image differences with MIS compensation and approximate BSDF PDFs

[mmp]

With

Sampler "halton" 
    "integer pixelsamples" 8

Film "rgb" "float iso" 200
    "integer yresolution" [ 800 ]
    "integer xresolution" [ 1200 ]

Scale -1 1 1
LookAt 3.69558 -3.46243 3
    3.04072 -2.85176 2.55
    -0.317366 0.312466 0.895346

Camera "perspective"
    "float fov" [ 38 ]

Integrator "volpath" "integer maxdepth" 1

WorldBegin


AttributeBegin
    LightSource "infinite"
AttributeEnd

AttributeBegin
    Material "coateddiffuse" "float roughness" .1
        "rgb reflectance" [ .2 .2 .2 ]
    Shape "bilinearmesh"
    "point3 P" [ -100 -100 0  100 -100 0   -100 100 0  100 100 0 ]
AttributeEnd

if UniformInfiniteLight is modified to ignore the allowIncompletePDF flag, then the image converges to a different result. (And worse, using the SimplePathIntegrator with light sampling only also seems to give something different!)

Everything matches if "coateddiffuse" is replaced with a "diffuse" material.

[sherholz-intel]

@mmp are there any updates on this topic?
I am facing the same problem when using coateddiffuse and trying to combine BSDF sampling with other sampling techniques.
I realized that explicitly using the outcomes of f() and PDF() for a sampled direction leads to a different result than using the outcome of Sample() directly.

[mmp]

No updates, unfortunately--I haven't had time to look at this in more depth. (Fixes welcome. :-) )