Improve file magic handling Before this change, the magic string was read by reading the number of bytes we're expecting to read based on our magic string. If ours happens to be a prefix of the real one (e.g. ours is 'odoc-3.1.0' and the real one is 'odoc-3.1.0-2-g12345789') then the check was succeeding and then the subsequent unmarshalling was failing. This fixes that by writing the length too, so we always unmarshal the right length.