Latency budget queries

A load test sweeps one service at a rising set of client counts. concurrency is strictly increasing and latency_ms holds the p99 measured at each of those counts, in the same order. Warm parallelism pulls the curve down to a single low point and contention pushes it back up, so latency_ms falls strictly to its minimum and rises strictly after it, and no two neighbouring readings are equal. Either side may be empty. For every entry of budget_ms return the largest client count whose p99 is at or below that budget, or -1 when no setting qualifies, in query order. Both lists are long, so rereading the sweep per query is too slow, and the readings are not sorted, so a plain bisect over them lands on the wrong side.