To summarize even further, I believe this is the issue: When a return packet is received by RouterOS, it is first processed in the right VRF context but after address translation the packet looses VRF context. Now, if a second lookup needs to happen (in this case due to vrf@vrf route instead of interface@vrf) RouterOS uses main VRF because VRF context was lost.
I am now sure this is the same underlying issue of the topic I linked above.
So the fix would be to make sure NAT back-translation doesn't loose VRF context of packets.
Amazing, thank you so much! I am happy this is acknowledged.
Just for reference, I believe this would address both SUP-219962 and SUP-221017 .