优化fix脚本

@@ -69,17 +69,18 @@ def should_remove(row: dict, main_durations: dict[str, float], threshold: float)
 def compute_summary(rows: list[dict], threshold: float, out_path: str) -> dict:
     """计算评测指标。"""
     tp = fp = tn = fn = 0
-    upload_ms = poll_ms = total_ms = []
+    upload_ms, poll_ms, total_ms = [], [], []
     by_variant: dict[str, dict] = {}
 
     for r in rows:
         pred = r.get("predicted_duplicate", "").lower() == "true"
         exp = r.get("expected_duplicate", "").lower() == "true"
         correct = r.get("correct", "").lower() == "true"
+        sample_class = r.get("sample_class", "unknown")
 
         variant = r.get("variant", "unknown")
         if variant not in by_variant:
-            by_variant[variant] = {"correct": 0, "total": 0}
+            by_variant[variant] = {"correct": 0, "total": 0, "sample_class": sample_class}
         by_variant[variant]["total"] += 1
         if correct:
             by_variant[variant]["correct"] += 1
@@ -137,8 +138,22 @@ def compute_summary(rows: list[dict], threshold: float, out_path: str) -> dict:
         "query_time_ms": _stats(total_ms),
         "upload_time_ms": _stats(upload_ms),
         "poll_time_ms": _stats(poll_ms),
+        "by_sample_class": {
+            cls: {
+                "by_variant": {
+                    v: {"accuracy": round(d["correct"] / d["total"], 4), "total": d["total"]}
+                    for v, d in sorted(by_variant.items())
+                    if d["sample_class"] == cls
+                }
+            }
+            for cls in ("positive", "negative")
+        },
         "by_variant": {
-            v: {"accuracy": round(d["correct"] / d["total"], 4), "total": d["total"]}
+            v: {
+                "sample_class": d["sample_class"],
+                "accuracy": round(d["correct"] / d["total"], 4),
+                "total": d["total"],
+            }
             for v, d in sorted(by_variant.items())
         },
         "out": out_path,
@@ -237,10 +252,14 @@ def main():
     print(f"Recall:    {summary['recall']}")
     print(f"F1:        {summary['f1']}")
     print(f"TP={summary['tp']} FP={summary['fp']} TN={summary['tn']} FN={summary['fn']}")
-    print()
-    print("按变体:")
-    for v, d in summary["by_variant"].items():
-        print(f"  {v:20s} acc={d['accuracy']:.4f}  n={d['total']}")
+
+    for cls in ("positive", "negative"):
+        variants = summary["by_sample_class"][cls]["by_variant"]
+        if not variants:
+            continue
+        print(f"\n[{cls}]")
+        for v, d in variants.items():
+            print(f"  {v:20s} acc={d['accuracy']:.4f}  n={d['total']}")
 
     # 打印被移除的样本明细
     if removed: