Seems like a nice way to artificially reduce the network hashing rate / temporarily split the network.
I use a number of nodes to send 2 transactions spending the same input to the network, but I make sure to send them as close as possible to the same time. Thus, the rest of the network see that there is a double spend, but there isn't a strong consensus as to which order the transactions arrived, so half the network mine a fork with one transaction and the other half mine a fork with the other. At some point in the future, one side decides it must have been wrong and one of the forks wins. Up until that time, each fork has been competing possibly with many reorgs.
That's for a single double spend. Now think about a collection of nodes which continuously pump out double spends. Each split halves the hashing power if perfectly timed, it doesn't take much to bring the network to its knees.
Great attack. One solution, as others have suggested, is to make the penalty time dependent. You apply no penalty when the race took a few seconds, than gradually diminish the Finney block's difficulty for higher delays as you are more and more sure most of the network agrees with your order. So rapid double spends will not halve the hashing rate (but will be detected at the store) while delayed broadcasts will shave <1% of the hashing power (the poorest connected miners, who will tend to be the same regardless on where you inject subsequent double spends).
This makes sense from a rationality perspective: it's irrational to extend a shorter chain if you don't know the ratio of altruistic nodes that share your view. For higher broadcast delays, when are you are sure altruistic nodes have detected the correct order, it's profitable for you to follow their penalty scheme, for fear they will tend not to extend your work otherwise.
But what happens if the Attacker Mines a Block at T=12 that contains Transaction 2?
In this case Miner group A will not accept this Block, but Miner group B will continue on the Block and the chain will Fork.
If both Mining Groups have the same Hashing power the fork can last for a longer Time.
It's not a good solution to put a discrete limit, like 10 seconds, after which penalty jumps, because you invite races against that limit. I propose a gradual penalty proportional with the delay. So that your scenario reduces to the classic Finney attack with a slight dispersion in penalty across the network, not enough to split the hashing power but enough to incentivize miners you not to attempt it and thus increase the cost of zero-conf double spends.
The correct Penalty[delay] function to achieve this needs to be experimentally or mathematically explored, as well as the resulting zero-confirmation payment you can "safely" accept. It's more than a tweak tho, it's a major modification to the blockchain extension rules...