I think this started when I upgraded the warnings-plugin to 4.6.
Some of our people want to fail builds when they get new warnings, so have thresholds set for that. The plugin is then supposed to compare the number of warnings to the last successful build. After upgrade to 4.6 it has instead chosen to compare to the last build before the upgrade, as if it doesn't recognise the newer builds.
A difference between before and now is also that we now have multiple graphs (not a problem, but maybe a contributing reason).

We mostly have jobs with multiple parsers, sometimes using a custom parser. Of the embedded parsers we use the "GNU Compiler" and "MSBuild" parsers.

I have seen other "funny" cases where build #N had 246 warnings and build #N+1 had 247, yet there were no new warnings according to the plugin. That job has 2 parsers working on 2 files each.

Let me know if you need to know more about our setup or the jobs?