AWS is always super closed mouthed about their infrastructure.

The info I have came from a conversation with an AWS Solutions Architect.

It's a couple of years old, but based on my experience it's still true today.

It's legal to send UDP packets with a zero checksum, indicating "no checksum." This can be set at a UDP socket level in Linux. I wonder if that would make any difference?

(Of course this assumes your protocol has some alternate method of verifying transferred data, which many do.)

Whenever people talk about behavior/treatment of UDP traffic I consider DNS as a special case. I have no idea how AWS handles UDP but I will never use DNS as a generalizable example of UDP traffic.

At least in the past, it really was true. I've been burned by this before using UDP in AWS. In AWS, I've learned to be skeptical of using protocols other than TCP. That said, it consequently has been a long time since I've tested UDP over AWS.

DNS was precisely the area where we got bit by it.

We were using a Non-AWS DNS resolver (aka Google) and we would often get dns resolution errors despite our NAT not being remotely taxed by the traffic.