Tools such as buildscripts/golden_test diff and last lock on to the most-recently-created golden test output directory.

If a golden test does not fail, those tools erroneously report a failure since they lock on the previous failure.

The fix is to always create an output directory even in the case the test succeeds.