I had to read it three times, but I now understand the issue.
I initialy though the first version was somewhat bugged :-)

Now I understand why the second one is better. This kind of
optimisation is very interresting (not capturing all the scope
when building a closure). I will look for it now.

And BTW, if I find the proper name for this technique in the
book that I'm currently reading (The Impl. of Func. Prog. Lang.)
then I will post it here to refresh your memory :)