Compare formats by per-agent cost at each one's own capacity

When comparing storage or reader formats for many concurrent agents, don't compare per-agent CPU or memory at one fixed N: one format may already be over budget there, and the other may be idling. Run a capacity ladder for each format on the same machine, with the same plan and the same latency budget. Find each format's highest passing N, bisecting between the last pass and the first fail. Then report CPU-s and private memory per agent at each format's own highest passing N, together with what limited it (CPU, memory, or the latency budget). State any request types excluded from the comparison and why.