fix(factors/engine): 修复 DataRouter 缓存 key 缺少 columns 导致的缓存污染

src/factors/engine/data_router.py

@@ -191,9 +191,8 @@ class DataRouter:
         Returns:
             过滤后的 DataFrame
         """
-        cache_key = (
-            f"{spec.table}_{spec.join_type}_{start_date}_{end_date}_{stock_codes}"
-        )
+        cols_key = ",".join(sorted(spec.columns)) if spec.columns else "*"
+        cache_key = f"{spec.table}_{spec.join_type}_{start_date}_{end_date}_{stock_codes}_{cols_key}"
 
         with self._lock:
             if cache_key in self._cache:
@@ -259,7 +258,8 @@ class DataRouter:
         Returns:
             过滤后的 DataFrame
         """
-        cache_key = f"{table_name}_{start_date}_{end_date}_{stock_codes}"
+        cols_key = ",".join(sorted(columns)) if columns else "*"
+        cache_key = f"{table_name}_{start_date}_{end_date}_{stock_codes}_{cols_key}"
 
         with self._lock:
             if cache_key in self._cache:

src/training/pipeline.py

@@ -374,8 +374,8 @@ class DataPipeline:
                     split_data[split_name]["X"] = split_df.select(feature_cols)
                     split_data[split_name]["y"] = split_df[label_name]
 
-        # 删除标签为 NaN 的行
-        for split_name in ["train", "val", "test"]:
+        # 删除标签为 NaN 的行（仅在 train/val 上执行，test 集保留用于预测）
+        for split_name in ["train", "val"]:
             if split_name in split_data:
                 y_series = split_data[split_name]["y"]
                 y_nan_count = y_series.null_count()