pbrt uses a scalar (shapes.h: Float tShapeHit = (height - Float(oi.z)) /
Float(di.z)). It produced NaN, and since every NaN comparison is false neither
the t range test nor the subsequent dist2 > radius^2 test could reject it. The
disk reported a hit on every ray tested against it, and the NaN reached dpdu/
dpdv, the shading normal, and the area-light MIS pdf.
NaN is absorbing in the film, so one bad sample poisoned pixel RgbSum permantenly. Wavefront now works