I'd like to see a definite profile of a bloated stdlib test process before we assume we know the issue - the usual leak I see in test code is used test objects, and I'm not sure we've ported the usual fix for that into unittest yet (we should).

As far as linecache is concerned, we shouldn't be triggering linecache population except when a traceback is actually rendered, and so on a passing test run there should be little hitting it at all outside of tests that actually test rendering of tracebacks.