Re: TEXT vs VARCHAR join qual push down diffrence, bug or expected?

Removing that entirely would be quite incorrect, because then you'd be
lying to the parent node about what collation your node outputs.

After thinking a bit more about the existing special case for non-foreign
Vars, I wonder if what we should do is change these code blocks to look
like

                collation = r->resultcollid;
                if (collation == InvalidOid)
                    state = FDW_COLLATE_NONE;
                else if (inner_cxt.state == FDW_COLLATE_SAFE &&
                         collation == inner_cxt.collation)
                    state = FDW_COLLATE_SAFE;
+               else if (collation == DEFAULT_COLLATION_OID)
+                   state = FDW_COLLATE_NONE;
                else
                    state = FDW_COLLATE_UNSAFE;

That is, only explicit introduction of a non-default collation causes
a subexpression to get labeled FDW_COLLATE_UNSAFE.  Default collations
would lose out when getting merged with a nondefault collation from a
foreign Var, so they should work all right.

                        regards, tom lane