[kacy]

Under write load every replica fell into a resync loop. Measured with 1.26M keys loaded and 50 clients writing at a pipeline of 16, on one 32-core machine:

• the replica applied one key per shard round trip and loaded the full sync in 13.5 s, at about 93k keys a second
• meanwhile 9.5M writes passed through a broadcast channel that holds 65,536 events, so the primary sent MSG_RESYNC
• the replica flushed its data and started another full sync, which overran in the same way. it ended with about half the keys

Changes:

• per-replica queue: a forwarding task moves events from the broadcast channel into each replica's own queue, capped at 256 MiB (REPLICA_QUEUE_LIMIT). writes during a full sync wait there. past the cap the replica gets MSG_RESYNC at once, without the stale backlog. the writer flushes only when the queue is empty
• one event per burst: shards collect records during a burst and send them as one ReplicationEvent, as the AOF does. every broadcast send locks the channel, which held the primary to about 600k writes a second with a replica attached. the wire protocol is unchanged
• batched applies: the replica reads through a buffer, collects records into one batch per shard, sends each batch as one message, and acknowledges once it is applied. snapshot keys go the same way
• WAIT: a new replica starts at offset 0 and acknowledges the snapshot offsets after loading them. WAIT counted it as soon as the snapshot was sent
• INFO replication: connected_replicas counts the replicas streaming from the primary. it counted those gossip had reported, which read 0 while one streamed. a new sync_full field counts full syncs

Same load, before and after:

	main	this branch
full sync of 1.26M keys	13.5 s	1.0 s
full syncs during 30M writes	2+, replica left with ~half the keys	1, replica matches
primary SET rate with a replica, P=16	928k (replica mostly resyncing)	1.12M

A cluster without replicas is unchanged or slightly faster: 2.43M SET at P=16, against 2.28M.

The new integration test shrinks the broadcast channel to 1,024 events and writes 100,000 INCRs during a full sync. Against main's server it fails: the replica is missing the counter after WAIT returns 1.