Future-proofing against a common mistake in attempts to optimize NOT IN.
We don't have such an optimization right now, but attempts to do so
are in the works, and some of 'em are buggy.  Add a regression test case
covering the point.

David Rowley

Discussion: https://postgr.es/m/CAKJS1f90E9agVZryVyUpbHQbjTt5ExqS2Fsodmt5_A7E_cEyVA@mail.gmail.com