fix: patch litellm.api_base in test and add branch-order guard comment

Addresses review feedback:
- Reset litellm.api_base to None in get_api_base tests for isolation
- Add comment clarifying why /projects/ branch must precede services.ai.azure.com

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

litellm/llms/azure_ai/responses/transformation.py

@@ -114,9 +114,11 @@ class AzureAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
             query_params["api-version"] = api_version
 
         # Build the responses endpoint path.
-        # Check /projects/ first: project-based endpoints under
-        # services.ai.azure.com need /openai/v1/responses, which differs
-        # from the non-project services.ai.azure.com path (/models/responses).
+        # IMPORTANT: The /projects/ check MUST come before the general
+        # services.ai.azure.com check (_should_use_api_key_header), because
+        # project URLs also contain that domain. If we checked
+        # services.ai.azure.com first, project URLs would incorrectly get
+        # the /models/responses path instead of /openai/v1/responses.
         if "/projects/" in api_base:
             new_url = _add_path_to_api_base(
                 api_base=api_base, ending_path="/openai/v1/responses"

tests/test_litellm/llms/azure_ai/responses/test_azure_ai_responses.py

@@ -125,8 +125,11 @@ class TestAzureAIResponsesURL:
         )
         assert url == "https://custom-proxy.example.com/v1/responses"
 
-    def test_missing_api_base_raises(self):
+    def test_missing_api_base_raises(self, monkeypatch):
         """ValueError raised when api_base is None and no env var set."""
+        # Patch litellm.api_base to None so the test is isolated from
+        # any global state set by other tests or config.
+        monkeypatch.setattr(litellm, "api_base", None)
         config = AzureAIResponsesAPIConfig()
         with patch.dict(os.environ, {}, clear=False):
             # Ensure env var is not set