You can fine Viresh's vhost-user implementation at
https://lore.kernel.org/qemu-devel/cover.1617278395.git.viresh.kumar@linaro.org/t/#m3b5044bad9769b170f505e63bd081eb27cef8db2 (local)

It looks OK so far; yet, it is not complete. But it might be bearable
in the end.

As you say, it does get a bit clumsy, but I think there is also a good argument
to be made that the clumsiness is based on the host Linux user interface
more than the on the requirements of the physical interface,
and that should not have to be reflected in the virtio specification.

Makes sense to me.

Right, this one has come up before as well: the preliminary result
was to assume that this probably won't be needed, but would be easy
enough to add later if necessary.

If adding support incrementally works for such an interface, this makes
sense as well.

So, where are we? As I understand, this v10 does not support I2C
transactions (or I2C_RDWR as you said). But you want to support all
clients. So, this doesn't match, or?