This fixes the regression, but there are two reasons we might want a different change:

1. Shallow still calls normalize-type. If soft-contract got ported to Shallow TR, it would be slow to typecheck.
2. Now that Deep and Optional don't call normalize type, Shallow might have a higher runtime cost for interacting with them. Take Optional for example. If Optional sends a function f to Shallow, then Shallow has to check that f is a function and maybe has to check the shape of results (f x). It depends on whether f is marked as trusted. Removing the normalize-type might remove the mark in some common cases.

Ideally we should either remove normalize-type in all cases (because it doesn't save enough runtime checks to justify the compile-time costs) or use it carefully but consistently across Shallow, Deep, and Optional.

I think the real question is why is normalize-type so slow. Contract generation involves multiple passes over types, so adding this one shouldn't have been an enormous slowdown.

I brought this up with Ben before: Asumu spent a lot of time wresting with types because they become very large. Sharing as much as possible solved his problem. Is something similar going on here?