Just at a quick glance this looks good, although I would have to check if all of the checks with lower and upper bound are propagated correctly to all branches.
The idea of having the bounds seems sensible at least 👍

Functionally, most of the changes look good to me.

For the state oracle, it may be worth considering to put the max_bound check a little bit earlier. As far as I can tell, the combination-loop is fully exhausted before any counterexamples are checked. So let's say you have a very big hypothesis but only want to test 200 TCs. The oracle would require a lot of time to compute all combination testcases just to throw away all but 200 of them.

On a more conceptional level (also with the mindset to re-integrate this into LearnLib), most of the padding consists of random sampling, correct? So you could also simulate this behavior via an EQO chain that first uses kway in its original form and then random walks/words afterwards. This would allow the oracle implementations to stick closer to their "true nature" without forcing the users to compromise on performance. Maybe this could be a solution for AALpy as well?