Your first option will be solved as a yes or no once someone tests dictation in airplane mode, since that shouldn't rely on Siri's natural language processing, only the speech recognition.
My guess? Locally, the audio signal gets processed and compressed into a form more easily sent over a 3G connection while simultaneously making sure to preserve the information relevant to speech recognition, and then this processed and compressed data is sent to the servers for Nuance's speech recognition and then Siri's natural language processing. My other guess is that this would run on an iPhone 4, but Apple deemed it too slow or otherwise dodgy to officially include it in iOS 5 across the board.
I'm pretty interested in machine learning and natural language processing myself, so even if she isn't perfect in responding to my requests, I'm quite excited about Siri. Hell, she'll still probably be better about it than my ex.