fix(image-generation): let LLM deliver images via message tool instead of runtime media attachment

The runtime media-attachment mechanism was broken for streaming channels
(e.g. WebSocket): the _streamed flag caused _send_once to skip the final
OutboundMessage that carried generated media, so images were never delivered.

Rather than adding complex coordination between streaming and media delivery,
delegate image delivery to the LLM: after generate_image returns artifact
paths, the next_step prompt now instructs the LLM to call the message tool
with the paths in the media parameter. This works uniformly across all
channels, streaming or not.

Remove generated_media from TurnContext, _assemble_outbound, and _state_save.
Update prompts in identity.md, SKILL.md, message tool description, and
artifacts.py to reflect the new flow.

tests/agent/test_context_prompt_cache.py

@@ -314,8 +314,8 @@ def test_system_prompt_keeps_message_tool_out_of_current_chat_replies(tmp_path)
     prompt = builder.build_system_prompt(channel="slack")
 
     assert "Do not use the 'message' tool for normal replies in the current chat" in prompt
-    assert "the runtime attaches those artifacts to the final assistant reply automatically" in prompt
-    assert "do not call 'message' just to announce or resend them" in prompt
+    assert "When 'generate_image' creates images" in prompt
+    assert "call 'message' with the artifact paths in the 'media' parameter" in prompt
     assert "Wait for the tool results, then answer once" in prompt