Python: Fix issue with orphaned tool/tool_calls (#12923)

### Motivation and Context

Tool/tool_calls are getting orphaned.

Issue : #12708

How to reproduce :

```python
from semantic_kernel.contents import (
    AuthorRole,
    ChatMessageContent,
    FunctionCallContent,
    FunctionResultContent,
    ChatHistoryTruncationReducer,
)

messages = [
    ChatMessageContent(role=AuthorRole.SYSTEM, content="sys"),
    ChatMessageContent(role=AuthorRole.USER, content="user A"),
    ChatMessageContent(
        role=AuthorRole.ASSISTANT,
        items=[
            FunctionCallContent(
                id="123", function_name="search", plugin_name="plugin", arguments={"q": "x"}
            )
        ],
    ),
    ChatMessageContent(
        role=AuthorRole.TOOL,
        items=[
            FunctionResultContent(
                id="123", function_name="search", plugin_name="plugin", result="RESULT"
            )
        ],
    ),
    ChatMessageContent(role=AuthorRole.USER, content="user B"),
    ChatMessageContent(role=AuthorRole.ASSISTANT, content="done"),
]

reducer = ChatHistoryTruncationReducer(target_count=3, threshold_count=0)
reducer.messages = messages

print("Before reduce:", [m.role for m in reducer.messages])
await reducer.reduce()
print("After reduce:", [m.role for m in reducer.messages])

# After reduce: [<AuthorRole.TOOL: 'tool'>, <AuthorRole.USER: 'user'>, <AuthorRole.ASSISTANT: 'assistant'>]
```

### Description

Move backward as long as we are in a tool/tool_call island when reducing
the messages.

---------

Co-authored-by: Allaoua Benchikh <[email protected]>
Co-authored-by: Evan Mattson <[email protected]>
Co-authored-by: Evan Mattson <[email protected]>

Author: 3 people

python/semantic_kernel/contents/history_reducer/chat_history_reducer_utils.py

@@ -96,12 +96,19 @@ def locate_safe_reduction_index(
     message_index = total_count - target_count
 
     # Move backward to avoid cutting function calls / results
-    # also skip over developer/system messages
+    # Stop if we encounter developer/system or a non-call/result message
     while message_index >= offset_count:
-        if history[message_index].role not in (AuthorRole.DEVELOPER, AuthorRole.SYSTEM):
+        msg = history[message_index]
+        if msg.role in (AuthorRole.DEVELOPER, AuthorRole.SYSTEM):
+            break
+        # If current is not a call/result, we've reached a safe boundary
+        if not contains_function_call_or_result(msg):
             break
-        if not contains_function_call_or_result(history[message_index]):
+        # Avoid stepping back past a user message boundary when current is a call/result
+        prev_idx = message_index - 1
+        if (prev_idx < offset_count) or not contains_function_call_or_result(history[prev_idx]):
             break
+
         message_index -= 1
 
     # This is our initial target truncation index

python/tests/unit/contents/test_chat_history_truncation_reducer.py

@@ -3,6 +3,8 @@
 import pytest
 
 from semantic_kernel.contents.chat_message_content import ChatMessageContent
+from semantic_kernel.contents.function_call_content import FunctionCallContent
+from semantic_kernel.contents.function_result_content import FunctionResultContent
 from semantic_kernel.contents.history_reducer.chat_history_truncation_reducer import ChatHistoryTruncationReducer
 from semantic_kernel.contents.utils.author_role import AuthorRole
 
@@ -18,6 +20,24 @@ def chat_messages():
     return msgs
 
 
+@pytest.fixture
+def chat_messages_with_tools():
+    return [
+        ChatMessageContent(role=AuthorRole.SYSTEM, content="System message."),
+        ChatMessageContent(role=AuthorRole.USER, content="User message 1"),
+        ChatMessageContent(
+            role=AuthorRole.ASSISTANT,
+            items=[FunctionCallContent(id="123", function_name="search", plugin_name="plugin", arguments={"q": "x"})],
+        ),
+        ChatMessageContent(
+            role=AuthorRole.TOOL,
+            items=[FunctionResultContent(id="123", function_name="search", plugin_name="plugin", result="RESULT")],
+        ),
+        ChatMessageContent(role=AuthorRole.USER, content="User message 2"),
+        ChatMessageContent(role=AuthorRole.ASSISTANT, content="Assistant message 2"),
+    ]
+
+
 def test_truncation_reducer_init():
     reducer = ChatHistoryTruncationReducer(target_count=5, threshold_count=2)
     assert reducer.target_count == 5
@@ -66,6 +86,22 @@ async def test_truncation_reducer_truncation(chat_messages):
     # We expect only 2 messages remain after truncation
     assert result is not None
     assert len(result) == 2
-    # They should be the last 2 messages
+    # They should be the last 4 messages while tool result is not orphaned
     assert result[0] == chat_messages[-2]
     assert result[1] == chat_messages[-1]
+
+
+async def test_truncation_reducer_truncation_with_tools(chat_messages_with_tools):
+    # Force a smaller target so we do need to reduce
+    reducer = ChatHistoryTruncationReducer(target_count=3, threshold_count=0)
+    reducer.messages = chat_messages_with_tools
+    result = await reducer.reduce()
+    # We expect 4 messages remain after truncation
+    # Tool results are not orphaned, so we expect to keep them
+    assert result is not None
+    assert len(result) == 4
+    # They should be the last 4 messages
+    assert result[0] == chat_messages_with_tools[-4]
+    assert result[1] == chat_messages_with_tools[-3]
+    assert result[2] == chat_messages_with_tools[-2]
+    assert result[3] == chat_messages_with_tools[-1]