Promptable segmentation (LibreSAM tier): point and box prompts, encode once and prompt many. Image inference; video is out of scope in v1.