make sabotage takes ~20 min and grows with every arm; run the arms in parallel #206

On 2026-10-27 batch PR #205 failed its required build and test twice at the 30-minute job limit: make sabotage (tests/sabotage.py) compiles and runs each arm one after the other, #204 already takes about 21 minutes and the planner and history arms add more. The job timeout was raised to 60 minutes as a stopgap (aa32a84c's commit on batch/planner-2026-10-28).

Better: run the arms in parallel (each arm works on its own copy of the tree, so a process pool is safe), keep the per-arm result lines in a stable order, and keep the 'text to break occurs N times' and 'survived' failures exactly as they are. Measure the wall time before and after on the CI runner. The job is the required check on main, so its length is every PR's latency.