Improve keepalive recovery handling

[peter-dolkens]

Summary

• ensure keepalive loops exit when the connection is not healthy so they can be restarted after reconnecting
• cleanly tear down transport state when reads fail so future reconnects start from a blank slate
• guard stop() cleanup to clear stored task references once they are cancelled

Testing

• pytest

---

https://chatgpt.com/codex/tasks/task_e_690ab45c96e8832aa1416d55a2d643fb