The control-plane Watch was a 250 ms database poll per subscription: an idle instance ran about 180 transactions a second (archied 14 streams, gateway 5, dashboard 26), almost all sequential scans of the small resources table. The State Store now keeps one shared view of every stored resource version, read once per 250 ms window and reused by every Watch stream; a stream reads its own resource only when its kind's version advanced, so idle database work is one read per interval for the whole process whatever the subscriber count.
The watch contract is unchanged: resume from AfterVersion, no missed update (versions are monotonic per org and kind), the immediate seed probe, the same 250 ms latency bound (streams wake on the same window boundary), and a store error still ends the stream with Unavailable so the clients' existing reconnect backoff applies.