Publish league baseline, G2 result, and PPO diagnostics

[jacklee12312]

Summary

• publish the current evolving PFSP/league branch and a concise, bilingual G2 experiment report
• record that all six G2 lineages completed 1.25M-to-1.50M training and evaluation, but G2 was not promoted because the worst-lineage gate failed
• add a compact machine-readable G2 summary with exact source-report and checkpoint SHA-256 hashes
• add PPO/critic diagnostics: explained variance, value/return/advantage statistics, approximate KL, policy clip fraction, and gradient clipping diagnostics
• document metric semantics and make clear that these new fields are unavailable retroactively for G2

Verification

• python -m unittest discover -s tests -v: 3032 passed, 1 skipped
• python -m compileall -q swb scripts tests: passed
• 16-agent-step standalone PPO smoke: 15 required diagnostic fields present and finite
• G2 summary arithmetic and all five local source-report SHA-256 hashes verified

Scope notes

• checkpoint binaries and full per-game generation reports remain excluded because of size
• existing local edits to the roadmap/checklist and untracked raw reports were not included in commit ca66bed
• the repository is currently private; merging this PR updates the default branch, while public visibility remains a separate repository setting