Three different, undocumented, cases that I think should be the same:

1. plot_surface(..., shade=True, lightsource=..., cmap=..., facecolors="garbage") - uses LightSource.shade, which does some messy and unnecessary trig in LightSource.hillshade, using the orientation of the lightsource.

2. plot_surface(..., shade=True, lightsource=..., cmap=...) - uses _shade_colors, which does a more sensible vector-based approach, but hard-codes a [-1, -1, 0.5] normal vector, rather than respecting the lightsource direction.

3. plot_surface(..., shade=True, lightsource=..., facecolors="garbage") - as above

Regarding (2), I think that there is a typo in axes3d.py, and this change should be made:

         # Shade the data
-        if shade and cmap is not None and fcolors is not None:
+        if shade and cmap is not None and fcolors is None:
             fcolors = self._shade_colors_lightsource(Z, cmap, lightsource)

It seems dumb to only calculate face colors if the user already asked for different ones.

Regarding (3) - I think it would be useful to unify the shading into a normal-vector based approach, which I might try in a later patch