Why doesn't coverage.py properly measure Django's runserver command?

I should know the answer to this, but I don't: if you try to measure the coverage of a Django project like this:

coverage run manage.py runserver 

you get coverage measurement that misses all of your actual code. Something early on in the process is stopping the measurement, or all the real work happens in a new context that doesn't get measured at all.

Can someone point me to the specific point in the process where the measurement breaks down, so that I can try to fix coverage.py so that it will measure it properly the way people expect?