One reasonable concern is that if there is no "block size pressure" transaction fees will not be high enough to pay for sufficient mining.
Here's an idea: Reject blocks larger than 1 megabyte that do not include a total reward (subsidy+fees) of at least 50 BTC per megabyte.
"But miners can just include a never broadcast, fee-only transactions to jack up the fees in the block!"
Yes... but if their block gets orphaned then they'll lose those "fake fees" to another miner. I would guess that the incentive to try to push low-bandwidth/CPU miners out of the network would be overwhelmed by the disincentive of losing lots of BTC if you got orphaned.
Since this locks in a minimum fee per transaction MB, what about scaling it with the square of the fees.
For example, every 2016 blocks, it is updated to
sqrt(MAX_BLOCK_SIZE in MB) = median(fees + minting) / 50
or 1MB if lower.
Also, a maximum change from the last value of +/- 25%.
So, MAX_BLOCK_SIZE would be
<=50 median gives 1MB (cost per transaction 1.0X)
100 median gives 4MB (cost per transaction 0.5X)
200 median gives 16MB (cost per transaction 0.25X)
This allows the fee size per transaction to decrease, but only as the total fees increase.
I also agree with the "public good" issue with a miner cartel. Trying to push up the MAX_BLOCK_SIZE might benefit the cartel, but each individual member has an incentive to defect.