For simplicity, we assume that the only value in the block reward is the mining fees.
For simplicity, we assume that there are no limits in block size, and that a reversing miner can take all fees in both blocks.

E(mining block) = sum of all fees on current block * p(mining block t)
E(mining previous block to realloc fees) = (sum of all fees on current block + prev block) * p(mining block t-1) * p(mining block t)
E(mining prev) > E(mining current) when:  (prev fees + current fees) * p(mining) * p(mining) > current fees * p(mining).
p(mining) := 0.5: E(mining prev) > E(mining current) when prev fees > current fees
p(mining) := 0.4: E(mining prev) > E(mining current) when prev fees > current fees * 1.5

So, if all transactions contribute to a 'fee pool', and the fee pool pays out exactly 33% of its contents every block, then this type of attack only makes sense if you have greater than 40% of the hashing power, regardless of how big the drop in new fees is from block to block.

Unfortunately, this attack makes sense to a short-term-greedy attacker with 50% hashing power regardless of what percentage of the mining pool pays out each block, assuming that all txns from both blocks will fit into a single reorg'd block.

Furthermore, a decay in the form that I was talking about (to incentivize txn propagation) is a bad idea, because it increases the incentive to reorg a block (by including a txn in an earlier block, you catch it at an earlier point-of-decay, which increases its value).