\subsection{The exact residual and its global norm} Take the model and parameters of Theorem~\ref{thm:finite-residual}. The following is its complete proof. \begin{proof} \par\smallskip\noindent\textbf{Exact conditional polynomials and parity.} Let $M$ be the number of particles with different initial and terminal states. Then $M\sim\Bin(n,p)$, and, conditional on $M$, \[ A=\sum_{j=1}^{n-M}\varepsilon_j,\qquad D=\sum_{j=1}^{M}\eta_j, \] where the two sets of signs are independent symmetric Rademacher variables. The correlated-Bernoulli construction is classical \cite[Theorem~1 and Section~6]{DiaconisGriffiths2012}. This is an exact endpoint representation; $M$ is not the actual jump count $K$. With $X=M-np$, differentiation at fixed physical grid gives \begin{equation}\label{eq:residual-latent-score} S=\gamma\E[X\mid Y]. \end{equation} Indeed the complete score for the endpoint representation in coordinate $p$ is $X/w$, and the log-clock derivative is $tr\partial_p$. Write $M=\sum_{i=1}^n B_i$ with independent Bernoulli indicators $B_i$. The subset polynomials below normalize the classical binomial Krawtchouk likelihood-derivative polynomials \cite[Eq.~(3.2)]{TakeuchiTakemura1987}. We derive their norms and conditional endpoint identities in the present notation. For $1\le j\le n$, define \[ H_j=j!\sum_{|I|=j}\prod_{i\in I}(B_i-p), \] and set $H_j=0$ for $j>n$. Expanding gives \begin{align} H_1&=X,& H_2&=X^2-rX-\nu,\notag\\ H_3&=X^3-3rX^2+ \{2r^2-(3n-2)w\}X+2nwr. \label{eq:residual-polynomials} \end{align} The last polynomial is zero on the binomial support when $n=2$, consistent with the empty-subset definition. Independence and centering show \begin{equation}\label{eq:residual-polynomial-norms} \E[H_jH_k]=0\ (j\ne k),\qquad \E H_j^2=j!(n)_j w^j. \end{equation} For distinct subsets some centered factor occurs once, so its expectation is zero; identical subsets give $w^j$. These observations prove both identities, where $(n)_j=n(n-1)\cdots(n-j+1)$ is the falling factorial, with $(n)_j=0$ for $j>n$. Set $\widehat H_j=\E[H_j\mid Y]$ and $L=s^2W_2-p^2W_1$. For fixed admissible endpoints $(A,D)=(a,z)$, the conditional weights of $M=m$ are proportional to \[ \frac{s^{n-m}p^m} {\left(\frac{n-m+a}{2}\right)! \left(\frac{n-m-a}{2}\right)! \left(\frac{m+z}{2}\right)! \left(\frac{m-z}{2}\right)!}. \] Only nonnegative factorial arguments are admitted: $|D|\le m\le n-|A|$. Feasible values of $m$ have fixed parity and are spaced by two. Their successive-weight identity is \[ s^2\{(m+2)^2-z^2\}\,\Pr(M=m+2\mid Y) =p^2\{(n-m)^2-a^2\}\,\Pr(M=m\mid Y). \] The omitted endpoint terms vanish at the feasible lower and upper bounds. Summing along this parity class therefore gives \[ p^2\E[(n-M)^2-A^2\mid Y] =s^2\E[M^2-D^2\mid Y]. \] Substitution of $M=np+X$ and \eqref{eq:residual-polynomials} yields \begin{equation}\label{eq:residual-parity-identity} d\widehat H_1=L-r\widehat H_2. \end{equation} Now $\E[W_1\mid M]=-X$ and $\E[W_2\mid M]=X$. Since $H_2$ is orthogonal to $X$, $\E[\widehat H_2W_1]=\E[\widehat H_2W_2]=0$. Thus \eqref{eq:residual-parity-identity} is an orthogonal decomposition: \begin{equation}\label{eq:residual-orthogonal-decomposition} \begin{aligned} \Pi_2S&=\frac{\gamma L}{d},\qquad S-\Pi_2S=-\frac{\gamma r}{d}\widehat H_2,\\ E_n(t)&=\frac{\gamma^2r^2}{d^2}\E\widehat H_2^2. \end{aligned} \end{equation} \par\smallskip\noindent\textbf{A differentiated likelihood certificate.