It's really hard to know how much variability is good and when training for consistency within variable patterns goes too far. The reality is that no movement is ever exactly replicated. At no point in your life will you even take two steps exactly the same at the biological, or likely even bio-mechanical level.
When we talk about creating players with better movement solutions for something like pitching, what we're really trying to do is teach them how to combine pieces of the pitching puzzle in different ways in order to compete in the cold, when they're sore, and in the rain. We need them to be able to make pitches when they gain mobility or lose strength. They need to be able to pitch when there's a hole in the mound and their Dad is in the stands.
We probably don't need an over the top guy to throw sidearm, and we'll never need someone to throw a baseball pitch with a softball. But doing these things allows the body to solve the problem of, "how do I pitch?" in unique ways, making it better at solving that problem when presented in a never before seen context. That's why we'd program these types of activities.
Still, there is a limit. I don't think it would be helpful to practice throwing while standing on your head, or to train to throw the baseball like a Frisbee. These are obvious.
The tricky stuff is in the middle - everything between throwing off the mound with a baseball and doing crazy things like behind the back headstand throws - and in the dosage of these things in the middle.
How often should our work be specific? How often and to what degree should it be variable?
In general, I think the further away the season's start, the more variable work you can do. As the season looms closer, the allocation of training economy needs to shift towards the specific.
Let's think of this from a macro sense. Early off season is for regaining mobility, strength and endurance lost in season. As the off-season progresses, lifting volume comes down as intensity increases and skill work is introduced. That skill work often begins with drills and low intensity, low volume throwing. As we get closer to the season, lifting volume is further reduced, drill work is reduced as throwing frequency and duration increase, and specific mound work begins.
In a micro sense, the most specific training you can do is in-game pitching. It is in this environment alone that a pitcher's execution matters. On the still very specific side we have bullpens. Throwing them on the game mound more closely simulates the game than throwing them in turfs to a command trainer, but all types of bullpens are more specific than non-mound throws. And all throws with a five ounce baseball are more specific than those thrown with over/under loaded/sized balls. Less specific still are plyo throws and constraint drills.
But all of these have value at different times and for different people. Finding the optimal mix of variability and specificity is half the art of coaching. (The other half is relationship building and communication.)
A coach needs to be able to identify each pitcher's most pressing need and communicate this idea in numerous ways. He needs to have a large toolbox of tactics to elicit change and know when and how to deploy them. Ultimately, he needs to expose each player to enough of these ideas that the player can learn how to coach himself.
The coach's job is to make himself obsolete, because at the end of the day the player is the only one who knows what it's like to be him. He's the only one making the pitches. Learning how to learn is ultimately the meta skill that a coach needs to teach his players. The skill of deciphering what information sources are valuable and how to utilize them is teachable.
So the questions remain: How much of the training load should be specific? How much should be variable? And how do we create a player with movement patterns that are stable within an appropriate bandwidth of variability?
And the answer: We test. We experiment. We re-test. We iterate. Ad infinitum.