Add broker-level integration test harness (pytest + paho-mqtt)

[TalkingJupiter]

Context

The hello → config → telemetry handshake has never been exercised end to end. The Radxa host does not exist yet, so a Python test harness that plays host against the real device is both the test rig for #3 and the executable contract for #11 — whatever MAC format and role names the harness uses becomes the spec the real provisioning service must match.

Scope

A pytest + paho-mqtt suite (suggested location: host/tests/) run against the Mosquitto broker with a real ESP32 on the bench.

Scenarios

1. Device boots unconfigured → hello appears on repacss/devices/<mac>/hello within HELLO_INTERVAL_MS, payload is valid JSON with the right MAC
2. Harness publishes a valid config → hello stops, telemetry starts on the assigned topic (configured && enabled gate opens)
3. Invalid / malformed config → telemetry stays off, device keeps helloing
4. Retained config + device reboot → config arrives immediately on subscribe (verifies subscribe-before-hello ordering for real)
5. Config payload larger than the 1024-byte MQTT buffer → dropped, device unaffected
6. Broker restart → device reconnects and re-subscribes, config delivery still works

Acceptance criteria

• Suite runs against a bench device with only the broker (no Radxa) and passes.
• Topic strings, MAC format, and config schema used by the harness match the decisions documented in Align config schema between firmware and host provisioning example #11.
• Failures print the device-side expectation clearly enough to debug without opening serial.

Relates to #3, #11.