[kacy]

Cluster mode served about 320k requests a second whether or not the client pipelined, against 870k (no pipeline) and 2.7M (pipeline of 16) standalone. Every command locked the migration state, a tokio mutex shared by every connection, and read the cluster state's async RwLock up to three times.

• the migration state is a std::sync::RwLock. no holder awaits while holding it, and the per-command check only reads it
• ClusterCoordinator::check_command makes the whole decision from one read of the cluster state: paused writes, replica redirects, the majority fence, CROSSSLOT and the slot check. it replaces is_replica/primary_addr_for_slot/reaches_majority_of_primaries/check_slot_with_migration/check_crossslot on the hot path
• unit tests call check_command, and a new one checks that a replica sends writes to its primary with MOVED and serves reads

One-node cluster owning every slot against standalone, 50 clients, 1M requests, 32 cores:

	standalone	cluster before	cluster after
SET, P=1	870k	327k	855k
GET, P=1	880k	319k	870k
SET, P=16	2.67M	318k	2.28M
GET, P=16	2.79M	316k	2.43M

With a pipeline of 16 cluster mode is still about 15% below standalone, from the one shared read lock each command takes.