Yeah  I’m mistaken about the protocol layer as the place where mutes of replies and mentions should happen. Based on some conversations today it makes the most sense that the clients and relays screen for these. 

In terms of filtering out the random jerks - other protocols are experimenting with mute lists - similar to what you suggest above and I think it’s a good option for Nostr as well. A user can subscribe to a mute list of their choice based on whatever criteria they have - it could be words / content or people. Another framing is feed curation. 

The beauty of Nostr at its basic level is that the user is in control of their feed, rather than an algorithm. Sorting through harassment challenges makes that control truly possible.