fix(agent): deliver LLM errors to streaming channels and avoid polluting session context

When the LLM returns an error (e.g. 429 quota exceeded, stream timeout),
streaming channels silently drop the error message because `_streamed=True`
is set in metadata even though no content was actually streamed.

This change:
- Skips setting `_streamed` when stop_reason is "error", so error messages
  go through the normal channel.send() path and reach the user
- Stops appending error content to session history, preventing error
  messages from polluting subsequent conversation context
- Exposes stop_reason from _run_agent_loop to enable the above check

tests/agent/test_hook_composite.py

@@ -307,7 +307,7 @@ async def test_agent_loop_extra_hook_receives_calls(tmp_path):
     )
     loop.tools.get_definitions = MagicMock(return_value=[])
 
-    content, tools_used, messages = await loop._run_agent_loop(
+    content, tools_used, messages, _ = await loop._run_agent_loop(
         [{"role": "user", "content": "hi"}]
     )
 
@@ -331,7 +331,7 @@ async def test_agent_loop_extra_hook_error_isolation(tmp_path):
     )
     loop.tools.get_definitions = MagicMock(return_value=[])
 
-    content, _, _ = await loop._run_agent_loop(
+    content, _, _, _ = await loop._run_agent_loop(
         [{"role": "user", "content": "hi"}]
     )
 
@@ -373,7 +373,7 @@ async def test_agent_loop_no_hooks_backward_compat(tmp_path):
     loop.tools.execute = AsyncMock(return_value="ok")
     loop.max_iterations = 2
 
-    content, tools_used, _ = await loop._run_agent_loop([])
+    content, tools_used, _, _ = await loop._run_agent_loop([])
     assert content == (
         "I reached the maximum number of tool call iterations (2) "
         "without completing the task. You can try breaking the task into smaller steps."