> Not just comment: we should document that every implementation of
> gui-get-selection should return either a valid timestamp (and say what
> is a valid timestamp) or nil, meaning 'CLIPBOARD 'TIMESTAMP is
> unsupported.

You are right that it is a dangerous assumption, and it's also hard to
test and probably unnecessary. So I have changed the code to check the
window-system variable instead and only use timestamps for X and haiku
(I don't actually know the list of systems that support it, I have
included haiku because it is also in the list of systems that support
reliably gui-backend-selection-owner-p, according to the code).

I also have defined the functions
gui--set-last-clipboard/primary-selection and
gui--clipboard/primary-selection-unchanged-p to just change the lines
(setq gui--last-selected-text-clipboard/primary text) and (equal text
gui--last-selected-text-clipboard/primary) with them. They behave
exactly the same for systems other than X and haiku, and are backwards
compatible with setting or reading directly the original
gui--last-selected-text-clipboard/primary variable in other places.

I have assumed this last thing (being backwards compatible) was the
reason why Po Lu suggested to use text properties instead of changing
the gui--last-selected-text-clipboard/primary variable, so I have taken
the liberty of moving away from that idea and using the new variables
gui--last-selection-timestamp-clipboard and
gui--last-selection-timestamp-primary instead, which simplify the code.

So here is the final, hopefully definitive patch, attached to the email
and commented below: