The tradeoff with this approach is that you lose any sort of agnosticism when you drop into the language specific tooling. So now if you work at a corporation and have to deal with multiple toolchains every engineer now needs to learn and work with new build tooling X times for each supported language. This always happens to some extent - there’s always going to be some things that use the language’s specific task runner constructs - but keeping that minimal is usually a good idea in this scenario.
Your complaint feels to me that it is about poorly implemented CI systems that heavily leverage container based workflows (of which there are many in the wild). If implemented properly with caching, really the main overhead you are paying in these types of setups is the virtualization overhead (on macs) and the cold start time for the engine. For most people and in most cases neither will make a significant difference in the wall clock time of their loop, comparatively.
my key concern whether tools to build project have to be in the same pool as project itself (that may or may not use tools to build/edit/maintain/debug it).
I think some of the euphoria around tracking tooling in the repo is "yes, now I can run a command in the repo and it's as if I spun up a docker container with the CI locally, but it's just regular software running on my machine!" This is a huge improvement if you're used to the minutes-or-hours-long CI cycle being your only interaction with the "real environment."
The reductio ad absurdum that you describe is basically "a snapshot of the definitions of all the container images in our CI pipeline." It's not a ridiculous example, it's how many large projects run.