Bumping up against the hard limit is probably wastefully expensive for this "attack"; you only need to produce blocks bigger than the (expected, since you don't know in advance) median to cause the new median to shift higher.
I'm not sure about this, since it depends on your objective. We can model this simply by viewing the calculation by epochs rather than a sliding window. If you create a set of blocks with size limit+10% in each epoch, the limit will only grow by 10% per epoch. If you double the size, the limit will double per epoch. So this allows
much faster growth (albeit at higher cost in penalty). But other objective functions are possible certainly. It would be a relatively simple matter to express them mathematically and maximize.
Hmm, the quadratic increase in penalties suggests to me that it'd be cheaper to do it over time, but maybe not. Either way, I don't see how you can raise it above the amount "allowed" by your hash rate.
Bumping up against the hard limit is probably wastefully expensive for this "attack"
What expense?
You're suggesting mining is (or can be) free? That's absurd. Even if it were free, this attack still costs you the reward.