\section{Population-comparison proofs}\label{app:population-proofs} \subsection{Bounds for the limiting compound-Poisson experiment} The following is the proof of Lemma~\ref{lem:Fbounds}. \begin{proof} The strict upper inequality follows because $D=0$ leaves both $K=0$ and $K=2$ possible. For the lower bound, $\Cov(K,D^2)=\lambda$ and $\Var(D^2)=\lambda+2\lambda^2$; projection onto the centered variable $D^2$ gives the result and strict positivity. For completeness we prove the coarse upper bound. Conditional on $|D|=d$, $K=d+2Z$ with weights proportional to $(\lambda^2/4)^j/[j!(j+d)!]$. Put $a(j)=\lambda^2/[4(j+d+1)]$. Index shifting gives $\E[Zf(Z)]=\E[a(Z)f(Z+1)]$. Since $a$ is decreasing, \[ \Var Z=\Cov(a(Z),Z)+\E a(Z)\le\E Z. \] The same weights give $\E[Z(Z+d)]=\lambda^2/4$, so $\E[K^2\mid |D|=d]=\lambda^2+d^2$. With $\mu_d=\E[K\mid |D|=d]$, Jensen's inequality and $\Var K\le2(\mu_d-d)$ imply \[ \sqrt{\lambda^2+(d+1)^2}-1\le\mu_d\le\sqrt{\lambda^2+d^2}. \] Thus $s(d)=\mu_d-\lambda$ satisfies $-1\le s(d)\le d^2/(2\lambda)$. Using $\E D^4=\lambda+3\lambda^2$ gives $\E s(D)^2\le7/4+1/(4\lambda)\le2$ for $\lambda\ge1$. Finally, $|D|$ recovers $K$ whenever $K\le1$, whence $1-F(\lambda)\le\lambda+1-e^{-\lambda}\le2\lambda$. Continuity follows by truncating the complete score at a fixed $K$ cutoff: on a compact intensity interval the Poisson second-moment tail is uniformly small, while each truncated finite sum is continuous. The coarse bound proves the remaining limit. \end{proof} \subsection{Uniform additive comparison} We use the exact fixed-grid stationary experiment of Proposition~\ref{prop:exact}. The two estimates below supply the proof of Theorem~\ref{thm:uniform}. \begin{lemma}[Clipped-score comparison]\label{lem:clip} Suppose $K\sim\Pois(\lambda)$ and two observations $Y,Y_*$ are coupled using the same $K$, with $\PP(Y\ne Y_*)\le\lambda\delta$. If $J_Y=\E(\E[K-\lambda\mid Y])^2$, then for every $h>0$, \begin{equation}\label{eq:clip} \frac{|J_Y-J_{Y_*}|}{\lambda} \le24h^2\delta+ 4\sqrt{\frac{\E[(|K-\lambda|-h)_+^2]}{\lambda}}. \end{equation} \end{lemma} \begin{proof} Write $S=K-\lambda$, let $C_h$ clip a real number to $[-h,h]$, and put $S_h=C_h(S)-\E C_h(S)$. Then $|S_h|\le2h$ and $\Var S_h\le\Var S=\lambda$. The last inequality follows by applying the Lipschitz property of $C_h$ to two independent copies of $S$. Also $\|S-S_h\|_2^2\le\E[(|S|-h)_+^2]$. For either observation, the contraction property of conditional expectation gives \[ |\|\E[S\mid Y]\|_2^2-\|\E[S_h\mid Y]\|_2^2| \le2\sqrt\lambda\,\|S-S_h\|_2. \] For the clipped score use \[ \|\E[S_h\mid Y]\|_2^2 =\sup_{|f|\le2h}\E[2S_h f(Y)-f(Y)^2]. \] On the coupling event $Y=Y_*$ the objectives agree for each $f$; off it their absolute difference is at most $24h^2$. Take the supremum and restore $S$ in both experiments. \end{proof} \begin{lemma}[Uniform Poisson score tail]\label{lem:poissontail} For $0<\lambda\le L$, $x\ge1$, and $h(L,x)=\sqrt{2eLx}+2x$, \begin{equation}\label{eq:poissontail} \frac{\E[(K-\lambda)^2\one_{\{|K-\lambda|>h(L,x)\}}]}{\lambda} \le2e(1+2x)e^{-x}. \end{equation} \end{lemma} \begin{proof} For $S=K-\lambda$ an exponential tilt gives exactly \[ \frac{\E[S^2e^{uS}]}{\lambda} =[e^u+\lambda(e^u-1)^2]\exp\{\lambda(e^u-1-u)\}. \] For $|u|\le1$, the right side is at most $(e+e^2Lu^2)\exp(eLu^2/2)$. Apply the exponential bound to the positive and negative tails with $u=\min\{1,\sqrt{2x/(eL)}\}$. In either case $eLu^2/2-u h(L,x)\le-x$ and $e+e^2Lu^2\le e(1+2x)$. Summing the two tails proves the claim. