Here's another version where I do skip searching for children twice, and rewrote the comments.
I also noticed that in child tables we were only looking for pg_attribute.attnotnull, and not whether the constraints had been validated or made inheritable. This seemed a wasted opportunity, so I refactored the code to instead examine the pg_constraint row and apply the same checks as for the constraint on the parent (namely, that it's valid and not NO INHERIT). We already check for these things downstream (alter table phase 2, during AdjustNotNullInheritance), but only after potentially wasting more work, so it makes sense to do it here (alter table phase 1) given that it's very easy. I added some tests for these things also, as those cases weren't covered.
if (conForm->contype != CONSTRAINT_NOTNULL)
elog(ERROR, "constraint %u is not a not-null constraint", conForm->oid);
I feel that using conForm->conname is more friendly than oid for users.
