Another possibly secure-ish system?
Generate a 'fake' email address, with the username being the recipient's username and some other identifier (so 'bob_jones' becomes 'bob_jones2934'). Then encrypt the name with the recipient's email server's public key so it's nonsensical (x0fj38keo435f325).
Encrypt the body of the message as well, using the recipient's public key and perhaps hiding that in an attachment with stenography. Or linking to an image containing the message.
Then spam it out to a million servers worldwide.
The receiving server will look at it's 'junk mail' list and try to decrypt any nonsensical addresses with it's private key. It'll find the one you've sent coming from a now-sensible address, look up it's address book to find the intended recipient and pass it along to them. There are now 1 million people who've received your email, 999,999 of whom have ignored it assuming it's a Viagra offer. Only one can read it.
By applying other obfuscation methods the NSA would have to try and crack every spam email. They'd have to track down every meaningless or meaningful combination of symbols that could possibly be combined in a way known only to the end user. Follow every single link, and parse those links looking for clues as to whether this even WAS a message hiding something. If they DID find it, they'd need to crack what should be an infeasible-to-crack sized key.
The only thing that could ruin this is behaviour analysis- if you find that 2 nodes are bouncing spam back and forward 10 times an hour it's possibly an email conversation. But again they'd need to know to look for at least one of the servers. With random email addresses at the destination you'd only be able to narrow it down to a per-server level, too. So multiple secure conversations could be taking place.
You know, given the number of spam emails I get anyway, this probably already happens.