It seems to me that the experiment would be more clear-cut if you tried using only standard libraries instead of frameworks when developing the applications, since the chosen framework is what accounts for the bulk of the overhead. For instance, if you used Java's HttpServer and the standard networking library - or a similar approach in PHP - I think they would yield better results.
I just took an old PHP site I used to run (no framework or libraries) and had codex transpile it to rust. Ran it behind identical nginx config, just switching out php-fpm's socket for the rust/tokio one. Rust was way worse, so I let codex time the performance resulting in it opting for micro-caching everything and then we got about comparable rps and latency - but those improvements would all help in PHP as well.
optimizing for the YT Algo, not the benchmark accuracy