DM-9953 created the SENSOR_EDGE mask to mark coadd pixels that were on or near the boundary of an input CCD (and hence should have INEXACT_PSF set as well). It also started the propagation of those EDGE regions to the coadd.
Propagating the EDGE regions to the coadd caused problems, however, because many of them were affected by otherwise-unmasked bad pixels (leaks from amps or somesuch), so that was disabled on
DM-12931. Contrary to comments on that ticket, this seems to have broken the propagation of SENSOR_EDGE, or perhaps something more recent broke it.
We should also consider whether to split the current CLIPPED flag into both CLIPPED and REJECTED, with the latter being used for pixels rejected due to mask values from the calexps rather than an explicit smart-clip algorithm (e.g. SafeClip or CompareWarps).