Thanks for the additional details, that’s really helpful! I remember now that this is why onion packets have been chosen to be 1300 bytes, so that an update_add_htlc would fit inside the 1500 bytes MTU.
So the conclusion is that we should assume that TCP packets will be fragmented in 1500 bytes chunks, so I’m curious to see the impact of sending 65kB lightning packets, which will require quite a reassembly…It would be interesting to run this in simulations (SimLN can probably help here?) or try it out with A/B testing on mainnet.