The really hard part about MineCraft isn't rendering, though; it's multiplayer.
Only good desktop or laptop machines can run it well. Runs like shit on average machines (lets not even mention mobile). This is after several months of dedicated optimization time to reduce resource usage and minimize build size.
Overall, you are faaaar better off writing it in JavaScript with one of the pure js engines if you are going to go this route.
Sometimes the language is less of a good choice. E.g. i did a small toy project in coffeescript about two years ago, and found to my horror, that if your function ended in a loop, it would return an array. Always. This isn't possible to accidentally do in the host language, so I'd say it is a bad fit (for game development, where you care about such things).
This is probably as a result of CoffeeScript's Ruby-esque syntax and semantics. It does mean you can write some very clear and succinct code at times, but the tradeoff is that it can have unforeseen performance implications.
You have to be explicit in Vanilla JS to return things or it just return null. Coffeescript explicitly returns the last line every time.
I tested several real world web sites, like google search, github, google docs - only around 30% of javascript loaded into browser is actually invoked at run-time. Maybe in your case the usage ratio is much lower?
Would it be helpful to remove unused code with a tool like this https://github.com/avodonosov/pocl ?
Google products (google docs for example) use Closure Compiler for dead code elimination, but anyway, after clicking almost every menu item in document editor, only around 45% of code was invoked.
If that 26mb web site was available online, I would be interested to measure it's code usage ratio.
BTW, the idea is not radically new. Java loads classes only when they are accessed first time.
Operating systems load native applications by mapping executable file into virtual memory. The pages are loaded into physical memory only when CPU tries to access them.
There is no inherent size advantage of doing things in Javascript vs. cross-compiled to asm.js. For instance, three.js is 100kByte minified and compressed, that's in the same ballpark as the (admittedly simple) demos here: http://floooh.github.io/oryol/
Once you are over the static overhead for the C runtime and web-API-wrappers of roughly 30..50kByte, emscripten demos don't grow much faster than minified and compressed Javascript code.
[edit: typos]
However, I remember that the first Unity HTML5/WebGL demos a few years ago were much smaller, a complete zombie FPS had a 3.5MB upfront download which is actually pretty decent.
There are other issues as well. The biggest offender being the inflexible heap size, e.g. when you launch an asm.js module, you pass in an array buffer as the heap which cannot be resized. If you need more memory than this, you are SOL.
This wouldn't be that bad, except in practice it needs to be ridiculously small if you want to run on most machines. 256MB, which not much memory for a game at all, will fail to allocate, or kill the tab on about 33% of machines (being generous. For us its probably more like 50% of machines, but we make educational games that need to run on university machines, which are uh, on the lower end).
Either way, as a result, a great deal of the optimization time I mentioned was spent getting all of our games to run with under 128MB heap.
That said it's definitely possible if you're careful (I'm not sure this is quite as true for a voxel game though, but I could be wrong). My complaint wasn't about the idea of using asm.js, it was the idea that Unity would produce a good result. It just doesn't.
There are other actively developed Web-friendly WebGL frameworks, like Goo Create, PlayCanvas, BabylonJS and many more.
I suspect Minecraft has a much harsher profile, since it's possible for every player to be loading a different piece of the world with minimal overlap, and world-pieces involve a lot more state that needs to be preserved/calculated/saved.
Think of two players running towards each other, and shooting at each other. By the time the player (or even the server) receives the data saying the other player has fired their gun, they've both moved to completely different positions.
(There is a good set of articles on this by a game developer, I can't find them now)
https://developer.valvesoftware.com/wiki/Latency_Compensatin...
https://developer.valvesoftware.com/wiki/Source_Multiplayer_...
The world pieces are a very specific set of permutations, which could easily be loaded in to a graphics card's memory at the start of the game. All you are really tracking are which pieces are where, and by the game's nature, there aren't that many pieces or positions they can be in...
https://github.com/SirCmpwn/TrueCraft
May give you some insight into how Minecraft multiplayer (or any other detail of Minecraft) works. The server code is in the TrueCraft directory.