The most useful line on the page for operators is easy to miss: after updating, `nginx -s reload` does not load the new binary, you need a restart. I'd expect a fair number of hosts to be "patched" on disk and still running the old worker for weeks.
The other one is the caveat about version strings. Distro packages and container images backport fixes without changing the version, so `nginx -v` can't confirm exposure or safety by itself. For anyone pinning an image tag like nginx:1.27-alpine, the question is whether that tag was rebuilt, and the only way to know is to pull again and compare digests.
On the 98.9% figure: the page is careful to call it version prevalence, which I appreciate. But the fourth condition in your checklist, external h2 negotiation, is measurable from outside with an ALPN probe. Did you run that against the same Shodan set? The share of affected-version hosts that actually negotiate h2 on a public port would be a much better estimate of real exposure.
I checked my own small server after reading this and it turned out not to matter, since the proxy in front is Caddy. What it did show me is that I couldn't have answered "which proxy, which version, is h2 on" from memory, and that's probably the common case for small deployments.
I hope anyone using nginx should know that 'nginx -s reload' reloads only config. And you don't necessarily need to restart - nginx can be upgraded gracefully without breaking existing connections: https://nginx.org/en/docs/control.html#upgrade on FreeBSD it can be done using /usr/local/etc/rc.d/nginx upgrade, may be other OSes have similar scripts.