An AI agent issuing a prose query like "how to make a type Sendable" gets appentity and systemcoordinator at top-3, NOT the Sendable protocol page. If the agent grounds on top-3 only, it doesn't see the right reference and may hallucinate.
Mitigation: agents typically chain queries. After a prose query returns unhelpful results, an agent can re-query with a more canonical form ("Sendable protocol") and the canonical-lookup baseline shows the right answer comes back at rank 1. So Criterion 2 impact is moderate: prose-query top-3 alone is not enough, but a multi-turn agent recovers.
The Phase 1.7 agent-end-to-end eval (docs/design/search-quality-eval.md §14.4) is where this trade-off gets measured end-to-end. Until that lands, the prose-baseline result here is the closest signal available.