Is it that baffling? After all, the core principle is that instead of representing a continuous spectra, you project the spectra to a few basis functions[1], so you can later reconstruct an approximation of the original spectra.
Our eyes have three different cone cells with different responses[2], so the fact that three basis functions[3] that roughly align with the cone cells work well as an approximation isn't that surprising to me.
That said, way back when I was doing rendering I was toying with the idea of trying to find a optimal basis functions for rendering, like if scene-specific ones could improve things.
[1]: https://www.psych.mcgill.ca/misc/fda/ex-basis-a1.html