I have the following minimal example of summing an array (taken from here):

#lib.cpp
template<typename T>
T arr_sum(T *arr, int size)
{
    T temp=0;
    for (int i=0; i != size; ++i){
        temp += arr[i];
    }
    return temp;
}

#lib_wrapper.pyx
cimport cython

ctypedef fused  float_t:
    cython.float
    cython.double

cdef extern from "lib.cpp" nogil:
    T arr_sum[T](T *arr, size_t size)

def py_arr_sum(float_t[:] arr not None):
    return arr_sum(&arr[0], arr.shape[0])

#setup.py
from setuptools import setup
from setuptools.extension import Extension
from Cython.Distutils import build_ext
import numpy as np

ext_modules = [Extension("lib_wrapper", ["lib_wrapper.pyx"],
                         include_dirs=[np.get_include()],
                         extra_compile_args=["-std=c++11", "-O1"],
                         language="c++")]

setup(
    name='Rank Filter 1D Cython',
    cmdclass={'build_ext': build_ext},
    ext_modules=ext_modules
)

applying python setup.py build_ext --inplace produces a 202K size shared object lib_wrapper.cpython-39-darwin.so. gcc -shared -fPIC -O1 -o lib.so lib.cpp
would produce a smaller object of ~4K size.
I assume that the redundancy in the file size comes from the C++- Python bridge, created by Cython.

Considering the numerous methods such as Numpy-C API, pybind11, etc, which one would allow the creation of this bridge without such a large file size overhead? Please exclude ctypes from the suggestions – it seems to bring a large addition to access time.