fix: shadow DOM scroll events — addGlobalScrollListener utility + close-on-scroll fix

[pzaczkiewicz-athenahealth]

Summary

Fixes #10093. The `scroll` DOM event has `composed: false`, meaning it does not propagate out of shadow roots — not even in the capturing phase. This silently broke two behaviours when `enableShadowDOM()` is in use:

1. Virtualizer / ScrollView — `document.addEventListener('scroll', onScroll, true)` never fires for elements inside a shadow root, so the virtualizer never updates which items are visible during scroll.
2. `useCloseOnScroll` (overlays) — `window.addEventListener('scroll', onScroll, true)` never fires for scrollable ancestors inside a shadow root, so combobox/popover overlays do not close when their ancestor scrolls.

Approach

Added `addGlobalScrollListener` to `DOMFunctions.ts` — the established home for shadow-DOM-safe DOM wrappers. When `shadowDOM()` is off the function attaches only to the global target, identical to the original code. When `shadowDOM()` is on it additionally walks the ancestor chain from a reference element, collects every `ShadowRoot` found, and attaches a capturing `scroll` listener to each.

Updated all affected callers to use this utility:

• `packages/react-aria/src/virtualizer/ScrollView.tsx` (replaced manual shadow-root-walking code from an earlier partial fix)
• `packages/react-aria/src/overlays/useCloseOnScroll.ts`
• `packages/@adobe/react-spectrum/src/menu/useCloseOnScroll.ts`

Callers that attach scroll listeners directly to a local element (`TabPanelCarousel.tsx`, `Pagination.tsx`) and the `visualViewport` listener in `useOverlayPosition.ts` are unaffected — they don't rely on event propagation crossing shadow boundaries.

`vitest.browser.config.ts` has a new `define: { 'process.env.VIRT_ON': '1' }` entry required for the Tree browser test: without it, the virtualizer treats `NODE_ENV=test` as a signal to use `Infinity` for viewport dimensions and skips virtualization, making the scroll test meaningless.

Pull Request Checklist

• Identified root cause: `scroll` events are `composed: false` and do not propagate out of shadow roots, even in the capture phase on `document`/`window`
• Audited all `addEventListener('scroll'/'scrollend')` usages in library code; confirmed which rely on global propagation vs. direct element attachment
• Created `addGlobalScrollListener` in `DOMFunctions.ts`, gated on `shadowDOM()` flag — light-DOM behaviour is unchanged when the flag is off
• Updated `ScrollView.tsx` to use the new utility (removed manual ancestor-walking code)
• Updated both copies of `useCloseOnScroll.ts` to use the new utility
• Types compile cleanly (`yarn check-types:tsc`)
• New browser tests written, confirmed to fail before the fix and pass after (see Test Instructions)
• 296 existing unit tests pass across Tree, VirtualizedMenu, and all overlay suites (1 pre-existing skip)

Test Instructions

Automated browser tests (Chromium)

`packages/react-aria-components/test/Tree.browser.test.tsx` — Virtualizer / ScrollView:

yarn vitest --config vitest.browser.config.ts run packages/react-aria-components/test/Tree.browser.test.tsx --project=chromium-desktop

Mounts a 50-item virtualized Tree inside a shadow root, scrolls the treegrid, and asserts Item 0 leaves the DOM while Item 20 appears.

`packages/react-aria-components/test/Select.browser.test.tsx` — useCloseOnScroll:

yarn vitest --config vitest.browser.config.ts run packages/react-aria-components/test/Select.browser.test.tsx --project=chromium-desktop

• Light DOM test: Opens a ComboBox inside a scrollable div, fires a scroll event, confirms the popover closes — regression guard, passes before and after.
• Shadow DOM test: Same setup inside a shadow root — fails before the fix, passes after.

Existing unit tests

yarn jest packages/react-aria-components/test/Tree.test.tsx packages/react-aria-components/test/VirtualizedMenu.test.tsx
yarn jest packages/react-aria/test/overlays/

🤖 Generated with Claude Code

[pzaczkiewicz-athenahealth]

Thanks for the review @nwidynski! I know you're not one of the maintainers, but it's best to get this PR in the best shape that it can be so that it's more likely to be approved. How did you even find this PR and realize that it touched on similar concepts as your in-progress work?

[nwidynski]

@pzaczkiewicz-athenahealth Yep, that was the whole idea. To answer your question - it is pretty simple - I've made it a habit to skip over just about every other PR in this repo and have been doing so for a while.

The ones I find interesting I comment on and most of the time try to pre-review, in the spirit of hopefully making time for the core team and also to stay up to date with changes in the hook layer, which we use at work.

I've also found it to just be an amazing learning tool that works for me, haha!

[snowystinger]

Opening Level 1 Item 3 then Level 2 Item 3 results in blank space at the end of the Tree. It was missing the rowHeight, so I've fixed that.

I also added a control to the top of the storybook to enable the shadow dom flag since once set, it cannot be unset and it would have leaked into every story you visit after that point with no indication that was what had happened.

I think this looks right. Thanks!

[snowystinger]

I think your intuition was correct for this one, we should allow it to attach to the window object. I'd go so far as to say that global scroll should always be attached to window, not document, and the other instances are wrong.

If you think otherwise, then this test is now a duplicate of the one before it.

For now I've updated it to default to window and added tests so that we know what the contract is