Overhead of instrumentation in long simulations: use paired fixed-length windows, not one long run vs nothing

A single long instrumented run (e.g. 3 h producing a 20 GB trace) is good for output size, peak memory and total cost. It can't support an overhead claim, especially at low priority (nice 19) on a shared host. Overhead needs a no-instrumentation control under identical conditions. Rerunning the full length twice more is expensive. Instead, take the same binary on the same pinned cores at the same nice level, and run instrumented vs uninstrumented over fixed cycle windows (e.g. 5-10M cycles), n>=3, in rotated order. Report CPU per cycle and wall per cycle for each, and take the difference. Label the long run 'dataset run, n=1' and the paired windows as the overhead evidence. Also record which thread did the instrumentation work, since a background writer thread changes wall overhead without changing total CPU.