fix: correct model fallback verification for custom providers

- Fix base_url lookup for custom provider: use model.base_url or custom_providers
- Remove colon heuristic: only ollama provider uses Ollama API path
  The colon check incorrectly routed custom models (hermes4:14b) to Ollama API,
  causing verification failures for llama.cpp servers.

Closes #514

bin/model-fallback-verify.py

@@ -121,7 +121,7 @@ def test_model(model, provider, api_key=None, base_url=None):
     Test a model with a verification prompt.
     Returns (success, response, error_message)
     """
-    if provider == "ollama" or ":" in model:
+    if provider == "ollama":
         # Local Ollama model
         ollama_host = os.environ.get("OLLAMA_HOST", "localhost:11434")
         try:
@@ -306,8 +306,16 @@ def run_verification():
         
         base_url = None
         if provider == "custom":
-            provider_config = config.get("auxiliary", {}).get("vision", {})
+            # Prefer base_url from primary model config if present
-            base_url = provider_config.get("base_url")
+            model_cfg = config.get("model", {})
+            if isinstance(model_cfg, dict) and model_cfg.get("provider") == "custom":
+                base_url = model_cfg.get("base_url")
+            if not base_url:
+                # Fallback: search custom_providers by matching model name
+                for p in config.get("custom_providers", []):
+                    if p.get("model") == model:
+                        base_url = p.get("base_url")
+                        break
         
         log("Testing [" + role + "] " + model + " (" + provider + ")...")
         result = verify_model(model, provider, api_key, base_url)