If the threat model is as zipslack describes, then I think a "send to self" mixnet would work.
Patch bitcoin so it generates, a random transaction every, oh, I dunno, eight hours or so, sending, oh, I dunno, 1/10'th or so of your bitcoins back to yourself via a newly generated address.
And to erase your trail in case your wallet gets seized, remove the source transactions and their (spent) keys from your wallet.
After a week of doing that you'll have put 21 extra transactions on the network and, on average, four extra transactions on the coins in your wallet (four because every send typically generates a change transaction). Do it constantly and you'll have an ever-churning wallet that aught to foil any attempt to connect incoming <-> outgoing transactions.
That is all assuming that you don't start with zero bitcoins in your wallet, get exactly 111.11 bitcoins from somebody, spend a couple weeks mixing them, and then pay exactly 111.11 bitcoins to somebody else. That transaction network graph is easy to analyze, and it would be insanely unlikely that you "just happened" to receive exactly those same 111.11-worth of coins that were given out.
Hmm, I wonder if the patch could detect a bad mix and warn you if you tried to do something stupid like that...
Somebody who knows a lot more about mixnets than I do can probably work out the math to know how much, and what type, of randomness to add to the eight hours and 1/10th to make statistical analysis as difficult as possible.