BTW, I just thought of an issue that might change some of these
conclusions: what about supporting deferred constraint checking,
as we just recently managed to do for straight UNIQUE constraints?
I don't say that this has to be in the first cut, but it's something
we ought to try to leave room for down the road.
The current infrastructure for deferred uniqueness requires that the
thing actually be a constraint, with an entry in pg_constraint that
can carry the deferrability options. So unless we want to rethink
that, this might be a sufficient reason to override my arguments
about not wanting to use CONSTRAINT syntax.
As far as implementation goes, I think there would be very little
choice but to use the insert-the-index-entry-first, check-later
approach; so your ideas involving extra state in shared memory
seem to fall to the ground anyhow.
regards, tom lane