\section{Fixed-generator details and other time scales} \label{app:fixed-generator-details} \subsection{Proof of the local classification} Take the fixed finite irreducible row-generator experiment and notation of Theorem~\ref{thm:classification}. \begin{proof} Let $\mathcal H$ be the ordered-state skeleton of Proposition~\ref{prop:holding} and set $Z_t=\E[S_t\mid\mathcal H]$. The endpoint observation $Y=(X_0,X_t)$ is a function of $\mathcal H$, and the variance decomposition gives \begin{equation}\label{eq:skeletonsplit} \mathcal M_Q(t)=\E\Var(S_t\mid\mathcal H)+\E\Var(Z_t\mid Y). \end{equation} For a fixed $k$-jump skeleton $i=i_0,\ldots,i_k=j$, integration over its jump times gives mass \[ \pi_i\frac{t^k}{k!}\prod_{\ell=0}^{k-1}q_{i_\ell i_{\ell+1}} (1+O_Q(t)), \] and $Z_t=k+O_Q(t)$. Both expansions are uniform among skeletons with a fixed bounded length. Indeed the holding factor lies between $e^{-a_{\max}t}$ and $1$, and $0\le\int_0^t a(X_u)\,du\le a_{\max}t$. Truncate at $K_t\le2$. The only different-length skeletons with the same endpoints have lengths $0,2$ on a return or $1,2$ on a shortcut. The pairwise variance formula gives the coefficient \eqref{eq:A2}. Same-length one-jump skeletons with fixed endpoints are unique, and a same-length two-jump discrepancy in $Z_t$ is $O_Q(t)$, contributing only $O_Q(t^4)$. Poisson domination by $\Pois(a_{\max}t)$ bounds the squared risk on omitted skeletons by $O_Q(t^3)$. The first term of \eqref{eq:skeletonsplit} is $O_Q(t^3)$ by Proposition~\ref{prop:holding}; this proves \eqref{eq:Mtwo}. If $A_2=0$, there is at most one supported length below three at each endpoint pair. Truncate instead at $K_t\le3$. Every different-length pair now contributing at order $t^3$ has lengths $m_{ij}<3$ and $3$. Their masses sum to $\pi_i(B^3)_{ij}t^3/6$, and their score differences have leading value $3-m_{ij}$. Thus the second term of \eqref{eq:skeletonsplit} is $A_3t^3+O_Q(t^4)$. Same-length differences contribute $O_Q(t^4)$ or less; the omitted squared risk is $O_Q(t^4)$. Restoring the first term supplies $H t^3$ and proves \eqref{eq:Mthree}. These risk bounds follow by using the retained conditional predictor on supported fibres and zero on unsupported fibres; its absolute value is bounded by $3+a_{\max}t$. Positivity of all stationary masses gives $A_2>0$ exactly when $h=2$, and, conditional on $A_2=0$, $A_3>0$ exactly when $h=3$. If rates are not constant, some allowed edge joins two states of different escape rate; otherwise irreducibility would make every $a_i$ equal. Hence $H>0$. Divide the missing information by $\bar a t$ to obtain the asserted powers. At constant escape the score is a centered Poisson count, so the previous graph alias theorem applies without a holding-time term. \end{proof}