Hi @automateme2 - thanks for all the context and I’m glad you’re surprising yourself with good results 
Project type is the same as task type, this can include Object Detection, Instance Segmentation, Classification, etc. In your case, it’s Keypoint Detection.
Unfortunately, Auto Label is not supported currently with Keypoint Detection (I’ve flagged this internally, we may be able to push out a quick fix).
For your labeling flow, you might have a much easier time using Workflows – we literally built it for cases like this. You can run a model and selectively upload images directly to Roboflow with the correct annotation format. You’re correct that you should be manually reviewing images before adding them to your dataset.
On the question of obscured keypoints, the rule of thumb is you should label images exactly how you would like your model to respond to the given input. If you want it to not give a keypoint, mark it obscured.