And (debugging some other code and rethinking about this for separate reasons), I noticed the side-comment on the reedbeta webpage:
“(There’s still a C1 seam between the quads, where the mapping derivatives jump; but that’s unavoidable as long as we insist that the texture completely fill the quad.)”
…which I believe is the explanation for remaining (subtle) jumps in normal map sampling. And I suspect there’s no way to get rid of them: you would need to have all the data for ALL adjacent quads when texturing one quad, so you could calculate the ddx/y across the borders (ouch!). So my current hack of disabling mipmapping entirely might be the only practical solution!