Hello David,
I have sent the dataset to Alexey and team, here's hoping they get back to me. I went through the same thing, with trying 0.017 by 0.017 pixel size, and the alignments came undone for sure. Funny enough, I ran the dataset on a different computer without changing pixel size, and the point cloud came out dimensionally correct, but the ortho is still elongated after the process. I also had no problem with the 19mm lens automation, that's the lens I'm running.
As far as getting water temperature goes, are you flying during the night or day? I would be very particular about getting yoru environmentals, as well as TReflect when you put the photos into FLIR tools. I always let my clients know "it depends" when it comes to determining absolute values. The ITC sUAS thermography course certainly answered a lot of questions, if you haven't attended I would suggest it when the COVID stuff blows over.
Now I'm looking for in depth Metashape training, for visual - multispectral and thermal mapping.