A language taking shape.

Seven randomized paired rounds on one affinity CPU; setup and teardown are outside timing except registration churn. Allocation runs are separate. Registrations are validity guards, not expression watchers or rendered row scopes. Unchanged controls also vary; this is not a CPU win or service-driven UI benchmark. No generated repetition, GPUI layout/paint, GPU or Rust parity claim.

Instrumented stage calls across all measured iterations · left / right. Layout, prepaint and paint count pane boundaries, not every descendant. Scroll horizontally for all stages on small screens.

All three policies use the SAME fine-grained native dependency machinery. This compares renderer, cache and invalidation choices, not alternative native graphs. Actual manually driven headless GPUI layout, prepaint and paint are included; OS wake scheduling, GPU work and the surrounding parent visual tree are excluded. Requested Rust heap is instrumented separately from timing; it excludes RSS, C allocations and GPU memory. No equivalent Rust application baseline or language parity claim.

Each sequence is one process per variant, not a timing distribution. These are actual headless GPUI frame allocations. Version 2 requires two clean readers and half the pane clean; version 3 retains caching with any clean reader. Both defer eviction until the third successfully published ineligible frame. Cold priming and the cost of rebuilding after eviction remain visible. Neither candidate is enabled.

Live memory subtracts each process baseline after model creation, before opening the window. It includes GPUI/font caches and excludes allocator metadata, RSS and GPU memory. The baseline is the corrected nonadaptive row cache, not the original direct renderer. Configured viewport counts are checked against actual mounted rows for the plotted non-cold phases. No automatic window scheduling, full-parent cost or Rust parity is measured here.

These timings measure the earlier cache: that version has an ancestor prepaint rollback defect. A separate corrected candidate passes retry tests, but its extra token and observer costs are not measured here. This cache remains experimental; corrected and adaptive performance acceptance is open.

51 matched service cases preserve measured row geometry, values and native cleanup. Sparse 4096-row granular updates use about 0.325× CPU with 19,900 extra warm retained bytes. All-visible churn regresses to 1.057× CPU; the one-row sparse case shows no established win. Larger panes mount eight visible rows each. Dense changes the first row in each pane; all-visible changes all eight.

Headless GPUI layout/prepaint/scene work; excludes GPU rasterization and OS scheduling. CPU and requested Rust heap use separate binaries. Nine missing font families precede the fallback in this environment. This is neither an enabled optimization nor a Rust-parity result.

Temporary Taffy layout arrays account for 84.6% of requested bytes in the large sparse granular case. Reserving smaller arrays reduces those bytes; it does not remove allocation calls.

All 42 service cases preserve geometry, values and native cleanup; 840 differential layouts agree. CPU results are mixed, including small-view regressions and noisy unchanged controls. No retained-memory win is demonstrated. Dense updates change the first row in both panes, not every visible row.

Headless GPUI layout/prepaint/scene work; excludes GPU rasterization and OS scheduling. CPU and requested Rust heap use separate binaries. Nine missing font families precede the fallback in this environment. This is neither an enabled optimization nor a Rust-parity result.