Currently, <numeric> includes <limits> in C++20 mode:

<limits> is a relatively large header (potentially affecting compiler throughput, i.e. slowing down builds), but our usage is very limited here - we just need floating-point min/max:

<numeric> always includes <xutility> which includes <climits>:

So <numeric> could directly use the classic macros:

Specifically, the changes in <numeric> I'm envisioning are:

1. Stop including <limits>.
2. Define a helper variable template _Floating_min<_Ty>, specialized for float, double, and long double.
3. Similarly define _Floating_max<_Ty>.
4. Replace the numeric_limits calls with uses of the variable templates. Note that because users could theoretically pass volatile values, the uses should say remove_cv_t<_Ty>; numeric_limits was previously handling that: