I found that Hexcodle (https://hexcodle.com/) had an interesting premise, but the guessing system with the hints just makes it too easy: you can just run a dichotomy and always win.
Also I don't think a 700kB bundle and 39 NPM dependencies are necessary: 300 lines of vanilla JS and 30kB do the job just fine.
I think eBPF could be interesting in you wanted to do something more advanced, like continuous profiling or things like that (I know Datadog uses eBPF for some of their tracing features), but it would be significantly more complex and would require an engineering time and cost that we do not have in my team. It's true that the proposed approach is quite manual and old-school, but I think it is also one of its biggest advantages: it is dead simple to understand and adapt to your use case.
If I'm not mistaken, metrics-server (which we also have deployed on our cluster) only gives information about the resource usage and state of the containers and/or pods of your cluster from a high level perspective. It won't give you precise and detailed info about the internals of your application process, ie which functions in your business logic code are responsible for how much CPU or RAM usage. So although metrics-server is very useful in its own right, it doesn't cover all needs and usages.