although in principle this was not necessary,
since we already publish a dependency on ocamlfind.

Looks nicer / more general. May save some allocations.

I used it in the branch bitset, which turns out to be incorrect (I think).

instead of leaving it to the caller. This offers more flexibility
to the implementor of [E], and already allows a speed gain (due to
the loop interchange, I think).

although the specification of [add] is more complex.

provided it appears in tail position in its caller and
its caller does not itself have a %prec annotation.