Thinking about this some more, I'm inclined to go the same way we did with issue 33932: classify it as an outright regression, work out the desired requirements for a missing embedding test case, and then fix the implementation to pass the new test case.
My suggestion for test commands would be:
../Lib/site.py
-m site
-c 'import sys; print(sys.argv)' some test data here
Once those work properly, we'd consider the regression relative to Python 3.6 fixed.
Those could either be 3 different test cases, or else we could run them all within a single test case, with a Py_Initialize() call before each one (since Py_Main() calls Py_Finalize() internally).