I am aware of the fact that this sounds like a duplicate of JENKINS-4152, but it goes one step further than the issue described there:

I have a POM with 2 modules:

– Project A
– Project B

I build the POM with the incremental build feature of Hudson's maven plugin.

If I check in something to Project A which breaks the build, Hudson will detect this correctly.
Afterwards, I check in something for Project B. Hudson remembers that something went wrong in the last build, and tries to build Project A again, which will fail. This is expected behaviour and as far as I can tell the result of bugfix JENKINS-4152.
But now, if I check in a fix for Project A, only Project A will be built, but NOT Project B.

Therefore, my changes to Project B which I made while Project A was broken were not build!

This results in build artifacts which do not contain the changes they are expected to.

My suggestion would be that Hudson does not only remember failed builds, but also all builds which could not be build due to a failed build.