However, the primary axiom is of a web server that is serving static content that is located on the device. That, in a sense, gives almost infinite time and knowledge to deciding prioritisation of work.
For HTTP proxies that have to fetch resources from some other server, the time and knowledge window becomes much shorter. Decision making needs to be done quickly in order to minimise the latency of returning anything. And the upstream server may response in unexpected ways.
This is the challenge for today's CDN deployments. And developing a solution that works for the huge variety of how websites are developed is the really interesting bit.
Even if the proxy does cache locally to memory or disk. Without infinite space, there will be some chance of cache eviction and a need to fetch from the upstream again.
Our scenarios are different ends of a broad spectrum, and there's a lot in between. Which is why getting prioritisation right is a hard nut to crack.
More dramatic improvement might be possible if streams can be interrupted and resumed (which is something protocols like RSocket support): https://www.cs.cmu.edu/~harchol/Papers/MIT-LCS-TR-767.pdf
Apologies if you're already familiar with these.