fix: guard against None model_name in OpenAI endpoint (#1700) (#2358)

Fix AttributeError: 'NoneType' object has no attribute 'lower' that
occurs when model_name is None (common with Azure OpenAI deployments).

The issue manifests when a None model_name propagates to langchain's
standardize_model_name() which calls .lower() on it without null-checking.

Changes:
- handle_wrapped_call: Check bindings.arguments as fallback, fall back
  to response.model when model is None
- _handle_response: Guard model_name to ensure it is always a string
- handle_generation_chunk: Guard response.model with 'or empty string'
- OpenAICostComputer.handle_response: Guard first_chunk.model

Add 11 unit tests covering both non-OTEL and OTEL code paths.

Co-authored-by: sfc-gh-jreini <sfc-gh-jreini@users.noreply.github.com>

Author: sfc-gh-jreini

src/providers/openai/trulens/providers/openai/endpoint.py

@@ -129,7 +129,7 @@ def handle_response(response: Any) -> Dict[str, Any]:
         if hasattr(response, "__iter__") and not hasattr(response, "model"):
             try:
                 first_chunk = next(response)
-                model_name = first_chunk.model
+                model_name = first_chunk.model or ""
                 response = prepend_first_chunk(response, first_chunk)
             except Exception:
                 logger.exception(
@@ -331,7 +331,8 @@ def handle_generation_chunk(self, response: Any) -> None:
                     if choice.finish_reason == "stop":
                         llm_result = LLMResult(
                             llm_output=dict(
-                                token_usage={}, model_name=response.model
+                                token_usage={},
+                                model_name=response.model or "",
                             ),
                             generations=[self.chunks],
                         )
@@ -455,6 +456,15 @@ def handle_wrapped_call(
         model_name = ""
         if "model" in bindings.kwargs:
             model_name = bindings.kwargs["model"]
+        elif "model" in bindings.arguments:
+            model_name = bindings.arguments["model"]
+
+        # Guard against None model_name (e.g. Azure OpenAI
+        # deployments may not set a model parameter). Fall back to
+        # response.model if available, otherwise default to "".
+        if model_name is None:
+            model_name = getattr(response, "model", None) or ""
+
         callbacks = [self.global_callback]
         if callback is not None:
             callbacks.append(callback)
@@ -474,6 +484,11 @@ def _handle_response(
         # types of calls being handled here, we need to make various checks to
         # see what sort of data to process based on the call made.
 
+        # Ensure model_name is always a string to prevent
+        # downstream crashes in langchain's standardize_model_name.
+        if model_name is None:
+            model_name = ""
+
         # Generic lazy value should have already been taken care of by base Endpoint class.
         assert not python_utils.is_lazy(response)