Harden Harbor iterative fallback for Frontier-CS 2.0

[joyemang33]

Summary

This PR hardens Harborized Frontier-CS evaluation around iterative submissions and timeout fallback, especially for Frontier-CS 2.0 directory-style tasks such as Vector DB ANN.

Main changes:

• Route Frontier-CS 2.0 final verification through the judge sidecar instead of importing evaluator code directly.
• Preserve judge-owned iterative submission records via the judge sidecar and mirror them into verifier artifacts.
• Fix best-submission fallback when an agent times out or when the final workspace scores worse than an earlier submission.
• Normalize CLI accounting for iterative submission rewards so raw 0-100 scores are reported as 0-1 rewards where appropriate.
• Clarify algorithmic Harbor artifacts are rebuilt from judge-owned submissions.
• Document Vector DB ANN’s resource budget and set it to 8 vCPUs / 16 GiB.

Please read CONTRIBUTING.md before submitting.

Type of Change

• New research problem
• New algorithmic problem
• New Frontier-CS 2.0 problem
• Bug fix
• Documentation update
• Other: Harbor adapter and CLI evaluation infrastructure

Testing

• git diff --check
• python3 -m py_compile on changed Python files
• Generated vector_db_ann Harbor task successfully
• Ran ANN Harbor trial and verified timeout fallback:
  • best iterative score: 31.3897
  • used_best_submission=1
  • recall_at_10=1.0
  • final reward: 0.31389732738629883

Checklist

• Code follows the project structure and conventions
• Self-review completed
• Documentation updated (if applicable)

CI Validation (for new problems)

N/A. This PR updates an existing Frontier-CS 2.0 task and Harbor evaluation infrastructure; it does not add a new problem.