fix(evals): 漏响应在所有表里都删；评分按句子重叠合并重复的标注单元

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

evals/check/score.py

@@ -64,6 +64,20 @@ def main() -> None:
         for s in u.sids:
             sid_units[s].add(u.id)
 
+    # 标注集里同一条要求有时记了两个单元（两个类别各一条，或表格一行拆成几条）：句子有重叠的单元合成一组，按组计
+    parent = {u.id: u.id for u in units}
+
+    def find(x: str) -> str:
+        while parent[x] != x:
+            x = parent[x]
+        return x
+
+    for ids in sid_units.values():
+        first, *rest = sorted(ids)
+        for o in rest:
+            parent[find(o)] = find(first)
+    group = {u.id: find(u.id) for u in units}
+
     per_unit: dict[str, list[dict[str, Any]]] = defaultdict(list)
     unmapped: list[dict[str, Any]] = []
     for it in res["items"]:
@@ -82,8 +96,13 @@ def main() -> None:
     fp_by_trap: Counter[str] = Counter()
     ok_units = trap_units = 0
     other_findings = 0
-    for uid, e in manifest["units"].items():
-        items = per_unit.get(uid, [])
+    seen_groups: set[str] = set()
+    for uid, e in sorted(manifest["units"].items(), key=lambda kv: kv[1].get("expect") != "defect"):
+        g = group.get(uid, uid)
+        if g in seen_groups:
+            continue  # 同组的单元只算一次（挂了问题的那个优先）
+        seen_groups.add(g)
+        items = [it for x in manifest["units"] if group.get(x, x) == g for it in per_unit.get(x, [])]
         found = [it for it in items if it["finding"]]
         if e["expect"] == "defect":
             if not e.get("reviewed") and not a.include_unreviewed: