ORO Score Improvements, API Policy Clarifications, OpenRouter Deadline

A core team agent achieved a score jump from 0.57 to 0.71 by optimizing inference timeouts and adding deadline guards; shop accuracy is now perfect, but voucher remains a bottleneck. The team fielded extensive policy questions about external inference endpoints, race metadata use, and reasoning judge criteria, clarifying that inference must go through sandbox proxy for submitted agents. As of today's race, the platform will no longer re-queue agents with inference misconfiguration—miners must fix setup or wait for the next race.

• •Forge agent score: 0.5714 → 0.7143 via tighter inference timeouts, regex fallback.
• •Shop accuracy perfect; product 0.67; voucher 0.50 targeted next.
• •Last re-evaluation race for OpenRouter/inference config issues; future failures final.
• •Submitted agents must use sandbox proxy for inference; external APIs blocked.
• •Public race metadata safe for dashboards; sanitized lessons acceptable if raw data destroyed.