Frontier cs 2.0 blackbox evaluator

[joyemang33]

Summary

• Add a black-box judge sidecar for Frontier-CS 2.0 Harbor tasks so agents can call submit.sh without seeing the evaluator implementation.
• Harden 2.0 evaluator execution by hiding evaluator source, running submitted solutions in an isolated temp file, dropping privileges when possible, and suppressing solution stdout/stderr and internal tracebacks from judge feedback.
• Add erdos_demo, a small N = 10 Erdos unit distance task for fast, visual sanity checks.
• Update 2.0 docs and Harbor adapter docs for the black-box workflow and demo task.

Testing

• Ran Python syntax checks for the updated 2.0 evaluators, adapter, submit helper, and judge server.
• Verified frontier list 2.0 shows both erdos_demo and erdos_unit_distance.
• Verified erdos_demo reference evaluates successfully with score 0.
• Verified a simple 10-point triangular construction scores 44.4444.
• Verified a malicious solution attempting to print evaluator contents does not leak evaluator source.
• Verified the 2.0 Harbor adapter generates both erdos_demo and erdos_unit_distance tasks.