It is both tempting and "neat" (as in "neato") to link against python libraries already installed on the system and call the interpreter through its API.
But some python libraries (COUGH numpy COUGH), which have their own DLLs, do not behave well when being unloaded/loaded/reloaded between successive invocations of the interpreter. As a result we have to run successive calls in the same Python instance, and that is Not Good.
We should probably just spawn a Python subprocess each time.