feat: add secure OJBench alpha adapters

[aabbdev]

Summary

• add separate ojbench_python and ojbench_cpp alpha tasks with the official 232-problem language tracks and eight epochs for Pass@1/Pass@8
• pin the official prompt/testdata revision, OJBench source commit, DMOJ judge commit, base image, and Python runtime dependencies
• download only the selected problem's LFS archive and validators into a host cache mounted read-only in the judge
• run submissions through DMOJ cptbox in a network-disabled, read-only, bounded Docker sandbox and keep per-case tests/results out of Inspect logs

Security boundary

The container drops all capabilities, restores only SYS_PTRACE as required by DMOJ to supervise its child, uses no-new-privileges, disables networking, and bounds memory/PIDs. Generated programs remain under DMOJ's syscall and filesystem policies; the mounted /problems tree is not readable by submissions.

A malicious Python submission attempting to read /problems/NOI/loj-2083/init.yml received DMOJ IR rather than test content.

Validation

• pre-commit run --all-files
• pytest — 474 passed, 11 skipped
• targeted mypy — success across all OJBench modules
• real arm64 DMOJ self-tests — CPP17 and PYPY3 passed
• real loj-2083 judging — Python and C++ wrong submissions returned WA over 20 cases
• lazy post-start download — newly cached loj-3758 became visible through the existing read-only bind mount
• Docker Compose profile inspection — network none, read-only rootfs, no-new-privileges, 3 GiB memory, 128 PIDs
• wheel and sdist — both contain the OJBench Dockerfile and compose definition

Alpha gates

The complete 7.85 GB problem corpus has not yet been exercised across all custom validators, and no credentialed model Pass@1/Pass@8 run has been completed. Both registry IDs intentionally remain alpha until those broader validation gates pass.

[github-actions]

✅ Benchmark documentation has been automatically updated.