Add Modal support: run the full benchmark on cloud GPUs

[arnavw]

Runs the full pipeline (agent → contamination judge → evaluation) on Modal rented H100s, for anyone without cluster access. Purely additive — everything lives in src/modal_adapter/ (same shape as the Harbor adapter in #8) and the condor path is untouched. Each phase script is a near line-for-line port of the corresponding section of run_task.sh, and results land in the exact EVAL_DIR layout, so scripts/collect.py works on downloaded results unchanged.

• One run = a chain of three Modal Functions: run_agent (standard image, 1–8×H100, ≤22h), run_judge (fresh CPU container, so agent tampering can't leak into the judge), run_eval (vllm_debug image, retry ladder as-is). Submissions are fire-and-forget from a laptop.
• Both .def files are translated to modal.Image chains; the HF cache is a Modal Volume seeded by running download_resources.py inside Modal, mounted read-only and exposed through a symlink farm (fuse-overlayfs can't mount in Modal's gVisor runtime — verified empirically).
• ~$55–60 per standard 10h run at list prices; full setup, usage, and a table of deviations from the cluster (preemption semantics, judge on CPU, run-id format, 22h cap) are in src/modal_adapter/README.md.

Verified so far: both images build and deploy; CPU smoke checks pass on both images (imports, pinned CLI versions, volumes, secret wiring, symlink farm); the repo's check_cuda.py/check_cuda_writing.py pass on a real H100. Not yet run: a paid end-to-end run — the README's smoke sequence (seed minimal cache, 1h claude run on gsm8k, ~$10–15) is the intended validation before this leaves draft.

Two notes for reviewers: standard.def installs inspect_evals from HEAD, which now requires Python ≥3.11 and would break a fresh cluster build too — the Modal image pins the same commit vllm_debug.def pins. And --torch-backend=auto needs a CUDA driver at build time, so the Modal images resolve vLLM from PyPI instead.

🤖 Generated with Claude Code