Yep, several reasons why HD would be a poor choice for putting resources in the loop.However, I’ve had discussion with @Discoducky about this off this thread and he thinks that Tesla won’t use HD maps even in the Loop where HD maps would be much more reliable and maintainable than surface roads. I’ll let him comment more on that if he wants, but I was convinced.
Training on HD maps would be a whole new code path that would essentially be throwaway work, a distraction for the AI teams, most likely postpone further FSD and be prone to corner case issues.
Training FSD with high likelihood images/video for specific needs of the loop makes much more practical and functional sense as it would add to the overall effectiveness of FSD. Eventually, all Tesla's will be allowed to drive in these tunnels so that is the path I think they will take.