Improve Go commentary with grounded move evidence and concise explanations

[DavidDing16]

Summary

Current-move commentary can confuse a geometric relationship with tactical intent, focus on individual stones while missing group development or whole-board priorities, and produce long explanations. This change supplies verifiable before-move evidence and asks the teacher to select one supported purpose for a concise explanation.

• Separate observed geometry, group relationships, intent hypotheses and verified move effects. Add friendly development, regional contact/separation and ko-connection evidence without inferring life/death or forced success from shape alone.
• Compare recommended and played choices from the same before-move root, with current-player perspective, search-quality checks and protection for near-optimal local alternatives. Keep specific played-move loss distinct from region-wide priority.
• Correct ownership parsing, knowledge-match overstatement, analysis reuse/deepening, tool-history retention, initial image delivery and evidence ordering under tool-message truncation.
• Default ordinary current-move requests to one short purpose sentence; retain detailed explanations when explicitly requested.
• Include reusable research/evaluation and local review tools, behavioral regression tests, and a concise methodology/limitations document.

Verification

• pnpm test (392 tests passed).
• pnpm typecheck.
• pnpm build.
• python -m unittest discover -s scripts/tests (11 tests).
• Local review UI: exact generated text, before/after boards, desktop/mobile layout and feedback persistence.
• Real ChatGPT GPT-6 Luna/KataGo development runs and human Go review; generated text was preserved rather than hand-edited.

Notes

For the best teaching results, we recommend a multimodal model that accepts images. This recommendation is documented in both Chinese and English READMEs. The current workflow validates board images alongside exact SGF coordinates and KataGo evidence. Structured board evidence itself does not require vision, but a validated text-only workflow and controlled image/no-image comparison are not part of this PR.

Intent candidates remain hypotheses. Ko counterfactuals validate board structure, not historical superko legality. A complete pipeline or passing tests do not establish semantic accuracy; reviewed positions used for fixes are development regressions. Some generated explanations still include unnecessary secondary observations.

The exploratory linear/CNN/probe routes are documented; no trained weights were integrated. Private SGFs, source comments, expert feedback files, local credentials, model binaries, generated reports and machine-specific evaluation artifacts are excluded. Selected local board/game/KataGo/knowledge evidence continues to be sent to the user-configured LLM, not web search. Packaging/release artifacts have not been validated or published.