LRM reasoning breaks down down past an unfamiliarity threshold, not "complexity"