Method

• Machines. Each server class gets a fresh server droplet and a loader droplet with more vCPUs, running oha, in one region, over the private network. The classes race at the same time; every droplet is deleted afterwards.
• Checks. Before racing, /plaintext must answer exactly Hello, World!, /echo the body it was sent, /menu the reference page, /sse the reference stream with one chunk per event (and 400 without signals), and an unknown path 404. A server that fails is a DNF.
• Measurement. 5 seconds of warmup (discarded), 20 seconds measured, three rounds in a seeded random order. The median counts, and only 2xx responses count.
• Noise. Shared-vCPU classes record CPU steal, so a noisy night shows up in the results.
• Builds. Each server is built and configured the way its users ship it. Every version is pinned and recorded with the results.

This is a reproducible nightly on hardware anyone can rent, not a best case on tuned bare metal. The same race runs on a laptop: dragrace race local.