The simplest form of this is literally just a token which replaces the universal identity. So you ban the token, you ban the user. This only applies for voting anyway, since commenting and posting follows the plaintext user agent.

A more robust trust model with rotating tokens would fully move ban enforcement to the home instance, which I actually believe is already the case in some situations. Eg, when I am banned from a specific community on another instance it seems as if my host instance knows not to even display a vote on the UI, which suggests that it has knowledge of my federated ban. With this trust model it would be possible to fully enforce cryptographically secure forward security as well.