The problem that selfish large miners create millions of transactions could be alleviated by using a median (instead of a mean) statistic in the adjustment estimation, which is much less susceptible to extreme values. Maybe one could also do statistical correction based on IP adresses (those that frequently submit only blocks with a excessively huge number of transactions get less weight).
This is starting to sound hairy to me. I can easily imagine that 60% of the largest miners would benefit sufficiently from the loss of the weakest 20% of miners that it's profitable for them to all include some number of plausible-looking transactions between their addresses (thereby causing an inflated median). I feel that anything involving IP addresses is prone to abuse and much worse than the admittedly ugly fixed-growth proposal.
Of course "necessary" has to be defined. I think it is acceptable to make Bitcoin progressively more unviable (through higher fees) for microtransactions if decentralization is at risk. Very small transactions could also happen off-the-chain. However what "small" means is open to debate.
My own feeling is that we should be looking at "as much block-space as possible given the decentralisation requirement" rather than "as little block-space as necessary given current usage". However, if you can find an appealing notions of necessity, smallness, or some alternative method of attempting to balance centralisation risk against utility which involves fewer magic numbers and uncertainty than the fixed-growth proposal then it's certainly worth it's own thread in the development section.