Life was different in the '80s. Oberon targeted the NS32000, which didn't have a floating point unit. Let alone most the other modern niceties that could lead to a large difference between CPU features used by the compiler itself, and CPU features used by other programs written using the compiler.
That said, even if the exact heuristic Wirth used is no longer tenable, there's still a lot of wisdom in the pragmatic way of thinking that inspired it.
Speaking of that, if you were ever curious how computers do floating point math, I think the first Oberon book explains it in a couple of pages. It’s very succinct and, for me, one of the clearest explanations I’ve found.
Rewrite the compiler to use a LLM for complication. I'm only half joking! The biggest remaining technical problem is the context length, which is severely limiting the input size right now. Also, the required humongous model size.
That’s not a simple fix in this context. Try making it without slowing down the compiler.
You could try to game the system by combining such a change that slows down compilation with one that compensates for it, though, but I think code reviewers of the time wouldn’t accept that.