Nifty!

On Sat, Jun 4, 2011 at 8:19 AM, Daniel Schoepe
<daniel.schoepe@googlemail.com> wrote:
> +        (minibuffer-completion-table (completion-table-dynamic
> +                                      `(lambda (s) (notmuch-query-completions
> +                                               (quote ,all-compls) s)))))
> +    ;; ^ emulate a closure to avoid recomputing the completion list each time

Dynamic scoping is obnoxious, but I think programmed completion is
steeped in the assumption that you'll use it.  This code would be much
simpler if notmuch-query-completions took only `string' and used the
dynamically-bound all-compls (which should probably be renamed
notmuch-completions or something if you do this).  Then this could be
just
  (minibuffer-completion-table (completion-table-dynamic
#'notmuch-query-completions)))
and there'd be no need for quasiquoting, comments, and fake lexical scoping.

> +    (define-key keymap (kbd "<tab>") 'minibuffer-complete)

This probably deserves a comment about why you're doing so much work
to avoid completing-read (which I assume is because it also binds SPC,
even if require-match is nil, which is unfortunate).