Ok first of all, NIP-05 is out of the question. It's not signed, it's not a contender here. That was just to make it easier to find someone, back when we didn't have better ways. It was my idea to put it into nostr.json in the first place, so I am shooting down my own idea.

My newer idea is expressed in NIP-65, which is just to spread that kind-10002 event to hundreds of relays (while only using a few relays for regular events).

I've always been in favor of relays or servers that specialize in serving kind-10002 events. I was going to suggest if someone does a DHT to put the kind-10002 events in there. Sounds like that is already part of your idea.

It seems that if all the clients move over to publishing into a DHT, then a DHT would have the advantage that all the data would be in it (the hash table being synchronized, I presume), whereas if you just ask a popular relay, it might not know, and you could never know for sure if that was because the relay didn't know or because the data wasn't there. BUT that property doesn't appear unless and until everybody moves to using the DHT. And I'm not sure people will because of the chicken-and-egg problem.

But a DHT seems like (even though it is distributed) a potential single source of failure. If there are "root servers", how many are there? Can they cheat?. Can things be removed from or replaced in a DHT, and if so, cant someone censor data out of it, and then replicate that censorship widely? If not, how is that enforced? I admit to my ignorance on this point.