3 improved formatting

As far as I know, the answer is no. We don't have an argument not based (directly or indirectly) on digonalization for even proving there is a language outside Dec (The set of Turing decidable languages).

Characterizing a class means having criteria for showing when a language is in the class and when a language is not. The second part would give us tools to prove lowerbounds on such classes.

For very small classes like $AC^0$ and $AC^0[p]$ there are combinatorial arguments (also for monotone-$AC^0$ circuits), but the current state of knowledge is consistent with $AC^0[6] = NP$. The Razbarov-Rudich (RR) Natural Proofs theorem is related to why the previous combinatorial arguments haven't been successful in separating larger classes. In simple terms, there is no "natural proof" for separating the larger classes unless there is no pseudo-random number generators (many believe they do exists), and all previous proofs were natural. Tim Gowers wrote a number of interesting blog posts on the topic: http://gowers.wordpress.com/2009/09/22/a-conversation-about-complexity-lower-bounds/

The reason that we have pumping lemmas for $Reg$ (regular languages) and $CFL$ (context-free languages) is that the machine models for them are very specific which gives the strings accepted by them special combinatorial structure. For classes defined by Turing machines this is not the case. If a complexity class contains some basic functions (say $AC^0$) and is closed under composition, proving lowerbounds for it becomes a rather difficult task. Note that $Reg$ does not contain $AC^0$ (${0^n1^n: n \in \omega }$ is in $AC^0$ but not in $Reg$.)

part 2. You are looking for a combinatorial property of sets of strings in a complexity class. The computational complexity of checking a property is a good measure. Any property stated in the first order logic (on a finite structure) is $AC^0$. The RR theorem applies to properties which are in $P$, i.e. can be checked in polynomial time given the truth table of the functions (characteristic functions of sets restricted to inputs of size n). Encoding Turing Machines (TM) is not difficult, and I think it can be done if one goes a little above $AC^0$, e.g. $TC^0$ is probably capable of checking if a string encodes a TM. You are looking for a property that holds for all functions in a class, but not for those outside it. The negation of it would be a property of the kind considered in RR's theorem. By these considerations, I think there are no properties of the kind you are looking for unless there are no PRNGs.
$AC^0$ is class of sets of strings accepted by (uniform) circuits of polynomial size and bounded depth, i.e. there is a constant $d$ and a polynomial $p(n)$ and a sequence of $\land$,$\lor$,$\lnot$ circuits ${C_n}_{n \in \omega}$ of using unbounded fan-in $\land$,$\lor$,$\lnot$ gates and of depth $d$ and size $p(n)$ that accepts them.