[PEAK] Re: trellis.Set.discard

[Sergey Schetinin]

The undo for tasks runs because Set._data first runs when we create CV
object because of .discard in maintain, which also changes .added
which Set._data depends on. Later, when processing .writes it is
rescheduled due to this dependency. When it runs for the second time
it adds savepoint to has_run, but the maintain is scheduled too,
because it depends on .v which is also registered in .writes, so when
that runs it schedules Set._data again, and that having run already
and being registered in has_run it gets added to to_retry, so the
savepoint is used. This gets to rollback maintain an .added which uses
its original savepoint rolling back the entire thing.