Does the Operating Envelope Survive an Upgrade? A Replication of Many-Core Trading-Infrastructure Measurements Across a Release, Kernel and Compiler Change

This paper replicates the measurements of "Predictable Trading Infrastructure on Many-Core CPUs" (doi:10.6084/m9.figshare.33972466), made on an AMD EPYC 9575F on 2026-09-13, on the same CPU model six weeks later. Three things changed together in between: the release (4.2.1 to 4.6.0), the operating system (Ubuntu 24.04 to 26.04.1) and the compiler (GCC 13.3 to 15.2.0). We state five of the earlier findings as claims that the new run could refute. All five held. The throughput plateau again extends to about one strategy per physical core: its peak reproduced within 0.6% and its level within 2% in two independent sweeps. Pinned run-to-run spread stayed within 1.03×. The ring-size curve reproduced within 3% up to the 256 MiB L3, with the product default again at its peak. The research engine's full closed-trade ledger was bit-identical over 20 runs. The end-to-end order path was 1.3–1.6 µs (11–15%) faster at the median and lower at every percentile through P99.9. The ABI boundary read 80 ns against 73 ns raw, a difference below the machine's 10 ns effective timer resolution. The research engine ran 6.2% more bars per second. Two results moved, both outside the operating envelope. Between 97 and 128 strategies, beyond one thread per physical core, throughput was 15–19% higher than before, with a worse tail. Above the L3 size, large rings cost 10–11% more. Because the three changes coincide, we attribute every delta to the combined change. New in this run are a concurrency sweep in steps of one strategy, a ring-size sweep with hardware counters, client pinning, and the measured effect of 4.6.0's position-state fix on a strategy that reverses. Raw outputs: https://doi.org/10.6084/m9.figshare.34785762