\section{Rigid operators and the original finite category}\label{op:section} Throughout Sections~\ref{op:section}--\ref{geo:factor-section}, the Dynkin diagram is of type ADE. Choose an orientation $Q$ and write $\overline Q$ for its doubled quiver, with reverse arrow $h^*$ for $h\in Q_1$. Let $e_i$ be the vertex idempotents and let $\Pi=\C\overline Q/(\sum_{h\in Q_1}(hh^*-h^*h))$ be the preprojective algebra. Give every doubled arrow degree $-1$, and put \[ B=\Pi[u],\qquad \deg u=2. \] The variable $u$ is central. All module maps and extension groups in these sections preserve the displayed grading unless a superscript specifies another degree. Write $Q_i=\Pi e_i$ and $J_i=D(e_i\Pi)$, where $D=\Hom_\C(-,\C)$; the generator of $Q_i$ and the socle of $J_i$ have degree zero. The shift $\langle s\rangle$ raises degrees by $s$. A finite-dimensional graded $B$-module whose restriction to $\Pi$ is projective is a pair $(P,U)$, where $P$ is graded projective over $\Pi$ and $U\in\Hom_\Pi^2(P,P)$. \subsection{An exact category of operators} Fix a height function $b=(b_i)$, so $|b_i-b_j|=1$ on each edge, and an integer $\ell\geq0$. Set \[ a_i=b_i-2\ell,\qquad \Omega=\{(i,b_i-2k):i\in I,\ 0\leq k\leq\ell\}. \] Let $\mathcal E_\Omega$ be the full subcategory of finite-dimensional graded $B$-modules whose restrictions to $\Pi$ are projective and whose graded $\Pi$-tops are supported on $\Omega$. Its conflations are short exact sequences of graded $B$-modules with terms in this subcategory. Every such sequence splits over $\Pi$. Conversely, an extension of two objects has a graded $\Pi$-splitting, so its top is the direct sum of the two endpoint tops. Consequently $\mathcal E_\Omega$ is extension closed and \begin{equation}\label{op:ext-one} \Ext^1_{\mathcal E_\Omega}(M,N)=\Ext_B^{1,0}(M,N). \end{equation} A nonzero map $Q_i\langle s\rangle\to Q_j\langle t\rangle$ is represented by a path of length $t-s$, and therefore requires $t-s\geq\operatorname{dist}(i,j)$, where $\operatorname{dist}$ is graph distance. The height inequality $|b_i-b_j|\leq\operatorname{dist}(i,j)$ gives, on the chosen parity lattice, \begin{equation}\label{op:boundaries} \Hom_\Pi^0(Q_i\langle s\rangle,M)=0\quad(s>b_i),\qquad \Hom_\Pi^0(M,Q_i\langle t\rangle)=0\quad(t0$, then $r\geq\xi_j-2\ell$. Every degree $s$ at color $i$ in $J_j\langle r+2\rangle$ therefore satisfies \[ s\geq\xi_j-2\ell+2+\operatorname{dist}(i,j) \geq\xi_i-2\ell+2. \] Thus $K$ has no support below the lower boundary of \eqref{geo:mutable-window}. Let $e$ be the dimension vector of a submodule. Suppose that some $e_{i,s}>0$ has $s>\xi_i$. Among these coordinates choose one maximizing the excess $E=s-\xi_i$, and among the coordinates with that excess choose $\xi_i$ maximal. For a neighbor $j$, a positive $e_{j,s+1}$ would have excess $E+2$ if $\xi_j=\xi_i-1$. If $\xi_j=\xi_i+1$, it would have the same excess $E$ and larger height. Both contradict the choice. Hence all $e_{j,s+1}$ vanish. The exponent of the out-of-window variable $Y_{i,s}$ in $m\prod_{j,t}A_{j,t-1}^{-e_{j,t}}$ is then \[ -e_{i,s}-e_{i,s+2}+\sum_{j\sim i}e_{j,s+1}<0. \] This monomial is deleted by the original truncation. Conversely, suppose that $e$ is supported on $W_{\ell,\xi}$. For $(i,s)$ in that set, the two same-color variables of $A_{i,s-1}$ have indices $s-2,s$ in the highest-monomial window. For a neighbor with $\xi_j=\xi_i-1$, its index $s-1$ lies between $\xi_j-2\ell$ and $\xi_j$; for a neighbor with $\xi_j=\xi_i+1$, the same inequalities follow from $\xi_i-2\ell+2\leq s\leq\xi_i$. Thus all neighboring variables also lie in that window. The monomial is retained. Applying this criterion to \eqref{geo:full-character-equation} proves \eqref{geo:finite-kernel}. \end{proof} \begin{corollary}\label{geo:bijection-and-character} The frozen-free real simple objects are in bijection with all rigid objects of the original finite category, including their decoration. For every real simple $L(m)$ in the original category, let $M_m$ be the unique rigid object with mutable degree $g(m)$. Then \begin{equation}\label{geo:all-real-character} \chi_q^{\leq\xi}L(m) =m\,F_{K(m)}\bigl(y_{i,s}=A_{i,s-1}^{-1}\bigr), \qquad K(m)\simeq\Hom_{\cC_A}(T,M_m). \end{equation} The corresponding full rigid operator has underlying module $P_w$ and satisfies \eqref{geo:full-character-equation}. \end{corollary} \begin{proof} Given a rigid categorical object $X$, take its normalized lift by Proposition~\ref{op:normalized-lift}. Theorem~\ref{geo:rigid-simple} produces the real simple $L(m)$ at exactly that column-minimum-zero dominant vector. Conversely, Theorem~\ref{ds:real-rigid} gives for a real simple the unique rigid object with its mutable degree vector and identifies its Hom module with the prescribed finite generic kernel. The two constructions are inverse: on the categorical side uniqueness is the rigidity/index statement in that theorem; on the dominant side the mutable differences and column minima determine $w$ uniquely by \eqref{op:column-differences}. This proves the frozen-free bijection. For an arbitrary real $L(m)$, put $d_i=\min_k w_{i,k}$ and remove $d_i$ full columns. By Lemma~\ref{ds:frozen}, the result is an actual simple $L(m_0)$ and $L(m)$ is its tensor product with the corresponding frozen simples. Multiplication by each frozen simple permutes the actual simple basis up to its specified quantum normalization. At $v=1$, this implies that a composition series for $L(m_0)^{\otimes2}$ remains a composition series after multiplying by the squared frozen factor, with distinct simple classes. Since $L(m)^{\otimes2}$ is simple, $L(m_0)^{\otimes2}$ has length one. Thus $L(m_0)$ is real. Lift its rigid categorical object to $K_0$, and add $\bigoplus_iF_{i^*}^{d_i}$. These are projective-injective in the same exact category, so \eqref{op:ext-one} shows that their direct sum $K$ with $K_0$ remains rigid. Its underlying projective type is exactly $P_w$. Theorems~\ref{geo:rigid-simple} and \ref{geo:full-character}, followed by Proposition \ref{geo:original-truncation}, give the asserted formulas for $L(m)$. The finite generic kernel and the Hom module are the same ones identified in Theorem~\ref{ds:real-rigid}. \end{proof} \section{Tensor factorization and indecomposability}\label{geo:factor-section} \begin{lemma}\label{geo:ext-extremum} Let $K$ be a rigid full kernel and let $M$ be a finite-dimensional graded $B$-module. Then \begin{equation}\label{geo:ext-extremum-equation} \max_{N\subset M}g_K^{\rm full}\cdot\dim N =\dim\Ext_B^{1,0}(M,K). \end{equation} \end{lemma} \begin{proof} Put $d=\dim\Hom_B^0(M,K)$ and form the evaluation map $M\to K^{\oplus d}$ using a basis of this Hom space. Let $U$ be its kernel and $V$ its image. From $0\to V\to K^{\oplus d}\to Q\to0$, rigidity and injective dimension at most one give $\Ext_B^{1,0}(V,K)=0$. Applying $\Hom_B^0(-,K)$ to $0\to U\to M\to V\to0$ now shows that $\Hom_B^0(M,K)\to\Hom_B^0(U,K)$ is surjective. Every map from $M$ to $K$ vanishes on $U$ by its definition, so $\Hom_B^0(U,K)=0$. The same long exact sequence, using $\Ext_B^{2,0}(V,K)=0$, gives $\Ext_B^{1,0}(M,K)\simeq\Ext_B^{1,0}(U,K)$. Equation \eqref{op:euler-pairing} therefore makes $U$ attain the right-hand side of \eqref{geo:ext-extremum-equation}. For any $N\subset M$, the exact sequence with quotient $M/N$ and injective dimension one make $\Ext_B^{1,0}(M,K)\to\Ext_B^{1,0}(N,K)$ surjective. Again by \eqref{op:euler-pairing}, \[ g_K^{\rm full}\cdot\dim N \leq\dim\Ext_B^{1,0}(N,K) \leq\dim\Ext_B^{1,0}(M,K). \] This is the required upper bound. \end{proof} \begin{lemma}\label{geo:vertex-coefficient} If $e$ is a vertex of the convex hull of the dimension vectors of submodules of $M$, then $M$ has exactly one submodule of dimension $e$, and $\chi(\Gr_B(e,M))=1$. \end{lemma} \begin{proof} For two such submodules $N,N'$, the equality $2e=\dim(N\cap N')+\dim(N+N')$ expresses $e$ as the midpoint of two points of that convex hull. Extremality makes both vectors equal to $e$. Thus $N=N\cap N'=N'$. The Grassmannian has one underlying point and Euler characteristic one. \end{proof} \begin{proposition}\label{geo:orthogonality} If $L(m_1)$ and $L(m_2)$ are real and their tensor product is simple, their full rigid kernels satisfy \[ \Ext_B^{1,0}(K_1,K_2)=\Ext_B^{1,0}(K_2,K_1)=0. \] The full kernel of their tensor product is $K_1\oplus K_2$. \end{proposition} \begin{proof} The pairwise-simple tensor theorem \cite[Theorem~1.1]{H10} implies that every mixed product $L(m_1)^{\otimes a}\otimes L(m_2)^{\otimes b}$ is simple and real: all pairwise factors are simple by realness and