Back-to-front permutation [count] — TRANSFERRED to the main thread.
Time spent inside the backend sort_splats_by_depth call, in ms.
For compiled WASM this INCLUDES the wasm-bindgen boundary copies
(centers copy-in, ordering copy-in/out, mallocs) — the shim performs
them inside the exported function; for the TS fallback it is the pure
kernel. Measured with performance.now() on the worker's clock.
Whole sortNode body duration in ms (registry lookup + output
allocation + kernel). workerMs - kernelMs is the worker-side
overhead around the backend call. Same clock as kernelMs, so the
difference is meaningful; both are durations, safe to compare with
main-thread-measured durations.
Generation the ordering was computed for (re-checked on the main thread).