fix: enforce fair benchmark budgets

[wq19901103wq]

Root cause

The official DeepSeek comparison showed direct mode stopping with empty patches. Its logs proved both runs ended at the inherited 100k cumulative-token cap, while SWE-agent used roughly 400k input tokens on the same task. This made the comparison invalid.

Fix

• explicitly set direct mode to 100 steps and a 1M cumulative-token budget
• restore user config after every task and record limits in report metadata
• abort instead of scoring budget exhaustion or runner timeouts as accuracy failures
• use fully isolated per-task venvs so editable installs cannot overwrite the main Python environment

Validation

• ruff and mypy pass
• 25 benchmark-related tests pass
• regression test verifies benchmark limits and config restoration