Clarify Agent0 evaluation scope and harden sandbox handoff

[noah-ing]

Summary

• present this repository as an independent OpenCompass evaluation harness and separate historical endpoint scores from Agent0 training claims
• tighten the SandFuzz handoff so child processes receive a small non-secret environment and add focused regression tests
• remove unused vendored package debris and add Apache-2.0 licensing, security guidance, and pinned read-only CI

Verification

• python3 -m unittest discover -s tests -v (3 passed)
• python3 -m py_compile src/tools/python_sandbox.py scripts/generate_iteration_report.py
• git diff --check

No hosted-model benchmark was rerun; doing so would incur API calls, and the README labels the checked-in scores as historical.