Stefan Monnier <monnier@iro.umontreal.ca> writes:
>> Or maybe instead of a buffer-local, we could have
>> minibuffer-hide-completions do (goto-char (point-min)) so no
>> completion is selected anymore.  Then it becomes harmless to reuse
>> that *Completions* buffer.
>
> Either way works for me.

OK, updated patch with this and other feedback: