Variance is not correctly converted to electrons

So I'm looking at the flux vs variance in the pfsArm, and when I look at the per fiber chi, it's not ~=1 but rather proportional to sqrt(gain)

After discussions with Paul, I looked at the calexp directly, for a quartz.
I plotted the image/variance and I took the median as a function of row, so you get the ratio for different signal level.

If the variance was perfectly estimated and in electrons, this curve should be unity, but you can see two problem :

• in the low flux regime, aka readout noise dominated, the value is slighly underestimated, I think it's because I've specified the value in adu instead of electrons
• when when transition to photon-noise regime, this ratio asymptotically converge to the gain value, which shouldn't happen if the variance is given in electrons.

I could