I'll admit it's unlikely that a single user / group is going to 51% the network (and in that case we're all hosed anyway), but you know that in that case, they could perpetuate a fork for a very long time in the darkness, and break it out later when they are ready to "unleash the fraud" on everyone, after the merchandise is received and they have permanently left the scene of the crime.
I am not an expert on Bitcoin but I think this kind of attack is possible. In that case as long as you spend your bad blocks before they can be invalidated by a longer chain, it becomes someone else's problem. If someone who knows more than me can enlighten us to why this attack would not work, I hope there is another reason than "it's unlikely that anyone would be able to 51% mighty Bitcoin now."
It's a matter of cost, not a possible vs impossible thing.
I think at 30TH/s, you'd need 500 new ASIC miners, and $500/day in power, to hit 50%. But that's just arbitrary - 2% would allow some attacks, 99% would allow more profitable ones.
At 50% of the network you'd theoretically solve 50% of the blocks, gaining 25BTC every 20m, or 72 times per day, for a "street value" of $80,000. During the day maybe $8M USD in BTC is traded.
You're talking about spending your money twice. Writing a transaction sending the BTC to someone and letting them record the transaction (transfer it to "the network") while trying to generate the next block containing which contains a matching transaction from that wallet to another wallet you control. If you fall behind you simply pretend you never tried and lose nothing by trying, except the costs of the transaction. If you win you can reveal your 1-unit chain to gain $1100 (your rightful reward for mining) or you can hide it to wait for a double-spend opportunity and try to lengthen your chain.
Meanwhile the merchant takes your transaction and sends it to the network and waits. If they're trusting they send you your merchandise now, but usually they wait for one or more blocks to be published to verify the transaction. Let's say that you can see the instant they do and their action is irrevocable. This is the best state for you. You then "simply" need to win more blocks than the rest of the network for however long until you see the transaction be finalized and reveal your chain with your blocking transaction at the head, causing the transfer to the merchant to be ignored by the rest of the network after renegotiation.
This simple plan won't work well though because when the network resyncs it's obvious which transactions are double-spent and your BTC will be known and people can refuse to deal with wallets that receive stolen coins - turning it into a painful single-spend, and loss of your mining earnings. So you need to actually coordinate a double-spend, getting something else irrevocably sent from another merchant with the same BTC. This is where it gets hard because each merchant is watching the same global pool for their confirmation and would see you spending the same BTC in another transaction.
Double-spends can only really happen when merchants are on separate blockchains already, as with the recent bug.
You're looking for something that can be converted into directly and efficiently back into BTC because if you don't manage your double-spend you're going to own this thing - this is like the house's cut. And something commodity and untraceable so you can unload it when hot. The best-case would be if someone was sending gold-bars via anonymous remailers - you'd only lose shipping costs.
A seller selling one-of-a-kind yachts which they special deliver to customers has nothing to worry about. In the months it takes to build and deliver a yacht the theoretical 51% enemy has earned enough for many yachts and doesn't need to double-spend. But a gold-bar dealer - they're justifiably worried. Luckily real-time gold shipments aren't a big deal. They can wait a day.
And when you double-spend you'd essentially lose your BTC from mining as the accounts they were paid to for the day would be blacklisted. So unless you bought more gold than your expected daily take in BTC you'd have been better off mining.
And remember that you're buying a bar of gold for every block you win and trying to get ahead enough at any point to manage to not pay so you've got to have a big bankroll and keep from going bust from transactional and infrastructure costs until you sell those incoming bars you ended up having to pay for (ie, all but the last one).
All this while competing with index funds and other far safer bets....
If you're dealing with an extremely technical adversary who you feel would spend $1M + $100K/day to mess with you, worry and plan appropriately to mitigate risk. General merchants are relatively safe though.
I am not an expert on Bitcoin but I think this kind of attack is possible. In that case as long as you spend your bad blocks before they can be invalidated by a longer chain, it becomes someone else's problem. If someone who knows more than me can enlighten us to why this attack would not work, I hope there is another reason than "it's unlikely that anyone would be able to 51% mighty Bitcoin now."