\section{A shared entropy lower bound}\label{a-shared-entropy-lower-bound} For a specified microscopic hidden excursion, its stationary occurrence-rate density divided by the density of its reversed excursion is the product of stationary flow ratios along its jumps. Holding-time factors cancel on reversal and stationary masses telescope. Summing the logarithmic ratio over all excursion paths recovers entropy production on edges incident with the hidden set. Direct visible jumps supply the first term of \eqref{eq:H1}. For each fixed finite irreducible chain these sums are integrable: hidden exit times have finite moments and there are finitely many nonzero rate logarithms. Log-sum after forgetting hidden paths and durations yields \(\sigma(Q)\ge B(d)\). This is the standard path-coarse-graining bound; see \cite{Roldan2010,Skinner2021,Nitzan2023}. The same $B(d)$ is the informed partial entropy-production rate of the marginal-observer framework \cite{PolettiniEsposito2017,Bisker2017}. Deleting only the direct visible pair gives a stalled stationary ratio $p_2^{\mathrm{st}}/p_1^{\mathrm{st}}=R_{12}P_2/(R_{21}P_1)$ by hidden-block elimination. With $J=F_{12}-F_{21}=R_{21}-R_{12}$, its rate is $J\log(F_{12}R_{21}/(F_{21}R_{12}))=B(d)$. The excursion data determine this ratio without requiring an experimental stalling intervention. Keeping the duration gives a useful stronger version. Let \(p=f_{12}/R_{12}\) and \(q=f_{21}/R_{21}\) be the conditional duration densities. Then \begin{equation}\label{eq:H5} \sigma(Q)-B(d)\ge R_{12}D(p\Vert q)+R_{21}D(q\Vert p). \end{equation} For this observation rule, the added duration term is the cross-excursion waiting-time divergence of \cite[Eq.~(11)]{Martinez2019}; optimization over compatible state budgets is a separate question. Indeed the cross-excursion term decomposes exactly into the count term and the two weighted relative entropies. Contributions from return-path irreversibility, which were discarded, are nonnegative. This does not require that fitted and true models have the same full path law. \section{Exact first moments with arbitrarily little extra dissipation}\label{exact-first-moments-with-arbitrarily-little-extra-dissipation} We first record an elementary channel construction. Suppose \(M=xy^T\) is a nonnegative nonzero rank-one \(2\times2\) matrix, \(y^T\mathbf1=1\), and \(x_i>0\) iff \(y_i>0\). Write \(s=\mathbf1^Tx\). Introduce one hidden state with mean residence time \(t>0\), stationary mass \(st\), entrance flows \(n_{ig}=x_i\), and exit flows \(n_{gi}=sy_i\). Its excursion matrices are exactly \(M,tM\). Its hidden-incident entropy contribution is \begin{equation}\label{eq:H6} (M_{12}-M_{21})\log(M_{12}/M_{21}), \end{equation} when the cross entries are positive, and zero for a diagonal-only channel. This follows since \(x_1-sy_1=M_{12}-M_{21}\) and the current at visible state 2 is its negative. A collection of these channels fits the data whenever its occurrence and moment matrices sum to \(R,L\): hidden balance holds separately, the consistency identities supply visible balance, and the total hidden mass is \(\mathbf1^TL\mathbf1\). Choose once and for all \[0<\delta<\min_{i,j} L_{ij}/R_{ij}.\] For a small positive rate \(\varepsilon\) with \(\delta\varepsilon<1\), define \begin{equation}\label{eq:H7} R_f(\varepsilon)=\frac{R-\varepsilon L}{1-\delta\varepsilon},\qquad S(\varepsilon)=\frac{L-\delta R}{1-\delta\varepsilon}. \end{equation} For all sufficiently small \(\varepsilon\), \(R_f\) is entrywise positive with positive determinant and \(S\) is entrywise positive. The exact identities are \begin{equation}\label{eq:H8} R_f+\varepsilon S=R,\qquad \delta R_f+S=L. \end{equation} Write \(R_f=\begin{pmatrix}A&B\\C&D\end{pmatrix}\). Its following two-channel factorisation is rational whenever its entries are rational: \begin{equation}\label{eq:H9} M_{f,1}=\begin{pmatrix}A&B\\C&BC/A\end{pmatrix},\qquad M_{f,2}=\begin{pmatrix}0&0\\0&D-BC/A\end{pmatrix}. \end{equation} The first matrix has reciprocal full support and the second only a visible-2 return. Give both channels residence time \(\delta\). Their entropy is exactly \((B-C)\log(B/C)\). The positive matrix \(S\) has a two-channel factorisation with full reciprocal support. To see this, select a fixed \(\eta\in(0,1/2)\) less than every entry of each row of \(L-\delta R\) divided by that row sum. Put \[Y=\begin{pmatrix}1-\eta&\eta\\\eta&1-\eta\end{pmatrix},\qquad X=SY^{-1}.