In my 4th attempt to test out the typ change to browser_test_runner, the gpu test were failing on the bot because the suite was leaking browser. (example log: https://chromium-swarm.appspot.com/task?id=342e79945fc3aa10&refresh=10&show_raw=1)

This was in facts caused by a bug in tear_down_process & fixed by https://codereview.chromium.org/2671383003/ (see the diff to the 1st patch). 

It was a bit hard to debug this because the failure in tear_down_process was completely silent: there is Exception output to stdout, it did not change return code of typ. If the swarming infra did not check for leaking processes, we wouldn't have been able to catch this bug.

I think it's a bit risky to move the new harness to typ while typ still handle this kind of failure silently.