bug#67196: M-: uses a wrong value of debug-on-error when it is nil.

[Stefan Monnier]

> Thinking about it more clearly, we're using one variable, debug-on-error,
> for two conflicting purposes:
> (i) The calling mechanism of execute-extended-command, including the
>   invocation of the debugger, should an error be signalled.
> (ii) Use in the form to be evaluated, where the variable's value outside
>   of execute-extended-command should endure.

Yes, there's a conflicting need here: we want to control the way `eval`
works "from the outside", i.e. we want the `debug-on-error` to apply to
the evaluator but not to the code being evaluated.

But `eval` doesn't distinguish between the context of the evaluator and
the context of the code being evaluated because all ELisp code uses the
same evaluator.  I can't think of any way to split the variable into two
to solve the problem.  Instead it gives me the impression we're trying
some impossible feat like watching oneself sleep.


        Stefan