Main question: What is the precise term for the smallest addressable memory block?
I'll add another answer to address that in a different way. In electronic hardware, we call it the bit - binary digit. That is an entity that can represent any two values. We usually think in terms of 0 and 1, but it could be 3 or 4, 365 or 266, -3 or -4, even 25 or 37.
Any signalling system can be used to represent these values - flag up, flag down, eyes open, eyes closed, +5v, -5v. That is not important.
What is important is that philosophically we are representing the smallest distinguishing amount of information. This could be on, off, or true, false, or up, down, or 0, 1 - anything that distinguishes two separate states. We can map these values onto any of the above signalling systems and many others.
Now the question is, how can we test and set such a small amount of information individually? As I said in previous answer, the B1700 chose to address that smallest amount of information directly.
However, most machines decided to only address larger amounts of information. Let's consider a group of four bits with a single address. So if we get the value of 1011 in our location, how do we test the second bit from the left. We use a mask: 1011 and 0100 tests just the second bit. So how do we set second bit to 1? A little CPU arithmetic says the value will be 15 or 1111, so that entire four bits is written back to memory, even though we have really only set one bit.
Now this is not useful for most applications. Most applications are representing data or information, up, down, true, false, open, shut.
We want to say things like:
if open then
...
else
...
end
or more likely apply that to a larger entity:
if door is open then -- most likely 'door.open'
...
else
...
end
'door is open' illustrates hierarchical addressing. The main system addressing gives the entity door, and door has its own addressing which gives access to open (and maybe other attributes).
Most sets also have more than two possible values (a set with one value never changes and therefore does not even need representation, so zero bits). For these we have enumerated sets, like (yellow, green, blue, purple haze, red). These define sets and types and the exact number of bits required is given by the number of values (log2 (number of values)).
Thus the optimal addressing really depends on the entity size used in the application - maybe even variable sized entities. But in most hardware such addresses must be translated to the fixed size the hardware defines. This of course could cost in terms of time. It should also be something that an automatic translator does (compiler or interpreter), not a programmer, just as such a system would generate code to test and set bits as above (if bits aren't directly addressable).
An important point here is not to think in terms of electronics - electronics is just a really good and fast way of processing computations. There is nothing magic about electronic computation that makes it possible to do computations you couldn't do otherwise. The magic is only in the speed. That is why low-level abstractions such as bit, byte, word, or hardware addressing mechanisms (pointers) are really not that useful.