@Elle has an article that talks about that, too, relevant to the white point:
I think the embedded profile is correctly classed, because the destination of the output image is nominally a display. The input profile was used internally by dcraw, as the camera profile created from the adobe_coeff lookup. The image was transformed from camera space to sRGB using the
adobe_coeff profile -> XYZ -> sRGB,
just like any regular transform.