Default system prompt for Arcana.Loop.

The prompt follows current best practices from Anthropic's tool use guidance, OpenAI's GPT-5 prompting guide, and the Self-RAG / CRAG literature:

1. Structured into named markdown sections (works across providers).
2. Heavy detail in tool descriptions, not the system prompt.
3. Explicit when-NOT-to-call rules to avoid over-eager tool use.
4. Tool budget is mentioned even though max_iterations enforces it.
5. No Thought:/Action: prefixes since native tool calling drives the loop.
6. Self-critique on retrieval quality before answering.
7. Soft language, not aggressive CRITICAL/MUST (these cause over-trigger).

Override with the :system_prompt option on Arcana.Loop.run/2.