test(debug): 添加因子回测一致性问题的调试测试套件

- 分析GTJA_alpha032等因子在不同LOOKBACK_DAYS下的差异来源
- 验证cs_rank嵌套和截面股票数量对结果的影响
- 测试ts_rank NaN处理和除法除零修复

src/factors/translator.py

@@ -184,7 +184,8 @@ class PolarsTranslator:
             "+": lambda l, r: l + r,
             "-": lambda l, r: l - r,
             "*": lambda l, r: l * r,
-            "/": lambda l, r: l / r,
+            # 【修复】除法处理除零，避免产生 NaN/inf 导致 EMA 永久感染
+            "/": lambda l, r: pl.when(r == 0).then(None).otherwise(l / r),
             "**": lambda l, r: l.pow(r),
             "//": lambda l, r: l.floor_div(r),
             "%": lambda l, r: l % r,
@@ -363,6 +364,7 @@ class PolarsTranslator:
         # 抛弃极慢的 rolling_map，借用 pandas 的 Cython 引擎
         def kurt_calc(s: pl.Series) -> pl.Series:
             import pandas as pd
+
             # pandas.rolling.kurt() 是用 Cython 编写的，速度比 pure python 快很多
             pd_series = pd.Series(s.to_numpy())
             result = pd_series.rolling(window).kurt().to_numpy()
@@ -499,11 +501,21 @@ class PolarsTranslator:
             # 当前值即为每个窗口的最后一个元素 (N - window + 1, )
             current_vals = windows[:, -1]
 
-            # 向量化广播比较，然后沿窗口轴(axis=1)求和，直接得出排名比例
-            ranks = np.sum(windows <= current_vals[:, None], axis=1) / window
+            # 【终极修复】使用窗口内实际有效数据个数作为分母
+            # 1. 统计小于等于当前值的个数
+            less_equal = np.sum(windows <= current_vals[:, None], axis=1)
+            # 2. 统计当前窗口内有效的非 NaN 数据个数
+            valid_counts = np.sum(~np.isnan(windows), axis=1)
+
+            # 3. 使用真实有效个数作为分母，避免分母陷阱
+            with np.errstate(divide="ignore", invalid="ignore"):
+                ranks = np.where(valid_counts > 0, less_equal / valid_counts, np.nan)
+
+            # 【修复】如果当前值是 NaN，则排名也必须是 NaN
+            ranks[np.isnan(current_vals)] = np.nan
 
             result = np.full(n, np.nan)
-            result[window - 1:] = ranks
+            result[window - 1 :] = ranks
             return pl.Series(result)
 
         return expr.map_batches(rank_calc, return_dtype=pl.Float64)
@@ -592,7 +604,7 @@ class PolarsTranslator:
             distances = window - 1 - argmax_indices
 
             result = np.full(n, np.nan)
-            result[window - 1:] = distances
+            result[window - 1 :] = distances
             return pl.Series(result)
 
         return expr.map_batches(argmax_calc, return_dtype=pl.Float64)
@@ -616,7 +628,7 @@ class PolarsTranslator:
             distances = window - 1 - argmin_indices
 
             result = np.full(n, np.nan)
-            result[window - 1:] = distances
+            result[window - 1 :] = distances
             return pl.Series(result)
 
         return expr.map_batches(argmin_calc, return_dtype=pl.Float64)
@@ -650,7 +662,7 @@ class PolarsTranslator:
             prods = np.prod(windows, axis=1)
 
             result = np.full(n, np.nan)
-            result[window - 1:] = prods
+            result[window - 1 :] = prods
             return pl.Series(result)
 
         return expr.map_batches(prod_calc, return_dtype=pl.Float64)