\section{Introduction} Dissipation inference from a partially observed trajectory depends on the class of models used to fit it. Minimizing entropy production over a class containing the microscopic model gives a lower bound on its dissipation. A hidden-state budget may exclude that model and raise the fitted minimum. The question is then whether the restricted class preserves the unrestricted dissipation floor. We answer this question for first excursion-duration moments with two visible states. The comparison concerns compatible models; a physically faithful coarse-graining has additional closure requirements \cite[Sections III.A--III.B]{EspositoCoarse2012}. The observed stationary label trace distinguishes two visible states and combines all hidden states into one instantaneous, reversal-even label. Its underlying model is a finite irreducible reciprocal continuous-time Markov chain. A hidden excursion runs from a departure into the hidden set to the next visible visit, including returns and excluding visible holding time. The data $d=(P,F,R,L)$ comprise stationary masses, observed flows, excursion occurrence rates and occurrence-weighted first duration moments. We compare the entropy infimum $I_2(d)$ over compatible models with at most two hidden states with $I_\infty(d)$ over all finite hidden dimensions. A nonempty reduced class is \emph{safe} when these infima agree. Nitzan--Ghosal--Bisker formulate the corresponding first-moment fitting problem \cite[Eq.~(17), Section III.B.1 and Appendix C]{Nitzan2023}; their simpler-topology experiments leave the general lower-bound validity of such fits open \cite[Section III.D and Conclusion]{Nitzan2023}. Skinner--Dunkel optimize general asymmetric two-step counts by hidden-edge rerouting and finite channel reduction \cite[p.~3; author-manuscript Appendix, Eqs.~(A12)--(A27)]{Skinner2021}, and give a closed-form symmetric negative-determinant optimum \cite[author-manuscript Appendix, Eqs.~(A28)--(A29)]{Skinner2021}. The new ingredients here are the explicit count formula for arbitrary asymmetric data together with exact restoration of prescribed first moments, and the resulting safety classification, finite attainment criterion and non-testability result. \paragraph{Results.} Section~\ref{sec:family} gives an explicit example: a true chain with three hidden states whose data admit exact two-hidden-state fits, all of which have entropy production above $7/100$ and more than sixty times the true value, which tends to zero as a parameter $\varepsilon\downarrow0$ (Theorem~\ref{thm:family-overestimate}). A computer-assisted step in exact arithmetic raises these bounds to $7/5$ and $1200$ times the true value (Theorem~\ref{thm:family-sharp}). The reduced minimum therefore overestimates the true entropy production, not only the unrestricted fitting floor. The remaining sections explain this phenomenon. Theorem~\ref{thm:H1} and Theorem~\ref{thm:all-infimum} give $I_\infty(d)=I_4(d)=\mathcal B(d)$ for every strictly positive consistent datum, where the sharp count functional $\mathcal B$ is defined in \eqref{eq:sharp-count-functional}. Thus exact first duration moments do not raise the unrestricted count infimum. Four hidden states suffice to approach it at exactly the prescribed data with uniformly bounded rates: rare excursions carry the required moments while their occurrence rates vanish. Theorem~\ref{thm:finite-attainment-all} classifies finite attainment in every determinant regime, and Corollary~\ref{cor:all-safety} classifies safety of nonempty two-hidden-state classes. For positive determinant, safety is equivalent to equal conditional cross-excursion means (Theorem~\ref{thm:H2}); unequal means force the quantitative bias of Theorem~\ref{thm:H3}. Rank-one data can be safe without a finite minimizer (Theorem~\ref{thm:rank-one}). Finally, Proposition~\ref{prop:fixed-bias-non-testability} fixes safe and positively biased unsafe data that share $P,F,R$, yet makes their stationary label-path laws arbitrarily close on any prescribed finite interval. Four hidden states and a common upper rate bound suffice. Rare long excursions carry the distinguishing moments, a constrained Markov realization of the classical rare-contamination mechanism in mean testing \cite{BahadurSavage1956}. \paragraph{Scope.} All results concern the stated positive-data, two-visible-state class and compatible fits. The entropy $\sigma$ is stationary path irreversibility; identifying it with total physical dissipation requires thermodynamically resolved transitions obeying local detailed balance. Safety is a statement about infima: it does not say that every individual fit has the same dissipation, identifies the true hidden network, or covers incorrect hidden topologies \cite[Section V]{Igoshin2025}. Four hidden states are a sufficient budget, not shown to be minimal, and the gap constant in Theorem~\ref{thm:H3} is not claimed to be optimal. Positive-determinant, rank-one and negative-determinant data are treated separately. \paragraph{Related work.} The entropy tools build on path irreversibility and coarse-grained relative entropy \cite{Roldan2010,Roldan2012,Seifert2012}, trajectory snippets \cite[Sections II.C--II.E]{Degunther2024}, and the cross-excursion waiting-time bound \cite[Eq.~(11)]{Martinez2019}. Waiting-time inference \cite{SkinnerWaiting2021} and sequence-versus-duration comparisons \cite[Sections II.B--II.C]{Kapustin2024} use observation protocols distinct from fixed first moments. Ehrich's reconstruction fixes the hidden-state count \cite[Section 4]{Ehrich2021}. Phase-type resolvents and realization theory provide classical algebraic ingredients \cite{TelekHorvath2007,BuchholzTelek2013,HorvathTelek2015}. Fixing moments also differs from preserving memory \cite{Teza2020}, resolving transitions \cite{vanderMeer2022}, or controlling hidden pumps \cite{Esposito2015}. Recent preprints study attainment with the entire stationary path law fixed \cite[Section 2 and Theorems A--B]{Aznagulov2026} and full excursion-kernel reconstruction under known event anchors and closure \cite[Eqs.~(4)--(5) and End Matter]{GuExcursion2026}; neither observation contract determines the fixed-first-moment infimum studied here.