The addressed issue is a regression from Blender 2.75, after the internal switch from double to single precision floating-point numbers in the Freestyle code base. Face normal calculations require the higher precision during the computations, even though the results can be stored as single precision numbers.