...
Wondering whether something had just railed, I poked a small hole in the foil over the window, and found the detector was still working more or less as expected even at -50C. Decided to take it all the way to 0C, expecting it to become unusable at some point. Not sure that ever happened, but at least the noise started to go up 
Went back through the data using a script written by Kaz that actually calculates the noise for the entire device. I believe it includes unit cells that don't have active pixels and may thus underestimate the noise. The results differ in detail but the general trend is similar to the 'box' proxy noise estimate. The noise from the script is plotted against the TempB reading.
Using 3.3e-/DN, the noise at -61C (TempB) is still below 9e-. Can't assume that the device temperature tracks TempA or TempB well.