fix: auto-generate lora_config.json in training script

This fix addresses the missing lora_config.json files in LoRA intent
classifier models, which caused Go tests to skip and model routing to fail.

Issue discovered by @yehudit1987 while investigating skipped classifier tests.

Changes:
- Updated merge_lora_adapter_to_full_model() in ft_linear_lora.py to
  automatically create lora_config.json when merging LoRA adapters

The lora_config.json file contains LoRA metadata (rank, alpha, dropout,
target_modules) required by the Rust router to detect LoRA models and
route them to the correct inference path.

The file will be manually uploaded to HuggingFace for existing models,
and future training runs will generate it automatically.

Fixes #627

Signed-off-by: Yossi Ovadia <[email protected]>

Author: yossiovadia

src/training/training_lora/classifier_model_fine_tuning_lora/ft_linear_lora.py

@@ -659,6 +659,27 @@ def merge_lora_adapter_to_full_model(
         )
         logger.info("Created category_mapping.json")
 
+    # Create lora_config.json for Rust router detection
+    # This file signals to the Rust router that this is a LoRA-trained model
+    # and should be routed to the LoRA inference path
+    logger.info("Creating lora_config.json for LoRA model detection...")
+    lora_config = {
+        "rank": 16,  # LoRA rank (r) - matches training configuration
+        "alpha": 32,  # LoRA alpha scaling factor
+        "dropout": 0.1,  # LoRA dropout rate
+        "target_modules": [
+            "attention.self.query",
+            "attention.self.value",
+            "attention.output.dense",
+            "intermediate.dense",
+            "output.dense",
+        ],
+    }
+    lora_config_path = os.path.join(output_path, "lora_config.json")
+    with open(lora_config_path, "w") as f:
+        json.dump(lora_config, f)
+    logger.info(f"Created {lora_config_path}")
+
     logger.info("LoRA adapter merged successfully!")