I'm confused by your comment, Romain. The only change I see so far is switching harbor to audio_video_any, which is a fixed content kind. So any variable audio test shouldn't work!
Also, to address my worry about generators being unsuitable for variable content kinds, there's no problem: the set_mode API is there to handle this cleanly. The only constraint is that a decoder doesn't change its content type without notifying the generator, but that's not new. In most cases, each new connection triggers the creation of a new decoder, which figures its content type and informs the generator about it.
I'll fix the issue for real tonight.