fix: restore terminalbench2_env.py from patch-tool redaction corruption

[teknium1]

Summary

Commit ed27b826 introduced patch-tool redaction corruption that destroyed terminalbench2_env.py — the file went from 925 lines to 516, losing the entire evaluation pipeline.

Corruption fixed

• max_token_length=*** → max_token_length=16000,
• api_key=os.get...EY", "") → os.getenv("OPENROUTER_API_KEY", "")
• tokenizer_name="NousRe...1-8B" → "NousResearch/Hermes-3-Llama-3.1-8B"

Code restored (409 lines)

• _run_tests() — test upload, execution, and verifier download
• _eval_with_timeout() — per-task wall-clock timeout wrapper
• evaluate() — main evaluation entry point (tqdm progress, concurrency, results aggregation)
• wandb_log() — metric logging
• if __name__ == "__main__" entry point
• Rest of rollout_and_score_eval() — result assembly, error handling, finally-block cleanup

Legitimate changes preserved

Re-applied from the two commits that landed after the corruption:

• eval_concurrency config field (ed27b826)
• docker_image registration alongside modal_image (ed27b826)
• ManagedServer branching for vLLM/SGLang backends (13f54596)

Closes PRs #1737 and #1740 (partial fixes from aydnOktay that spotted the syntax errors — credit to them for surfacing the issue).

[github-actions]

⚠️ Supply Chain Risk Detected

This PR contains patterns commonly associated with supply chain attacks. This does not mean the PR is malicious — but these patterns require careful human review before merging.

⚠️ WARNING: exec() or eval() usage

Dynamic code execution can hide malicious behavior, especially when combined with base64 or network fetches.

Matches (first 20):

230:+                self.rollout_and_score_eval(item),
253:+        Runs all tasks through rollout_and_score_eval() via asyncio.gather()

---

Automated scan triggered by supply-chain-audit. If this is a false positive, a maintainer can approve after manual review.