The only information that miners absolutely must keep on hand is the set of unspent outputs (UTXO set) because it's possible to allow them to discard all the rest of the historical transaction data using optimizations that have been discussed but not yet implemented. The UTXO set is what some people refer to as unprunable data.
Right now that set is a few hundred megabytes, but if we imagine a billion Bitcoin users all playing Satoshi Dice it would get very large unless there is an incentive for users to limit the number of outputs in their wallet. Especially when the transaction rate is very high miners are probably going to want to keep the UTXO in RAM to speed up verification, which means as the set gets larger their capital equipment costs go up.
So if they effectively subsidize transactions which reduce the UTXO set by requiring a lower fee in the number of outputs is less than the number of inputs they lose a bit in fee revenue but gain in terms of lower hardware requirements.
Once miners start doing this, client developers will have an incentive to include dust collection into their coin selection algorithms and the dust will be cleaned up automatically.
I understand now. Here's the problem, a very long time could pass before the dust carrying cost crosses miners' comfort threshold. And, every miner will have a different threshold so they will start prioritizing at different times. By the time it is in enough miners' best interests to prioritize the transactions in the manner you suggested, we will have incurred a significant and permanent penalty to CPU burden. If only some of the miners prioritize these transactions, they still end up paying the cost for any blocks produced by miners that don't prioritize.
There is too little coupling of the financial incentive to the desired behavior. When I first read your description I thought that there would be some active system of reward. Ideally we would think of something that rewards miners
today for taking transactions that collect outputs. Although it's not clear how that could be done, if at all.