The action is sometimes a saddle, but it is a minimum over small enough regions, but the reason it is a minimum instead of a maximum is due to a convention.
The action depends on the endpoints, the path and the Lagrangian. So the stationary action(s) and the physical path(s) depend on the endpoints and the Lagrangian.
How do you know which Lagrangian to use? Firstly, it should generate the correct Euler-Lagrange equations. This is a hood ohtsical reason to use a Lagrangian compared to another because this criteria focuses on testable experimental predictions. Secondly, you can use some pure conventions. Because if you multiply a Lagrangian by a nonzero scalar it will generate the same solutions to the Euler-Lagrange equations, and make the same predictions about observations and experiments. So which of the continuum many Lagrangians that are physically the same you choose to use is determined by convention.
Thus you could multiply all the Lagrangians you've seen by a negative one and you'd get the same physics and now all the minimum actions would be maximum actions. So we get minimum actions instead of maximum actions because we make an arbitrary choice unrelated to the observation that is literally chosen to give the least action (this is because the whole game was inspired by the idea of a ray taking the least time, so if you select the action that has the term for the kinetic energy be positive then the free particle action gives a minimum).
But even if you decided to break convention and multiply your Lagrangian by $-1$ you still couldn't say greatest action instead of least action because the action can still be a saddle. So the only terminology that is always correct is stationary action.
When can the action be a saddle? Remember how the action depended on the Lagrangian and the endpoints? Whether you are a saddle or a minimum depends on the endpoints.
Specifically, the action can be a saddle when the endpoints are far away. See for example When action is not least by C.G. Gray and Edwin F. Taylor in the American Journal of Physics Volume 75 Issue 5 pages 434-458 (2007); http://dx.doi.org/10.1119/1.2710480
A free source is, I think available on one of the author's websites www.eftaylor.com/leastaction.html and the other author has a peer-reviewed article on least action at Scholarpedia at http://www.scholarpedia.org/article/Principle_of_least_action
Keep in mind that if you write a Lagrangian as a sum of a bunch of terms, it is only the overall nonzero scalar that doesn't matter, the relative sign and magnitude of each term are totally physical.
Now, let's say you have a Lagrangian and some endpoints and you wonder how to tell whether the stationary action is a minimum or a saddle. In one-dimensional calculus you can take a second derivative and if it is strictly positive you have a minimum. There is a version of the second derivative test in variational calculus and so you can use that. It involves taking a second variation and checking to see if it is strictly positive.
It won't change the physics. The physics won't change in the sense that you get the same Euler-Lagrange equations regardless of whether the action was a minimum or a saddle. Now it is possible to have multiple paths between the same endpoints. You already knew that. You could send a billiard ball from the center of one side to the center of the other side in two different ways that take the same amount of time (shoot left or shoot right) and the same thing happens in physics, you can have two different velocities that take you on two different paths that lead to you starting in the same place and ending up in the same place and taking the same amount of time.
That's life. It happens when that time interval is long enough.
