Usage grouped by key, hour or endpoint — errors and p50/p95 latency

From the hourly rollup (every outcome, not only billed calls). errorRate is (clientErrors + rateLimited + serverErrors) / observed, null when nothing was observed. Latency percentiles are read off a fixed histogram (edges 25, 50, 100, 200, 400, 800, 1600 ms). Reachable with the dashboard session or an API key that names the usage scope.