Avoid repeated selector walks and live UIA pattern probes - #450

Summary

• Cache each parent's children once for the duration of a grouped selector query. The existing path-based clause matcher still determines document order, :nth, and deduplication; later clauses reuse the same query-local snapshots instead of repeating provider calls. A unit test counts provider fetches across five clauses, including a child combinator and :nth.
• Include the six common UIA patterns in the existing bulk cache request and read their interfaces from the cache. This removes six live pattern-acquisition probes per element. Window and transform patterns remain role-gated and live. Public APIs are unchanged.

Motivation and measurements

sfmtool #647 reports that a GitHub-hosted Windows UIA tree walk took about 18.6 seconds; avoiding most accessibility walks cut its UI test from hundreds of seconds. sfmtool #653 reports a direct UIA walk of 52 nodes taking about 35 ms locally, supporting a large host/runner effect. This PR removes two xa11y costs that we can isolate locally, while the hosted-runner gap still needs an A/B run on that runner.

On this Windows machine, with the same 69-node AccessKit fixture and four warm runs:

The final tree output was identical across all 76 CLI lines before and after pattern caching, including actions and states. A separate direct UIA probe on 69 nodes found six live pattern acquisitions cost 40–62 ms, versus 9–11 ms when those patterns were prefetched; the number of supported patterns stayed at 37.

Validation

• cargo clippy --workspace --all-targets -- -D warnings — passed.
• cargo test -p xa11y-windows — 81 passed.
• Core selector tests — 96 passed, including the new fetch-count test.
• scripts/run_integ_tests_windows.ps1 selector_group_ — 6 passed.
• Full Windows integration — 152 passed, 9 failed twice. All nine require foreground focus or simulated input in this desktop session: SetForegroundWindow was denied by the foreground lock, SendInput returned error 5, and no process was reported foreground. Action, selector, value, and state tests passed.
• cargo xtask test-integ reports that the command is unsupported on Windows; the Windows integration script above is the repository's runnable equivalent.
• cargo xtask check passed the Rust, JS, formatting, and parity checks, but could not finish its Python checks because this host Python lacks pip and pytest.

The deterministic provider-fetch assertion guards against another clause-count multiplier without a timing threshold that would be flaky on shared CI runners. The hosted Windows slowdown itself remains unreplicated locally and should be measured in CI before claiming an end-to-end fix for that environment.