That's very interesting. I always assumed averaging (maybe with varying weights across frames and time), or applying a few rounds of other steganography, would degrade the encoded message into noise.

Could you link to some papers? If you try to encode lots and lots of messages with different keys into the video, does it ruin the quality before ruining previous messages?

I'm not sure in how much details I can go since a lot of this is behind patents and NDAs and whatnot, but just googling for "video watermarking" should already yield quite a few interesting details for you.

the algorithm could tell exactly which sources are used in the final video?

impressive!

Are you being sardonic? There are extremely obvious ways to do this particularly since they've got both space, color, sound and particularly time to play with. Unmixing signals isn't rocket science, especially if you've provided the signals that are getting mixed.