I've attached an audio file demonstrating the issue using the latest version from the libsidplayfp-experimental branch; compare it to this recording from a real sid with a pretty neutral filter. Commit r2664 seemed to introduce this bug, and on top of that, setting the 6581 filter to 1.0 to match the sound of the filter prior to that commit reintroduces the octave drops in songs like Wizball that don't occur on real hardware, even on chips with a comprable filter brightness.
Moved to Github: https://github.com/libsidplayfp/libsidplayfp/issues/2