Code debt and shrinking use. Considering it's only on desktop and only on old videos, I'd assume the number of uses of annotation per view is probably under 0.01% in terms of traffic, and I can also assume it adds quite a lot of complexity to their player code.

Personally, I think maybe they could flatten all annotations into a single non-interactable layer and keep that instead. That would require a lot less code and do most of the job it needs to.