} Let $f_y(p)$ denote the positive joint mass of an admissible endpoint pair. Its finite Bernoulli mixture, whose conditional signs have no $p$ dependence, gives \[ \widehat H_j=w^j\frac{f_y^{(j)}}{f_y},\qquad j=1,2,3. \] To see this directly, differentiate the product of $n$ linear Bernoulli probabilities. The $j$th derivative sums ordered choices of $j$ distinct factors and hence gives precisely the subset polynomial $H_j/w^j$. For $n=2$, both the third derivative and $H_3$ are zero. Equation~\eqref{eq:residual-parity-identity} becomes \[ rw^2f_y''+dwf_y'-Lf_y=0. \] Differentiate this identity at fixed endpoint pair, and put $\beta=(dw)'=r\{1+4(n-2)w\}$. Since $dw+(rw^2)'=we$, the result is \begin{equation}\label{eq:residual-differentiated-identity} \begin{aligned} e\widehat H_2&=(L-\beta)\widehat H_1+wL'-r\widehat H_3,\\ L'&=-2sD^2-2pA^2-n(r^2-2w). \end{aligned} \end{equation} The projection statistics in the previous paragraph are evaluated at the given parameter with their centers held fixed. Here $L(p)$ is a moving coefficient of an identity valid throughout the parameter family; its derivative, including that of its centers, must be kept. Define the latent polynomial $Z=(L-\beta)X+wL'-rH_3$. Then \eqref{eq:residual-differentiated-identity} says $\E[Z\mid Y]=e\widehat H_2$. Conditional Jensen gives the global estimate \begin{equation}\label{eq:residual-jensen} \E\widehat H_2^2\le\frac{\E Z^2}{e^2}. \end{equation} \par\smallskip\noindent\textbf{Exact second moment.} Conditional on $M$, $\E[L\mid M]=hX$ and $\E[L'\mid M]=-nh-2rX$. Consequently \[ \E[Z\mid M]=hH_2-4(n-1)wrX-rH_3. \] Its three terms are orthogonal by \eqref{eq:residual-polynomial-norms}. For a sum of $k$ independent Rademacher signs, the variance of its square is $2k(k-1)$. The coefficients of $D^2$ and $A^2$ in $Z$ are respectively $s^2(X-2p)$ and $-p^2(X+2s)$. Conditional independence therefore gives \begin{align*} \Var(Z\mid M)&=2s^4(X-2p)^2M(M-1)\\ &\quad+2p^4(X+2s)^2(n-M)(n-M-1). \end{align*} The binomial factorial weighting identity $\E[(M)_2 g(M)]=(n)_2p^2\E[g(M'+2)]$, with $M'\sim\Bin(n-2,p)$, yields \[ \E[M(M-1)(X-2p)^2] =n(n-1)p^2\{(n-2)w+4r^2\}. \] Selecting two zero indicators gives the horizontal counterpart with $s^2$ in place of $p^2$. Thus \[ \E\Var(Z\mid M)=2n(n-1)w^2h\{(n-2)w+4r^2\}. \] The squared norm of the conditional mean is \[ \E\bigl(\E[Z\mid M]\bigr)^2 =2n(n-1)w^2\{h^2+(11n-14)wr^2\}. \] Variance decomposition proves $\E Z^2=\mathcal V_n(p)$, including $n=2$. Together with \eqref{eq:residual-orthogonal-decomposition} and \eqref{eq:residual-jensen}, this proves $E_n(t)\le T_n(t)$. This calculation includes every endpoint cell. It requires neither a Gaussian approximation nor a lower bound on a rare endpoint probability. \par\smallskip\noindent\textbf{Uniform constants and relative width.