When the user wants to write a Coq proof, she needs to run Coq on the goal,
wait five seconds for it to fail (it will fail, otherwise there is no point
in running Coq on this goal: another prover would have succeeded already),
and finally edit it. This is a waste of time. So goals run with an
interactive prover are now marked as unknown until their file is edited.

Interactive provers could have been detected by a nonempty "editor" string,
but there are interactive provers that don't have dedicated editors, and
there might be automated provers with dedicated user interfaces. So a new
field was added to prover descriptions.

TODO: actually run the editor when there is only one selected goal,
rather than keeping the current three-click method of editing proofs.

t_fold already iters on the subterms of a term, so calling t_fold on a
subterm actually iters over grandchildren of the term, which may well be
under a quantifier. So t_fold should be called on the term itself.

This explains why the transformation was trying to create declarations with
unbounded variables.

kudos to François for spotting the problem

fallback on colorizing the identifier.

otherwise we lose the beginning of the transformation chain