I hoped lowering the fillfactor will improve this, but fillfactor=75 had pretty much no effect in this case. Is that expected for this kind of workload? I see the previous discussion talked about random updates, not inserts, so maybe that's the culprit?
Yes. Since posting trees are ordered by item pointers, you can get benefit of fillfactor only if you use some item pointers lower than item pointers already in use. You can still get benefit in the insert case but you should have already some free space in the heap (perhaps do some deletes and vacuum).
Actually, this is narrowing benefit from GIN fillfactor. Probably, that means that we should still have default value of 100. But I think GIN fillfactor still might be useful.