Or maybe we should use a middle-ground solution for example your suggestion #2 & #3, and maybe also set a high value for typo_tokens_threshold?