So, I can't reproduce that issue unfortunately. When I set the UI to have both chars walk at each other they do it forever regardless of what window is on top. I worry it may be a config issue that will be hard to reproduce on my end. I'm not sure what your current workflow is, but I personally recommend using a workflow similar to what I do in this video (timestamped start at 8:46, or I tried to, the embed doesn't like to use the timestamps).
Basically, create a recording sequence that starts with a training reset button press, do the inputs you want to have in the combo, then fiddle with the timings in the editor and press play to test if your timings work. I've never considered a workflow where you pause and step every frame or use the UI to set your inputs every frame. The UI input was mainly to have a dummy hold block or crouch block while trying to record a sequence, and the frame stepping was intended to break down and data mine moves, not input them one frame at a time. Let me know what your intended workflow is for what you are doing.