fix(normalize): 原子化解析容错+产出率兜底——LLM输出no为字符串/浮点时isinstance(int)全过滤致46文档仅6原子;int()转换+范围校验;批LLM失败用标题作原子;原子覆盖<50%文档时缺者标题兜底(众包标题即浓缩语义,保证覆盖率可算)

pipeline_opportunity/opp_normalize.py

@@ -128,17 +128,33 @@ async def _atomize(sor, snaps, cfg, ctx):
                 arr = json.loads(raw)
                 items = []
                 for it in arr if isinstance(arr, list) else []:
+                    if not isinstance(it, dict):
+                        continue
                     no = it.get("no")
                     atom = str(it.get("atom") or "").strip()[:255]
-                    if isinstance(no, int) and 1 <= no <= len(chunk) and atom:
+                    # 容错：LLM 可能输出 "1"/1.0/True，统一 int() 转换后范围校验
-                        items.append((chunk[no - 1]["id"], atom))
+                    try:
+                        no_i = int(no)
+                    except (TypeError, ValueError):
+                        continue
+                    if 1 <= no_i <= len(chunk) and atom:
+                        items.append((chunk[no_i - 1]["id"], atom))
                 async with lock:
                     results.extend(items)
             except Exception as e:
                 logger.debug("atomize chunk failed: %s", e)
+                async with lock:
+                    results.extend([(s["id"], s["title"][:255]) for s in chunk])
 
     chunks = [snaps[i:i + bs] for i in range(0, len(snaps), bs)]
     await asyncio.gather(*(_one(ch) for ch in chunks))
+    # 产出率兜底：原子数 < 文档数 50% 时，缺原子的文档用标题作原子
+    # （众包标题本身即浓缩需求语义；保证覆盖率统计可算，2026-09-12 实测 46 文档仅 6 原子）
+    covered = set(sid for sid, _ in results)
+    if len(covered) < len(snaps) * 0.5:
+        for s in snaps:
+            if s["id"] not in covered:
+                results.append((s["id"], s["title"][:255]))
     return results