VulcanBench developer hits cost estimation bug

Pre-run cost estimation is crucial for teams running expensive LLM evaluation sweeps, but scaling logic must account for already-computed costs in local run histories. In VulcanBench's case, the estimator triple-counts the judge tokens because historical logs already store the full cost.

• –The bug stems from `_task_judge_mult` unconditionally multiplying the indexed cost by 3.0 when judges are enabled, even though the stored `cost_usd` already includes the judge fee.
• –Storing the base agent-only cost in the local runs index and scaling it dynamically would resolve the discrepancy.
• –Reliable spend estimation is a key developer experience requirement for SWE benchmarking tools, as sweeps across massive reasoning models can quickly drain API budgets.
• –This highlights the difficulty of maintaining accurate telemetry across compound AI workflows where execution costs are split between the main agent and verification judges.