The original code uses the similary score
1 - len(diff(after, before)) / len(after)
The diff can at most be the size of the 'before' file, so any small
'before' file would be considered very similar. Removing an empty file
would cause all files added in the same revision to be considered
copies of the removed file.
This changes the metric to
bytes_overlap(before, after) / len(before + after)
i.e. the actual percentage of bytes shared between the two files.