So I tried it, and it works perfectly (and without using raytracing or other bs ) See attached screenshot. The reflection is now blurred and the water surface is still crisp. The DOF for the water reflection is just the reflection texture downsampled twice and then blended in with the original reflection texture with CoC per pixel calculated from depth.
I think the reason I didn't think of this in the first place is that I just thought it would be too expensive. However, if you just assume there will be other things happening, tonemapping, bloom, motionblur, dof, color correction passes and whatever else you might have going on, then adding a simplified DOF filter at half res on top of the reflection rendering isn't that big of a deal.
Here's a photo I found after a quick google, just for reference. Notice how the image in the reflection matches the CoC of the real world.
EDIT: Sorry for spamming like this, but I just realized that I could just run the DOF post processing (or a simplified variant of it) on the reflection texture when I render it. Since the reflection texture is about a quarter of the main backbuffer, maybe performance won't be horrible. I'll try it and post my findings.