bug#73584: 29.3; read-key

[Eli Zaretskii <eliz@gnu.org>]

> From: Stefan Kangas <stefankangas@gmail.com>
> Date: Wed, 2 Oct 2024 21:47:19 +0000
> Cc: Emacs-hacker2023@jovi.net, 73584@debbugs.gnu.org
> 
> Eli Zaretskii <eliz@gnu.org> writes:
> 
> >> From: Stefan Kangas <stefankangas@gmail.com>
> >> Date: Wed, 2 Oct 2024 10:02:44 +0000
> >> Cc: 73584@debbugs.gnu.org
> >>
> >> It would arguably be a bit nicer to signal the error early to avoid
> >> having to mash C-[:
> >
> > Maybe.  But what exactly is wrong with signaling the error from
> > read-key-sequence-vector, which read-key calls?
> 
> I just observe that I have to mash C-[ to get back control of Emacs, and
> that it would probably have been better if I didn't have to.  I imagine
> situations when this happens because of a bug in some library, and users
> won't know what to do in that situation.  It clearly confused OP.

The OP shot himself in the foot by invoking by hand a low-level
function with a wrong argument.  Interactive functions and high-level
APIs must check their arguments right away, but low-level functions
like this one do not.

What I could understand is if we'd arranged for the temporary keymap
to stop being in effect even inside the debugger.  Because the
scenario of this bug reveals some more general issue: when we signal
an error from a function that sets up some limited keymap, the
debugger is invoked with that keymap still in effect, or so it seems.

Stefan Monnier, do we have a good way of doing that in cases like
this?  Perhaps the debugger should define some minimal key bindings
when it starts or something?

> I meant to ask if we have something similar to `cl-check-type' that does
> not come from cl-lib.el.  I'm asking that precisely because we can't use
> cl-lib in subr.el.

Sorry, my misunderstanding.

> It would have been convenient if we had something like:
> 
>     (defmacro check-type (form predicate)
>        `(unless (,predicate form)
>            (signal 'wrong-type-argument (list ,predicate form))))
> 
> We have this pattern in quite a few places in our code, so maybe we
> should consider adding it if we don't have it already.

If we want something like this, I think we should support a list of
allowed types, not just one type.