Consider an attacker who wants to publish N blocks. They compute N blocks at base pow. If there is no pow scaling, they publish N blocks, N blocks are confirmed, all done.
Now if pow scales to 8x, they have to spend on average 8x more effort finding pow for those blocks. However, the pow solutions they find along the way at base pow are still potentially useful to them as spam, not for their intended purpose but as 'bonus' spam for a marginal cost near zero. They can't guarantee specific transactions will be published at 8x, but they know that on average 1 in 8 of the base pow they find will have 8x pow. For the purposes of the attacker, this is good enough. So they publish roughly 8N blocks, approximately N of which have the desired 8x pow but the remaining 7N of which come as low priority spam for nearly zero marginal cost to the attacker. Though nodes may still prioritize blocks based on pow perfectly, the attacker has not been deterred from posting blocks relative to the average user. In fact, the attacker has been encouraged to post otherwise-unnecessary low pow blocks, potentially making the end-ledger state more unprunably bloated than if there were no pow scaling at all (because the attacker could have achieved their goal with only N base pow blocks without it, and has now published 8N blocks)
In summary, PoW scaling seems to actually perversely encourage an attacker to cause more ledger bloat than they otherwise would have.
@NanoOrca identified this problem, though perhaps not the implications for the attacker, in this thread and provided one solution, and @PlasmaPower provided another improved solution. Basically, add the intended difficulty to the nonce. It now seems necessary to me that some sort of such system be implemented if PoW scaling is to serve its intended purpose.