Put a mutex around query dispatch. One query at a time through the coordinator. Felt wrong immediately -- this is a search system, it should handle concurrent queries. But the epoch system is new and I don't trust it yet. Concurrent queries during an epoch transition would be a nightmare to debug if something went sideways. Easier to reason about one at a time and remove the mutex later. "Later" turned into weeks because it turns out worker computation is the actual bottleneck, not dispatch concurrency. The mutex doesn't matter when each query takes 200ms on the workers.
