U+000D is converted to U+000A in the pre-tokenizer, so it must be serialized as a character reference.