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:uniform}] Construct the actual embedded jump chain and a fair-sign walk with common uniform random variables at each event, and the same independent $K\sim\Pois(\lambda)$. Start the actual chain in $\pi_\Omega$. At state $n$, the actual upward probability differs from $1/2$ by $|n/\Omega-1/2|$. The embedded chain is stationary because its total escape rate is constant. A union bound at the first disagreement therefore yields \begin{equation}\label{eq:coupling} \PP(D_{\rm actual}\ne D_{\rm fair}) \le\lambda\E_{\pi_\Omega}|N/\Omega-1/2| \le\frac{\lambda}{2\sqrt\Omega}. \end{equation} Include the same $N_0$ in both observations. It is independent of the fair walk, so its inclusion does not alter $F$. Lemmas~\ref{lem:clip} and \ref{lem:poissontail} imply \begin{align} \sup_{0<\lambda\le L}|R_\Omega(\lambda)-F(\lambda)| &\le\mathcal E_\Omega(L,x),\label{eq:explicit}\\ \mathcal E_\Omega(L,x)&= \frac{12h(L,x)^2}{\sqrt\Omega} +4\sqrt{2e(1+2x)e^{-x}}.\notag \end{align} To treat larger intensities, subdivide a bin into $m$ equal parts. Information cannot decrease when these extra observations are retained. Stationarity and orthogonality of the transition scores give \begin{equation}\label{eq:refinement} R_\Omega(\lambda)\le R_\Omega(\lambda/m),\qquad m\in\NN,\ m\ge1. \end{equation} For $\lambda\ge L$, choose $m=\lfloor\lambda/L\rfloor$; then $L\le\lambda/m\le2L$. Both $F(\lambda)$ and $R_\Omega(\lambda)$ are bounded above by $2/L+\mathcal E_\Omega(2L,x)$, using Lemma~\ref{lem:Fbounds}. Thus \begin{equation}\label{eq:globalexplicit} \sup_{\lambda>0}|R_\Omega(\lambda)-F(\lambda)| \le\frac2L+\mathcal E_\Omega(2L,x). \end{equation} Take $x=\log(e\Omega)$ and $L=\Omega^{1/4}/\sqrt{x}$. The elementary inequality $h(L,x)^2\le4eLx+8x^2$ proves \eqref{eq:uniform}; the additional $x^2/\sqrt\Omega$ term is absorbed in its universal constant. Finally, $1-\tau/T\le\Delta/T$ and $F(\lambda)\le2/\lambda$, so the unobserved terminal remainder costs at most $2/(\Omega\theta T)$. \end{proof} \subsection{The relative window from finite residual control} The following is the complete proof of Theorem~\ref{thm:relative}. \begin{proof} Write $n=\Omega$, $\ell=\log(en)$, $\rho=n^{-1/6}\ell^{1/3}$, and $M=1/\rho=n^{1/6}\ell^{-1/3}$. For $n\ge16$, $M\ge1$: the inequality $\sqrt n\ge\log(en)$ holds at 16 and its difference increases for $n>4$. Also $M/n\le n^{-5/6}<1/43$ by the cubic Taylor lower bound, and $12\sqrt6<30$ bound the constant by $1962$. For $M\le\lambda\le L$, Corollary~\ref{cor:finite-poisson-comparison} instead gives \[ |R_n/F-1|\le9L/n+57/(8M)=9L/n+(57/8)\rho. \] Join the two intervals; if $L0$. If $L_n/n\to0$, \eqref{eq:relativewindow} applies eventually and its right side vanishes. Conversely, if this ratio does not vanish, choose $\delta>0$ and a subsequence with $L_n/n\ge\delta$; fix $0