fix: step commands see shadow DOM and custom controls; refs no longer go stale

[karngyan]

Why

The reins do benchmark showed Claude, driving reins step by step, failing tasks because of the step commands themselves, not because of its reasoning:

• reins snapshot didn't look inside shadow DOM, so MDN's search and Open Library's sort menu were invisible.
• Clicks on wizard cards and settings toggles failed with "element has zero size". The cause was stale refs, not hidden inputs.

What

• Stale refs (the real "zero size" bug). reins snapshot never cleared the previous snapshot's data-reins-ref tags. After a wizard step or tab panel changed, a now-hidden element kept its old e8. The new snapshot gave e8 to a visible element too, and querySelector returned the hidden one first. Each snapshot now clears the old tags before tagging.
• Shadow DOM in snapshots. reins snapshot lists elements inside open shadow roots and names them the way reins do does: through aria-labelledby, labels, slots and composed text. The helpers moved from jev-snapshot.ts into a shared page-dom.ts, so both snapshots use one implementation. --ref reaches shadow elements in click, type, fill, hover, select, scroll, text, wait and upload. --selector stays light-DOM CSS, and that's documented. Closed shadow roots stay invisible.
• Controls with no box. A native radio or checkbox shrunk to 0×0 under a styled card is clicked through its visible <label> or its role=radio/switch/checkbox/option ancestor. A display: contents wrapper is clicked through its first child that has a box. The disabled, covered and moving checks apply to the element actually clicked. Hit testing now follows slotted content to its slot.
• release.yml installs Chrome and sets REINS_TEST_CHROME the same way ci.yml does. The release run on feat(extension): jev_observe / jev_act for reins do (2/3) #42 failed on the browser suite's setup without this.
• Benchmark fixture. In wizard.html, .backdrop { display: flex } overrode hidden, so the modal was open on load in every run. That's fixed. The benchmark report and the /docs/benchmarks page now note which step-by-step failures these fixes address.
• Changeset: patch for @karnstack/reins and @reins/extension, written without em dashes so the website's changelog prose check passes.

Behaviour change to know about: the snapshot now filters with checkVisibility instead of offsetParent. It now includes fixed-position elements and drops visibility: hidden ones, so ref numbers can shift on some pages. Switches, labelled inputs and selects now get their real names instead of their text content.

Testing

• pnpm build, pnpm lint and pnpm typecheck are clean. Tests pass: protocol 95, CLI 509, extension 325.
• The browser suite has 7 new cases: shadow DOM snapshot, click and type by ref, a nested shadow root with slotted names, a 0×0 radio card, a role=switch, a stale ref, and display: contents. The 5 cases for stale refs, 0×0 and display: contents fail on the old code.
• Live on a logged-in Chrome:
  • MDN: the snapshot lists button "Search the site" from inside mdn-search-button's shadow root. Clicking it by ref opened search, and typing by ref returned results.
  • wizard.html: clicking the Team card by ref checked it, and the review step showed Team.
  • settings.html: clicking the switches by ref after a tab change toggled the right ones, and Save stored them.

🤖 Generated with Claude Code