Multipass rendering -- buffer changes vs. texture changes

I'm going back and writing a multipass path for my terrain engine, and as I see it, there are two ways to structure it:

• For each patch, set the first texture, render. Second texture, render. Third texture, render. This means a lot of texture switches but minimal buffer changes.
• For each texture, render every patch. This means a lot of buffer changes, but minimal texture switches.

I'm inclined to think that the former method is going to make way better use of the post-TnL cache on-chip, but what I know of graphics leads me to think that I should strive to minimize texture state changes rather than buffer changes. Another thing is that I'm using offsets into a single buffer, so I never change buffer, only offset. I'm choosing the second method for this reason, because I don't think the offset changes are costing much...but I seriously doubt I'm putting the cache to good use. Any insights?