Harden safety, privacy, and SWE-bench comparisons

[wq19901103wq]

Summary

• remove the SWE-bench environment-variable shell bypass and restore safe defaults
• harden URL fetching, local persistence permissions, backup retention, safety logs, retries, and token budgets
• make the three-system comparison use issue-only input and equal timeouts
• add endpoint preflight checks and preserve infrastructure errors instead of recording false benchmark failures

Validation

• full suite: 346 passed, 3 skipped
• final security regression suite: 187 passed
• benchmark/LLM targeted suite: 42 passed
• ruff and mypy pass

Benchmark status

A fresh three-system smoke run was intentionally aborted before scoring because the shared direct/SWE-agent endpoint returned insufficient_quota. Claude proxy connectivity is healthy. The new preflight now prevents this condition from being recorded as a model failure.