In the pattern-match check we are looking for a proof of
*unsatisfiablity* among a bunch of givens.  The unsat-ness might be
hidden in the superclasses, so we must expand them.  But in the common
case where the constraints are satisfiable, we don't want to expand
a recursive superclass forever.

This is all a bit arbitrary, but then the whole question is
undecidable anyway.

The bug in Trac #10592 comment:12 was that I expanded superclasses
forever.  This patch fixes it.