Interesting.
I was going to do an analysis what using _ready.appendleft() for adding selector events would do for that scenario. The idea being to consistently juxtapose exiting callbacks, selector events and new callbacks. However I think this just moves the pawn in this ioloop halting problem.
Is it worth investigating a change to the stop mechanism instead? Instead of raising an exception in the middle of run_once, it could set a flag to be seen by run_forever(). This may avoid this class of problem altogether and ensure run_once is a fairly simple and predictable.