I don't disagree with the statement regarding shorter cycle times. However, I am not so sure about how linuxcnc processes this. It has been my understanding that closed loop in linux cnc is part of the servo thread and thus usually defaults to 1 ms - faster is certainly possible, but becomes more difficult. Furthermore, ethercat with cycle times below 1 ms is also asking for trouble.
What I do not know is how much of a time offset results from using ethercat to input glass scales. I would image that between t0 (encoder position is read) and the processing of the information within the servo thread a certain time (number of cycles passes). I do not know how many.