1. task x is started on w1
2. task x is cancelled
3. task x calls distributed.secede()

There's no such thing as a ("cancelled" "long-running") transition.

1. task x is started on w1
2. task x is cancelled
3. task x is started on w2 and terminates successfully
4. task y, which depends on x, is started on w1 -> task x is transitioned to resumed(fetch)
5. task x calls distributed.secede()

There's no such thing as a ("resumed" "long-running") transition.

In both cases, the worker will kill itself off with @fail_hard, losing all data stored on itself.
This is tightly related to #6685.