I think a sub-$500 per-station cost would be wonderful, but this is all just a pipe dream...

Are there established algorithms to determine what seismic data is 'interesting' as opposed to streaming it all in real time (and keeping the radio on) constantly?

Almost all digitizers I've seen support the same STA/LTA (short term average ÷ long term average) triggering mechanism, where data is declared interesting if the energy over a short time window divided by the energy over a long time window exceeds some configurable threshold. If you only send triggered data, it's a great way to trigger repeatedly on local noise and miss all/parts of events you actually want to record.

Sending continuous data from stations to a central processing site is greatly preferred, especially since the data rate is so low. Three channels of 20-bit, 100 samples/s data from a low- to moderate-noise site that's losslessly compressed by the digitizer fits comfortably in 9600 bits/s.

Why 802.11n instead of cell phone networks - don't you need to be away from traffic vibrations, and hence roads and homes?

Siting seismograph stations is a tradeoff. Too far away from civilization and you have no way to get data back home except via (expensive, power hungry) VSAT or high power radios. Too close to civilization and you are subjected to civilization's noise (but you can use civilization's communications infrastructure to send your data home, sometimes for free).

The higher a site's noise level, the higher your event detection threshold gets. In other words, the noise consumes the signal from weak and/or distant earthquakes. You can make up for this somewhat by deploying a more dense network that pushes stations closer to where the earthquakes are happening...