I've added a new feature to the RT module in gcl ansi-tests. It can now do 'extended random regression' testing. This involves repeatedly executing tests drawn randomly (with replacement) from the set of tests that normally pass. It stresses the system, looking for dependencies that the tests in their original order didn't exercise, for memory leaks, race conditions, and so on.
Bugs found by ERR can be very hard to track down, so it's not suitable for early debugging. On the other hand, it's been shown (see page 7) that ERR can expose defects that are very difficult to find otherwise.
The first thing I found when I ran these tests were many order dependencies in the test suite, particularly in package-related tests. I think they're all fixed now. SBCL has been running ERR for about ten hours on my machine with no failures or leaks. Clisp, on the other hand, showed an unexpected failure in one of the DEFCLASS tests. I'll try to track that down by doing ERR on limited subsets of the test suite.