I ran the rc2_subset processing twice, using identical configurations. Indeed, the metric values come out slightly different (PA1: 6.985124378559791 mmag vs. PA1: 6.774156219385405 mmag). The attached image shows a comparison of the deepCoadd_calexp images for a randomly selected patch – the two separate processing runs are the left and middle panels, and the right panel shows the difference between those images. There are clear differences between the coadd images.
The same is true of the deepCoadd images, implying that the relative calibrations are the same, and that the coadded images are intrinsically different.
I compared two randomly selected visit images (calexps) of the same dataId, and they are identical. Thus the differences between runs must have occurred after single-frame processing.
Another suggestion was that the random number seed for FGCM may not have been fixed, so that FGCM would return different results for consecutive runs. I confirmed that randomSeed from both fgcmFitCycle_configs is the same (89234, as in obs_subaru/DRP.yaml).
I checked the FGCM calibrations for the two runs (for the same dataId), and they differ slightly, so my guess is that FGCM is deriving slightly different solutions based on the ordering of inputs.