If you comment out line 65 you will see what I mean.
Thin Instances:
Instances:
Its like the specular is borked?
For performance reasons (to avoid calculating the inverse of a matrix per thin instance), we use a simplified normal calculation in the case of thin instances that does not support non-uniform scaling, only uniform scaling.
We could add a flag to switch to the exact calculation, if you need it (the best solution would be to avoid non-uniform scaling altogether, though 
)?