This argument has a couple of iffy points, but I believe it does work.

In [this][1] paper, Shelah introduced a logic $\mathcal{L}(Q_{\mathrm{Brch}})$ which is fully compact, has the property that any countable theory with infinite models has models of every uncountable cardinality, and is strictly stronger than FOL for countable models. That paper is pretty awful to try to read. There's a slightly better exposition of this logic in [this][2] paper of Mekler and Shelah's, but it's still not great. (Technically I have not verified that this logic is natural under your amusingly precise definition of naturalness, but it probably is, and I think you'd probably accept this as an answer anyways.)

To specify the logic, we need some preliminaries. A tuple of formulas (possibly with parameters) $(\varphi_t(x),\varphi_r(x),\psi_t(x,y),\psi_r(x,y),\rho(x,y))$ *defines a level tree* if 

 - $\psi_i(x,y)$ defines a partial orded on $\varphi_i(x)$ for $i \in \{t,r\}$, which I will now write as $<_t$ and $<_r$,
 - for every $a$ satisfying $\varphi_t$, the set of $<_t$-predecessors of $a$ is linearly ordered by $<_t$,
 - $<_r$ is a directed partial order on $\varphi_r$ with no largest element, and
 - $\rho(x,y)$ defines a function from $\varphi_t$ to $\varphi_r$ which is strictly order-preserving.

So basically, the idea is that this is the first-order version of a tree with a rank function. (By the way, there seems to be a typo in the published version of this paper. They have $\rho$ as a formula with three variables. There's also what seems to be an incorrect statement in the remark after Subclaim 3.6: They say that every countable level tree has a branch (which I will define in a second, but you can guess what it means). This seems to contradict the existence of countable well-founded trees.)

A *branch* of a level tree is a maximal linearly ordered subset of $\varphi_t$ whose image under $\rho$ is unbounded in $\varphi_r$. The logic $\mathcal{L}(Q_{\mathrm{Brch}})$ allows monadic second order variables and has this quantifier
$$Q_{\mathrm{Brch}}b(\varphi_t(x),\varphi_r(x),\psi_t(x,y),\psi_r(x,y),\rho(x,y))\Theta(b),$$
which is interpreted as saying that if $(\varphi_t(x),\varphi_r(x),\psi_t(x,y),\psi_r(x,y),\rho(x,y))$ defines a level tree, then there is a branch $b$ of the level tree such that $\Theta(b)$ holds (and $\Theta(b)$ is any $\mathcal{L}(Q_{\mathrm{Brch}})$-formula). We are allowed to freely form formulas using this quantifier and standard quantifiers and connectives.

I'll call the version of $\mathsf{PA}$ that has induction for $\mathcal{L}(Q_{\mathrm{Brch}})$-formulas, $\mathsf{PA}(Q_{\mathrm{Brch}})$. By full compactness, we immediately get that $\mathsf{PA}(Q_{\mathrm{Brch}})$ has models other than $\mathbb{N}$. (Although, I do not know whether it has any non-standard countable models.)

Now we come to the point where if I knew more computability theory or reverse mathematics, I could probably have done this proof in a shorter way.

> **Proposition.** $\mathsf{PA}(Q_{\mathrm{Brch}})\vdash \mathrm{Con}(\mathsf{PA})$, so in particular $\mathcal{L}(Q_{\mathrm{Brch}})$ is intermediate for induction.

*Proof.* I will prove this by showing that if $M \models \mathsf{PA}(Q_{\mathrm{Brch}})$, then $(M,\mathrm{Def}(M,\mathcal{L}(Q_{\mathrm{Brch})}))$ is a model of $\mathsf{ATR}_0$, where $\mathrm{Def}(M,\mathcal{L}(Q_{\mathrm{Brch}}))$ is the collection of all $\mathcal{L}(Q_{\mathrm{Brch}})$-definable-with-parameters subsets of $M$. I will call the induced second-order structure $M_2$ to avoid having to type that out again. Since models of $\mathsf{ATR}_0$ always satisfy $\mathrm{Con}(\mathsf{PA})$, the result clearly follows.

First note that it's pretty easy to see that $M_2$ is a model of $\mathsf{RCA}_0$. (More than this is also easy, but this is enough.) Now we will use the fact that $\mathsf{ATR}_0$ is equivalent to the comparability of any two well-orderings (over $\mathsf{RCA}_0$).

