Claims verification (issue #1250)

@@architect attaches the external facts its plan depends on to the @@context hand-off via AgentCall.verify.

How it works

1. While planning, @@architect writes each dependency as a falsifiable VerificationClaim (id, claim, kind, category, if_wrong).
2. The engine validates the set: only a hand-off to context may carry verify, the count is capped by complexity (low 3, medium 5, high 8), at least one third of the claims must be kind="refute", duplicate IDs are rejected and every if_wrong must be at least 10 characters.
3. ClaimTaskRenderer renders the claims as a fixed ### CLAIM VERIFICATION block prepended to the callee’s task, so @@context fills in a verdict (confirmed / refuted / contested / inconclusive) per claim with evidence.

Metrics

The engine counts claims per run and the verdict distribution; a refuted verdict that changes the plan is the signal that the cost paid off.