Drew Adams <drew.adams@oracle.com> writes:

> (elisp) `Introduction to Minibuffers' says this:
>
>  The text in the minibuffer always starts with the "prompt string",
>  the text that was specified by the program that is using the minibuffer
>  to tell the user what sort of input to type.  This text is marked
>  read-only so you won't accidentally delete or change it.  It is also
>  marked as a field (*note Fields::), so that certain motion functions,
>  including `beginning-of-line', `forward-word', `forward-sentence', and
>  `forward-paragraph', stop at the boundary between the prompt and the
>  actual text.
>
> So I would expect that `backward-word' and `backward-sexp' would stop at
> the field boundary, which is the end of the prompt.  `beginning-of-line'
> does indeed do this, as the doc suggests.  But `backward-word' and
> `backward-sexp', at least, do not - they move backward into the prompt.
>
> Seems like this is the wrong behavior, and the doc describes the right
> behavior.  But perhaps it is the other way around and this is a doc
> bug.

The attached patch let `forward-word' does what the docstring says.
It uses `t' as the value for the ESCAPE-FROM-EDGE parameter within
the call to `constrain-to-field'.

I agree this seems the correct behavior in the minibuffer.

But I'm often frustrated by such constraints for `beginning-of-line'
(e.g. hit G c on a Gnus group and get stuck in a non-selected field
anyone?).

So perhaps the nil value for ESCAPE-FROM-EDGE is intentional here.

In any case, this is just to bring attention to the problem and
a possible solution: I don't know C so I won't commit this myself.