Why not a adjustable limit based on a previous median timeframe?
It depends on what purpose you want the limit for. If it's to protect from unusually high loads as the original limit was for then something along the lines you suggest would work (and I've shown support for similar schemes). If, for some reason I can't fathom, you think 1 million bytes per block is a magically correct number for all time then it's not even open for consideration (and you might also not see a problem with foot-binding).
I have no numbers to present. I am just spitballing and am trying to figure out the root of why this hasn't been fixed as it seems to be a pretty academic problem.
I would think that having an agreed upon median would allow the algo to stabilize itself depending on market forces.
This is the first time I have known that this was the issue and appreciate Marcus boiling it down so nicely.