I suspect what the parent author meant is that the lambda calculus is a very simple language. If a new language feature can be defined as a small extension to the lambda calculus, then creating a toy implementation is should be easy because the new language is so small.

More importantly, since your new language is small, you can eyeball the implementation and be fairly confident that the implementation "matches" your formal definition (about which you might prove theorems such as type safety).

Contrast this with defining a new feature as an extension of C++, Haskell or Java. You'll invest a lot of time, and in the end might not even be very confident that your implementation matches your theory. So you might write a few hundred line program in your new programming language, get a wonky result, and not be sure if the formal definition is flawed or if the implementation is buggy.

The "implement almost for free" phrase takes on a different meaning if you decide to implement computer-checkable proofs about your language.