diff --git a/logic_state_coverage.py b/logic_state_coverage.py
index 7a5a920..34d2ae3 100644
--- a/logic_state_coverage.py
+++ b/logic_state_coverage.py
@@ -99,7 +99,19 @@ def snapshot_of(day: str) -> dict:
     except Exception as e:  # noqa: BLE001
         print(f"  （行业观点快照表读取失败，乙路整体缺席: {e!r}）")
         return {}
-    return {str(r["cluster_key"]): dict(r) for _, r in df.iterrows()} if not df.empty else {}
+    if df.empty:
+        return {}
+    # pandas 把空值读成 NaN，而 NaN 是真值——直接往下传会让"没有上一版倾向"看着像有值。
+    # 全部归一成 None，判据那边只认 None。
+    import math
+
+    def _n(v):
+        if v is None or (isinstance(v, float) and math.isnan(v)):
+            return None
+        return v
+
+    return {str(r["cluster_key"]): {k: _n(v) for k, v in r.items()}
+            for _, r in df.iterrows()}
 
 
 def main(ds: str | None = None) -> None:
@@ -121,8 +133,11 @@ def main(ds: str | None = None) -> None:
 
     # 乙路按环节名对上主题：候选卡按环节，产业研判按主题聚簇，两者不在一个命名空间，
     # 这里只做同名匹配，对不上的票乙路就是缺失。这一路的天花板本来就低（实测 1.4%）。
-    by_seg = {(r.get("segment_name") or r.get("subject_name") or "").strip(): r
-              for r in snaps.values()}
+    by_seg = {}
+    for r in snaps.values():
+        name = str(r.get("segment_name") or r.get("subject_name") or "").strip()
+        if name:
+            by_seg[name] = r
     seg_of = {}
     try:
         d = db.read_pg("SELECT ts_code, target FROM v_factor_transmission_moved "