R08 4.5: every subscription operation that awaits I/O and the failover's
teardown + reconnect + replay share one async gate, and every read/write of the
subscription fields (failover's null-outs included) is under _subscriptionLock.
A subscribe in flight when a failover starts is recorded first and then replayed
onto the new session instead of landing on the dying session's subscription. The
dead session is disposed before the gate is taken so a blocked call fails fast.
A DisconnectAsync during a failover is no longer undone by it. ConnectAsync on a
connected or reconnecting client throws instead of overwriting (and leaking) the
live session.
Refs #747
Claude-Session: https://claude.ai/code/session_01Ub9hGmYLQvn9XETzw7CMsT
Joseph Doherty
2d8bb037f5