>>> Please note that you have to handle not only format-strings of
>>> ‘message’, but also ‘error’ and even more low-level ‘format’, i.e. all
>>> these (error STRING &rest ARGS) (message FORMAT-STRING &rest ARGS)
>>> (format-message STRING &rest OBJECTS) (format STRING &rest OBJECTS)
>>>
>> I expect that 'format' won't translate its first argument, whereas
>> 'error', 'message', and 'format-message' will. This will be for the same
>> reason that 'format' does not translate quotes.
>
> Then it should be sufficient to add a gettext call to 'format-message' only,
> because all other related functions 'message', 'error', 'tramp-message',
> 'tramp-error', etc. all they use 'format-message' directly or indirectly.

Maybe I'm too stupid to comprehend the complexity of this task in its entirety,
but I tried to install gettext infrastructure in Emacs with gettextize,
and then tried to run xgettext on source code, and see no technical problems.
What I tried is to run this command, and it extracts all messages:

  xgettext --from-code=UTF-8 -kformat-message -kmessage -kerror -ktramp-message -ktramp-error *.el

then this command extracts all Gnus messages into a separate file:

  xgettext --from-code=UTF-8 -kformat-message -kmessage -kerror gnus/*.el -o gnus_messages.po

this command extracts all menu items:

  xgettext --from-code=UTF-8 -kmenu-item *.el **/*.el -o menus.po

and this extracts all docstrings:

  xgettext --from-code=UTF-8 -kdefcustom:3 -kdefvar:3 -kdefun:3 *.el **/*.el -o docstrings.po

The size of docstrings.po is about 9MB, so perhaps it should reside in
a separate catalog defined by e.g.

  (defdomain emacs-docstrings

with semantics similar to defgroup, but I have no opinion about this.

I think this project urgently needs a coordinator: to negotiate with
package authors and translation teams about how to better split
translations to message catalogs.  So there are not so much technical
problems, but mostly organizational ones.

> IIUC, using standard gettext functions this would rather correspond to
>
>   (message (ngettext "Replaced %1$d occurrence%s"
>                      "Replaced %1$d occurrences%s"
>                      replace-count)

It seems better to start with this standard function
and add more optimizations like ‘nmessage’ later.

Other Lisp implementations use ‘ngettext’ as well, e.g.:
https://clisp.sourceforge.io/impnotes.html#ggettext

So I'm going to start with more obvious parts of the task
by fixing the current bugs of incorrect English syntax
in a forward-compatible way: