I haven't tested the parallel case, but I think we should sort out (3) get_cheapest_fractional_path_for_pathkeys as mentioned above.
I am lost about the comment regarding startup_new_fractional. Could you elaborate what you mean by that?
Apart from that, I'd argue for a small test case. I attached a slimmed down case of what we were trying to fix. It might be worth to integrate that with an existing test, since more than a third of the time seems to be consumed by the creation and attachment of partitions.