The calculation of the pure function is still used for any other combination of input values.
Supercompilation looks at the unevaluated symbolic recursive code and replaces the code in some cases with simpler, specialized code tuned to the actual input values, so does a memoizing not only of the result, but the code used to calculate the function to begin with.
> deep program transformation
It doesn't have to be deep, does it? You can't transform the entire program, so there's an arbitrary cut-off anyway.
Here is some discussion here on why a pretty serious for GHC stalled: https://www.reddit.com/r/haskell/comments/2s97d0/the_state_a...
In fact, supercompilation subsumes many common optimizations, such as function inlining, constant propagation, list fusion, etc. So we can say that it's actually used "to a small degree", although I wouldn't really call it "supercompilation" per se.
The reason that full-blown supercompilation is not used is that it's too unpredictable -- the direct consequence of being too smart. I've elaborated on it in the final section of the blog post.