The mantra of optimization is you get what you optimize for. It's amazing how that seemingly innocuous phrase is something every person doing optimization has at some point been bitten in the ass by at least twice. Once when you do something stupid as you are learning and once later when the optimization produces some completely perplexing result leaving you in awe of the power of that mantra.
There was likely no conspiracy precisely because of the difficulty of maintaining the conspiracy at this scale. A much simpler explanation is that they had the system train itself. There's no reason to leave out certain features in the input vector so all the sensors go in. The car learns that when there's no frequent steering input and the cost function is dominated by emissions then you minimize the emissions. And later on the test track, where there is no emissions term in the cost function, the car learns to anticipate accelerations when there is steering input, so the cost function optimizes for performance and fuel economy not emissions. and so on.
One can see how this could happen so easily. And even if one group thought about it they didn't control the whole cost function and were exploring one part of it. Component manufacturers might notice this too but assume it's fixed in the full system. indeed one report said that there was some internal review of some odd issues.
But if you aren't expecting this and you are relying on the model training to integrate many different team testing one can see how this could accumulate.
It's also easy to see how this could even be seen and not noticed. For example, shutting down emission controls and air conditioning and other things is completely the norm in perfromance tuning. When you stomp on your accelerator the clutch in your Air conditioner disengages to give more power. THe exhaust gas recrculator shuts off. You want those things to happen, just as you want the turbo to kick in before you need it and to kick out when it won't be needed. Thus cars that anticipate these changes rather than wait for then feel much more responsive yet can get much better emissions and fuel milage.
But one can see that these traits could accidentally "cheat" when ever two different optimization features come into conflict.