One of the tests wasn't restoring the changes it made to the config.
As the test was difficult to track down and the problem would have
become only worse in the future, the issues was solved starting with
a new config before each run.

The performance impact seems negligible and the testing environment is
more predictable.