API loops and judge examples lack cost / rate-limit safeguards

Labels: security, safety, cost, examples, orchestration

Multiple example scripts and the researcher orchestration loop invoke paid LLM APIs repeatedly with no budget caps, token limits, or rate-limit handling:

• examples/interleaved-thinking/reasoning_trace_optimizer/loop.py
• examples/llm-as-judge-skills/src/config/index.ts
• researcher/scripts/loop_step.py

Impact: A user running the examples could unexpectedly exhaust API credits. A misconfigured orchestration loop could repeat expensive calls indefinitely.

Suggested fix:

1. Add max_cost_usd, max_total_tokens, max_requests_per_minute, and max_retries controls.
2. Implement a lightweight budget tracker that aborts the loop when limits are reached.
3. Add exponential backoff and HTTP 429 handling.
4. Display a clear cost warning in every affected README and default new users to dry-run or stub mode.