In the Windows world, you often find desktops with ten different copies of the "Windows C++/.Net redistributable" (the windows version of the C++/CLR standard library dynamically loaded artefacts) installed because each individual app have their own specific dependencies and it's better for them to bundle/specify it rather than rely on the OS to figure out what to load. The JavaScript, Julia, Rust, Go ecosystems all have first party support for pulling in driver binaries that may be hundred of gigabytes in size (because Nvidia is about as cooperative as a 3 year old child). You don't waste time fiddling with autotools and ./configure and praying that everything would run. Just run `npm install` and most if not all of the popular dependency heavy libraries would work out of the box.
As a C++ game developer myself, I make sure that my dependencies are part of my repo as submodules so that I can update/pull and build the version I need to from git tag versions.
So if you are tagging your releases, your final outputs, in your git source tree, then going back to a version from 20 years ago is just as simple as git checkout v0.0.1
Vcpkg for C++ dependencies is another option (my preferred if you don't go git submodule route) and ALWAYS USE CMAKE! Don't opt for some crazy build setup, or some internal build tool used by <insert FAANG here> that they force you to use (V8 team, if you're reading this, fix your build pipeline).
KISS. Keep it simple slick. If your package isn't available in the OS package manager, it's time to adopt a package manager or adopt a devops practice that allows you to revert to any version of the code you need (git submodule route).
Maybe because you're spending half a mil a year on folks doing matrix and vector math? I agree the reasoning is flawed. In order to have a decent ML shop, you need devops. DevOps will make sure your builds work and can be deployed into the architecture of whatever product you're building. Not saying this is a person, but you should be practicing good software engineering first, before trying to AI things. You can grow both skill sets at the same time but one definitely comes before the other. Any company who is just deploying ML models with the kitchen sink and throwing everything over the fence is going to need to brush up on their investigative skills because they will be hacked.
IMO it's mostly because of a culture of laziness among system programmers. They like to kick the ball down the road and make dependency management the distro's problem. Too many embedded engineers not being paid enough. When you are building a product you shouldn't be wasting your time on this sort of "make work" that reinvents the wheel multiple times. Come to the machine learning world and everything including the kitchen sink and graphic drivers are bundled in the build because otherwise your deployment won't work and you can't afford to pay engineers that cost half a mil a year to spend their time fiddling with deployment binaries.