Fix connection and subscription stability bugs

[FZambia]

Summary

This PR fixes several bugs that could cause stuck reconnect loops, missed resubscriptions, crashes, and silent event loss under realistic concurrency conditions.

---

Bugs fixed

1. Client stuck in connecting state after transport closes mid-connect

When the WebSocket closed while a ConnectRequest was in flight, the transport's
_onDone callback ran synchronously (via Completer.sync()), calling
_processDisconnect + _scheduleReconnect before the catch block in
_connectInner resumed. The catch then called _processDisconnect a second
time, which cancelled the already-scheduled reconnect timer without rescheduling
one, leaving the client permanently stuck in connecting.

Fix: Return early from _connectInner's catch when the error is
ClientDisconnectedError; the reconnect timer is already running.

---

2. Server-initiated temporary unsubscribe did not trigger resubscription

When the server sent an Unsubscribe push with code ≥ 2500 (temporary, e.g. code
2502 for state invalidation), _handleUnsubscribe called moveToSubscribing to
transition the subscription but never attempted a resubscribe. Unlike a full
disconnect, the client remains connected in this case, so waiting for a reconnect
means the subscription never recovers.

Fix: Call resubscribeOnConnect() after moveToSubscribing() for all
temporary server unsubscribes (code ≥ 2500 and the existing code 2502 path).

---

3. _resubscribe lacked a concurrency guard

_resubscribe() had no mutex, so concurrent calls (e.g. from a pending retry
timer firing at the same moment as resubscribeOnConnect() on reconnect) could
result in two simultaneous SubscribeRequests for the same channel.

Fix: Added _resubscribing boolean flag (set/cleared in try/finally) that
causes any concurrent call to return immediately.

---

4. State not checked after getToken await in _resubscribe

After await _config.getToken!(event) yielded, a concurrent unsubscribe() or
disconnect could change the subscription state. The code proceeded to build and
send a SubscribeRequest regardless.

Fix: Added if (state != subscribing || client.state != connected) return
immediately after the getToken await.

---

5. _addUnsubscribe StateError on concurrent close + unsubscribe

When client.close() ran concurrently with subscription.unsubscribe(), the
moveToUnsubscribed future could resume after subscription.close() had already
closed _unsubscribedController. Calling .add() on a closed broadcast
StreamController throws StateError.

Fix: _addUnsubscribe now checks !_closed before adding to the stream.

---

6. subscription.close() did not clean up state before closing streams

close() closed the stream controllers but did not set _closed = true, cancel
pending timers, error out pending ready() futures, or set state to unsubscribed
first. This left a window where timer callbacks and async continuations could still
try to interact with the subscription after it was destroyed.

Fix: close() now sets _closed = true, cancels _resubscribeTimer and
_refreshTimer, calls _errorReadyFutures, and sets state = unsubscribed before
closing any controllers.

---

7. subscribe() did not guard against calling on a closed subscription

After client.close(), calling subscription.subscribe() would proceed past the
state check (state was unsubscribed but _closed was unset), producing confusing
errors downstream.

Fix: subscribe() now throws ClientClosedError when _closed is true.

---

8. Missing UnsubscribedEvent when unsubscribe cleanup send fails

In moveToUnsubscribed, if the cleanup UnsubscribeRequest to the server failed
and prevState was subscribed, the code triggered a reconnect and returned
early — but did so without emitting the UnsubscribedEvent to the caller. The
caller's await unsubscribe() would complete silently with no event on the stream.

Fix: Added _addUnsubscribe(UnsubscribedEvent(code, reason)) before the early
return in that error path.

---

9. _errorController.add after client.close() in _connectInner (4 sites)

If client.close() ran while _connectInner was awaiting getToken, transport.open,
or getData, the continuation could fire after _errorController was closed,
causing a StateError: Cannot add event after closing.

Fix: Added if (_closed) return guards at all four affected sites:

• getToken exception catch
• transport.open onError callback
• transport.open exception catch
• getData exception catch

---

10. Transport decode errors crashed the stream listener

A malformed incoming frame caused _replyDecoder.convert() to throw, which
propagated uncaught through the WebSocket stream onData handler, killing the
listener and silently stopping all further message processing.

Fix: Wrapped _replyDecoder.convert() in a try/catch; decode errors are now
forwarded to the onError callback so the connection handles them gracefully.

---

11. backoffDelay panicked with minReconnectDelay: Duration.zero

When minReconnectDelay was zero, val.toInt() evaluated to 0 and
Random.nextInt(0) threw a RangeError.

Fix: Return minDelay immediately when val <= 0.

---

12. _refreshToken missing state check after getToken await

_refreshToken() on the client checked state before calling getToken but not
after. A disconnect during the getToken await could leave the client in a
non-connected state, yet _refreshToken would still proceed to send a
RefreshRequest against the (now-null) transport.

Fix: Added if (state != State.connected) return after the getToken await,
before the RefreshRequest is built.

---

13. send() used wrong protobuf message type

Client.send() was constructing protocol.Message instead of
protocol.SendRequest, causing a runtime ArgumentError from the transport's
command encoder on every call.

Fix: Changed to protocol.SendRequest.

---

14. onError callback called transport.close() redundantly

The onError handler inside transport.open(...) called transport.close(), but
by the time a stream error fires the transport is already in an error state and the
onDone path handles cleanup. The extra close caused a double-close and surfaced
as a spurious error event.

Fix: Removed the transport.close() call from the onError handler.

---

Tests added

• close() errors pending ready() futures instead of leaking them
• removeSubscription after close() does not throw StateError
• unsubscribe() during getToken does not create a server-side subscription (validates _inflight cleanup path)
• subscribe() on a closed subscription throws SubscriptionUnsubscribedError
• Zero minReconnectDelay does not crash on retryable disconnect