So you will end up with a bunch of choices that are good in isolation, but an incoherent whole.

That's a plausible outcome. It may not be possible to design a best-in-class language without some kind of top-down planning, or even without an auteur like Matz or Hejlsberg managing the vision part. "Benevolent dictators" are also common in large and successful software projects, so maybe there's something to that.

However, it's also possible that relying on formal empiricism for language design will lead to at least a few breakthroughs. It might require another auteur to synthesize them into a great language, but the discovery would still be valuable.