feat: add training preprocessing and model routing

src/lmpm/training/routing.py

@@ -0,0 +1,51 @@
+"""Map each experiment outcome to its initial model family."""
+
+from typing import Literal
+
+from pydantic import BaseModel, ConfigDict
+
+from lmpm.training.targets import TargetTask, target_specs
+
+ModelFamily = Literal["gaussian_process_regressor", "random_forest_classifier"]
+
+
+class TargetModelRoute(BaseModel):
+    """The default estimator family for one target and learning task."""
+
+    model_config = ConfigDict(frozen=True)
+
+    target_name: str
+    task: TargetTask
+    model_family: ModelFamily
+
+
+def _default_route(target_name: str, task: TargetTask) -> TargetModelRoute:
+    if task == "regression":
+        return TargetModelRoute(
+            target_name=target_name,
+            task=task,
+            model_family="gaussian_process_regressor",
+        )
+    return TargetModelRoute(
+        target_name=target_name,
+        task=task,
+        model_family="random_forest_classifier",
+    )
+
+
+DEFAULT_TARGET_MODEL_ROUTES = tuple(
+    _default_route(spec.name, spec.task) for spec in target_specs()
+)
+
+
+def model_route_for(target_name: str) -> TargetModelRoute:
+    """Return the default compatible model route for a known outcome field."""
+    for route in DEFAULT_TARGET_MODEL_ROUTES:
+        if route.target_name == target_name:
+            return route
+    raise ValueError(f"unsupported training target: {target_name}")
+
+
+def default_model_routes() -> tuple[TargetModelRoute, ...]:
+    """Return every stable target-to-model mapping in dataset column order."""
+    return DEFAULT_TARGET_MODEL_ROUTES