I looked into this a bit.  It seems like this is because with the patch, the call to "loader.loadTestsFromModule(the_module)" inside regrtest comes before the try-finally:

http://hg.python.org/cpython/file/fcdb35b114ab/Lib/test/regrtest.py#l1277

whereas with the current code, the analogous test-loading code is part of test.support.run_unittest() and so is protected by the try-finally inside test_main().  Apparently, simply discovering/loading tests from test_concurrent_futures.py is enough to modify threading._dangling (e.g. when finding the tests to pass to load_tests).

I'm not sure yet what the right solution is, but it doesn't seem like test discovery should have that side effect.  It could be because of how test_concurrent_futures is written, or because of certain initialization code in one of the modules it depends on.