You've stumbled on an actual, practical use case: encoding watermarks in the grain. I can imagine a watermark hidden in the grain, spread over long enough time (say couple seconds to a minute), so it can survive reencoding with heavy compression and containing enough bits to identify the source of the video on a per-copy/per-user basis.
Modern video codecs almost entirely remove grain and replace it with a lightweight description of what it looked like so they can generate a new and random grain that looks pretty much the same.
