This question is motivated by Vaananen's paper Generalized quantifiers in models of set theory. Say that a (set-sized, regular) logic $\mathcal{L}$ is
unbounded if there are $\mathcal{L}$-sentences $\varphi,\psi$ each with arbitrarily large models but with disjoint classes of cardinalities of models, and
strongly unbounded if there is an $\mathcal{L}$-sentence $\theta$ in a language containing a unary relation symbol $A$ and a binary relation symbol $<$ such that $(i)$ whenever $\mathcal{M}\models\theta$ the subreduct $(A,<)^\mathcal{M}$ is a well-ordering and $(ii)$ there is no bound on the lengths of well-orderings of the form $(A,<)^\mathcal{M}$ for $\mathcal{M}\models\varphi$.
It turns out that strongly unbounded logics are indeed unbounded. The proof is a simple consequence of the nonexistence of a descending sequence of cardinalities. A linear order $I$ is a well-order iff there is a family of sets $X_i$ ($i\in I$) of sets such that $i<_Ij \implies \vert X_i\vert<\vert X_j\vert$, and so if $\mathcal{L}$ is strongly unbounded via $\varphi,\psi$ we can take as our $\theta$ an appropriately-constructed sentence describing a linear order $(A,<)$ with a collection of pairs of sets $X_a,Y_a$ for each $a\in A$ such that $X_a$ is (the domain of) a model of $\varphi$ and $Y_a$ is (the domain of) a model of $\psi$, and if $a\le_Ab$ then there are injections from $X_a$ to $Y_a$ and from $Y_a$ to $X_b$.
I'm curious whether this trick is avoidable. Or, put another way:
Question 1: Over $\mathsf{ZF+DC}$, does "Every strongly unbounded logic is unbounded" imply "There is no descending sequence of cardinalities"?
(See Asaf Karagila's comment below for why I've included Dependent Choice here. Meanwhile, note that it is currently open whether "There is no descending sequence of cardinalities" is equivalent to $\mathsf{AC}$, see Howard/Tachtsis.)