I thought as much, but maybe there is some driver which can offload
whole i2c_transfer to HW, and has to pass the addresses of the buffers
to the HW, and the HW can have problems if the buffers overlap
somewhere...

Well, sure, you can never know what crazy HW is out there :) But that
shouldn't prevent us from doing legit I2C transfers. The likeliness of
such HW is low enough; They must process the whole transfer in one go
(rare) AND have the limitiation with the buffer pointers (at least I
don't know one) AND have no possibility of a fallback to a simpler mode
where they can handle the transfer per message. If such a controller
exists, it would need a new quirk flag, I'd say, and reject the
transfer. But that shouldn't stop you from fixing your issue.

Thanks for thinking thoroughly about drawbacks! Much appreciated.