"I'm sure anyone spending more than $750 on amazon is probably happy enough waiting for two or more confirmations."
Maybe, but if I can do it without waiting by just sticking with credit cards, why wouldn't I?
"Given that double spends are reasonably expensive"
I see this being tossed around quite a bit. I am not entirely sure where this idea came from. Can you prove it, the way that ECDSA (on which Bitcoin is based) can be proved secure as a signature system? When I hear this sort of statement, it sounds a lot like, "Nobody has published any way to double spend cheaply, and I cannot think of one; therefore it must be an expensive thing to do!"
I do not mean to be insulting with that, I really would like to know why people think double spending attacks are necessarily expensive. On top of that, how do we know that a double spending attack would be more expensive than the payoff -- especially when you are dealing with a complex marketplace, where a double spending attack may be just one component of a larger scheme. Maybe the point will just be to deny another person the ability to buy something, by double-spending until the merchant is sold out of the item, or as shills in an auction to try to inflate the price (should Amazon revive its auctions system).
This reminds me of the sort of argument I had with Hushmail users back before the infamous DEA incident. Your secret key is sitting there, on their servers; they could read your mail if they wanted to or if someone had a gun to their head. People said it couldn't happen, that the key would only be in RAM for a split second, that using their Java applet solved the problem, etc. Then the DEA turned up in court with a mountain of decrypted email from Hushmail (and then, as if the entire world just wanted to aggravate me, people starting saying that PGP had security problems). Cryptography tends to be black-and-white when it comes to security: you are either secure, or you are not secure, and there is rarely anything in-between (with multiparty computation, that might be changing; people are now talking about the "covert model" and "one-bit leakage" notions of security, which are in that middle area).