ci: MG x86-main guard-cargo p50 2635s and guard-tree p50 1895s; 7 of 18 runs hit the 3600s section cap #5739

Measured 2026-10-19 over the last 18 completed merge_group x86-main runs (job logs, lines 'section X: ... in Ns'), plus run 39214997865 (both 3600s).

For comparison sov.test 638s, sov.lint 657s in the same run. The guards are on the critical path of every merge-group pass (the Sections step is the longest).

Split CPU vs wait: not_measured. The job logs carry no load or PSI lines, and the runner host is not this box. Indirect evidence for wait/contention: the same hermetic script check_crux_inference_judge.sh was 508s+ in passing run 39196197124 and killed at 2400s in 39214997865 with identical code; 12 sections run concurrently on the one runner. #5709 alone passed (39214996476: guard-cargo 2602s, guard-tree 1800s), so the failure is not that PR.

Proposed: (1) have fat_driver log /proc/pressure/{cpu,io} and loadavg at each 300s 'still running' line so the split is measurable; (2) per-section concurrency cap or move guard-cargo off the shared runner. Refs flow rows; filed by aprender-w5032.