The mutation checker judged a range one mutant at a time. For modules with slow tests a range of about a hundred mutants took hours, and it was the last step before every push. Today mutate_check.py --jobs N landed: it runs the plan as N parallel shares on one machine and judges them with the same merge code CI uses. Measured on the 28th: 98 mutants with --jobs 4 in 5394 s (about 90 minutes) against roughly 6 hours sequentially; 101 mutants of the graph package with --jobs 2 in about 12 minutes against about 25 for one process.

The first CI run of the PR lost mutation shard 1. Exit 143 at minute 27 of 50, about 60 orphaned Python processes cleaned up by the runner, and an empty log. Nothing in the diff looked wrong, so I reproduced it locally instead of reasoning about it.

The cause was a mutant. The checker mutates its own code, and one of the mutants turned if args.jobs > 1 into >= 1. A share never receives --jobs, so it has the default of 1, and with the mutation every share believed it was a parent and started its own set of shares. They lived in new sessions, outside the kill of the test run, and multiplied while the mutated file was in place: 10 mutators in 4 seconds on my machine, hundreds on a runner. The tests did not fail; the machine just filled with mutators.

The fix is one refusal at the top of run_jobs, before the lock is taken: if --shard is set, this process is a share and must not be a parent. The parent always gives a share --shard, and the argument checker never lets a parent have one, so the condition cannot be true for a legitimate parent. I considered a marker in the environment and rejected it, because the test run passes the whole environment to pytest and the test_jobs_* tests would be refused inside a share. Checked on the same mutant: the test fails in 0.64 seconds, no more than one mutator at a time, no orphans. The CI step also runs with PYTHONUNBUFFERED=1, so a shard that dies still leaves its lines in the log.

The review count tells its own story. The entry circle closed on round 9. Six rounds went into a hand-made hand-off of an exclusive lock to the children; the sixth replaced it with a shared flock, and three more rounds polished that design. The invariant I would carry to the next such tool: a process that can spawn copies of itself needs a structural reason it cannot be one of the copies, and a test of that reason has to run under the mutant that breaks it.