\section{First-moment realization details}\label{app:first-moment-details} \subsection{Attainment at fixed nonsingular data} \begin{proof}[Proof of Proposition~\ref{thm:H4}] Every hidden rate is at most \(\tau\), since each hidden total departure rate is a positive diagonal entry of \(K\) and their sum is \(\tau\). Visible-to-hidden rates are bounded by \(k_i/P_i\), where \(k_i=\sum_jR_{ij}\) is fixed. Direct rates are fixed by \(F,P\). Thus all rates lie in a fixed compact box. The hidden masses have a uniform positive lower bound. Let \(k_{\max}=\max(k_1,k_2)\). Since each row of \(C\) has sum at most \(\tau\), \(|\det C|\le\tau^2\). For either hidden state \(h\), stationarity gives \(\sum_i A_{ih}\le\pi_h K_{hh}\le\pi_h\tau\). Bounding the other column entries of \(A\) by \(k_{\max}\) gives \(|\det A|\le\pi_h\tau k_{\max}\). From \(\det A\det C=\det R\det K\) we obtain \begin{equation}\label{eq:H14} \pi_h\ge\frac{|\det R|\Delta}{\tau^3 k_{\max}}>0. \end{equation} Take a minimizing sequence of finite-entropy fits and a convergent subsequence of their rates and stationary masses. Equations \eqref{eq:H13} keep \(K\) uniformly nonsingular, so the excursion data remain equal to \(R,L\) in the limit. The limiting stationary masses remain positive by \eqref{eq:H14}. Extend \((x-y)\log(x/y)\) to the nonnegative quadrant by assigning zero at \((0,0)\) and \(+\infty\) when exactly one argument is positive. This function is lower semicontinuous. A finite-entropy sublevel limit therefore has reciprocal support. It is irreducible: the direct visible edge connects the visible states, and a hidden component disconnected from them would make the killed block \(K\) singular. Entropy lower semicontinuity now gives a minimizing admissible model. This proves the proposition. Notice that a general nonclosed reciprocal-support class is \emph{not} being declared compact; compactness is used together with finite entropy and the nonsingular-data constraints. \end{proof} \subsection{Equal means admit a saturating two-state realization}\label{equal-means-admit-a-saturating-two-state-realization} Assume now \(\mathcal C_2(d)\ne\varnothing\), \(\det R>0\) and \(\Delta m=0\). Set \[c=(R_{12}/R_{21},1)^T,\qquad D_c=\operatorname{diag}(c),\qquad G=D_c^{-1}R,\quad H=D_c^{-1}L.\] The cross-mean condition makes both \(G,H\) symmetric. Their diagonal entries are positive; \(\det G>0\), and \(\det H>0\) because \(\det L=\det R/\det K>0\). Consequently both are positive definite. Their unique symmetric positive square roots define \[S=H^{-1/2}GH^{-1/2},\qquad Z=H^{-1/2}G.\] Both \(S\) and \(H\) are positive definite and \(Z\) is invertible. The forced symmetrised Laplace kernel is \begin{equation}\label{eq:H17} D_c^{-1}\mathcal K(s)=G(G+sH)^{-1}G=Z^T(S+sI)^{-1}Z. \end{equation} Its time-density is \(\mathcal F(t)=Z^Te^{-St}Z\). Because \eqref{eq:H17} agrees with the kernel of an actual feasible model, uniqueness of the Laplace transform of finite exponential matrix functions makes \(\mathcal F(t)\) entrywise nonnegative. In particular \(\mathcal F_{12}(0)\ge0\). We need the following elementary two-dimensional realization lemma. \begin{lemma}[Positive two-mode rotation]\label{lem:rotation} If \(S\) is a \(2\times2\) positive-definite symmetric matrix, \(Z\) is invertible, and \([Z^Te^{-St}Z]_{12}\ge0\) for every \(t\ge0\), there is an orthogonal change of hidden coordinates giving matrices \(M,B\) with \begin{equation}\label{eq:H18} M>0,\quad M_{12}\le0,\quad B\ge0\text{ entrywise},\qquad B^T(M+sI)^{-1}B=Z^T(S+sI)^{-1}Z. \end{equation} \end{lemma} \begin{proof} Diagonalise \(S\) with eigenvalues \(0<\lambda_1\le\lambda_2\). For distinct eigenvalues write the transformed rows of \(Z\) as \(u,v\). Positivity at large \(t\) implies \(u_1u_2\ge0\), and at zero implies \(u_1u_2+v_1v_2\ge0\). Choose the sign of \(u\) so \(u\ge0\). If \(u_1u_2=0\), the zero-time inequality gives \(v_1v_2\ge0\), so choose its sign to have \(v\ge0\); the diagonal coordinates already work. The same applies when \(u>0\) and \(v_1v_2\ge0\). In the remaining case choose the sign of \(v\) so \(v_1<0