llama.cpp

9506fc0e - vlib-video: skeleton + tool-call parser (Qwen2.5 JSON + Qwen3.5 XML)

vlib-video: skeleton + tool-call parser (Qwen2.5 JSON + Qwen3.5 XML) Hand-rolled state-machine parser ported from mlx-vlm-continuous' parse_tool_call (continuous_analyzer.py:98-146). 21 synthetic fixtures cover both formats, malformed cases, multi-tool, empty/nested args, and truncated/streaming inputs. Lib is project-agnostic by design so the same files cherry-pick into ik_llama.cpp's examples/vlib-video/ later. See VIDEO_CONV3D_CPP_DESIGN.md §7.1 S1+S2.