agentHost: correct reconnect handling found in review

- Keeps the connection entry when a transport drops during the handshake. The protocol client
  moves to `reconnecting` and schedules its own retry before it rejects the original connect,
  so disposing the entry cancelled that retry and discarded the client replay state.
- Settles a pending connection wait when a soft reconnect restores the transport. The wait
  previously ran to its timeout even though the host was reachable again.
- Counts reconnect attempts before recording the next one, so a policy of `maxAttempts: n`
  performs n attempts. The Dev Container policy of 3 attempts performed 2, and a policy of 1
  performed none.
- Follows an in-flight dial in `waitForConnection` instead of a fixed timeout. An SSH host can
  install the remote CLI first, which takes longer than the WebSocket-scale limit, and the caller
  received a timeout while that attempt continued and later succeeded.

(Commit message generated by Copilot)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>