I think part of the problem is that the common format used is somewhat of a compromise between being human readable and parseable by tools. I can sort of see where the author of this tool is coming from with trying to address some of that with metadata, but I feel like the better way might be to come to with a format that isn't reliant on plaintext but instead can be rendered to something more readable. Coming up with a way to calculate reasonable diffs on files like you mention that doesn't generate worse diffs than we have now for existing stuff would be challenging, but it doesn't feel like it would be impossible to solve.