bug#12456: 24.2.50; Completion in `Info-goto-node' (cross-manual jump)

[Stefan Monnier]

>> IIRC the core of the problem is that Info's completion table does not
>> (yet) know how to do completion after something like "(elisp)", so it
>> basically returns some "dummy" completion data (because the way
>> completion tables are defined currently, they have no way to say "I
>> don't know").

> Can the completion tables say no completion candidates at all?  That
> would be the best short-term solution; the stand-alone Info reader
> does just that.

So that we always get a [No completions] message?
We can try.  It's all decided in the following three lines of
Info-read-node-name-1:

   ((string-match "\\`(" string)
    (cond
     ((eq code nil) string)
     ((eq code t) nil)
     (t t)))

As long as we don't know the actual list of nodes, it's important to
return non-nil when `code' is `lambda': that's the `test-completion'
case, which determines whether RET will accept the user's input since
the completion is called with `require-match'.
     
>> So there are 2 ways to fix the above problem:
>> - Extend minibuffer.el so a completion table return "don't know" (at
>> which point it could put a message like " [No completion info]").
>> - Extend info.el so that it does provide actual completion by opening up
>> the "elisp" info file and gathering the corresponding node names.
> The latter sounds like the best long-term solution to me.

Agreed.

> Of course, next we will see a bug report with "g (eli TAB"...

Try it: that one already works.


        Stefan