We reported this issue back in 2018 to AWS! In our case it was exacerbated because of a NAT through which traffic to the cross-AZ loadbalancer would flow. As mentioned in the article, the client side is free to reuse ports as long as the destination in the tuple is different - as would be the case for cross-AZ NLB.

We reduced the severity of this by randomizing port mappings on the NAT but that just reduces the probability. They claimed that it was a high priority issue for them back then but seems this is still in the wild - insane!