> But in the given example, shouldn't the compiler be able to see, via a simple analysis, that this function is only being applied to lists of numbers?

No, that's the beauty and curse of dynamic languages such as the Scheme mentioned here. Of course, you could do special optimizations, such as inlining bindings that never get used out of scope or defining special functions that don't operate on real Lisp data structures but plain C structures, but they add extra complexity and duplicated logic to the compiler. I'm sure those kind of optimizations would be out of scope of this article.

> in practice I have not seen any optimisation that would take that Scheme-compiled C program and turn it into one a C programmer would write

Obviously Scheme code is going to look different than C code, they're different languages. I think the point being made was that it's okay to emit some C code that's a little extra verbose, doing things like defining extra vars and such, knowing that the compiler will simplify many such things.

> There's also the question "why even generate 'superfluous stuff' if it's going to disappear anyway?"

To keep the compiler's implementation simple and understandable. The more complex output it has to produce, the harder it is to read/write/debug.

> That's a huge assumption

I am not talking about optimizations that require understandings of the source language. But, for a simple example, a Pascal-ish language may have the statement

  x := (a+b)*(c+d)

compile into something like:

  _t1 = a + b;
  _t2 = c + d;
  _t3 = _t1 * _t2;
  x = _t3;

in order to simplify translation from an internal representation such as an AST. Putting such a piece of code back into human readable form makes the backend more complicated, while C compiler know how to optimize away the extraneous variables well.

Note that this is just one of many things that contribute to the verbosity of compiler output, not even necessarily a major one in this case.