Phase2 complete ablation (added missing variants):
- Coverage-only: 20% contamination rate (confirms Gate is critical)
- Gate-only: +5.2 tokens vs Full (coverage optimization marginal on clean data)
- -Recency: 0 effect on clean data
- -IDF: 0 effect on clean data
Phase4 end-to-end quality evaluation:
- CGK vs Last-5 across 5 queries:
* CGK: 42.2 tok, purity=1.000, anchor_recall=0.638, term_cov=0.380, contamination=0
* Last-5: 67.6 tok, purity=0.280, anchor_recall=0.066, term_cov=0.080, contamination=5
- All quality metrics CGK >> Last-5 on synthetic clean data
Known honest limitations:
- Still no real dialogue data (synthetic 4-topic only)
- No real LLM calls (quality is rule-estimated)
- Parameter sensitivity only on clean data, not noisy real data
- Add ANCHOR_STOPWORDS set in anchor.py (真正通用的疑问pattern)
- Filter Chinese n-grams against stopwords in extract()
- Update sparse.py content_words extraction to use stopword-filtered query
- Diagnosis: 'Git rebase vs merge' query now correctly excludes Redis/asyncio blocks
- Phase1 results: Full CGK 42.6 tokens avg, 0% contamination (vs Last-5 67.6 tokens, 100%)
- Phase2 ablation: Gate-only accounts for most of the benefit
- Phase3 sensitivity: OVERLAP/NEW_RATIO thresholds insensitive on clean data;
RECENT_WINDOW is the primary token budget control
Known honest limitations:
- Test set is clean 4-topic synthetic data (no real dirty dialogue)
- No strong baselines (BM25 ablation incomplete)
- No answer-level evaluation (only retrieval blocks measured)
- No parameter sensitivity on noisy real-world data
- Zero contamination on 5 queries is not generalizable
