So I’ve been racking my brain over this problem for a while, and I think your problem is that you should be running a synchronous latch.

By doing this asynchronously, the CO pulse will be an extremely narrow glitch. As soon as you reach zero, the latch will set, which will load the data bits immediately, which will suddenly bounce CO back as fast as signaling allows. Remember that PE is asynchronous. As soon as it goes high, it will quickly become no longer true that we have a carry out, on the order of nanoseconds. This could be shorter than the minimum pulse width requirement of your latch gates which might not yet have had time to finish stabilizing. I could see that causing the latch to misbehave.

I don’t know if having such a skinny pulse could violate timing constraints on your gates.