Resume TerminalBench Harbor runs from checkpoints

[neubig]

Summary

• periodically archive completed TerminalBench Harbor trials to the configured results bucket
• restore that checkpoint when the same workflow run receives a replacement pod
• skip already verified selected tasks and merge results across timestamped Harbor attempts
• remain a no-op outside the existing RESULTS_BUCKET/MODEL_SLUG/GITHUB_RUN_ID environment

Motivation

Long TerminalBench jobs can lose hours of completed work when Kubernetes replaces the evaluation pod. Harbor stores each attempt under a timestamped directory, but the benchmark wrapper previously converted only the newest attempt and always restarted every selected task.

This change is independent of Perplexity and applies to any TerminalBench Harbor agent.

Validation

• pytest -q tests → 546 passed
• focused TerminalBench/Harbor tests → 36 passed
• Ruff check and format check passed
• Pyright → 0 errors
• added coverage for reward-based completion detection, malformed/incomplete trials, cross-attempt merging, latest-attempt deduplication, checkpoint opt-in, and URI construction

This PR was created by an AI agent (OpenHands) on behalf of Graham Neubig.