\section{Gluing the conjugate residue components} \label{sec:conjugate-gluing} We use the actual transpose pencil from the preceding reduction. To make the hypotheses of this last gluing step explicit, let $V=E\oplus K$ be an orthogonal decomposition with $\dim E=\dim K=8$, and let $\iota:\R^8\to K$ be an isometry. The coefficient space is another copy $A$ of $\R^8$. We have a linear pencil \[ M(a):\R^8\longrightarrow V,\qquad C(a)=M(a)M(a)^T, \qquad a\in A, \] of rank eight for every nonzero real $a$. There are a fixed nonzero vector $e\in\R^8$ and an isomorphism $L:A\to E$ such that \begin{equation}\label{eq:cg-actual} M(a)e=-La,\qquad \iota^T M(a)\text{ is skew-symmetric}. \end{equation} The map $L$ need not be an isometry: a change of coefficient coordinates may have been used to normalize a positive quadratic form. \subsection{A determinant obstruction on six coefficient directions} \begin{lemma}\label{lem:cg-two-losses} Let $W\subset A$ have dimension six, and let $q_W$ be positive definite on $W$. There do not exist constant real symmetric operators $D,H$ on $V$ and a nonzero real quadratic form $s$ on $W$ such that \begin{equation}\label{eq:cg-two-forms} C(x)=q_W(x)D+s(x)H\quad(x\in W), \end{equation} where $\rank s\le2$ and $H$ has rank two and signature $(1,1)$. \end{lemma} \begin{proof} The range of $s$ on the ellipsoid $q_W=1$ is a nondegenerate closed interval, since $s$ and $q_W$ are independent. Every matrix $D+tH$ on this interval is positive semidefinite of rank eight. For positive semidefinite matrices, the kernel of a strict convex combination is the intersection of their kernels. Equality of the ranks therefore shows that these matrices have a common kernel, including at the endpoints. Thus $M(x)$ has a fixed image $F$ of dimension eight for every real $x\in W\setminus\{0\}$. The radial identity in \eqref{eq:cg-actual} gives $L(W)\subset F\cap E$. Consequently $R=\iota^T F$ has dimension at most two. For every such $x$, the skew operator $\iota^T M(x)$ has image exactly $R$. If $R$ were nonzero, its dimension would be two. Every skew operator with image in this fixed two-plane is a scalar multiple of a fixed operator. A linear map from the six-dimensional space $W$ to this one-dimensional space has a nonzero kernel, contradicting its constant positive rank. Hence $R=0$ and $F=E$. Independence of $q_W,s$ now implies that $D,H$ are supported on $E$. The radical of $s$ has dimension at least four. At a nonzero vector in that radical, \eqref{eq:cg-two-forms} shows that $D|_E$ is positive definite. The two nonzero eigenvalues of $D^{-1/2}HD^{-1/2}|_E$ are therefore $\alpha>0>\beta$. Taking determinants on $E$ gives \begin{equation}\label{eq:cg-determinant} \det(C(x)|_E)=\det(D|_E)q_W(x)^6 \bigl(q_W(x)+\alpha s(x)\bigr) \bigl(q_W(x)+\beta s(x)\bigr). \end{equation} Both last quadratic factors are positive definite: each is positive on the radical of $s$, and neither vanishes on the connected set $W\setminus0$, because their product is positive there. They are irreducible real quadratics of rank six. Independence of $q_W,s$ and $\alpha\ne\beta$ makes the three displayed quadratic factors pairwise nonassociate. The last two thus occur to odd multiplicity. On the other hand, in orthonormal coordinates on $E$, the matrix $M(x)$ is square, so $\det(C(x)|_E)=(\det M(x))^2$. This contradicts unique factorization in the real polynomial ring on $W$. \end{proof} \subsection{The real balanced form} Split $A=W\oplus Z$, where $\dim W=6$, $\dim Z=2$, and write \[ q(x,z)=q_W(x)+q_Z(z)>0\quad\text{for }(x,z)\ne0. \] Let $\rho$ be a definite quadratic form of rank two on $Z$. Over $\mathbb C$ it is a nonzero scalar multiple of $\ell\bar\ell$ for independent conjugate linear forms. Set \[ U_+=\ker\ell,\quad U_-=\ker\bar\ell,\qquad A_\pm=V(q)\cap\mathbb P(U_\pm),\qquad \Sigma=V(q_W)\subset\mathbb P(W_{\mathbb C}). \] Each component quadric has rank at least six. Its homogeneous coordinate ring, and that of $\Sigma$, are factorial. For the latter, for example, write $q_W=x_1x_2+x_3x_4+x_5x_6$. The element $x_1$ is prime, and inverting it gives a Laurent polynomial ring. Nagata's criterion applies \cite[Theorem~6.3 and its Corollary, pp.~17--18]{Samuel1964}; alternatively, \cite[Theorem~8.2, p.