Switch all Ollama models to gemma3n:e4b on node-197 GPU

Bake-off results: gemma3n:e4b (80.6% cmd match, 100% safety, 5.9s)
outperforms qwen3-coder:30b on all metrics. Updated paper, shrink,
and langgraph gateway configs. Frees steel141 for LoRA training.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

mc_langgraph_gateway.json

@@ -1,9 +1,13 @@
 {
-  "ollama_url": "http://192.168.0.141:11434",
-  "message_model": "gemma3:12b",
-  "command_model": "qwen3-coder:30b",
-  "tool_model": "qwen2.5:1.5b",
+  "ollama_url": "http://192.168.0.179:11434",
+  "message_model": "gemma3n:e4b",
+  "command_model": "gemma3n:e4b",
+  "tool_model": "gemma3n:e4b",
   "session_ttl_seconds": 21600,
+  "rcon_host": "127.0.0.1",
+  "rcon_port": 25577,
+  "rcon_password": "REDACTED_RCON",
+  "world_observation_enabled": true,
   "session_persistence_enabled": true,
   "session_db_path": "/var/lib/mc-langgraph-gateway/sessions.db",
   "knowledge_base_dir": "/var/lib/mc-langgraph-gateway/knowledge",