...
The argument is that unless there is a hard block size limit, miners are incentivised to include any transaction no matter how small its fee because the cost of doing so is practically zero (less than a microdollar, according to Gavins calculations). Therefore if a bunch of transactions stack up in the memory pool that pay a smaller percentage than "normal", some miner will include them anyway because it costs nothing to do so and maximizes short term profit. Hence, you get a race to the bottom and you need some kind of hard network rule saying you can't do that. We already have one in the form of block byte size, so the debate becomes "let's keep the size limit" vs "let's remove it".
Hang on a second. Am I missing something? I don't think miners need a hard block size limit to have incentive to stop accepting transactions. They will do so because there is always a time limit.
The difficulty target is adjusted to regulate time between blocks, and results in a target with a probability a correct hash will be found within a certain time (regardless the total hashing power of the network). Every second a miner waits to include more transactions in their block the probability is increased a competing miner will find a correct hash for their own block.
When the network receives more than one valid block version within close time proximity it holds both and waits to see which is extended longest breaking the tie. Again, every second a miner waits to announce a block it increases the chance their found block won't be permanently regarded as valid. Physical block size is not a factor in that, and miners will naturally stuff their block with most profitable transactions first.