About the prompt used in the code

Hello, I have some questions about the prompt used by SAM3. SAM3 supports a rich assortment of prompt types: legacy point, bounding-box and mask prompts inherited from SAM2, exclusive text prompts newly introduced in SAM3, as well as box-annotated exemplar prompts. While reviewing the source file sam3_image.py, I noticed a variable named visual_prompt_embed that falls outside all the aforementioned prompt categories. Could you elaborate on the exact definition of this specific prompt, feasible approaches to leverage it for my research innovation, and clarify whether it represents image-level contextual exemplar information or CLS-token-style global feature embeddings? Thank you very much！

Start by reading sam3_image.py and tracing the definition and use of visual_prompt_embed. Clarify what this prompt represents, how it relates to SAM3's other prompt types, and whether it carries image-level contextual exemplar information or CLS-token-style global features. Done looks like a documented explanation that addresses the research-use questions.

Written by the indexing model from the issue text.