Again, you haven't used it. You're working with a theory, and really have no idea what you're going on about.
Latency increases somewhat under load (as it must), though not appreciably enough to affect any of the things we do with it. Jitter is very low as well. No matter how hard people or things hit the network, the user experience remains very responsive for interactive tasks...perceptibly the same as it is with an unladen connection.
This, as opposed to hundreds of torrent peers hammering away, one or more Netflix stream soaking up as much as it can get its hands on, and et cetera: Without QoS (and I didn't name it that, such terminology has been in place for quite a long time as relating to this sort of technique), this network was essentially unusable.
And now, it works reasonably. Individual TCP or UDP sessions are placed into groups with other similar sessions, and those groups have their own assigned priorities. This can be done by port, IP/MAC address, or deep packet inspection, or the amount of data the session has used, or even DiffServ flags.
It even has a fancy GUI that actually works.
You sound a lot like people used to sound back in the day, proclaiming that NAT (or ipmasq as it was more-commonly known at the time) could never successfully allow FTP, ping, or traceroute to seamlessly work. They'd list a lot of seemingly-logical reasons as to why it can't work and never will work, and then go on a long-winded rant about why either proxy servers or public IP assignments or at least one-to-one NAT is the only way.
Fast forward, and those people have STFU because -- gosh -- NAT works and does these things. They were ignorant of the possibilities of creative people making creative solutions.
I mean, sure: "Proper" QoS (ie: DiffServ and sensible routers with sensible queues and routes from end to end) might be nice. Maybe it even works on a private network. It doesn't work on the greater Internet, though, as you yourself say.
So rather than say "fuck it, I give up, there's nothing to do," I've simply solved the contention issues of my own grossly-overburdened last mile. And I've done it all from one side of the pipe.
If that seems impossible, then you're the ignorant one. There is a world of things that you did not learn in school, and some of them actually solve real problems that people experience. Nothing of this universe is so rigidly-defined that it cannot be adjusted in some useful way.
If you want to learn about it, though, download Tomato and spend an hour or two playing with it on a relatively-saturated network. Then read the source code if you still think it can't work that way.