I'm concerned about the buildbots, where we'd
really rather get full output on the first try (since we don't in
general have access to the machine for debugging).
What we do for the lua testsuite is, we have a hacky shell script
line for all the buildslaves that says 'if make check failed, cat
tester_dir/tester.log'. One possibility would be to simply use this
same mechanism -- if the unit_tests always wrote their full output to
a log file, then we could teach the buildslaves to dump this logfile
into their output whenever the tests failed.