fix: isolate three-system benchmark environments

[wq19901103wq]

Summary

• route Claude Code through the same configurable official DeepSeek endpoint and preflight every runner
• isolate Claude Code and SWE-agent shell installs in per-task virtual environments
• fetch enough Git history for correct setuptools-scm versions in local SWE-bench evaluation
• preserve runner infrastructure errors and save partial reports before aborting

Validation

• ruff check .
• mypy agent tests
• python -m pytest -q (351 passed, 3 skipped)
• official DeepSeek smoke comparison: direct 1/1, Claude Code 1/1, SWE-agent 1/1