GPLv2 was stricter in that regard, in that the entity shipping the system libraries couldn't be the one shipping the GPLed software; GPLv3 drops that and allows anyone to ship GPLv3ed software for a proprietary platform, as long as you still meet the other requirements of the GPLv3.
I'm largely thinking about the current situation with ZFS on Linux. IMO, this benefits no one, and so I'd generally like to see less draconian interpretations of what code is allowed to link together—but more licenses which protect other user freedoms, like the ability to run your own code.
What would arguably not be fine, under my interpretation, would be if Windows would not function without a GPL component, that's built-in to the OS and distributed on the install disks and updated through Windows Update. Even if it's a self-contained binary, and you can navigate to it in the filesystem.
The standard is - would a reasonable person view the combination of GPL and non-GPL code as a single piece of software?
I think reasonable people would disagree with each other.
macOS used to come with iTunes preinstalled, but you could remove it by turning off SIP and running `sudo rm -rf /Applications/iTunes.app` in the Terminal. I always did so, because iTunes sucks, and the rest of the OS always continued to work fine.
Is iTunes a part of macOS, or is iTunes a separate piece of software? Would it matter if you could remove iTunes through Finder, or without turning off SIP? Would it matter if iTunes wasn't included with macOS, but was automatically downloaded and installed a day later? Would it matter if iTunes was created by a third party, or if macOS couldn't output any audio without it?
What about Google Play Services? DirectX? ZFS? I could certainly make a case for how I think each of these should be classified, but I don't think it's at all obvious.
Software engineers consistently underestimate the degree to which law is not like coding.
The even more interesting case is when there is more than one option for that binary, all with the same interface (but possibly different features), and only some of them are GPL. This was famously the hypothetical presented by the author of CLisp (which had a dynamic dependency on GNU Readline) to Stallman.
I think the battle was lost by not pushing the advantage hard enough.