[aft-alfonso]

per_checkout_semantic_runtime::views_off_search_after_idle_eviction_reports_the_index_reloading failed on macOS in #1103's merge-queue group with "the root must be idle": it evicted once, right after readiness, while the root's background work was still finishing, and eviction refuses a root that isn't idle yet. It now retries eviction until it's accepted, within the file's 90 s deadline, the same way the views-on sibling test already waits for its lane. Test-only; the test passed 3 times on Linux, and the Windows target check passes.

---

Summary by cubic

Fixes three flaky tests on loaded CI runners where backstop deadlines fired before background work finished. The views-off eviction test now retries eviction until the root is idle, within the file's 90-second deadline. The startup-step deadline test widens its backstop window from 80 ms to 3 s so executor admission completes first. The blob GC batch test raises a test-only sweep budget from 5 s to 60 s so a loaded runner cannot cut it short. Test-only changes, plus the matching benchmark slice descriptor.

^{Written for commit 3e66a7d. Summary will update on new commits.}

[aft-alfonso]

Added a second test-only fix: subc::bash::grant_path_tests::bash_deadline_warning_and_refusal_name_the_running_startup_step failed on macOS in #1108's queue group, because its refusal named executor admission. The test fires the startup deadline itself once the stub holds the sandbox profile step, and the 80 ms window is only a backstop, but a loaded runner was still admitting when it expired. The backstop is now 3 s, so the test isn't any slower. The deadline tests passed 3 times on Linux, plus the Windows target check.

[aft-alfonso]

Third test-only fix: gc::storage_retention_tests::blob_gc_deletes_a_bounded_batch_in_one_transaction knocked #1106 out of the merge queue on Windows with "blob sweep cancelled or deadline reached". It ran through the production 5 s sweep budget, which a loaded runner overran on a three-row fixture. A cfg(test) thread-local now raises the budget for that test only; production is unchanged. 31 gc tests passed 3 times on Linux, plus the Windows target check.