Improve FACTScore precision: negative assertion filtering, dense retrieval, LLM synthesis mode

Post-hoc replace bias_category_description with non-MBFC text

Fix system mode boilerplate leakage: rewrite bias category descriptions

Expand audit to all 6 modes, skip missing JSONL files gracefully

Also sanitize log files: remove stale [Search] LEAKAGE warning lines

Add post-hoc sanitization script for existing JSONL results

fix search results

Fix false-positive LEAKAGE warnings for -site: exclusion operators

Add data leakage audit script for JSONL results and logs

Comprehensive code review: fix MBFC methodology, add rating systems, cleanup