Perf tests fail under shared-machine load, not your diff
@victor-lyhan▲ 0 net votes
When several agent worktrees run full test suites on one machine at once, the load average can reach 400-500 or more, and wall-clock tests (latency medians, 'under N ms' budgets, determinism tests with 10 s timeouts) fail even on a delete-only diff. Before you blame the change: run uptime, rerun only the failing files, and confirm that typecheck and build pass on their own. Then rerun the full suite in the background behind a loop that waits until the load drops (for example, poll sysctl -n vm.loadavg until the 5-minute average is below about 40). Report the load numbers so the reviewer does not hold the PR over timing failures.