Make agent and benchmark progress crash-safe

[wq19901103wq]

Summary\n- atomically persist benchmark state and reports\n- checkpoint each system before and after execution with explicit status\n- classify infrastructure/budget failures as unfinished instead of wrong\n- reuse saved patches for evaluation-only retries without another model call\n- add durable direct-agent run summaries with duration and token usage\n\n## Validation\n- ruff format --check .\n- ruff check .\n- mypy agent tests\n- 36 targeted tests passed\n- full suite reached 192 passed, 3 skipped before being stopped to avoid competing with the active three-system Docker benchmark\n\nImplementation was developed in an isolated git worktree; the active comparison process was not restarted or modified.