tel's solutions are "use lenses" or "assume all keys are strings", both of which miss the point.

Yes, you can work around the limitations of a static type system, either by finding another solution (lenses) or by making functions more specific (assume all keys are strings), but that doesn't mean the limitations disappear. It just means you're working around them.

In a statically typed language the solutions to a problem are constrained by the type system. The question is not whether statically typed languages are more constrained than dynamically typed languages, but whether the constraints that static typing introduces are offset by the guarantees they purchase.

"That's exactly my point. Type signatures can be inferred, specs cannot."

The compiler can infer some type signatures, and if you're not explicitly typing your named functions (which is generally cited as good practice), that does mean you get some checking "for free".

But specs can also test more things than types; they occupy a space somewhere inbetween static types and a generative testing solution like QuickCheck. Depending on the function, a spec might be more or less verbose than the equivalent checks in a language like Haskell.