Patch looks basically OK on an eyeball scan. A couple of suggestions:

- use test.script_helper.run_python rather than rolling your own run_command (or, at least, just make run_command a thin wrapper around run_python). I've been trying to trim down the number of different ways the test suite launches interpreter subprocesses for testing purposes, and script_helper also does a bit of extra cleanup to try to eliminate false alarms regarding reference leaks. That said, I do like the way you're checking that the output went to the correct stream, so perhaps instead add a run_python variant to script_helper that returns a 3-tuple containing the exit code, stdout and stderr and make your run_command a wrapper around that (then create an issue to migrate the rest of the script_helper.run_python based tests to switch to the new version so they will also check that the output is on the correct stream rather than lumping the two streams together).

- update the description of gc.DEBUG_UNCOLLECTABLE in the docs to specifically mention that the complete list of uncollectable objects will also be printed at interpreter shutdown.