There's a chunk of bogus code in the inflate output filter: I think it was originally cut&paste (by me) from the inflate input filter. Since 2.2 we're getting flush buckets in real life, and they're tickling the bug causing it to die with "Inflate error -5 on flush". The apache-modules post I've referenced explains it.
Fixed in trunk: revision 416165
Proposed a patch for backport as r437674 (http://svn.apache.org/viewvc?rev=437674&view=rev).
Backported to 2.2.x as r488817 (http://svn.apache.org/viewvc?rev=488817&view=rev).