Add test for trailing assistant+system ordering, improve docstring

Add test exercising the step 3 ordering where stripping a trailing
assistant exposes a system message that then gets converted to user.
Move the reasoning about when a trailing system message can occur
into the docstring.

src/pipecat/adapters/services/perplexity_adapter.py

@@ -84,8 +84,11 @@ class PerplexityLLMAdapter(OpenAILLMAdapter):
            "assistant", remove it. OpenAI appears to silently ignore trailing
            assistant messages server-side, so removing them preserves equivalent
            behavior while satisfying Perplexity's "last message must be
-           user/tool" constraint. If the last message is "system" (e.g. the
+           user/tool" constraint. If the last message is "system", convert it
-           context only contains system messages), convert it to "user".
+           to "user". A trailing system message can only occur when the context
+           consists entirely of system messages (possibly followed by assistant
+           messages that were just removed), because step 1 converts any system
+           message that appears after a non-system message to "user".
 
         Args:
             messages: List of message dicts with "role" and "content" keys.
@@ -144,8 +147,7 @@ class PerplexityLLMAdapter(OpenAILLMAdapter):
             while messages and messages[-1].get("role") == "assistant":
                 messages.pop()
 
-            # If the last message is "system" (e.g. the context only contains
+            # If the last message is "system", convert it to "user".
-            # system messages), convert it to "user".
             if messages and messages[-1].get("role") == "system":
                 messages[-1]["role"] = "user"
 

tests/test_get_llm_invocation_params.py

@@ -1152,6 +1152,25 @@ class TestPerplexityGetLLMInvocationParams(unittest.TestCase):
         self.assertEqual(params["messages"][1]["role"], "user")
         self.assertEqual(params["messages"][1]["content"], "Always be polite.")
 
+    def test_trailing_assistant_removed_then_system_converted(self):
+        """Test that trailing assistant is removed, exposing a system message that becomes user.
+
+        This exercises the ordering of step 3: strip trailing assistants first,
+        then convert a trailing system to user.
+        """
+        messages: list[LLMStandardMessage] = [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "assistant", "content": "Sure thing."},
+        ]
+
+        context = LLMContext(messages=messages)
+        params = self.adapter.get_llm_invocation_params(context)
+
+        # Trailing assistant removed → [system] → system converted to user → [user]
+        self.assertEqual(len(params["messages"]), 1)
+        self.assertEqual(params["messages"][0]["role"], "user")
+        self.assertEqual(params["messages"][0]["content"], "You are helpful.")
+
     def test_consecutive_assistants_merged_then_trailing_removed(self):
         """Test that consecutive assistant messages are merged, then trailing assistant is removed."""
         messages: list[LLMStandardMessage] = [