I know that a manual from the KRC1/KRC2 times is hardly up to date, but since the behavior described in the manual aligns with what I observed up until software version 8.3, I thought that it was fair to assume that the behavior would still be the same in 8.5. That it is no longer mentioned in the manual I find less important, since KUKA sometimes seems to omit some things in some manuals.
I have emailed my contact at KUKA asking whether they have change-logs, so let's see what that brings.
But anyways, I managed to solve my problem using an interrupt that is triggered by a cyclical flag that is defined like this:
$cycflag = ($PRO_STATE1 == #P_ACTIVE) AND (STOP_REQUESTED == TRUE)
The STOP_REQUESTED variable is then set from the SPS.
Because we use $PRO_STATE1 (program state) in the cyclical flag, the rising edge that triggers the interrupt will never come while the program is stopped, only when it is started again. I have now implemented it and tested it multiple times, and it works!
I am pretty satisfied with that solution, unfortunately I didn't come up with it myself, one of my colleagues suggested it.