There are two things to discuss here:
The optimal method
As far as the optimal method is concerned: eyeballing from the hoods will result in a shorter stem, compared to eyeballing from the drops.
With modern "compact" handlebars (and huge hoods of modern STIs) that difference is very small, but it is still worth mentioning and can be noticeable.
For riding on the road (especially if the rider is fit and flexible), the drops method will be better. For riding off road/gravel (or for the less flexible riders) the hoods method will be better (for most riders). It boils down to whether you want the stem to be on the shorter side (better for off road riding in my opinion and experience).
Another caveat is the handlebar height. A vast majority of road bikes and cyclists have the handlebars set too low (for the rider's flexibility and riding style), so riders almost never ride in the drops. Such riders may also prefer the hoods eyeball method (it will make the bars more comfortable for them).
Fork shape
Old forks were made of steel and had the lower section
curved towards the front and
thinner. Why?
The lower section was
thinner to minimize weight. With rim brakes, you could afford to use weaker lower fork section because the loads and the torques were not huge in the lower fork leg section (see
disc brake forks section of the frame design article for more details).
The reason for curvature was to achieve the
optimal steering (it affects trail to put it briefly). With stiff disc brake rigid forks, and with suspension forks, this is achieved by making the entire fork legs start at an angle, right from the crown. Old rim brake steel forks started in line with the
steering column and then curved at the bottom section.
So, those are two ways to achieve the same wheel position. That is why the fork leg shape does not affect the optimal stem length.
Final notes
This "eyeball" method is a rough guide to get the ballpark stem lenght. It often happens that a +-1 cm (or even a bit more) stem length ends up with the wheel's axle blocked from view (i.e. being optimal according to this method).
So, the method is not a be all end all, but more of a way to get a decent starting point, and do fine adjustments from that point.
Relja