps already load lazily.

What the traces do say is that both paths are script-bound, and that the expensive one is open, not the animation: 1331 ms at 6x throttle with the main thread busy essentially throughout, against a drag that idles. If more work goes into performance, that is where it belongs — the promise-actor chain, focus-trap's tabbable scan, and the aria hider's body > * walk all run before the sheet is visible.

Test plan

• npm test: lint, 25 unit tests, library build, docs build
• Header handle and both divider lines verified pixel-identical after the position: relative change, which was the risk
• Sticky fixture in Chrome: open, drag, snap, content resize, close
• Packed and installed into a clean Vite consumer on React 19: open, drag to dismiss, unmount, scroll lock released
• Real iOS Safari and Android Chrome