feat: 训练框架适配 DualStream_MNF + 部署矩阵 .npz 导出

modeling_batch.py:
- preprocessing_methods 新增 DualStream_MNF
- save_model 新增 C++/Rust 部署导出逻辑:
  自动检测 Pipeline 中的 MNFTransformer + SVR，
  提取 mean_/W_mnf_/dual_coef_/support_vectors_/
  intercept_/gamma/kernel 等纯量矩阵，
  保存为 *_deploy.npz 文件
- 新增 _find_mnf_in_pipeline / _find_svr_in_pipeline /
  _find_physical_extractor_in_pipeline 辅助方法，
  支持 FeatureUnion 嵌套递归查找

src/core/modeling/modeling_batch.py

@@ -223,7 +223,8 @@ class WaterQualityModelingBatch:
 
         # 预处理方法列表
         self.preprocessing_methods = [
-            "None", "MMS", "SS", "CT", "SNV", "MA", "SG", "MSC", "D1", "D2", "DT", "WVAE"
+            "None", "MMS", "SS", "CT", "SNV", "MA", "SG", "MSC", "D1", "D2", "DT", "WVAE",
+            "DualStream_MNF"
         ]
 
         # 样本划分方法列表
@@ -716,7 +717,7 @@ class WaterQualityModelingBatch:
         """
         # 清理目标列名，移除可能的特殊字符
         safe_target_name = "".join(c for c in target_column_name if c.isalnum() or c in ('-', '_')).rstrip()
-        
+
         filename = f"{safe_target_name}_{preprocess_method}_{model_name}.joblib"
         filepath = self.artifacts_dir / filename
 
@@ -732,6 +733,112 @@ class WaterQualityModelingBatch:
         joblib.dump(save_data, filepath)
         print(f"模型已保存: {filepath}")
 
+        # ═══════════════════════════════════════════════════════════
+        # ★ C++/Rust 部署导出：提取 MNF + SVR 纯量矩阵
+        # ═══════════════════════════════════════════════════════════
+        if not isinstance(model, Pipeline):
+            return
+
+        _mnf = self._find_mnf_in_pipeline(model)
+        _svr = self._find_svr_in_pipeline(model)
+        _phys = self._find_physical_extractor_in_pipeline(model)
+
+        if _mnf is not None and _svr is not None:
+            _deploy = {
+                'mnf_mean': np.asarray(_mnf.mean_).ravel(),
+                'mnf_W':    np.asarray(_mnf.W_mnf_),
+                'mnf_n_components': int(_mnf.n_components),
+            }
+            # SVR 部署矩阵（rbf kernel 需要 support_vectors_）
+            _deploy['svr_dual_coef'] = np.asarray(_svr.dual_coef_)
+            _deploy['svr_intercept'] = np.asarray(_svr.intercept_)
+            if hasattr(_svr, 'support_vectors_'):
+                _deploy['svr_support_vectors'] = np.asarray(_svr.support_vectors_)
+            _deploy['svr_gamma'] = float(getattr(_svr, '_gamma', 1.0))
+            # kernel type for C++ dispatch
+            _kernel = str(getattr(_svr, 'kernel', 'rbf')).lower()
+            _deploy['svr_kernel'] = _kernel
+            if _kernel == 'poly':
+                _deploy['svr_degree'] = int(getattr(_svr, 'degree', 3))
+                _deploy['svr_coef0']  = float(getattr(_svr, 'coef0', 0.0))
+
+            # 物理特征索引（C++ 端据此定位波段列）
+            if _phys is not None:
+                _deploy['physical_wavelengths'] = np.asarray(
+                    _phys.wavelengths, dtype=np.float64
+                )
+                _feat_map = {}
+                for name, ia, ib in _phys._feat_cols_:
+                    _feat_map[name] = {
+                        'wl_a': float(_phys.wavelengths[ia]),
+                        'wl_b': float(_phys.wavelengths[ib]),
+                        'idx_a': int(ia),
+                        'idx_b': int(ib),
+                    }
+                _deploy['physical_feature_map'] = _feat_map
+
+            npz_path = str(filepath).replace('.joblib', '_deploy.npz')
+            np.savez_compressed(npz_path, **_deploy)
+            print(f"[部署导出] MNF+SVR 纯量矩阵已保存: {npz_path}")
+        elif _mnf is None and _svr is not None:
+            print("[部署导出] 仅检测到 SVR 未检测到 MNFTransformer，跳过 .npz 导出")
+
+    # ═══════════════════════════════════════════════════════════
+    # 部署矩阵提取辅助方法：从 Pipeline 中定位特定 Transformer
+    # ═══════════════════════════════════════════════════════════
+
+    @staticmethod
+    def _find_transformer_in_pipeline(pipeline, target_class_name):
+        """在 Pipeline 的 'preproc' 步骤中递归查找指定类型的 Transformer。
+
+        支持 FeatureUnion 嵌套：若 preproc 是 FeatureUnion，
+        则遍历其 transformer_list 逐个子变压器查找。
+        """
+        from sklearn.pipeline import FeatureUnion
+        preproc = pipeline.named_steps.get('preproc')
+        if preproc is None:
+            return None
+        if isinstance(preproc, FeatureUnion):
+            for name, trans in preproc.transformer_list:
+                found = WaterQualityModelingBatch._find_transformer_in_step(
+                    trans, target_class_name)
+                if found is not None:
+                    return found
+            return None
+        return WaterQualityModelingBatch._find_transformer_in_step(
+            preproc, target_class_name)
+
+    @staticmethod
+    def _find_transformer_in_step(step, target_class_name):
+        """在单个 Pipeline 步骤（可能是 Pipeline 自身）中查找 Transformer。"""
+        if step.__class__.__name__ == target_class_name:
+            return step
+        # 处理 Pipeline 嵌套
+        if hasattr(step, 'named_steps'):
+            for sub_name, sub_step in step.named_steps.items():
+                found = WaterQualityModelingBatch._find_transformer_in_step(
+                    sub_step, target_class_name)
+                if found is not None:
+                    return found
+        return None
+
+    @staticmethod
+    def _find_mnf_in_pipeline(pipeline):
+        return WaterQualityModelingBatch._find_transformer_in_pipeline(
+            pipeline, 'MNFTransformer')
+
+    @staticmethod
+    def _find_svr_in_pipeline(pipeline):
+        model_step = pipeline.named_steps.get('model')
+        if model_step is not None and model_step.__class__.__name__ == 'SVR':
+            return model_step
+        return None
+
+    @staticmethod
+    def _find_physical_extractor_in_pipeline(pipeline):
+        return WaterQualityModelingBatch._find_transformer_in_pipeline(
+            pipeline, 'PhysicalFeatureExtractor')
+
     def train_models_batch(self, csv_path: str, feature_start_column: Union[int, str],
                           preprocessing_methods: Union[str, List[str]] = "None",
                           model_names: Union[str, List[str]] = "RF",