I agree with your analysis that the conditions for it to occur don’t require a hostile adversary, a large enough number of mildly strategic users trying to save some gas could lead to this sort of instability causing unexpectedly high state growth and/or unexpectedly low base fees. However it isn’t difficult to see that the base fee drop across a single cycle of the base fee oscillation is substantially smaller than the amplitude of such oscillation (assuming that the adversary controls a relatively small share of transactions), so I was thinking that a strategic user would have a perverse incentive to delay transactions and send them as batches of double-full blocks in order to both profit off the periodic gas price drops due to the base fee oscillation and amplify any preexisting oscillation, even if they have no interest in depressing the base fee in a long-term attack as you were describing in your post.
This has the potential of negating any usability improvements from this proposal, because: 1/ The oscillation would increase the chances of users overpaying above the “fair” market-clearing price, giving them a perverse incentive to bid below their marginal utility (which would reinforce any preexisting oscillation). 2/ Because it would potentially create periodic times of congestion where the pricing mechanism reverts to a first-price auction. 3/ Because of the rather unpredictable state growth I was complaining about earlier in this thread.
I agree that making the base fee update formula path-independent as in EIP-3416 is a move in the right direction: It would certainly fix the state growth unpredictability and the long-term base fee drop you were describing. However it wouldn’t remove the incentive for users to cause such oscillations, because, if my understanding is correct, in EIP-3416 the base fee paid by the transactions of a block is only dependent on the gas usage of the previous blocks (just as is the case for this proposal), so the submitter of a double-full block doesn’t pay any penalty whatsoever (only the subsequent blocks get to pay the price…), and this short-term attack would still be profitable with either proposal.