\] Every entry of \(X\) is positive, since its row entries are \((S_{ij}-\eta\sum_jS_{ij})/(1-2\eta)\). If \(S_g=X_{\cdot g}Y_{g\cdot}\), then \(S=S_1+S_2\) and both \(S_g\) are positive rank one. Give occurrence matrices \(\varepsilon S_1,\varepsilon S_2\) residence time \(1/\varepsilon\). The four channels fit \eqref{eq:H8} exactly. Each connects to at least one visible state; the direct visible edge and the positive first channel ensure irreducibility. Every nonzero transition has its reverse. All stationary masses are positive and sum to one. The hidden-incident entropy is \begin{equation}\label{eq:H10} (R_{f,12}-R_{f,21})\log(R_{f,12}/R_{f,21}) +\varepsilon\sum_{g=1}^2(S_{g,12}-S_{g,21})\log(S_{g,12}/S_{g,21}). \end{equation} The first term tends to the cross-count term of \eqref{eq:H1}; the second tends to zero. The channel matrices are continuous and have the required positive limits, so the excess is \(O(\varepsilon)\). The direct term is fixed. The lower bound in Section~\ref{a-shared-entropy-lower-bound} proves \eqref{eq:H2}. The fast departure rates are \(1/\delta\), the slow departure rates are \(\varepsilon\), and visible departure rates are fixed by \(F,P\). Hence all rates are bounded above uniformly. The slow stationary masses remain positive while their entrances and exits vanish. The endpoint is reducible and its excursion first moments are not obtained by naively substituting \(\varepsilon=0\). No invalid interchange of an unbounded duration moment and a parameter limit is made. \section{Nonsingular two-hidden-state fits cannot escape to infinity}\label{nonsingular-two-hidden-state-fits-cannot-escape-to-infinity} For any nonempty hidden set, irreducibility implies eventual exit: a supported path to the visible set gives a common positive exit probability in a bounded time from every hidden state. Iteration makes the killed survival tail exponential. Thus its negative killed generator is invertible, and integrating the killed semigroup gives the excursion identities below. In particular, $f(t)=Ae^{-Kt}C$, $\int_0^\infty e^{-Kt}dt=K^{-1}$ and $\int_0^\infty te^{-Kt}dt=K^{-2}$. These are the standard phase-type resolvent and moment identities \cite[Section 2]{TelekHorvath2007}, here with stationary entrance flows and endpoint-specific exits. The square-factorization consequence is a continuous-time relative of the reconstruction principle in \cite{Ehrich2021}. For a two-hidden-state model, let \[A=\operatorname{diag}(P_1,P_2)Q_{VH},\quad C=Q_{HV},\quad K=-Q_{HH}.\] The hidden block is transient, \(K\) is invertible, and \begin{equation}\label{eq:H11} R=AK^{-1}C,\qquad L=AK^{-2}C. \end{equation} These are the marked phase-type resolvent and moment identities, with stationary entrance rates in $A$ rather than a normalized initial law \cite[Section 3.1, Eqs.~(17), (19)]{Lindqvist2023}. The nonsingular square-matrix consequences needed here follow directly below. Nonsingularity of \(R\) forces \(A,C\) invertible and consequently \(L\) invertible. Straight multiplication gives \begin{equation}\label{eq:H12} R(R+sL)^{-1}R=A(K+sI)^{-1}C, \end{equation} and \begin{equation}\label{eq:H13} \operatorname{tr}K=\operatorname{tr}(RL^{-1})=:\tau, \qquad\det K=\det R/\det L=:\Delta. \end{equation} Both \(\tau,\Delta\) are positive. The two hidden eigenvalues are real and positive: the discriminant is \((K_{11}-K_{22})^2+4q_{g_1g_2}q_{g_2g_1}\ge0\), and trace and determinant are positive. \begin{proposition}[Attainment at fixed nonsingular data]\label{thm:H4} If \(R\) is nonsingular and a reciprocal irreducible two-hidden-state fit exists with positive visible probabilities and positive direct visible flows, its entropy-production infimum is attained. This proposition does not require \(\det R>0\). \end{proposition} Proofs of Proposition~\ref{thm:H4} and the sufficient direction of Theorem~\ref{thm:H2} are given in Appendix~\ref{app:first-moment-details}. \section{A computable strict gap from unequal cross means}\label{a-computable-strict-gap-from-unequal-cross-means} Apply \eqref{eq:H5} and the randomized map that, conditional on a duration \(t\), outputs 1 with probability \(e^{-st}\). Log-sum gives the Bernoulli relative-entropy lower bounds. For fixed \(q\in(0,1)\), the second derivative of binary relative entropy in its first argument is \(1/[p(1-p)]\ge4\), and its value and first derivative vanish at \(p=q\). Hence each relative entropy is at least twice the squared difference of the corresponding Laplace expectations. Thus \begin{equation}\label{eq:H15} \sigma(Q)-B(d)\ge2(R_{12}+R_{21})\left( \frac{\mathcal K_{12}(s)}{R_{12}}- \frac{\mathcal K_{21}(s)}{R_{21}}\right)^2. \end{equation} For a two-hidden-state fit, \eqref{eq:H12} fixes \(\mathcal K(s)\). The \(2\times2\) adjugate is linear in the entries; therefore \[\mathcal K(s)=R-s\, \frac{(\det R)L+s(\det L)R}{\det(R+sL)}.\] Dividing its cross entries by their occurrence rates gives exactly \begin{equation}\label{eq:H16} \frac{\mathcal K_{12}(s)}{R_{12}}- \frac{\mathcal K_{21}(s)}{R_{21}} =-\frac{s\Delta m}{1+as+bs^2}. \end{equation} Here \(a=\operatorname{tr}K^{-1}>0\) and \(b=1/\det K>0\), by \eqref{eq:H11}. The derivative of \(s/(1+as+bs^2)\) has sign \(1-bs^2\); its maximum over \(s>0\) is \((a+2\sqrt b)^{-1}\) at \(s=b^{-1/2}\). Equations \eqref{eq:H15}, \eqref{eq:H16} and Theorem~\ref{thm:H1} prove Theorem~\ref{thm:H3} and the necessary direction of Theorem~\ref{thm:H2}.