[mbreiser]

Summary

The rig acceptance harness (rig/tests/test_ros2_poc.py) now checks the host monitor's own CPU against the mean over the run, with a limit of 3 % of one core. It used to check the median against 2 %.

• Why: on macOS/Linux, rig_nodes 0.3.1 (0.3.1: host monitor — quantized typed HostStats, cheaper macOS/Linux process sampling (LAB-338) rig-nodes#6) reads the whole process table only every 5 s. Those windows are a minority, so a median hides them; the mean is the honest cost.
  • Lab Mac: median 0.56 %, mean 1.9 %, max 7.4 %.
• Why 3 %: it leaves room over the Mac's 1.9 % and keeps the 5 s process-table read, so a process that starts using CPU shows up within 5 s. It is still well inside the 5 % budget.
• Reporting: the harness info line now also prints self_cpu_mean; the median and max are still printed.

Follow-up to #282 (released 2026-10-28).

Tier

Next candidate, alongside rig-nodes 0.3.1 (reiserlab/rig-nodes#6).

Release notes

none (only the acceptance harness changes)

Test plan

• ww18 (Windows 11, 24 CPUs), released main 59a377b + this change, rig-nodes 0.3.0 from the lock: cd rig && pixi run test gives 60 / 60.
  • Host monitor CPU: mean 0.0151, median 0.0085, max 0.035 cores.
  • /rig/host: 4.7 % of the compressed bag.
• Mac, after rig-nodes 0.3.1 is pinned: the harness host monitor cost check passes (measured standalone: mean 1.9 %).

🤖 Generated with Claude Code