I guess what I'm really saying is that giant simulations have two main purposes-- one is to try and guarantee that the key physics is included (the "fishing net" philosophy of fishing), and the other is to try and obtain highly quantitative results when the situation is very well constrained by the physics and idealized assumptions are either not necessary or well supported in fact. But in neither case are we "finished" when the simulation achieves these ends-- yet that is often how they are treated.

In the first case, the work has just begun-- we have insured that the dominant physics is present in each situation the simulation covers, but we have not yet identified what that physics is. More often than not, in my experience, giant simulations can be broken down into smaller pieces where in each subset there is actually something quite simple happening, something that does not require the full simulation to understand (though it may have required it to find). It is often the case that in hindsight, the simple physics makes perfect sense and feels like it should have been anticipated prior to the simulation (though in practice it often is not). But my point is, if this followup analysis never occurs, as is all too often the case in the literature and at meetings, then this great promise is never actualized. Instead, people think the simulation has accomplished its goals, and nothing more is needed from it. It is a "complete" simulation-- but is it completed?

In the second case, where quantitative accuracy is the motivation for the complexity of the simulation, there is no simpler subset that can achieve that end. However, we are still not done, because even when quantitative accuracy is the goal, there is still a role for approximation. One role is in anticipating how the results will vary as you change the parameters. The "kitchen sink" approach to that issue is a brute-force variation of the input parameters (a so-called "grid of models"), and then interpolate to any actual desired situation.

But a simplified approximate understanding of the dominant physics also informs the process of understanding the sensitivity to input parameters, and the inaccuracy introduced by taking that approach is well compensated by the insight gained. Once the approximate dependences are understood, one can anticipate what combination of parameters will achieve some desired end, and then a full simulation can be run for the new parameters as a final check on that prediction. But the simple fact is, astronomers are forever undertaking calculations that are more accurate in form than they are in substance-- they are grinding out that third decimal place, yet invariably some new physics discovery comes along and changes the first decimal place. Examples are countless.

Granted, this is certainly a good example of how a "complete simulation" can make predictions that observers can then attempt to test, as a check on the assumptions of the simulation. But what I'm saying is, does this really complete the relationship that the theorists and observers should be having? Is it enough for theorists to say "I predict this, never mind why, you can picture this cartoon if you like, but what matters is that you look for it"? Put differently, can you tell me (in physical terms, not cartoon terms-- we get plenty of cartoons) why the Millennium Simulation produces so many dwarf galaxies? What is the key aspect of that simulation that gives you that, and once you understand that key aspect, how could you have obtained that same result without the simulation? As I said, hindsight is better than foresight, which is the main purpose of simulations if you ask me, so I'm not saying we never needed the full simulation. I'm merely saying it should not be viewed as the completed pinnacle of the theoretical effort.

In some situations, the accurate quantitative analysis is all that has value, because you are not trying to understand how something is working-- you already know how it works, and you are pressing for greater and greater detailed accuracy in your model inputs. Such situations are pretty rare in astronomy! But this is indeed one.
Transit searches are observational, I'm talking about the interaction between theorists and observers. So your last point is more relevant-- about what we can theoretically infer about the atmospheres. But this is a perfect example of what I'm talking about-- when we get information about the atmosphere of a transiting planet, do we need a giant "black box" simulation of all the things that might possibly be happening in that atmosphere? Do we really expect to be able to anticipate the possibilities so completely? I'd say we need a simplified understanding of what kinds of observed features map roughly and approximately into what general kinds of physical phenomena, long before we'll be in any position to apply meaningfully a black box simulation to making quantitative predictions (indeed we may never be in that position-- we can hardly even do that in our own atmosphere, just witness global climate simulations).
Another perfect example of what I'm talking about. The discovery was observational, and grossly disagreed with the current best models. One did not need a gigantic "black box" simulation to tell us that acceleration meant gravity was doing something weird! Even now when we do quantitative simulations, we are forced to use a rough treatment of dark energy, via the simplest possible equations of state we can think of. Again we're having at that third decimal place-- yet do we really think that physics 100 years from now will not have discovered something that significantly changes the entire picture of what dark energy is? Only if we are no students of history!

I take a different message-- I claim that if you analyze those elaborate models, you will generally find that something much simpler actually constrains any particular observation you want to understand. It can be different for different scales and different questions, so it's nice to have it "all in one place", i.e., in the kitchen sink somewhere. But that's no reason to ignore that oftentimes physics involves a simple interaction embedded in a much more complex milieu-- and isn't it our job to find that out? We end up with less egg on our face when some new physics means the "kitchen sink" simulation is no longer reliable. (No matter, one nice things about codes is they are as easy to change as a four-day weather forecast...)
I would view these as observational details. Yes many observations require sophisticated data reduction and analysis-- but that's in an effort to isolate and separate the physics of interest, it is not necessarily part of why you did the observation in the first place-- that's usually done to try and isolate a particular physical effect that is embedded in whatever is the complete simulation you are using. Extraction and isolation is a key element of doing physics, that's basically all I'm saying.
This is mixing in issues about analysis of complex datasets. I am certainly not suggesting that reality is not complex-- I'm suggesting that our job is become adept at finding ways that allow us to pretend, in some isolated context, that it isn't.

Right!

__________________
Logic is the grammar of truth.

Meaning and absolute certainty are incompatible, and profound meaning and absolute certainty are profoundly incompatible.

The only thing intelligence is capable of is recognizing itself.