I can think of two big improvements to desktop GUIs:

1. Incremental narrowing for all selection tasks like the Helm [0] extension for Emacs.

Whenever there is a list of choices, all choices should be displayed, and this list should be filterable in real time by typing. This should go further than what Helm provides, e.g. you should be able to filter a partially filtered list in a different way. No matter how complex your filtering, all results should appear within 10 ms or so. This should include things like full text search of all local documents on the machine. This will probably require extensive indexing, so it needs to be tightly integrated with all software so the indexes stay in sync with the data.

2. Pervasive support for mouse gestures.

This effectively increases the number of mouse buttons. Some tasks are fastest with keyboard, and some are fastest with mouse, but switching between the two costs time. Increasing the effective number of buttons increases the number of tasks that are fastest with mouse and reduces need for switching.

[0] https://emacs-helm.github.io/helm/

All of the other examples you gave are products constrained by physical reality with a small set of countable use-cases. I don't think computer operating systems are simply mature appliance-like products that have been optimized down their current design. I think there is a lot of potential that hasn't been realized because the very few players in the operating system space have been been hill-climbing towards a local maxima set by path dependence 40 years ago.

> Take any other praxis that's reached the 'appliance' stage that you use in your daily life from washing machines, ovens, coffee makers, cars ...

I wish the same could be said of car UX these days but clearly that has regressed away from optimal.