You could also use something like Zerotier-one to recover from network being shortly disrupted. This service connects your machines into a virtual network with static IPs, so the IP addresses through which your machines talk will stay the same and TCP connections will recover and ssh will keep working if a device briefly goes offline.
I don't think you're giving the Dota client enough credit there. Game clients don't wait for network to pretend to respond to user input, because it absolutely is noticeable at 100ms latency. Instead, they predict as much as they can client side and display it as fast as possible. For this reason, the client is not at all thin. It knows how to run the whole program by itself. Not so with remote X11. With remote X programs, the client doesn't know what happens on click for anything, so the user really does get hit with a 100ms delay for every interaction with the program.
Although if I connect to something that's 20ms away (within the same country), the experience is really good.
Dota is fine because it has network prediction and a finite and well defined set of inputs.
And that's also ignoring video compression artefacts which are that you introduce if you send a modern GFX pipeline across the network. Computers screens can no longer be described with just filled rectangles and text.
That is how we have been giving development machines since 2011, Windows workstations on whatever cloud vendor the project is taking place.
The best is always going to be to use a local copy of the logic and data needed most frequently/immediately. E.g. in the VS Code example, they keep the buffer you're editing locally, the logic for cursor movements, the errors you're flipping through, and maybe even some precached autocompletion/highlight stuff. You don't have to wait for a round trip to see your keystrokes show up. That's the advantage of splitting the system in the right place.
But that's never where I want the split to be when working across the network.
Remote storage? Sure, sign me up. (The POSIX APIs are horribly unsuitable for network filesystems, but I'm speaking about the concept more than the current implementations.)
Remote heavy computations, like AI workloads and compilation? Definitely.
Remote GUI code? No ugh. Compare the experience of VS Code Remote vs just running VS Code remotely with ssh X11 forwarding, using a high latency connection.
You are fighting the speed of light. It won't go well.