Didn't need to bisect, it's very easy to find the problematic commit, since writelines doesn't change that often:

https://hg.python.org/releases/2.7.11/rev/db842f730432

The old code was buggy in a sense that it always called PyObject_AsCharBuffer due to the way the condition is structured, but this bugginess was what allowed it to work correctly with unicode objects. After the commit unicode objects are treated like any other buffer, and that's why internal UCS2 or UCS4 representation gets written to the file.