so from 21s down to 1s via 3 PRs - good team work @theimowski!

LGTM, pity there are no tests to be sure there's no regression.
Can we name the nested rec function paths' to better distinguish from outer fun?

@theimowski my question is: are we ver going back here? In my optimization we only add to the set and never reuse old sets.

Hm crap you're right. Now it might skip some chains ...
E.g. when we have a diamond:

A
  B
B
  C
  D
C
  E
D
  E
E

Then it would output just A->B->C->E and skip A->B->D->E because B was already added to set
So it changes behavior - but aside maybe such behavior is ok?

so revert?

That's up to you - behavior changed, but I'm not sure whether it's ok or not.
By that I mean the output shown when --details is specified might get very long for bigger trees

@matthid ?

Do we know how does yarn handle such cases?

Dunno. But I'm fine with revert. 3s is still usable.

Ok then how about revert and create issue to cover the code with some tests and optionally then improve perf