Proposed resolution:

This wording is relative to N4950.

1. Modify [alg.fold] as indicated:

   template<input_iterator I, sentinel_for<I> S,
            indirectly-binary-left-foldable<iter_value_t<I>, I> F>
     requires constructible_from<iter_value_t<I>, iter_reference_t<I>>
     constexpr see below ranges::fold_left_first_with_iter(I first, S last, F f);
   template<input_range R, indirectly-binary-left-foldable<range_value_t<R>, iterator_t<R>> F>
     requires constructible_from<range_value_t<R>, range_reference_t<R>>
     constexpr see below ranges::fold_left_first_with_iter(R&& r, F f);
   

   -9- Let U be

   decltype(ranges::fold_left(std::move(first), last, iter_value_t<I>(*first), f))
   

   -10- Effects: Equivalent to:

   if (first == last)
     return {std::move(first), optional<U>()};
   optional<U> init(in_place, *first);
   for (++first; first != last; ++first)
     *initinit.operator*() = invoke(f, std::move(*initinit.operator*()), *first);
   return {std::move(first), std::move(init)};
   

[ 2023-11-03; Reflector poll ]

Many votes for NAD. "Yuck, can we just use .value() instead?" "The example is not good motivation, but we should ADL-proof to avoid attempting to complete incomplete associated classes."

The following program is currently ill-formed, because [alg.fold]/10 requires evaluating *init, where init is an object of an optional specialization, which triggers ADL and finds unwanted overloads.

#include <algorithm>
#include <optional>

namespace myns {
  struct Foo {};
  
  void operator*(std::optional<Foo>&);
  void operator*(const std::optional<Foo>&);
}

int main()
{
  myns::Foo x[1]{};
  std::ranges::fold_left_first_with_iter(x, []<class T>(T lhs, T) { return lhs; });
}

I think only the member operator* overload is intendedly used.