Propagate unexpected API disconnects to the ESPHome async transport

[balloobbot]

Problem

ESPHomeSerialTransport connects with APIClient.connect(login=True) — no on_stop callback — and _call_protocol_connection_lost() is only ever dispatched from a local close(). When the ESPHome device restarts (or the API connection drops for any other reason), aioesphomeapi cleans up its connection internally and nobody tells the transport:

• protocol.connection_lost() never fires, so a StreamReader blocks in read() forever.
• Writes keep "succeeding": serial_proxy_write is scheduled with call_soon_threadsafe, and the APIConnectionError it raises inside the client loop is swallowed by the loop exception handler.

Consumers see a healthy port that never produces data again. For request/response protocols on top (e.g. the samsung-exlink / lg-rs232-tv Home Assistant integrations), every command just times out — indistinguishable from a device in standby — and their reconnect logic never triggers, even after the proxy comes back.

Fix

Pass on_stop to APIClient.connect() for connections the transport owns, and translate a stop into connection_lost() on the transport's loop (marshalled cross-loop the same way as _on_data):

• unexpected drop → connection_lost(SerialException(...)), and the serial is marked broken — matching how the descriptor transport reports fatal errors;
• clean device-initiated disconnect (ESPHome sends DisconnectRequest before a reboot, e.g. OTA) → connection_lost(None), i.e. EOF.

The handler is a no-op when close() already ran (_closing guard), and close() after a remote stop still works as before. Externally-passed APIClient instances are untouched: their owner called connect() and owns the lifecycle.

The sync ESPHomeSerial API is intentionally unchanged (the hook defaults to doing nothing there).

Testing

New test test_daemon_death_propagates_connection_lost kills the host daemon mid-connection and asserts the reader unblocks (EOF or SerialException) and the port reports closed. It times out after 10s of a blocked reader without the fix, and passes with it. Existing ESPHome transport tests (18) pass; prek (ruff, mypy, pylint, etc.) is clean.

🤖 Generated with Claude Code

[puddly]

I believe #112 handles the same problem but more directly.