Proposed resolution:

This wording is relative to N5054.

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

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

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

   -2- ReturnsEffects: Half the sum of `a` and `b`. If `T` is an integer type and the sum is odd, the result is rounded towards `a`If `T` is a floating-point type, equivalent to `lerp(a, b, T(0.5))` ([c.math.lerp]). Otherwise, 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.

While `std::midpoint` accepts operands of floating-point type, it is unclear what the midpoint is if any operand is an infinity, if both operands are differently signed infinities, if any operand is a NaN, etc. The specification merely says:

Returns: Half the sum of `a` and `b`. If `T` is an integer type and the sum is odd, the result is rounded towards `a`.

Not only is there no description of the result, there is also no description of error handling because the blanket wording in [library.c] does not apply to <numeric>, only to <cmath> functions.