Eink works by physically moving bits of material around in a viscous liquid. You can refine that (use stronger electric field, stronger reacting pigments, less-viscous fluid, thinner layer for smaller travel distance) but never get around it.
I suspect that the current technology works by "setting" in place whole regions of the screen at once, and that this kind of process is inherently slow. Maybe in a (not so distant) future it'll be possible to have individual pixels changing, in a similar way as an mp4 doesn't update all the image when not necessary.
Or not. So many technologies just embank somewhere and find their niche.
It's all tradeoffs, improving one number only to a limited extent and only at the expense of other numbers. Ie, faster action at the expense of worse image. No free lunch or getting around it by handwaving 'advancement'.
Trade-offs are a problem, but in this case i don't see any physical reason for this to be a theoretical trade-off. More like a limit of the current technology.
Why? Nobody said that the cell is "reset". The cell is modified. The resulting state depends on the former state. This causes ghosting.
To achieve what you are suggesting, the procedure would be e.g. to first set the pixel black, then white (reset to a base state), then the greyscale value of the intended content. This would slow down rendering dramatically, and make regions flicker.
And, note, this may be something that some "EPD waveforms" do - but there are tradeoffs... E.g. Fast? Ok, lose the greyscales. Detailed? Ok, get lots of flashing at update. Mid-way? Ok get some degree of ghosting.
E-ink displays are hardly new and lots of people have been thinking about/working on the slow refresh rates for over 2 decades. Do you really think that anything you can come up with on HN in 5 minutes without any knowledge of the tech is something people hadn't thought of before?
EPD displays already have that capability.
When you modify the dots you use a modality that is sensitive to former states. What happens in full-screen refresh is a cleanup to redraw on the best former state for the cleanest output.
Which means that the resetting operation kinda works globally anyways, masking out everything between the xmin, xmax and ymin, ymax.
I'm really speculating here, but this was my impression every time I saw individual areas changing without a gobal refresh.