Automated hardening of agent environment verifiers

A new scalable approach uses a hacker-fixer loop in an isolated sandbox to automatically harden programmatic verifiers against reward hacking. This prevents agents from exploiting verifiers to receive high rewards without completing intended tasks.