What I messed up is the 'clipping' of the output values back to the input value range. This over-ranging occurs because 3dmerge uses FFTs for blurring, and you can get the Gibbs artifact at sharp edges (brain-air, say).
So the upshot is that very few voxels will be different -- if you subtract the two datasets (new and old blurred ones), you'll see very very few voxels where anything is changed, and the changes are all outside the brain. So the results of 3dDeconvolve will be only VERY slightly affected, in regions you don't care about anyhoo.