I'm guessing that having every book loaded into a password cracking database, subdivided and indexed by each leading phrase word, is still computationally infeasible for non-government actors.
If I walk into a library, pick a floor, aisle, shelf, book, and page at random (just walk, don't think about it), and use a phrase that is a minimum of 12 words long -- is that more random than what I presume happened here, where someone knew that their target liked that style of poetry and was able to concentrate their search on that genre? ( a "crib" in Bletchley Park terms)
The comments about English grammar are correct - classes of words (nouns, verbs, adverbs, etc) do fall in certain positional order and frequency analysis becomes important. A brute-force attacker would have to work through four types of passwords - the commonly used passwords like "12345" and "letmein", language-based phrases (like my not-great idea), language-based phrases with letter substitution (leet-speak, etc), and then truly random letter sequences.
What's happening is that people collect endless phrases and alter them with a ton of standard manipulation schemes, compute the corresponding private and public keys & addresses for all the variations, create a lookup table for the addresses and private keys, and as soon as they see a known keypair in use then they use the corresponding private key to swipe the funds.
See my comment above - unless I'm mistaken, taking all 2 to 24 word quotes from the most popular 1 million novels gives you about 40 bits of entropy (less than a password of length 7), and can easily be stored on one hard drive. In other words, feasible even for some script kiddie in mom's basement.
No need to have every book loaded, only the top 50000~ read by people who would use that method of passphrase generation should work fine (and be feasible for almost everyone). Cryptonomicon would probably be in that list.