This post is about building untrusted servers, which is somewhat orthogonal to "the snapchat problem."

The best a privacy preserving protocol can hope to achieve is to match the characteristics of its non-privacy preserving equivalent.

Asymmetric PIR (the server transmits an unencrypted bloom filter) falls short of the non-privacy preserving equivalent, since the user makes offline queries that can't be rate limited by the server. Symmetric PIR (the server transmits an encrypted bloom filter) provides identical properties to the non-privacy preserving equivalent, since the user needs to make queries that can be rate limited by the server, even though the queries leak no information to the server.

How that rate limiting (or unlisted numbers, or whatever) works is outside the scope of the protocol, just as it is in the non-privacy preserving case. For TextSecure, you'd need to verify 1 million phone numbers in order to perform your attack. Whether you can do that or not isn't really the point: we simply want to do as least as well as everyone else, without revealing information to the server.