Flake batch 23: watcher overflow-burst fixture, first callers retry
Test-only. watcher_filter::tests::busy_folder_that_is_not_ignored_is_never_excluded failed on Windows in #1131's queue group because its helper record_overflow_burst raced the filter's 250 ms flush window: the burst was dispatched as ordinary Paths before the overflow marker arrived.
The helper now queues the whole burst and the overflow before starting the filter, and arms its flush deadline past the 30 s receive budget, so the burst and overflow arrive back to back as in a live overflow. It also asserts that the overflow recorded the entire burst (diagnostic prefix counts sum to the burst length), so the exclusion assertions in all three callers can't pass without an observation. Callers and production code are unchanged.
Production note: overflow attribution reads a 512-entry ring of recent occurrences, not only pending paths, so an early flush on a slow machine does not lose evidence. Verified: watcher_filter tests three times on Linux (79), native Windows (68), and the Windows target check.
Summary by cubic
Fixes two test-only flakes on Windows: the watcher_filter overflow burst fixture no longer races the filter's 250 ms flush window, and the first callers query in callgraph_workspace_package_cache_refreshes_after_reconfigure retries while the background build finishes.
- The helper queues the entire burst and overflow before starting the filter, extends the flush deadline past the 30 s receive timeout, and verifies overflow counts sum to the burst length.
- The
callersquery now retries for up to 30 s, matching the post-reconfigure query in the same test. - Adds a search descriptor for the flake batch slice. Production code and all caller assertions are unchanged.