Detect dead connections via a read timeout and reconnect

[ptr727]

Why

The Host Command connections (event stream and command client) read with no timeout and rely on the peer closing the socket to surface a lost connection. On a hard power-off there is no TCP RST, and the event stream's ECHO keepalive writes only fail once the OS abandons the connection (about 15 minutes). Until then:

• the event stream read blocks forever, so Connected / Disconnected / Reconnected never fire and consumers keep treating the controller as online;
• the command client reads time out (it already passes read_timeout) but the timeout does not close the socket, and _get_connection only reconnects when the connection is closed, so every later request reuses the dead socket and keeps timing out.

Net effect: a powered-off or crashed controller is not detected for about 15 minutes, and per-object data does not recover until then.

How it is implemented

• Event stream (_command_client/events.py): read with a timeout of 2 * KEEPALIVE_INTERVAL. The ECHO keepalive round-trips every 60s on a healthy link, so 120s of silence means the link is dead. ClientTimeoutError subclasses ClientConnectionError, so it already routes into the existing reconnect path.
• Both clients: on ClientConnectionError, close the connection so the next _get_connection opens a fresh socket instead of reusing the stale one. Without this, a read timeout leaves the socket open and the event stream re-emits Reconnected on a dead socket, flapping every timeout and never recovering.

Two small follow-on commits keep CI green: a Generator (not Iterator) return type on a @contextmanager example to clear the only pyright reportDeprecated error, and Python 3.14 added to the lint matrix.

How it was tested

Validated against a real InFusion controller by correlating Home Assistant entity state with ground-truth ICMP reachability across several power-off / power-on cycles:

• Connection loss detected about 108-120s after the link drops (previously not detected within 8+ minutes).
• Connection state no longer flaps while the controller is unreachable (single latched transition).
• Event stream and per-object command data both recover within a few seconds of the controller returning (previously stuck until the ~15 min OS timeout).
• Confirmed ECHO round-trips every 60s (ECHO -> R:ECHO), so a healthy but quiet link never reaches the read timeout.

ruff check, ruff format --check, and pyright 1.1.410 pass across the 3.10 - 3.14 matrix.

[Copilot]

Pull request overview

This PR improves resiliency of the Host Command event stream and command client by detecting “dead” TCP connections via read timeouts and forcing a reconnect, reducing the time-to-detect for hard power-offs/crashes.

Changes:

• Add an event-stream read timeout tied to the keepalive interval so stalled reads surface disconnects promptly.
• Ensure both event stream and command client close the underlying socket on ClientConnectionError so subsequent operations create a fresh connection.
• Update an example type annotation and extend the lint workflow matrix to include Python 3.14.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
src/aiovantage/_command_client/events.py	Adds a read timeout to detect dead event-stream connections and explicitly closes the socket on connection errors to enable clean reconnects.
src/aiovantage/_command_client/client.py	Closes the command client connection on ClientConnectionError so future requests don’t reuse a stale socket.
examples/loads/control_load.py	Updates the @contextmanager return annotation (but currently introduces an invalid Generator[...] parametrization).
.github/workflows/lint.yml	Expands the lint matrix to run on Python 3.14.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated no new comments.

[ptr727]

Superseded by upstream PR loopj#373 (same branch). This fork-internal PR was only for the Copilot review pass.