FWIW, I'm inclined to think that pg_config should be marked as stable
not immutable. Aside from the minor-version-upgrade issue, what if
you install new binaries that are the same version but built with
different configure flags?
The pg_config outputs are roughly as stable as initdb-inserted
catalog data, and we've never judged that functions that return
catalog data can be marked immutable.
There seems no reason it can't be parallel safe, though, since
worker processes should get the same answers as their parent.
regards, tom lane