More specifically, if I also recall correctly, they need to be built with the oldest version of NumPy they want to support at runtime, and then can be run with any backward-compatible later version of NumPy. Much like, as @brettcannon said, an API target (e.g. macOS, Android, Windows SDK targets, and also like HPy is doing and sort of like the min version in the limited API/stable ABI for CPython.
EDIT: See @rgommers 's just-posted message on another thread:
https://discuss.python.org/t/allow-package-references-as-version-specifiers/19226/18?u=cam-gerlach
which goes into more detail about this.
And also a whole infrastructure around Conda-Forge, conda-build
, conda-smithy
, recipies, feedstocks and CI designed around those requirements, and taking care of providing a consistent set of build dependencies such that the various packages (at least built around the same time) will be ABI-compatible with each other.