Stop publishing system stats for offline robots

[b-Tomas]

Problem

__publish_pending_system_stats() runs once per execution loop iteration and publishes system stats for every robot with an active session — including robots the connector reports as offline.

That drives a loop in the platform:

1. stats arrive for a robot InOrbit has marked offline
2. InOrbit sends get_state to resolve the discrepancy
3. the edge-sdk's _handle_get_state calls our online status callback, gets False, and publishes a retained offline status message
4. InOrbit timestamps it

So the robot's offline timestamp keeps moving forward for as long as the connector runs, and an offline robot never looks like it has been offline for any length of time.

Fix

Gate the publish on _is_fleet_robot_online(robot_id) inside __publish_pending_system_stats, which is where the stored-stats path and the default-values path converge. One guard covers both, so stats explicitly stored via publish_robot_system_stats() are dropped for an offline robot too — not just the framework's zeroed/host defaults. Guarding only the public setter would have left the defaults path still refreshing the timestamp.

Nothing is lost by skipping: the forced state request only helps when the robot is online but InOrbit believes otherwise. For a robot we already consider offline, the reply just re-asserts offline.

Notes for review

• The online check now runs once per robot per loop iteration, where before it only ran when InOrbit sent get_state. It is subclass code, so it is wrapped in try/except and falls back to publishing on error — matching the edge-sdk's own get_state fallback (robot.py:777-781), so a broken health check keeps today's behaviour instead of silently muting a robot. The _is_fleet_robot_online() / _is_robot_online() docstrings and the docs now require the implementation to be cheap and non-blocking.
• Pre-existing gap, not addressed here: nothing pushes an online→offline transition. The edge-sdk only sends robot status on connect, on the MQTT will, and in response to get_state. InOrbit still detects offline through its own data timeout, and a connector that keeps publishing pose or key-values for a dead robot will keep it looking online regardless.
• Behaviour change, so this wants a minor version bump on release.

Testing

• test_publish_pending_system_stats_skips_offline_robots — offline robot gets no publish, its stored stats are dropped, pending dict still cleared
• test_publish_pending_system_stats_publishes_when_online_check_fails — raising online check still publishes for every robot
• Full suite: 237 passed. flake8 and black clean.

🤖 Generated with Claude Code