*Claim 1.* If $T$ is an $\mathcal{L}(Q_{\mathrm{Brch}})$-definable tree in $M^{<M}$ (i.e., $M$'s version of $\omega^{<\omega}$) with a branch, then it has an $\mathcal{L}(Q_{\mathrm{Brch}})$-definable branch.

*Proof of claim 1.* We can define a branch inductively. For each $n$, given $\sigma \in T$ of length $n$, with the property that there exists a branch in $T$ extending $\sigma$, find the smallest $m$ such that there exists a branch in $T$ extending $\sigma \frown m$. Having such a branch is an $\mathcal{L}(Q_{\mathrm{Brch}})$-definable property, so a smallest such $m$ must exist and by induction we can build a path this way. This branch is definable since a $\sigma$ is an initial segment of it if and only if every initial segment of it satisfies this inductive definition. $\square_{\text{claim 1}}$

*Claim 2.* If $\varphi(x,y)$ is an $\mathcal{L}(Q_{\mathrm{Brch}})$-formula (possibly with parameters) that defines an internal well-ordering on $M$ (in the sense that the tree of internally finite descending sequences in $M$ has no branch), then for any $\mathcal{L}(Q_{\mathrm{Brch}})$-formula $\psi(x)$ (possibly with parameters), either $\psi(x)$ does not hold for any elements of $M$ or there is a $\varphi$-least element of $M$ satisfying $\psi(x)$.

*Proof of claim 2.* We will prove the contrapositive. Assume that $\varphi(x,y)$ defines a linear order on $M$ and suppose that $\psi(x)$ is a formula that defines a non-empty set with no $\varphi$-least element. Then we can inductively build a descending sequence by choosing the $<$-least element that is $\varphi$-less than whatever we have chosen so far, where $<$ is the standard order on $M$. $\square_{\text{claim 2}}$

Now fix two formulas (possibly with parameters) $\varphi(x,y)$ and $\psi(x,y)$ which define internal well-orderings (i.e., linear orders with no definable-with-parameters infinite descending sequences). For notational simplicity, call the linear orders defined by $\varphi(x,y)$ and $\psi(x,y)$, $(A,<)$ and $(B,<)$. (They both have universe $M$, but it's easier to state things this way.)

Let $T$ be the tree of all (internally) finite strictly order-preserving maps from $(A,<)$ to $(B,<)$. Let $\chi(x,y)$ be a formula (which we're thinking of as a relation between $A$ and $B$) that holds if and only if there exists a branch on $T$ whose domain is an initial segment of $A$ and whose range is an initial segment of $B$ and which maps $x$ to $y$. 

*Claim 3.* $\chi(x,y)$ is a maximal order isomorphism between initial segments of $A$ and $B$. That is to say, it is either an isomorphism between $A$ and $B$, a mapping of $A$ onto an initial segment of $B$, or the inverse of a mapping of $B$ onto an initial segment of $A$.

*Proof of claim 3.* (This is pretty much the standard argument for the comparability of well-orders. Mostly we're just checking that it all goes through in $\mathsf{PA}(Q_{\mathrm{Brch}})$.) First note that $\chi(x,y)$ must be a one-to-one relation. If there were some $a\in A$ and $b_0,b_1 \in B$ such that $\chi(a,b_0)$ and $\chi(a,b_1)$ both hold, then by composing some maps we'd get a definable strictly order-preserving map of an initial segment of $B$ onto a smaller initial segment $B$, whence we could get a definable infinite descending sequence of elements of $B$. (Take the first element of $B$ not in the range of this map and iterate, which we can do by Claim 2.) Also it's fairly clear that $\chi(x,y)$ must actually be a bijection between initial segments of $A$ and $B$. Assume that the domain of $\chi(x,y)$ is not all of $A$ and the range of $\chi(x,y)$ is not all of $B$. By Claim 2, there is a smallest $a \in A$ not in the domain of $\chi(x,y)$ and a smallest $b \in B$ not in the range of $\chi(x,y)$, so we can build a branch in $T$ witnessing a bijection extending $\chi(x,y)$ to $a$ and $b$, which is a contradiction. $\square_{\text{claim 3}}$.

So since we can do this for any two definable internal well-orderings in $M$, we have that well-orderings are comparable in $M_2$, so since it satisfies $\mathrm{RCA}_0$, it is a model of $\mathrm{ATR}_0$. Therefore every model of $\mathsf{PA}(Q_{\mathrm{Brch}})$ satisfies $\mathrm{Con}(\mathsf{PA})$. $\square$

  [1]: https://www.sciencedirect.com/science/article/pii/0003484378900098
  [2]: https://arxiv.org/abs/math/9301204