Wikimedia memory growth comparison data

This derived dataset supports the memory growth plot. It combines observations from two verified replays on naamah of the same 1,483,700,913-command numeric trace. All endpoints contain 80,798,328 members and the same page-ID-to-score mapping.

The earlier packed and standard Goblin curves and all four incumbents come from the September 8 six-engine replay. The RLE curve comes from the September 22 twelve-variant replay. These are separate concurrent runs, with different Goblin builds and numbers of clients. The plot compares observed memory growth; it does not compare throughput between runs. Packed curves use INT32/FLOAT32; standard Goblin and incumbents retain their general member representations.

No original observation is smoothed, removed, or replaced. Some incumbent final samples differ by 12 KiB from the separate final reading in summary.tsv; both records are retained. Five-minute samples cannot establish peak RSS. Labels and report savings use the final summary readings.

Regenerate the data and standalone SVG from the repository root:

python3 benchmarks/wikimedia_history_2026_08/plot_rle_comparison.py

The generator requires Gnuplot and verifies the recorded statuses, final mapping digests, common input hash, and monotonic member counts before plotting. The two original evidence bundles remain unchanged.