Romain, you might just want to check 471263 where there is disagreement over how this lintian test should or should not behave towards generated files. In particular, if a package contains a patch in debian/patches that causes upstream files to be modified indirectly (e.g. because upstream is old / quiet and hasn't updated the autotools stuff for years, any patch that affects Makefile.am or configure.in|ac is going to cause changes in generated files and these changes should not be wrapped into yet more patch files because of the inevitable build failures when any of the tools used to generate those files are updated.
Your stats may also be out because of this problem - it is hard to see how lintian can resolve the problem cleanly without carrying a long list of "possibly generated files" and risking the list going stale.
There is more to the contents of .diff.gz than may meet the eye.