Exactly. I've never had the time to dig into it, but these LIFO-optimality results always strike me as being non-physical or introducing some implicit hard-to-model effects.[1]
Like you say, people can just re-enter the queue, or (in meatspace) form a meta-queue that vies for entering the moment it has an open slot, reproducing FIFO all over again.
Furthermore, you get a lot of pro-cooperation effects (necessary for queues to work at all) by giving people "skin in the game" in the form of valuing their place in line. Once people lose nothing by inventing a new name and re-entering the line, that's all gone, and they no longer have an incentive not to be disruptive, which throws off disproportionate negative utility onto the rest of the system.
What's wrong with degrading from LIFO to a random queue due to client retry, when you're starting from a FIFO that is already failing client expectations?
Nothing, if it's just used for that (narrow, least-bad, degraded) situation. I was addressing the more general point that LIFO queues have some actionable superiority over FIFO in the general case because of the theoretical results cited in the literature.
Like you say, people can just re-enter the queue, or (in meatspace) form a meta-queue that vies for entering the moment it has an open slot, reproducing FIFO all over again.
Furthermore, you get a lot of pro-cooperation effects (necessary for queues to work at all) by giving people "skin in the game" in the form of valuing their place in line. Once people lose nothing by inventing a new name and re-entering the line, that's all gone, and they no longer have an incentive not to be disruptive, which throws off disproportionate negative utility onto the rest of the system.
One day, I promise, I will unpack this result.
[1] My comment from 2015 expressing similar reservations: https://news.ycombinator.com/item?id=10182781