According to the reviewer's comments on PIPE2D-114:

You have a single giant function which tests a number of different calls to PolyFit. Rather than doing that, you ought to have a number of short unit tests, each of which tests a single call.

Ideally, each of the unit tests would be written in Python. However, I understand that the full PolyFit interface is not exposed to Python, so this is impractical. That leaves you with two choices:

Use a C++-based test harness. The one in use in LSST is Boost::Test. For example, see these tests in afw.
Write individual test functions in C++ then expose each one to Python by wrapping it separately using Swig.
Either of those approaches is adequate. What you've got now isn't.