But let's make the crazy assumptions that we are effectively in a world where end-to-end crypto is massively used, virtually by everyone.

I guess it would be stupid to assume that the message are encrypted but not signed.

This means that it would be easy to have a list of identities (i.e., keys) which are sending spams, for instance using a web of trust, without users having to disclose any other informations that "I trust these identities, not these ones" (which could be as easy to do as clicking "spam" or "not spam" for the users).

Now that means that the spammers not only have to encrypt every single email for every single recipient but also to generate new key-pairs for almost each encryption.

Of course it is also very easy to mark as spam any email signed with a key that is considered too small (i.e., too quick to generate).

Now if you tell be that still won't do it without a "centralized spam system with global knowledge", I have to seriously rethink a lot of my assumptions about the cost of some computations.