Proposed resolution:

This wording is relative to N5054 on top of LWG 4614.

1. Modify [numeric.ops.overview] as indicated:

   […]
   // [numeric.ops.midpoint], midpoint
   template<classintegral T>
     constexpr T midpoint(T a, T b) noexcept;
   template<floating_point T>
     constexpr T midpoint(T a, T b) noexcept; // freestanding-deleted
   template<class T>
     constexpr T* midpoint(T* a, T* b);
   […]
   

2. Modify [numeric.ops.midpoint] as indicated:

   template<classintegral T>
     constexpr T midpoint(T a, T b) noexcept;
   

   -1- Constraints: `T` is an arithmetic type other thannot cv `bool`.

   -2- Effects: If `T` is a floating-point type, equivalent to `return lerp(a, b, T(0.5));` ([c.math.lerp]). Otherwise, returns half
   Returns: Half the sum of `a` and `b`; if the sum is odd, the result is rounded towards `a`.

   -3- Remarks: No overflow occurs. If `T` is a floating-point type, at most one inexact operation occurs.

   template<floating_point T>
     constexpr T midpoint(T a, T b) noexcept;
   

   -?- Effects: Equivalent to: `return lerp(a, b, T(0.5));` ([c.math.lerp])

   -?- Remarks: At most one inexact operation occurs.

The standard library typically does not provide freestanding floating-point numerics, but floating-point `std::midpoint` is freestanding, presumably by accident. LWG 4614 makes this pre-existing issue readily apparent by making floating-point `std::midpoint` (freestanding) equivalent to `std::lerp(a, b, T(0.5))` (not freestanding).