diff --git a/sanguo_factor/fast_ops.py b/sanguo_factor/fast_ops.py
index b618049..7fc5ff9 100644
--- a/sanguo_factor/fast_ops.py
+++ b/sanguo_factor/fast_ops.py
@@ -61,7 +61,12 @@ def fast_ts_corr(feature1: DataProxy, feature2: DataProxy, window: int) -> DataP
 
     Uses correlation identity: corr(x,y) = cov(x,y) / (std_x * std_y)
     """
-    df_merged = feature1.df.join(feature2.df, on=["datetime", "vt_symbol"])
+    # 两操作数源自同一父df且全算子保长保序(rolling/shift/cs_rank 均保长)——
+    # 按位拼接等价于按键join,免 10M 行 hash join(py-spy 实锤分钟级)
+    h1, h2 = feature1.df.height, feature2.df.height
+    if h1 != h2:
+        raise ValueError(f"ts_corr 操作数长度不一致: {h1} vs {h2}(行序对齐前提被破坏)")
+    df_merged = feature1.df.with_columns(feature2.df["data"].alias("data_right"))
 
     # Use native polars rolling correlation for exact match
     corr_result = pl.rolling_corr(
@@ -92,7 +97,11 @@ def fast_ts_cov(feature1: DataProxy, feature2: DataProxy, window: int) -> DataPr
     corr_result = fast_ts_corr(feature1, feature2, window)
 
     # Extract std deviations from the merged dataframe
-    df_merged = feature1.df.join(feature2.df, on=["datetime", "vt_symbol"])
+    # 两操作数源自同一父df且全算子保长保序——按位拼接等价于按键join
+    h1, h2 = feature1.df.height, feature2.df.height
+    if h1 != h2:
+        raise ValueError(f"ts_cov 操作数长度不一致: {h1} vs {h2}(行序对齐前提被破坏)")
+    df_merged = feature1.df.with_columns(feature2.df["data"].alias("data_right"))
 
     std_x = pl.col("data").rolling_std(window, min_samples=1, ddof=0).over("vt_symbol")
     std_y = pl.col("data_right").rolling_std(window, min_samples=1, ddof=0).over("vt_symbol")
@@ -417,7 +426,12 @@ def fast_ts_corr_v2(feature1: DataProxy, feature2: DataProxy, window: int) -> Da
     Uses: corr = (E[xy] - E[x]E[y]) / (std_x * std_y)
     All statistics computed via shift-based mean/std
     """
-    df_merged = feature1.df.join(feature2.df, on=["datetime", "vt_symbol"])
+    # 两操作数源自同一父df且全算子保长保序(rolling/shift/cs_rank 均保长)——
+    # 按位拼接等价于按键join,免 10M 行 hash join(py-spy 实锤分钟级)
+    h1, h2 = feature1.df.height, feature2.df.height
+    if h1 != h2:
+        raise ValueError(f"ts_corr_v2 操作数长度不一致: {h1} vs {h2}(行序对齐前提被破坏)")
+    df_merged = feature1.df.with_columns(feature2.df["data"].alias("data_right"))
 
     x = pl.col("data")
     y = pl.col("data_right")
@@ -476,13 +490,17 @@ def fast_ts_cov_v2(feature1: DataProxy, feature2: DataProxy, window: int) -> Dat
     std_y_result = fast_ts_std(feature2, window)
 
     # Merge and compute covariance
-    df_merged = corr_result.df.join(
-        std_x_result.df.select(["datetime", "vt_symbol", pl.col("data").alias("std_x")]),
-        on=["datetime", "vt_symbol"]
-    ).join(
-        std_y_result.df.select(["datetime", "vt_symbol", pl.col("data").alias("std_y")]),
-        on=["datetime", "vt_symbol"]
-    )
+    # 三操作数源自同一父df且全算子保长保序——按位拼接等价于按键join
+    h_corr, h_x = corr_result.df.height, std_x_result.df.height
+    if h_corr != h_x:
+        raise ValueError(f"ts_cov_v2 corr与std_x长度不一致: {h_corr} vs {h_x}")
+    h_y = std_y_result.df.height
+    if h_corr != h_y:
+        raise ValueError(f"ts_cov_v2 corr与std_y长度不一致: {h_corr} vs {h_y}")
+    df_merged = corr_result.df.with_columns([
+        std_x_result.df["data"].alias("std_x"),
+        std_y_result.df["data"].alias("std_y")
+    ])
 
     df = df_merged.select(
         pl.col("datetime"),