The carry propagation hardware is the hardest to make work right. The low-digit add has to power the entire chain of carries. As the number of digits increases, it gets harder to do that, because so much mechanism has to be pushed.
There are carry mechanisms which use an external power source for carry propagation. Babbage's Difference Engine has one.[1] All the pending carry values are stored in a latch for each number wheel. Then a cam system applies the carries one at a time. This scales to large numbers of wheels.
Yeah carry propogation is tricky. Matthew Jones (the historian who wrote the book I discuss in the video) calls it the "Sufficient Force" problem. Babbage certainly did solve the problem, but so did Thomas de Colmar in his calculator and all the derivatives it spawned (including mine). This is because the mechanism that performs the carry is driven by the central drivetrain and furthermore each column is offset in the cycle (by 1 tooth of the drive sprockets) so that carries ripple, instead of happening all at once. See 20:35 in the video. This means that a low-digit add does not directly power the entire chain of carries, it just triggers them. And furthermore, you can basically keep adding columns without worrying about the force increasing drastically: Thomas de Colmar even made a "piano Arithmometer" with a 30 digit capacity.
While carry propagation is certainly a hard problem to solve (just ask Leibniz), I had much more difficulty getting the zeroing mechanism to work smoothly - in a way it's a similar issue because you need to move a bunch of parts all at once, which from a force perspective is difficult.
Oh I should mention that Pascal also solved the sufficient force carry propagation issue in the Pascaline with his "sautoir" mechanism.
You can definitely feel it when doing subtraction on a Curta, there's significantly more drag involved in it both because you're generally adding a larger number so more teeth interact but also the wave of the carries going around. However the low digit doesn't have to power all the carries though on a Curta because all the carry does it shift a gear up that then interacts with a single tooth (or 9 during subtraction) on the drum that performs the carry for the next digit up.
There's a whole page of Curta info [0] and a 3d simulator [1] where you can see how similar the setup is and some of the ingenious tricks to fit all of the functions of this machine into a little larger than a grenade sized package.
Yes, as the number of wheels scales up, powered carry becomes necessary.
Another mechanism that's been used is sort of analog - differential gears, with two inputs and one output. Race track totalizators used that to add multiple unsynchronized inputs. Here's one from Adelade.[1] The machines were huge and heavy, but reliable.
(It is a tradition and a contract term in the gambling industry that gambling equipment companies are strictly liable for errors. As a result, that industry builds unusually reliable equipment. GTech once mentioned in an annual report that they paid out about 3% of revenue in error payments.)
There are carry mechanisms which use an external power source for carry propagation. Babbage's Difference Engine has one.[1] All the pending carry values are stored in a latch for each number wheel. Then a cam system applies the carries one at a time. This scales to large numbers of wheels.
[1] https://youtu.be/vdra5Ms__9s?t=247