https://gitlab.freedesktop.org/wayland/wayland/-/issues/159
I have probably chased ghosts deep inside GTK from this one alone for multiple full time weeks, to no avail. There will eventually appear a rather harsh teardown post of some of the major mistakes in the implementation (and that alone, no xml business), wayland-server might not improve, but future designers should be able to stumble upon it.
I think to get further with the discussion we need a sort of list of what mechanisms we consider mechanisms, and how policies project over them. Not to be disrespectful, but my budget for engagement is thin, pandemics, sorry :-/
The politics stuff - so I have some logs that precedes the formation that are quite incendiary and outright vile, but there is enough shit flinging to suffice to stop at there being good reason why I wouldn't consider getting closer than bayonet range, and why the post was obtuse enough to discourage most #metoo and discard me as an idiot who knows nothing. The tech stuff matters though and that's only what I want to see. A protocol and not marriage as if from 'the war of the roses'(1989).
There is a fourth option that might materialise, but that's for then.
On that bug, that problem seems to go deeper than Wayland and I don't know if there is any way to fix that in all cases on Linux due to the nature of SIGSTOP. I know you may not answer this, but wouldn't that also affect X too? Or does it have some other way it deals with it?
The way the disconnect happens is interesting though, I am not certain enough about the threaded-IO implementation in regards to hotplug but a quick glance looks like if a hotplug occurs at the same time a client is disconnected the sparse allocation requirement of open() could get an input device confused as a client as it is cleaned up. The odds of that happening though :D
In the wayland code it looks like they tried to have a buffer strategy at some point, but then got too clever for their own good (epoll designs + asynch-OO projected over C is... yeah) and somewhere in 'closure/write/buffering to lower #syscalls/callbacks' landed in a long chain of error propagation and immediate kill.
The best way I found (hence using it) is to have memory mapped ring buffers and just have an atomic head-tail indicator. This makes file-descriptor transfers (DuplicateHandle from win32 fame is just so much prettier) messier, but it means that you can have a recovery strategy and see the backpressure buildup and rate-limit. Since it is mostly input events or 'drag-resize' (and those can be merged) in that direction, silently dropping them and have mouse cursor 'jank' is better than losing your work. If all else fails, the watchdog handle is pulled and the client is set to migrate to a fallback server (if set).
I see what you mean with that document, but if you have been watching then none of that stuff is new or unique to Wayland. What you may have missed is that this document is a reaction to the politics that already were around in the X days. (I was alluding to this earlier when I was talking about WM atoms, the popular desktops already have their own sets of policies there that are necessary and incompatible and in some ways do limit the expressive power of clients)
Not working with networked clients is a minor detail. It's trivial to have a user upload a compositor script to a different machine. If it becomes annoying to do it manually then a client that needs more features simply needs an automated way to do this. Still no need to touch the Wayland portions there either. I actually do appreciate the efforts to fix these issues with the whole shmif setup and I sympathize, but it's the same uphill political battle if you want to push that out to the rest of the ecosystem.