I was totally wrong yesterday. I blamed the spurious test failures on taint mode, but it was really autoflush that caused problems. Maybe taint was a problem too, but it was not the core problem. CGI::Compress::Gzip disables itself if autoflush mode is on, because that implies that the programmer wants HTML sent to the user NOW, not buffered and sent later via gzip compression.
I rolled out a 1.00 release this evening which I hope will work around the test failures. My thanks go out to Slaven Rezic (and others) for prompt smoke testing of new releases!