The six jobs have no data dependencies, so they now schedule concurrently on the
capacity-2 runner. Guard against the two memory-heavy jobs (mutation + verify-stack)
co-scheduling and re-triggering the e2e OOM (#126): order verify-stack after mutation
via needs, with if: !cancelled() so it still runs when the ratchet fails. Add a
concurrency group so a new push supersedes the previous run instead of wasting a slot.
closes#127
not