Add GPT-5.6 Luna high TensorCircuit benchmark results

[QingyunQian]

Summary

Archive the single-trial GPT-5.6 Luna/high TensorCircuit benchmark across all 12 ORBIT-Q challenges, independently audited by GPT-5.6 Sol/high.

• 9 / 12 final expert-adjudicated valid solutions; failures on Tasks 01, 04, and 08
• Functional and static checks pass on all 12 tasks
• Docker-isolated, sequential execution with one valid model outcome per task
• Includes selected solutions, rewards, functional/static/audit outputs, logs, configs, provenance, normalized summary data, and reproducible figures
• Preserves the raw Task 08 verifier pass while recording the final human expert adjudication as failure

Validation

• Verified solver model gpt-5.6-luna with reasoning effort high
• Verified audit model gpt-5.6-sol with reasoning effort high
• Verified one non-infrastructure outcome for every task
• Archive regeneration and consistency checks pass

See results/gpt56luna-high/README.md for the complete result table and protocol.

Result figures

Final task outcomes

Agent-side resource use

[refraction-ray]

I think problem 8 can be taken right as the LLM audit indicates