At least on Windows, i'm reasonably sure all the API does is tell the OS which image to render as cursor and the cursor is rendered entirely by the OS; meaning there's no control over whether the cursor is being rendered partially outside the viewport or not.

Solutions where the browser would have enough control would likely require rendering the cursor by itself, which would impart it noticable input latency.

Edit: Quick addition of a bounding box to the demo: https://wchristian.github.io/cursory-hack/