Good breakdown of why bundle size and allocation rate are separate levers, that distinction gets lost a lot. On a dense data grid we worked on, most of our INP wins actually came from cutting re-render fan-out rather than the network layer. TanStack Query's caching solved the fetch side, but the click-to-paint gap was still row-level subscriptions. Curious if useLegacyTable adds meaningful overhead of its own during the bridge period.