UnityWebGL-LoadingTest VS PeLib

Maybe, but Godot does particularly poorly in this area. Unity can go below 3MB out of the box and that is with bunch of engine systems in use.

https://github.com/JohannesDeml/UnityWebGL-LoadingTest

[0] https://github.com/JohannesDeml/UnityWebGL-LoadingTest

Just as a reference, Unity with PhysX and URP will amount to about 5mb. You can relatively easily create something 3D that is appropriate for browsers within 30mb, which is what you probably want to target as a max initial download to boot. You can ship additional content with Addressables at runtime.

You can find the template here: https://github.com/JohannesDeml/UnityWebGL-LoadingTest/tree/master/Assets/WebGLTemplates/Develop

I'm usually working with the third edition of the ISO 23271 standard, which corresponds to ECMA 335 5th Edition. There is also the book ".NET IL Assembler" (2014) by Serge Lidin which gives some additional context and a good structure to learn the details. My compiler is written in C++, not depending on the .Net framework; I don't use the reflection and emit features of .Net, but https://github.com/rochus-keller/Pelib/ to generate assemblies; also the generated code only requires minimal mscorlib.dll features, so I can use a minimal CLR.

Most of my languages have VM backends; see e.g. https://github.com/rochus-keller/Oberon; I implemented different backends generating LuaJIT bytecode; a year ago I switched to Mono which is based on ECMA-335; here is a discussion why I switched: https://github.com/rochus-keller/Oberon/releases/tag/IDEv0.9.0; I implemented utility libraries for both LuaJIT and CIL bytecode; see https://github.com/rochus-keller/LjTools/, https://github.com/rochus-keller/Pelib/ and https://github.com/rochus-keller/MonoTools/. I evaluated many VMs and think the mentioned ones are best suited. There were a lot of challenges with both technologies, what is to be expected, and too much to describe here.

And as I understand now achieving optimal performance with CoreCLR requires inside know-how of RyuJIT and Roslyn (which I don't have anyway), i.e. ECMA-335 alone is not sufficient; the actual compilation target to achieve optimal performance with CoreCLR is therefore Roslyn, not ECMA-335; of course I could generate C# instead of IL, but then I would depend on the framework and slow down the compiler. My current code generator based on https://github.com/rochus-keller/PeLib is about 30 times faster than ILASM.