the given simple product, and the same test applies to its square. Its highest monomial is $m_1^am_2^b$. Suppose that the maximum of $g_2^{\rm full}\cdot\dim N$ for $N\subset K_1$ is positive. A linear function on the finite submodule polytope attains its maximum at a vertex $e$. Lemma~\ref{geo:vertex-coefficient} and \eqref{geo:full-character-equation} give coefficient one for the monomial of exponent $e$ in $L(m_1)$. Taking the highest monomial from $L(m_2)^{\otimes b}$ gives a positive coefficient at the same root exponent $e$ in the simple product of highest monomial $m_1m_2^b$. Corollary~\ref{geo:bijection-and-character} identifies this coefficient with a Grassmannian coefficient of its full rigid kernel. Since it is nonzero, that Grassmannian is nonempty. Proposition~\ref{op:generic-and-strict} consequently gives \[ (g_1^{\rm full}+b g_2^{\rm full})\cdot e<0\qquad(b\geq1). \] The positive coefficient of $b$ contradicts this inequality for large $b$. The maximum is therefore zero, since the zero submodule attains zero. Lemma~\ref{geo:ext-extremum} gives $\Ext_B^{1,0}(K_1,K_2)=0$. Reversing the two modules proves the other vanishing. Their direct sum is rigid and has underlying projective type $P_{w_1+w_2}$. The full kernel of the real tensor product has the same type and is rigid by Corollary \ref{geo:bijection-and-character}. Both operators lie in the open orbit of the same irreducible affine operator space, so they are isomorphic. \end{proof} \begin{lemma}\label{geo:F-product} For finite-dimensional graded $B$-modules, $F_{K_1\oplus K_2}=F_{K_1}F_{K_2}$. \end{lemma} \begin{proof} For a submodule $N\subset K_1\oplus K_2$, record $N_1=N\cap K_1$ and its image $N_2\subset K_2$. Once these are fixed, $N/N_1$ is the graph of a unique $B$-map $N_2\to K_1/N_1$. Conversely, each such map gives a submodule with that intersection and image. The fibers are therefore the affine spaces $\Hom_B^0(N_2,K_1/N_1)$. On the product of the two quiver Grassmannians, this Hom space is the kernel of a morphism of vector bundles: write a homogeneous linear map between the universal graded bundles and impose the finitely many equations of commuting with the arrows. Stratify by the ranks of those equations. On each stratum its kernel is a vector bundle, so the fibers have Euler characteristic one and Euler characteristic is multiplicative. Summing over strata and over $\dim N_1+\dim N_2=e$ gives the coefficient of $y^e$ in the claimed identity. \end{proof} \begin{proof}[Proof of Theorems~\ref{thm:rigid-real-main} and \ref{thm:character-main}] The bijection and exact character formula are Corollary \ref{geo:bijection-and-character}. It remains to identify prime objects. Suppose a frozen-free real simple factors as a tensor product of two nontrivial simples $L(m_1)$ and $L(m_2)$. Applying the only-if direction of \cite[Theorem~1.1]{H10} to the four factors in the square of this product shows that each factor is real. Column minima show that both are frozen-free: a zero entry of $w_1+w_2$ is a zero entry of each nonnegative vector. Proposition \ref{geo:orthogonality} identifies its full kernel with $K_1\oplus K_2$. Both summands are nonzero, since their highest monomials are nontrivial, and neither contains a full interval. Proposition~\ref{op:normalized-lift} then gives a nontrivial decomposition of the associated rigid categorical object. Conversely, decompose a normalized rigid lift as $K=K_1\oplus K_2$ with two nonzero summands. Each summand is rigid. Its underlying projective multiplicities are nonnegative, and the zero column minima of their sum force zero column minima for each. Theorem~\ref{geo:rigid-simple} supplies real simples $L(m_1)$ and $L(m_2)$ with $m=m_1m_2$. Their full characters and Lemma~\ref{geo:F-product} give \[ \chi_q L(m_1)\,\chi_q L(m_2)=\chi_q L(m). \] Injectivity of the $q$-character map \cite{FR99} identifies the corresponding Grothendieck classes. A tensor product with the class of one simple has length one, so $L(m_1)\otimes L(m_2)\simeq L(m)$. Both factors are nontrivial because $K_1,K_2$ are nonzero. Indecomposability of the categorical object is equivalent to that of its normalized lift by Proposition~\ref{op:normalized-lift}. Thus primality is equivalent to indecomposability, completing both theorems. \end{proof}