Yes, page-level predicate locking should happen only when fast update is off.
Actually, I forgot to put conditions in updated patch. Does everything else look ok ?
I think that isolation tests should be improved. It doesn't seems that any posting tree would be generated by the tests that you've provided, because all the TIDs could fit the single posting list. Note, that you can get some insight into GIN physical structure using pageinspect contrib.
I have created new isolation tests. Please have a look at
