relcache sometimes initially ignores has_generated_stored

Hi,

I think it's probably not relevant, but it confused me for a moment
that RelationBuildTupleDesc() might set constr->has_generated_stored to
true, but then throw away the constraint at the end, because nothing
matches the
/*
* Set up constraint/default info
*/
if (has_not_null || ndef > 0 ||
attrmiss || relation->rd_rel->relchecks)
test, i.e. there are no defaults.

A quick assert confirms we do indeed pfree() constr in cases where
has_generated_stored == true.

I suspect that's just an intermediate catalog, however, e.g. when
DefineRelation() does
heap_create_with_catalog();
CommandCounterIncrement();
relation_open();
AddRelationNewConstraints().

It does still strike me as not great that we can get a different
relcache entry, even if transient, depending on whether there are other
reasons to create a TupleConstr. Say a NOT NULL column.

I'm inclined to think we should just also check has_generated_stored in
the if quoted above?

Greetings,

Andres Freund