Rejecting blocks based on observed double spends also seems problematic. It would let me freeze the block chain by generating lots of double spends and sending them directly to major miner nodes in random order. Every miner would then generate a block that contained some transactions other nodes would perceive as double spends and so every node would reject the block, allowing me to catch up with the head of the chain.
I think it is a reasonable assumption that major miners will be well-connected with each other. There is certainly a strong incentive for miners to be well-connected in general (better connected == more likely to win 'block races').
So I don't see how you could freeze the block chain-- if you generate lots of double-spends, the miners will quickly see both of spends and will drop those transactions like hot potatoes. The "finney attack" only works if the first double-spend is generated by a miner that finds a block and includes it in the block without transmitting it.
Also, my proposal was to only reject blocks containing 'suspicious' transactions that you hadn't seen transmitted that have a double-spend attempt before the next block.