previously built into Phil's test was clashing with the runtests.sh
framework.  I think one of the core problems might be that scripts
which don't explicitly "exit 0" have return codes which vary depending
on the particular shell.  (c.f. the autoconf portability guide.)

This might break the Debian test suite but I'm afraid I don't know
enough about how it works yet.  Hopefully eventually it can become
just a wrapper around our 'make check'.