Moving to CHOLMOD was a mistake:  After all, the brittle-fracture
increment functional is non-convex, and hence second derivatives
can be indefinite.