Right.
By present-day standards, the Manabe and Wetherald model is very simple. This was indeed the criticism at the time-- "but the model doesn't account for XXX effect"-- and all of the present-day models basically work on adding in the various feedback effects you mention.
The simplicity is both a flaw, but also a virtue. Basically, Manabe and Wetherald is a model of the greenhouse effect, and nothing but the greenhouse effect: while there are a hundred more sophisticated models these days, now all of the criticism is "but how do you know that you didn't get XXX feedback effect wrong?" Well, Manabe and Wetherald didn't have all the bells and whistles-- it was the first real greenhouse effect model that incorporated real-world, measured infrared absorptions, accurate radiative transfer, and convective equilibrium, but that's all.
This is typically the way science is done. First you make the back of the envelope models, then the simple models, then you progressively add more refinements.
Surprisingly, the other effects matter less than you might think. Clouds was the first criticism made (and all modern models have cloud effects)-- but clouds aren't actually a huge effect. If clouds blocked visible and infrared light equally well-- and to first order they do-- cloud cover would have little effect on average temperature: the infrared radiation scattered downward heats the planet, the albedo scattering cools the planet, and in the simplest model the two balance out. Of course, the real world isn't the simplest model, but in some places clouds can actually increase the temperature (you see this in models of carbon dioxide clouds on early Mars.) What clouds mostly do is tend to equalize the daytime and nighttime temperatures. This is actually a good way to separate cloud effects from infrared absorption effects. (Another way is to look at vertical profiles).
As for the constant relative humidity assumption-- well, what would you suggest would be a better input assumption? Again, it's a good simple assumption. It does implicitly include a feedback effect, but it's pretty much the most transparent way to incorporate it. Most importantly, note that by assuming constant humidity, there aren't any adjustable parameters. If you worry that models have been "tweaked" to make the output match the data, well, there isn't any feedback to tweak here.
With the advent of supercomputers in the 70s, models got more detailed. The next good summary of models would be the U.S. National Academy of Sciences report 1979, by which time the report could look at and compare several models. The '79 NAS report is a good go-to reference for models of the '70s; and is still a bit before the politically-motivated attacks started muddying up the conversation. That still gives 35 years of data that can be compared to prediction -- a long enough run to average out some of the year-to-year variation and compare the models to reality. I graphed (but haven't yet added the most recent data, 2014, yet) and, yes, the measured temperatures fit inside the error bars of the NAS models.
After that, models got much more sophisticated very quickly (and so did the attacks on the models, resulting in a fast evolution as ever more sophisticated models addressed ever more complicated critiques). Today there are hundreds, and probably thousands of models being run. Comparing them to measurements is more like looking at statistics than looking at an individual model. There's some good graphs comparing models to reality in the IPCC working group 1 report, if you're interested in tracking them down.