fix(llm): inject system prompt into user message for OpusAgent

OpusAgent topic system_prompt is used for summarization context but
not enforced as a binding instruction to Claude. Prepend [SYSTEM]
block to the first message of each session so Claude sees the full
JSON schema and grammar specification.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/multi_swarm_core/multi_swarm_core/llm/client.py

@@ -163,9 +163,11 @@ class LLMClient:
         model = self._tier_map[genome.model_tier]
         topic_id = self._get_or_create_topic(system)
 
+        prompt = f"[SYSTEM]\n{system}\n\n[USER]\n{user}" if session_id in (None, "new") else user
+
         body: dict = {
             "topic_id": topic_id,
-            "prompt": user,
+            "prompt": prompt,
             "model": model,
         }
         if session_id is not None: