I have some doubts that this would provide a noticeable benefit in practice. The type information in the result set is fairly compact (though it's true we could save the full metadata in practice). I'm not sure reading the message is much of a bottleneck in practice for small messages (and for big ones, the metadata size is negligeable anyway). And there is compression too.
On the other side, this does complexify client drivers. Currently, you can fully decode a result message without any external information. This is a nice property implementation wise, and is somewhat safer. And I'm not sure requiring too much state from the client driver to do basic things is ideal.
I can be wrong, but my intuition is that neither MySQL nor PostgreSQL does that because they don't consider it worth the complexity. And that's my intuition too.
So I'm fine doing some benchmarking to see if this can make a measurable difference in practice, but I'm -1 on going ahead with this without concrete evidence of the benefits since there is known drawbacks. And I kind of feel it's too late for 2.0.