Wrong prediction for absolute vs relative points

When using relative points in the sam3_for_sam2_video_task_example Notebook the prediction for the first frame is fine. When using absolute points and rel_coordinates = False, the prediction is somewhere else (not anywhere on the object with the prompt).
I also noticed it for my example that I implemented without the notebook, and for the box prompts it is the same problem (first prediction not even in proximity to the box).

Start with the sam3_for_sam2_video_task_example Notebook and reproduce the first-frame prediction using relative points, absolute points with rel_coordinates = False, and box prompts. Compare where each prompt is placed against the resulting prediction; done means absolute point and box prompts produce first-frame predictions on or near the prompted object.

Written by the indexing model from the issue text.