Natural-language planning

In agent and server modes, the LLM planner translates objectives into tool calls.

• Model: claude-sonnet-4-6 (Anthropic), configurable.
• Context: the genome's capabilities and the registered tool vocabulary are injected into the prompt, so the model only proposes actions the robot can perform.
• Dispatch: the planner emits a tool name + arguments; the NavigationEngine routes it to the right layer (camera / GPIO / MAVLink) with the BCM pin pre-filled and direction inferred.
• Vision: scene description uses the Anthropic vision API.