An attacker can rely on this (the transaction ID from their perspective never confirmed) and ask them to resend the funds. The attacker doubles their money and suddenly Gox has outputs they think aren't spent, but really have been in a transaction ID they don't know is their own. When they roll these "unspent" outputs into new transactions they fail to broadcast, and then we are in the situation we are in today with a number of backlogged transactions.
Fortunately for them if any money has been stolen using this, they probably have the ID of the person in question.
Your system would work as a hack, but for somebody as big as Gox they would have a significant impact on the blockchain size for no good reason.
MtGox made a very bad decision in its choice of wording in this press release. It shouldn't have framed this as a "design flaw." The only reason this happened was because of MtGox's custom software. No one else was affected. By definition, that's not a design flaw.
Of course that's much harder, both in terms of algorithmic and code complexity, than just remembering a hash.
And, it doesn't require global agreement to deploy such a workaround.
Now, any transaction coming from that temporary address (which can also be told to the user as "expect the money to come from this address", which might be separately useful for purposes other than transaction proofing) can be considered to be the transaction in question: the computed hash of the transaction is irrelevant.
Of course, I should not be able to solve this problem after two minutes of thinking about it; so: anyone mind teaching me what I'm missing? I can't imagine I could come up with a solution this simple so quickly to a problem that is apparently so well known and so problematic to such an established player in this space ;P.