Ideal Attribution and Faithful Watermarks for Language Models