Also, it's only "by hand" if you, the developer, writes it. In this case:

"The frontend — lexer, parser, semantic analysis (it’s called mkf) — hands off to two of its own IRs before writing WASM-GC by hand."

So those frontend tools do the writing, not the developer?

That's not "by hand". One might say "writes WASM-GC directly". But no that also doesn't happen, it goes through an IR first?

Personally I wouldn't put too much trust in a developer that even can't get their terminology correct. Well either that or I foobar'd my understanding of the writeup?

Anyway, looks like an interesting project.

'by hand' is probably the AI anthropomorphizing itself

Prompt regurgitation. Model was likely instructed to create the app to do this piece "by hand" (or a phrase with similar interpretation) because the author wanted to avoid pulling in a dependency of some kind.

Thank you.

I agree. "One pass" is wrong, and I shouldn't have written that. This is a normal multi-stage pipeline: AST -> HIR -> MIR -> WASM-GC, meaning each reduction is one pass.

I know, Kotlin has forward references and actual type inference, so name/type resolution requires multiple passes, which is precisely why there's explicit forward reference processing there. So, a single pass is never the case. My statement was incorrect.

What I meant was that it goes from source code to WASM-GC in one go, entirely in the browser (no LLVM, no Binary, no round trip to the server, and the WASM-GC backend is handwritten in C).