Or a coalition of pools will play it the other way around.
They would insert only fees with low transactions into their blocks thus making the next block smaller but if filled correctly would trigger the next block to be larger and they could stuff also higher fees in this, you can't force miners to propagate a block they find if they don't want to. Of course, it will need also quite a bit of luck alongside hashing power but they could still try to game it, this will turn ugly once the fees will be the dominant reward, miners will have an incentive for selfish mining.
In the end, it's normal if you open even the smallest door of opportunity someone will try to make more money than the others.
What you're suggesting here is risky for the miner (it's like taking a gamble to earn more money). I'd agree that we can remove the volume of transactions out of equation and just use the average transaction fee to calculate the limit on the next block but I don't buy that having an adjustable blocksize cap would make it easier for anyone to "game the system".
In my opinion, the ideal solution is to adjust the blocksize cap in real time, not to fix a limit based on future predictions. To anyone talking about an off chain solution, just so you know the L1 network is the barrier to entry to the L2 network.