When the auto-reverter reverts a commit that was flagged for a correctness failure, its PR still runs benchmarks_sep. If benchmarks_sep fails for any reason on the revert PR (including an unrelated regression or noise), the auto-reverter's PR is blocked from merging and the correctness fix does not land. Either the auto-reverter PR should be allowed to bypass benchmarks_sep failures, or benchmarks_sep should not run on auto-revert PRs at all.

Skipping benchmarks_sep on auto-revert PRs is the simpler option and matches the intent: correctness reverts are time-sensitive and do not need a perf signal.