Analogous to how we buffer tokens, we need to delay reporting these errors until we have committed to a parse.