feat(ai-services): AVM v2 residential — expanded features, training pipeline, model versioning

Add neighborhood_score, developer_reputation, floor_level, direction premiums
to the multi-model ensemble. Implement real Optuna-based training pipeline
for XGBoost/LightGBM/CatBoost with grouped train/val/test splits. Add
file-based model registry with rollback and list-versions endpoints.
23 Python tests covering all new features.

Co-Authored-By: Paperclip <noreply@paperclip.ing>

libs/ai-services/app/routers/avm_v2.py

@@ -1,6 +1,6 @@
 """AVM v2 ensemble router — residential property valuation."""
 
-from fastapi import APIRouter
+from fastapi import APIRouter, HTTPException
 
 from app.models.avm_v2 import (
     ABComparisonRequest,
@@ -8,6 +8,7 @@ from app.models.avm_v2 import (
     AVMv2ModelInfo,
     AVMv2PredictRequest,
     AVMv2PredictResponse,
+    AVMv2RollbackRequest,
     AVMv2TrainRequest,
     AVMv2TrainResponse,
 )
@@ -30,7 +31,9 @@ def predict_v2(req: AVMv2PredictRequest) -> AVMv2PredictResponse:
 def train_v2(req: AVMv2TrainRequest) -> AVMv2TrainResponse:
     """Trigger model retraining with Optuna hyperparameter optimization.
 
-    Requires training data pipeline (Phase 3). Currently returns scaffold.
+    Loads training data from the model directory, runs Optuna for each
+    model in the ensemble, saves versioned artifacts, and registers
+    the new version in the model registry.
     """
     return avm_v2_service.train(req)
 
@@ -49,3 +52,22 @@ def compare_v1(req: ABComparisonRequest) -> ABComparisonResponse:
 def model_info_v2() -> AVMv2ModelInfo:
     """Get current active ensemble model information."""
     return avm_v2_service.get_model_info()
+
+
+@router.get("/versions", response_model=list[AVMv2ModelInfo])
+def list_versions() -> list[AVMv2ModelInfo]:
+    """List all registered model versions with their metrics and status."""
+    return avm_v2_service.list_versions()
+
+
+@router.post("/rollback", response_model=AVMv2ModelInfo)
+def rollback(req: AVMv2RollbackRequest) -> AVMv2ModelInfo:
+    """Rollback to a previously trained model version.
+
+    Copies the target version's artifacts to the active model directory,
+    reloads models, and updates the registry.
+    """
+    try:
+        return avm_v2_service.rollback(req.target_version)
+    except ValueError as e:
+        raise HTTPException(status_code=404, detail=str(e))