My limited experimentation seems to be consistent with what others have noticed -- exactly as bobw stated, MJPEG generally produces better looking output for comparable file sizes, and also encodes faster. But, this comes at the cost of considerably more complex decode. Since (M)JPEG is DCT-based, decode requires performing inverse DCTs, which is computationally expensive. Vector quantization basically requires just a lookup table to do the decode (but a massive search operation to generate the lookup table during encode, hence the slowness). Cinepak evolved with the specific purpose of easy decode, so they apparently made concessions to allow older CPUs to keep up.

I'd be very interested in the results of any experiments you run. In particular, if there's a way to make Cinepak perform better, that would be terrific. In the meantime, I'm going to take the clips I've transcoded and see how well they'll play on some of my older macs, to get calibrated on what can play what.