An AI agent issuing find async throws functions returning Result gets top-10 results that don't constrain on async/throws/return-type. The agent then has to read those pages and hope they're relevant; for most of the 8-zero-P@5 queries above, none would be useful.
Mitigation in current state: the agent would need to issue more concrete queries (Result<Success, Failure> plus async throws separately and intersect mentally) rather than rely on cupertino to filter. This is feasible but inefficient.
The Phase 1.7 agent-end-to-end eval (docs/design/search-quality-eval.md §14.4) is where the practical impact of these attribute-query limitations gets measured end-to-end. Until that lands, this 25% baseline is the closest signal.