With Fastboot, somehow an additional, non-interfering layer of computation on the server does the same loop (without having to download anything and without significantly slowing down overall client app JS initialization)
Right. Typically, the way most client-side JavaScript applications (or "SPAs") work is by having a small, static HTML file that doesn't contain much content (beyond maybe a loading page). It contains tags that point at your JavaScript payload.
First the browser downloads the HTML, then the JavaScript, then the JavaScript runs and fetches the data via XHR. Only then does the user see the content they were after in the first place.
This actually works surprisingly well for "workspace" apps where the user is using it throughout the day (Gmail, Google Docs, etc.) On modern devices with good broadband, the difference is negligible.
But the thing this sucks for is content sites, where you aren't using an "app" but you're just clicking a link in Twitter or something. If it doesn't load within a second or so, you aren't that invested that you don't just close the tab. That has been the biggest source of pushback on frameworks like Angular and Ember for sites like this.
FastBoot bends the curve by replacing that static HTML file. Rather than serving an empty document that just points to JavaScript assets, we keep your Ember app running in Node.js on the server. When an HTTP request comes in, we direct it to Ember's router, where it figures out what models to load and components to render. When it finishes, it sends the document back to the browser.
You can think about this is as effectively outsourcing the JavaScript runtime to the server for the first load, but then the browser can take over again on subsequent navigations so it's very fast.
I think FastBoot is a great option for search crawlers, Facebook and Twitter embedding (it supports Open Graph and Twitter Cards), and supporting JavaScript-less clients. Most importantly, it's a way to get content quickly to users with a cold cache.
That said, we are planning to aggressively take advantage of App Cache and Service Worker, so ideally any second-time visitor to your site only has to fetch the raw data to see what they're after.
I'm guessing that somehow the client's processing is disabled on that first "pageview" to avoid double calculation. Is that correct? (if it is, it's very cool :)
Currently it does a full rerender once it loads, but one of the motivating features for writing Glimmer 2 is the ability to quickly "rehydrate", so that rerenders are imperceptible to the user assuming nothing has changed. We'd also love to automatically serialize the backing models of the app so you don't have to double fetch.