Gendaylit has implemented also a luminance efficacy model from Perez, not only the Perez sky luminance distribution model. In gensky you feed-in irradiance only in the visible range, that means you need to apply a luminance efficacy in advance (and that value typically varies with the sky type and its turbidity). Typical measured irradiance (or provided in weather files like in EPW) are for the full solar spectrum and not only in the visible range. That might explain also your deviations.
If you have 0 direct irradiation, the sun disk has also zero radiance, see below the output of gendaylit. Epsilon is calculated to 1 and therefore pure overcast sky. For your sky probably delta is calculated as “very bright” which means the sky gets also brighter in sun direction, even the sun itself is not visible. In general this makes sense, although we all know that sky models can differ for individual timesteps dramatically (especially for intermediate sky that depends on the cloud distribution on the sky), but over the coarse of the year things balance out and for many climates it fits pretty well. Be also aware that the Perez model also depends on the dew-point and that you can feed the dew-point temperature with the option -d Dewtemp - that increases the accuracy of the luminance efficacy calculation.
Here the output of gendaylit when applying it with 0 direct-irradiation:
gendaylit -ang 20 -45 -W 0 228
gendaylit_dew -ang 20 -45 -W 0 228
Local solar time: 0.00
Solar altitude and azimuth: 20.0 -45.0
epsilon, delta, atmospheric precipitable water content : 1.0000 0.4673 2.0000
void light solar
3 0.0 0.0 0.0
solar source sun
4 0.664463 -0.664463 0.342020 0.533000
void brightfunc skyfunc
2 skybright perezlum.cal
10 2.544e+01 5.664e+00 0.091877 -0.241465 5.424682 -2.333236 0.183553 0.664463 -0.664463 0.342020