The article now uses one constructed workload through CUDA Graphs, local fusion and a megakernel that retains tiles across row-wide dependencies. The final study is 20261020T023055Z; the initial grid-wide study 20261020T022150Z is retained separately in the full appendix.

This website carries a selected evidence bundle. Full-size binary arrays, raw profiler captures, sanitizer logs and complete experiment manifests remain in the verified research archives; they are not all included here. No new GPU work was needed for this editorial update.