I don't understand why it would have to compromise in either way.
If you trully want censorship resistance, having an option to transfer any nostr event without relays is the way to go. Especially if the event is private.
Discussion
For your use case you need ephemeral pubkeys sending encrypted messages.
The design of nostr builds its censorship resistance on the fact that events are not bound to any one relay. They can be re-submitted to different relays. But relays have to filter spam, which is hard with ephemeral keys sending encrypted messages.
This dilemma already led to paid relays that only let paying pubkeys store events there or to throttling and time-constraints, making it hard to re-submit all my events to another relay.
I think, we need relays that charge sats first thing when opening the websocket and then credit that balance with what gets requested and sent. Then, relays could charge for example x3 for encrypted events and x2 for ephemeral authors, all without the need for long-lived pubkey-relay relationships. With these things in place, GiftWraps are ok. Without DM and GiftWraps, maybe we don't need these things. (I think we should have these things regardless.)
I agree with most of this. But I don't think GiftWraps make spam filtering impossible. It's just a different way to do it. Because the event is private, the relay must rely on Reports (or Deletion events) from the author of the p-Tag that the GiftWrap is sent to. Clients can help relays filter these events out when they are in-fact spam.
I think people are waaay to stuck on the current way of filtering spam. Also, I expect to see relay operators that will only deal with GiftWraps. And I would go further to say that they will become the most used relays because most of the information traveling through Nostr will be private in the future.
To me, P2P is now a must. If we truly care about decentralization, P2P should be an option in the toolkit. Relays will always be there as a fallback, but they don't need to be the only game in town.