test(mention): fail loudly when background search never settles

[srothgan]

Summary

run_search, the test helper that waits for the background file index in src/app/mention.rs, polled a fixed 200 × 5 ms (1 s) and then returned silently when the budget ran out. Callers then asserted against whatever half-scanned state existed.

• Replaced the fixed iteration count with a deadline, raising the budget to 30 s
• Turned an exhausted budget into an assertion failure that reports index_status, search_status and pending_match_sequence
• Drained events before sleeping, so an already-settled search returns without waiting out a poll interval

Test-only change; no production code is touched.

Why

The Nightly run 34098364599 failed on both attempt 1 and attempt 2, only in Test (macos-latest); Ubuntu, Windows and every build and smoke job passed.

• Attempt 1: completed_search_includes_shallow_and_deep_matches (search_status != Ready) and query_change_refilters_from_cache_without_restarting_walk (left: 2, right: 1)
• Attempt 2: only the latter, same assertion

src/app/mention.rs and src/app/file_index.rs were both last changed by dd12688 (2026-08-23), and the Aug 24 and Aug 31 nightlies passed on that same code, so this is not a regression. The shared cause is the silent timeout: on a loaded macOS runner the walk does not settle inside 1 s, and the tests then assert against a partial index. The current assertions cannot distinguish "the index never settled" from "the matcher settled and still returned the wrong candidates", which is why the failure reads as a confusing 2 vs 1 mismatch.

This does not by itself prove the extra candidate is only a test artifact. apply_match_result replaces mention.candidates wholesale and for_each_candidate skips entries whose strip_prefix(root) fails, so neither duplicate accumulation nor macOS /var → /private/var path pollution explains it; the remaining suspect is a stale nucleo snapshot around rebuild_stream(false) (src/app/file_index.rs:510). After this change, a recurrence reports which of the two it was.

Closes #

Validation

• Automated: cargo fmt --all -- --check and cargo test --lib app::mention:: pass locally on Windows. That is weak evidence — the failure is a macOS-only timing flake and did not reproduce locally before the change either. CI is the real check.
• Manual: N/A
• Screenshot/video (if UI changed): N/A

Notes

• Breaking changes: N/A
• Docs updated: N/A
• Governance/release impact: N/A
• cargo clippy --all-targets --all-features -- -D warnings was not run locally; left to CI.