I would look at the raw network (Drive) values for position demand (command) and position actual, to see how those compare to the values in the RSI Axis. Take note of these values in the failing case, as well as the case after re-issuing the S-curve move from RapidSetup.
Since you mentioned drive-based homing, it is possible that the drive has adjusted internal positions by some drive origin position? (In this case it seems like it may be adjusting just command or actual, not both.)
The MultiAxis should not be an issue.