There is a reason, although I suspect you won't agree with me if I try to tell you it's a good one.
The metered dataset actually comprises three sources of data from three different suppliers - the embedded generation files, the unit-level IR files, and everything else. The underlying databases that this data is loaded into are not pretty, and we publish all three datasets as a single collection that we call 'metered data'. But we don't actually join them up into a single coherent dataset. This is all legacy stuff that I am keen to see get superseded with a collection of wholesale volume datasets that is infinitely more useful and more easily maintained. We have a pretty good idea of what we want to do and we have all the data we need to do it. What we haven't been able to do thus far is find the developer time to make it happen. That said, I'm hopeful we'll get to it this later this calendar year.
The legacy process that creates the Generation_MD dataset simply maps points on the grid to a list of plant names, i.e. it infers that injections from the metered grid data are generation. This process does not make use of the embedded generation dataset except in a couple of hard-coded cases. It's high maintenance and clumsy, and we just aren't resourced sufficiently to provide the necessary upkeep.
Mill Creek and Te ahi o Maui are from the embedded generation dataset and I'm guessing they came along after the Generation_MD mapping was created.
I hope that makes sense - and we will eventually get this sorted.