🎬 ToolMerge — Long-Video Keyframe Retrieval

Decomposes a question into SigLIP visual-search tool calls combined with AND/OR boolean logic, scores every frame, merges the ranks, and picks the top-K keyframes with temporal non-maximum suppression.

Faithful port of the SigLIP path of ToolMerge (paper) using the released text-only planner michalsr/toolmerge-planner-grpo (Qwen3-VL-8B, GRPO-tuned) + google/siglip2-giant-opt-patch16-384.

The paper's T-REN region-text tool and OCR-judge stage need offline-built per-video caches and are omitted here; the boolean merge runs SigLIP-only, exactly as the reference code does when other tools are disabled.