It looks like fetch_minimal_work is not per physical CPU, but rather CPU thread. So it works the same as project resource zero in this scenario. Dang.