Can't perform PCS with bounding box prompts on videos

Hi and thank you for the amazing work.

I've been trying to perform video PCS with just bbox prompt on the provided test video, but the model is not propagating the predictions through frames.

I'm adding box promt to first frame as:
`
out = video_predictor.handle_request(
request=dict(
type="add_prompt",
session_id=session_id,
frame_index=0,
text=None,
points=None,
point_labels=None,
bounding_boxes= np.array([[0.3953, 0.2667, 0.1164, 0.5736]]) ,
bounding_box_labels=[1],
)
)
`
and If I display this ouput it is actually doing the job and all 4 people are actually segmented:

If, instead, I do propagate across the entire video, it seems that no instance is detected at all:

`outputs_per_frame = {}
for response in predictor.handle_stream_request(
request=dict(
type="propagate_in_video",
session_id=session_id,
)
):
outputs_per_frame[response["frame_index"]] = response["outputs"]`
same frame after propagation:

I do not want to do PVS SAM2-style, and I'm understanding that SAM3 should be able to perform PCS retrieving the semantic from visual prompts (also because within the prompted frame it is working), so am I doing something wrong, or is PCS on video not possible without text?

Reproduce the behavior with the provided test video using video_predictor.handle_request for add_prompt with bounding_boxes, then handle_stream_request with propagate_in_video. Compare the prompted frame outputs with the propagated frames and inspect the request handling around these entry points. Done means either bbox-only propagation retains the detected instances across frames or the limitation and required prompt format are clearly documented.

Written by the indexing model from the issue text.