In order to have an ML, you take a subset of Common Lisp and specialize/extend it in a very specific direction. The part that are removed from Lisp are the one you don't care about. I care about being able to use different equivalence classes, including the ones the compiler is going to require, like identity equality, because this matters when I implement abstractions myself.

> However, there's no way to portably manipulate compiler internals in Haskell.

Yes, why not, that would be damn useful. The alternative is to hack each implementation in its own way.

> Heck, Haskell syntax, unlike Lisp syntax, can actually be represented in multiple ways: [...]

Do you get the difference between internal and external representations? When you see (lambda (a) (+ a 1)), you don't know how the code is represented internally and how its value, when executed, is represented internally (when my compiler performs data-flow analysis, it uses a specific internal representation I don't need to know). Yet, you have access to a uniform API, defined at the language level, to manipulate data. That's why you can have the same source code working on the JVM, interpreted using a C/C++ runtime or directly expressed as assembly. In that sense, there is a separation between the language and its implementation. But part of the language specification is also there to guarantee a uniform way of building new extensions and integrate them with the compiler. That is a good thing. What is not enforced is how and when each feature is used. If that matters, dump a specialized Lisp image that you call the "compiler" and call it with Make on a file expressed in your custom DSL.

> There's no dichotomy between extensibility and abstraction enforcement, even if you try to set up one.

I am not trying to set up one. Just read more carefully.