} On $02\nu$. For $n\ge2$, all terms in the bracket defining $\mathcal V_n$ are nonnegative and their sum is at most $5+12\nu$. Hence \[ \mathcal V_n\le2\nu^2(5+12\nu),\qquad T_n(t)\le\gamma^2r^2 \left(\frac3{2\nu}+\frac5{8\nu^2}\right). \] Using $\nu=\lambda r/\gamma$ gives \[ T_n(t)\le\frac{3\gamma^3r}{2\lambda} +\frac{5\gamma^4}{8\lambda^2} \le\frac3{2\lambda}+\frac5{8\lambda^2}. \] For the relative interval, the quadratic information formula is $\lambda B_n(t)=\gamma^2\nu h/d$. Thus \[ \frac{T_n(t)}{\lambda B_n(t)} \le\frac{r^2}{h}\left(\frac3\nu+\frac5{4\nu^2}\right) =\frac{3\gamma r}{h\lambda} +\frac{5\gamma^2}{4h\lambda^2} \le\frac3\lambda+\frac2{\lambda^2}, \] where $r\le h$, $h\ge5/8$, and $\gamma\le1$. Equation~\eqref{eq:residual-E} proves both information bounds; the simplifications with constant $17/8$ use $\lambda\ge1$. \end{proof} \subsection{A finite comparison with the compound-Poisson experiment} \begin{corollary}[A finite compound-Poisson comparison] \label{cor:finite-poisson-comparison} For $n\ge2$, $\lambda\ge1$, and $t=\lambda/n\le a_0$, \begin{equation}\label{eq:finite-half-information} \left|\lambda R_n(\lambda)-\frac12\right| \le3t+\frac{17}{8\lambda}. \end{equation} For every $\lambda\ge1$, \begin{equation}\label{eq:finite-F-half} \left|\lambda F(\lambda)-\frac12\right| \le\frac{17}{8\lambda}. \end{equation} Consequently, on the preceding $n,t,\lambda$ domain, \begin{equation}\label{eq:finite-poisson-relative} \left|\frac{R_n(\lambda)}{F(\lambda)}-1\right| \le\frac{9\lambda}{n}+\frac{57}{8\lambda}. \end{equation} In particular, this relative error is at most $1/10$ throughout the nonempty continuous region \begin{equation}\label{eq:finite-poisson-ten-percent} n\ge25650,\qquad285/2\le\lambda\le n/180. \end{equation} \end{corollary} The following is the proof of Corollary~\ref{cor:finite-poisson-comparison}. \begin{proof} Let $x=n(e^{4t}-1)$ and $a=x/(x+2)$. Then \eqref{eq:finite-gaussian-identity} gives $\lambda B_n(t)=g(t)a$, where $g(t)=\gamma^2h/2$. Since $\gamma=4tr/(1-r^2)\ge r$ by $1-e^{-4t}\le4t$, and $h=(1+r^2)/2$, \[ 0\le\frac12-g(t) \le\frac{(1-\gamma^2)+(1-h)}2\le3t. \] Also $x\ge4\lambda$ and $g(t)\le1/2$, so \[ 0\le\frac12-\lambda B_n(t) \le3t+\frac1{2(2\lambda+1)} \le3t+\frac1{4\lambda}. \] Use the sign and bound $0\le E_n(t)=\lambda(R_n-B_n)\le17/(8\lambda)$ to obtain \begin{equation}\label{eq:finite-half-one-sided} \frac12-3t-\frac1{4\lambda} \le\lambda B_n(t)\le\lambda R_n(\lambda) \le\frac12+\frac{17}{8\lambda}. \end{equation} This proves \eqref{eq:finite-half-information}. For each fixed $\lambda\ge1$, let $n\to\infty$ with $t=\lambda/n$; the preceding domain is then satisfied eventually. The already proved uniform additive convergence in Theorem~\ref{thm:uniform} gives $R_n(\lambda)\to F(\lambda)$. Passing to this limit in \eqref{eq:finite-half-one-sided} gives \begin{equation}\label{eq:finite-F-one-sided} \frac12-\frac1{4\lambda}\le\lambda F(\lambda) \le\frac12+\frac{17}{8\lambda}, \end{equation} which also proves \eqref{eq:finite-F-half}. This is a limit of the information fractions already defined with fixed physical designs, not a derivative of a parameter-dependent grid. Subtracting opposite endpoints of \eqref{eq:finite-half-one-sided} and \eqref{eq:finite-F-one-sided} in both directions gives \[ |\lambda\{R_n(\lambda)-F(\lambda)\}|\le3t+\frac{19}{8\lambda}. \] For $\lambda\ge1$, Lemma~\ref{lem:Fbounds} gives $\lambda F(\lambda)\ge\lambda/(1+2\lambda)\ge1/3$; division proves \eqref{eq:finite-poisson-relative}. In \eqref{eq:finite-poisson-ten-percent}, $t\le1/180