Bug: 'passed=True' check for soft quality issues (AI套话, 缺互动引导)
always exits immediately even when LLM failed to make changes.
Score 85 stays 85, content unchanged.
Fix:
1. polish_with_llm: retry once on exception (was: single attempt)
2. main loop: only accept fix when score actually improves AND
LLM produced output (opt_logs non-empty). If LLM returns same html
unchanged, treat as failed attempt and retry.
Yuzhiran Dev
004698cf98