It worked, but the main issue was related to object file generation.
Basically, you have to allocate one big i8 array to represent all of the QVM's data segments (including its bss segment). This doesn't translate well once compiled with Emscripten, as bss relocation is on a per-variable basis. For one of the Quake 3 Fortress QVMs which was only a few hundred kb it output a ~50mb .bc file and a ~70mb .js file of mostly 0s.
I experimented with storing each segment as its own i8 array (which could be 0 initialized for the bss segment) in a packed struct, which significantly lowered the .bc file size. However, while I was looking into what I'd need to change in Emscripten to support this, I instead decided to go the route of writing the runtime compiler.
Fun fact: Quake 3 mods didn't have access to malloc / free, so it wasn't uncommon to have large amounts of static, 0 initialized data.
If the answer is "less than 100K" and "no, of course not," then you can probably get away with:
1. compiling all the QVM files to their Javascript equivalents ahead-of-time;
2. stuffing them all on a CDN (with each JS blob named after the md5 of the relevant QVM source);
and then 3. having a "compiler" in your client that just hashes the source it's about to "compile", and requests the blob with that hash from your CDN.
Of course, if the amount of generated JS is small enough, you could even just serve it all with the client. I'm doubting that one, though; it's probably at least 50MB of code. (Though it might be heavily redundant code... it could compress very well!)
Because that's how quake3 works. What you're describing would be incompatible with all other quake3 clients. And part of the point of implementing the QVM in the first place is compatibility.