Add snapshot options, browser_dom_query, and unit test coverage

[BryceBarbara]

Summary

• extractPageSnapshot now accepts maxDepth, maxChildren, and selector options; browser_snapshot and browser_get_page_structure expose these as optional inputs via snapshotSchema
• PageSnapshotNode replaces the role/name fields with a generic attributes map; hiddenChildCount and hiddenAttrCount signal truncation to callers
• formatNode (used by formatPageStructure) updated accordingly: attributes are now rendered as key="value" for every entry in the map, with a trailing […N more attrs] token when the cap is hit and a […N more children] line appended when child nodes were omitted — previously only role and name were surfaced
• generateLocatorCandidates removes the .first() call so the evaluate callback runs against all matching elements rather than only the first
• New browser_dom_query tool returns element presence, count, and state (visible, enabled, checked, value) without waiting — useful for conditional logic without implicit timeouts
• buildNode is now unit-tested by having the page.evaluate mock call the serialised callback directly in Node.js with a vi.stubGlobal document, covering SKIP_TAGS filtering, depth/child capping, attribute capping and priority ordering, text truncation, whitespace normalisation, innerText fallback, and selector scoping

Test plan

All 41 tests pass (pnpm test). Manual testing of browser_snapshot params and browser_dom_query behaviour confirmed locally.

[BryceBarbara]

@etairl Happy to make any changes needed or remove anything that isn't wanted.

[etairl]

The big one: removing .first() from generateLocatorCandidates is a regression, not an upgrade. Playwright's Locator.evaluate runs in strict mode and throws on multi-match — for "all elements" you'd need evaluateAll. The unit test passes because the mock doesn't enforce strict mode, but browser_generate_locator exists specifically to disambiguate broad selectors like button or .btn, which will now throw instead of picking the first match. Please restore .first().

Otherwise: top-level node truncation when no selector is set silently drops children without reporting hiddenChildCount (worth fixing while you're in there), and extractPageSnapshot does await evaluate() then await page.title() sequentially — Promise.all them. The buildNode test infra and the attribute-priority/cap work are great; happy to merge once the .first() thing is sorted.

[BryceBarbara]

Addressed the .first() regression and tried to look for other concerns in the changes.

[etairl]

Nailed it — and the URL race catch plus the 500-node cap and the split hidden-reason hints are real improvements I didn't ask for. Merging.