Human-in-the-loop approval for autonomous agents. Asymmetric timeouts, per-argument risk, and a gate that cannot fail open.