feat: add deterministic no-LLM code graph

[tkhjp]

Summary

• add an /understand-no-llm skill for deterministic code-graph generation
• reuse the existing scanner, import resolver, Tree-sitter extractors, and KnowledgeGraph validation
• emit file/function/class nodes and contains/imports/calls edges
• add Japanese documentation covering architecture, file responsibilities, usage, security boundaries, and limitations

Why

Some environments cannot send source code to an LLM but still need a navigable structural code graph. This adds a narrower execution path that performs local static analysis only while keeping the normal KnowledgeGraph JSON shape for dashboard compatibility.

Behavior and scope

• does not invoke LLM agents or external AI APIs
• does not execute target repository build, test, application, or shell scripts
• resolves calls conservatively only when caller and callee are unambiguous
• writes code-graph.json and a coverage/report file
• leaves layers and tour empty and does not claim semantic parity with /understand
• can optionally copy the result to knowledge-graph.json for dashboard use

Validation

• node --check understand-anything-plugin/skills/understand/no-llm-code-graph.mjs
• pnpm --filter @understand-anything/core test: 35 files, 753 tests passed
• smoke-tested against soongenwong/claudecode with --no-scripts: 119 code files, 1,800 nodes, 2,578 edges, 0 skipped structure files

[Lum1104]

Hi, the idea is great. How about we add this to the /understand skill, and let agents pause after the non-llm graph generation? Or, at the very beginning, during user's interaction with .understand ignore?