Interesting questions Marc.
My longest integration to date is this one:
https://www.astrobin.com/364378/G/?nc=user
Although I did not think of creating separate masters of smaller batches, I experimented with calibration frames (all done in PI).
I found that for my camera and for current location, adding even a significant number of darks (40 or so) adds a bit of noise to the master, but only when integrating up to about 120-150 15-minute dithered subs.
For 200 15-minute dithered subs, there was a slight decrease in mostly background noise when calibrating with darks (in addition to using Bias and flats). I think this is due to fixed patterns that camera introduces, and this fixed pattern becomes measurable when stacking hundreds of subs. Visually I could not tell the difference, no matter how close and hard I looked...
Measurements of noise were performed with PI using statistics process for various areas in the integrated images.