[Appliance Core] Build async Q-Learning inference/learning loop with MQTT reporting

[saviornt]

Description

Create the main asynchronous event loop for the edge Q-Learning agent: observe state → select action → (dummy) execute → store transition → periodic learning → publish episode stats via MQTT.

Why: This is the heart of the edge autonomy — must be non-blocking and observable.

Type

• Feature

Focus Area (pick one)

• Appliance Core (Pi edge) / MQTT & Comms

Priority

• Critical

Acceptance Criteria

• Async loop using asyncio that runs indefinitely (with graceful shutdown)
• Publishes episode summaries to MQTT (e.g. network-chan/edge/qlearn/stats)
• Integrates shared replay, agent, and safety stubs
• Logs key metrics (epsilon, avg reward, steps)
• No blocking I/O in the main loop
• Uses paho-mqtt async client

Blocker / Dependencies

• [Appliance Core] Implement dummy environment & basic Q-Learning agent

Notes / Links

• Later: replace dummy observe/execute with real telemetry/Netmiko