I am somewhat suspicious that #3466 may have broken the fork_signal_to_child test, as UV_RUN_ONCE is a very risky thing to test, and we always initialize the loop with pending work.

Originally posted by @vtjnash in #3466 (comment)