Rough transmission with dispersion is too dark under direct light (pdf differs between sample() and eval()) #329

Summary. A rough refractive material with use_dispersion renders about 15% darker when light and BSDF samples are combined than with either alone. Without dispersion all three agree. Version 2026.29.1, scalar mode.

Reproduce. A rough glass sheet (DwaBaseMaterial, transmission 1, roughness 0.3, use_dispersion true, dispersion_abbe_number 20) lit by a disk light behind it, max_depth 1:

Cause. TransmissionCookTorranceBsdfLobe::sample() (lib/rendering/shading/bsdf/cook_torrance/BsdfTransmissionCookTorrance.cc) picks a colour channel and multiplies the returned pdf by the channel probability:

Color result = eval(slice, wi, etaT, &pdf);
// Dispersion PDF
pdf *= spectralPDF;

eval(), which the integrator calls for light samples, returns the pdf at mEtaT without that factor. The integrator's power heuristic then sees 0.35 p on the BSDF-sample side and p on the light-sample side, and the two weights sum to less than one.

Suggestion. Have eval() and sample() report the same density for a given direction (for instance the mixture over the three channels in both).