refactor(training): 简化 LightGBM 模型参数处理

- 重构 LightGBM 和 LambdaRank 模型，移除参数提取逻辑
- 模型类只保留 params 属性，符合 LightGBM 设计规范

src/experiment/learn_to_rank.py

@@ -293,6 +293,7 @@ SELECTED_FACTORS = [
     "net_profit_yoy",
     "revenue_yoy",
     "healthy_expansion_velocity",
+    "ebit_rank",
     # ================= 6. 基本面估值与截面动量共振 =================
     "EP",
     "BP",
@@ -304,12 +305,11 @@ SELECTED_FACTORS = [
     "pe_expansion_trend",
     "value_price_divergence",
     "active_market_cap",
-    "ebit_rank",
 ]
 
 # 因子定义字典（完整因子库）
 FACTOR_DEFINITIONS = {
-    "turnover_rate_volatility": "ts_std(log(turnover_rate), 20)"
+    # "turnover_rate_volatility": "ts_std(log(turnover_rate), 20)"
 }
 
 # Label 因子定义（不参与训练，用于计算目标）
@@ -341,7 +341,7 @@ MODEL_PARAMS = {
     "max_depth": 4,
     "min_data_in_leaf": 20,
     "n_estimators": 2000,
-    "early_stopping_round": 300,
+    "early_stopping_round": 100,
     "subsample": 0.8,
     "colsample_bytree": 0.8,
     "reg_alpha": 0.1,
@@ -372,7 +372,7 @@ def stock_pool_filter(df: pl.DataFrame) -> pl.Series:
     )
 
     valid_df = df.filter(code_filter)
-    n = min(1000, len(valid_df))
+    n = min(500, len(valid_df))
     small_cap_codes = valid_df.sort("total_mv").head(n)["ts_code"]
 
     return df["ts_code"].is_in(small_cap_codes)