Right, but store forwarding is definitely not 2 cycles. 4 and 5 cycles is common for store forwarding although other values are possible. So I think Agner measured 4 cycles and divided it as 2 for the load and 2 for the store. For pure loads the latency is easy to measure and often widely reported and it's 4 cycles minimum for Intel.
He does say he's reporting the minimum latency; I was thinking that no case would be faster than reading from L1 cache, but it makes sense that store-forwarding can be faster.