fix: anchor stopwords - remove generic question patterns causing cross-topic contamination

- 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

experiments/phase1_baseline_results.json

@@ -0,0 +1,157 @@
+{
+  "Last-3": {
+    "avg_tokens": 43.6,
+    "contamination_rate": 100.0,
+    "raw": [
+      {
+        "label": "问PG",
+        "pt": 42,
+        "cont": true
+      },
+      {
+        "label": "问Git",
+        "pt": 44,
+        "cont": true
+      },
+      {
+        "label": "问Redis",
+        "pt": 43,
+        "cont": true
+      },
+      {
+        "label": "问asyncio",
+        "pt": 43,
+        "cont": true
+      },
+      {
+        "label": "再问Git",
+        "pt": 46,
+        "cont": true
+      }
+    ]
+  },
+  "Last-5": {
+    "avg_tokens": 67.6,
+    "contamination_rate": 100.0,
+    "raw": [
+      {
+        "label": "问PG",
+        "pt": 66,
+        "cont": true
+      },
+      {
+        "label": "问Git",
+        "pt": 68,
+        "cont": true
+      },
+      {
+        "label": "问Redis",
+        "pt": 67,
+        "cont": true
+      },
+      {
+        "label": "问asyncio",
+        "pt": 67,
+        "cont": true
+      },
+      {
+        "label": "再问Git",
+        "pt": 70,
+        "cont": true
+      }
+    ]
+  },
+  "Last-10": {
+    "avg_tokens": 137.6,
+    "contamination_rate": 100.0,
+    "raw": [
+      {
+        "label": "问PG",
+        "pt": 136,
+        "cont": true
+      },
+      {
+        "label": "问Git",
+        "pt": 138,
+        "cont": true
+      },
+      {
+        "label": "问Redis",
+        "pt": 137,
+        "cont": true
+      },
+      {
+        "label": "问asyncio",
+        "pt": 137,
+        "cont": true
+      },
+      {
+        "label": "再问Git",
+        "pt": 140,
+        "cont": true
+      }
+    ]
+  },
+  "BM25-5": {
+    "avg_tokens": 70.6,
+    "contamination_rate": 60.0,
+    "raw": [
+      {
+        "label": "问PG",
+        "pt": 68,
+        "cont": true
+      },
+      {
+        "label": "问Git",
+        "pt": 74,
+        "cont": true
+      },
+      {
+        "label": "问Redis",
+        "pt": 70,
+        "cont": false
+      },
+      {
+        "label": "问asyncio",
+        "pt": 67,
+        "cont": true
+      },
+      {
+        "label": "再问Git",
+        "pt": 74,
+        "cont": false
+      }
+    ]
+  },
+  "Full CGK": {
+    "avg_tokens": 42.6,
+    "contamination_rate": 0.0,
+    "raw": [
+      {
+        "label": "问PG",
+        "pt": 18,
+        "cont": false
+      },
+      {
+        "label": "问Git",
+        "pt": 59,
+        "cont": false
+      },
+      {
+        "label": "问Redis",
+        "pt": 19,
+        "cont": false
+      },
+      {
+        "label": "问asyncio",
+        "pt": 56,
+        "cont": false
+      },
+      {
+        "label": "再问Git",
+        "pt": 61,
+        "cont": false
+      }
+    ]
+  }
+}