(I'm the author of Tern -- I'm obviously biased.)

It compares very favorably. The Orion approach, as I understand it (I only did some quick messing around with scripted, am not an expert) is very local, whereas Tern tries to do whole-program analysis. This should allow it to pick up much more information.

Is it able to track all dependencies in a web page for autocomplete? I mean, keywords defined in dependent files - do they automagically work?

Not magically. That part is, for now, in the scope of the IDE/editor plug-in. You can feed Tern a set of files to analyze, but Tern doesn't, for example, scrape HTML files to determine inclusions.

The one exception is requirejs-style dependencies. Those, if you give Tern some information on where to look, it will resolve and automatically include. I'll do something similar for node.js-style require calls in the future.

"whereas Tern tries to do whole-program analysis."

I suppose you're not doing live type feedback? Given a test suite for the code being edited, that would seem like a natural thing to do.