~26]{Samuel1964} applies directly. The same argument allows the additional transverse variable on either component. The constant-column exclusion and homogeneous factorization in Propositions~\ref{prop:m6-constant-column} and~\ref{prop:m6-linear-extraction} leave two complex linear columns on a component. The next proposition is the precise gluing statement used here. \begin{proposition}\label{prop:cg-real-gluing} Assume that on $A_+$ there are complex linear columns $p,r$ with \begin{equation}\label{eq:cg-component} C=pr^T+rp^T. \end{equation} Assume that $C$ has generic rank two on $\Sigma$ and that its induced binary pairing there is hyperbolic over $\R(\Sigma)$. Then there exist real linear columns $P,R$ on $A$ and constant real symmetric operators $D_0,D_1$ such that \begin{equation}\label{eq:cg-glued} C=qD_0+\rho D_1+PR^T+RP^T. \end{equation} \end{proposition} \begin{proof} Write $p_0=p|_W$, $r_0=r|_W$. They are nonzero and independent over $\mathbb C(\Sigma)$, since $C$ has rank two there. Their lines are the two isotropic lines of the induced binary pairing. Hyperbolicity over $\R(\Sigma)$ makes each line individually invariant under conjugation. Let $S_0=\mathbb C[W]/(q_W)$. Suppose first that $p_0$ is primitive over the factorial ring $S_0$. Its conjugate is primitive and spans the same fraction-field line. Their proportionality factor and its inverse have nonnegative valuation at every height-one prime, so that factor is a unit of $S_0$, hence a nonzero constant. Conjugating twice shows that its modulus is one. Rescale $p$ by a constant and $r$ inversely so that $p_0$ is real. Reality of $C$ on $\Sigma$ gives \[ p_0(r_0-\bar r_0)^T+(r_0-\bar r_0)p_0^T=0. \] For a nonzero vector $u$ over a field of characteristic different from two, the map $v\mapsto uv^T+vu^T$ is injective, as follows by using a nonzero coordinate of $u$, its diagonal entry, and its row. Thus $r_0=\bar r_0$ in $S_0$. Since the ideal $(q_W)$ has no degree-one part, both restrictions are real linear columns on $W$. The argument is symmetric if $r_0$ is primitive. It remains to exclude the case where neither restriction is primitive. Homogeneity then gives \[ p_0=f c,\qquad r_0=g d, \] with nonzero complex linear forms $f,g$ and independent constant vectors $c,d$. Conjugation invariance of each constant line permits us to take $c,d$ real by rescaling $f,g$. Put $H=cd^T+dc^T$; this real symmetric matrix has rank two and signature $(1,1)$. Reality of $C$ on $\Sigma$ implies \[ fg-\overline{fg}\in\mathbb C q_W. \] The left side has rank at most four, while $q_W$ has rank six. It follows that $s=fg$ is a real polynomial, not merely a real class modulo $q_W$. It is nonzero and has rank at most two. Lifting the quadratic identity on $\Sigma$ gives \[ C(x,0)=q_W(x)D+s(x)H \] with $D$ constant real symmetric. This contradicts Lemma~\ref{lem:cg-two-losses}. At least one restriction is therefore primitive, and the preceding paragraph makes both restrictions real. Choose a nonzero transverse vector $\zeta\in U_+\cap Z_{\mathbb C}$. There is a unique real linear column $P$ with \[ P|_W=p_0,\qquad P(\zeta)=p(\zeta),\qquad P(\bar\zeta)=\overline{p(\zeta)}. \] Indeed $\zeta,\bar\zeta$ are independent and the prescriptions commute with conjugation. Define $R$ similarly. These columns restrict to $p,r$ on $U_+$ and to their conjugates on $U_-$. Hence the quadratic matrix $C-PR^T-RP^T$ vanishes on both component quadrics. In the factorial ring $\mathbb C[A]/(q)$, the distinct prime hyperplane elements $\ell,\bar\ell$ have intersection of principal ideals equal to their product. The reduced union consequently has quadratic ideal $\operatorname{span}_{\mathbb C}\{q,\rho\}$. Reality and symmetry give \eqref{eq:cg-glued} with real symmetric coefficient matrices. \end{proof} In the actual transpose-pencil application both hypotheses hold. By Lemma~\ref{lem:cp-complex-ranks}, $\rank C=2\rank M-8$ is even at every point of $\Sigma\subset A_+$ and is at most two. Since $\Sigma$ is irreducible, either $C$ vanishes identically on $\Sigma$, the case of Proposition~\ref{prop:m6-zero-intersection}, or $C$ has generic rank two on $\Sigma$. In the latter case the hyperbolicity hypothesis has already been supplied by the kernel identity. At the generic point of $\Sigma$, the space $K_0=\ker M$ is a totally isotropic three-space for the standard eight-dimensional form. The induced binary pairing is the nondegenerate quotient $K_0^\perp/K_0$. Removing three hyperbolic planes leaves determinant square class $-1$, which makes the binary quotient hyperbolic over $\R(\Sigma)$. This argument fixes the two isotropic lines individually; it does not assume that a chosen pair of polynomial generators is already real. Finally, writing \[ u=(P+R)/\sqrt2,\qquad v=(P-R)/\sqrt2 \] turns the last term of \eqref{eq:cg-glued} into $uu^T-vv^T$. Since $\operatorname{span}\{q,\rho\}=\operatorname{span}\{q,h\}$, this is the real balanced form required by Proposition~\ref{prop:balanced-alignment}; its hypothesis \eqref{eq:balanced-nonsingular} is the injectivity of $N_g$ in Lemma~\ref{lem:cp-L20}, which a real change of coefficient coordinates preserves. For this conjugate configuration the sixfold relative eigenvalue is extreme, so Proposition~\ref{prop:balanced-spectral-applications} supplies the strict interval $\lambda_2<\lambda_7$.