Yes, you're right. Expression trees are especially interesting from an integrated DSL perspective, since you can do ridiculous things like turn lambda expressions into a syntax for hashes and so on. I omitted them because I believe that really did require a runtime modification to support.

But I would argue that since DSLs are all about affordances, it's not sufficient to say that similar functionality would be "possible". If the new approach doesn't represent significant semantic compression (which your hypothetical LINQ+C#2 would not), no one will use it. By that measure, the syntactic sugar added to C# 3.0 was absolutely a necessary precondition for LINQ.