synthetic differential geometry and other alternative theories

Perhaps I can make the implications of what Harry said a bit more explicit. A well-adapted model of SDG embeds smooths manifolds fully and faithfully. This in particualar means that the SDG model and the smooth manifolds "believe" in the same smooths maps between smooth manifolds (but SDG model contains generalized spaces which do no correspond to any manifold), and moreover precisely the same equations hold in the SDG model and in smooth manifolds. In this sense SDG is conservative: the model will never validate an invalid equation involving smooth maps between smooth manifolds.

The situation is really quite similar to other situations where we have to distinguish between truth and meaning inside a model and truth and meaning outside the model. For example, there are models of set theory which violate the axiom of choice, but these models are built in a setting where the axiom of choice holds. This is no mystery or magic, as long as we remember that a statement can have a different meanings inside the model and outside. The same applies to SDG: when the internal meaning of statements in the model is appropriately interpreted on the outside, nothing can go wrong (that's what a model is, after all).

From n-lab:

"A topos T modelling the axioms of synthetic differential geometry is called (well) adapted if the ordinary differential geometry of manifolds embeds into it, in particular if there is a full and faithful functor Diff →T from the category of ordinary smooth manifolds into T."

The main point here is that we can develop the theory abstractly, then apply the theory to the category of differentiable manifolds, which is equivalent via this functor to a subcategory of this nice topos where we're working. In particular, all of the abstract results restrict to the classical results.