Traced two huge plans. They differ. The fast one does use Materialize and Memoize (the psql). Is there something in JDBC 42 that blocks these algoruthms?
Directly blocking those is not likely. Maybe the way the drivers fetch partial results is different, such that with one the planner knows to expect only partial results to be fetched and with the other it does not. So in one case it chooses the fast-start plan, and in the other it doesn't. But it will be hard to get anywhere if you just dribble information at us a bit at a time. Can you come up with a self-contained test case? Or at least show the entirety of both plans?