It's generic for a little while as long as things don't change too much, but eventually they do.
See WWDC talk about watchOS migration to 64 bits.
With the end of support for the Apple Watch Series 3 (armv7k abi), I guess it was discontinued by Apple as they no longer needed Bitcode. Is there something I misunderstood?
> I may compile a C code to Bitcode and have LLVM generate a running binary for x86 CPUs in the end. If I save the Bitcode, however, I can later on tell LLVM to also create a running binary for an ARM CPU from that Bitcode, without having to compile anything and without access to the original C code. And the generated ARM code will be as good as if I had compiled to ARM from the very start.
There is some other stuff in the answer that is interesting (eg. that Apple apparently never actually used Bitcode for iOS, despite what they implied)
My guess is that they planned to use Bitcode to do minor optimisations for new generations of their chips, but it turned out to not work in practice, and so they scrapped it.
With llvm-mca you can check how each micro architecture will behave differently for the same machine code: https://llvm.org/docs/CommandGuide/llvm-mca.html
However, "agnostic" usually means that behavior doesn't vary. Performance is orthogonal to behavior in this case. I definitely don't expect code optimized for Haswell to be as fast on a Broadwell as code optimized for Broadwell, but I do at least expect it to run.