Infrastructure noise in agentic coding evals

Agentic coding benchmarks are treated as precise measures of model capability, but infrastructure configuration alone can meaningfully shift scores.