\subsection{Moment inequalities from the escape recursion} To estimate $D_\lambda$, we compare the moments of $X/(c+X)$ with an explicit sequence. This comparison gives a positive lower bound for $\Psi$ and bounds on the tails of the series used below. Put \begin{equation}\label{eq:archive-parameters} \begin{gathered} W=\frac1{c+X},\quad w=\E W,\quad h=w^{-1}-c,\quad d=1-h,\\ U=\frac{X}{c+X},\qquad u=\E U=\frac{h}{c+h}. \end{gathered} \end{equation} Here $0\rho_{n+1}>0\quad(n\ge1),\qquad \rho_n\longrightarrow0. \end{equation} In particular, $a_n/a_{n-1}<\beta_n/\beta_{n-1}$ for $n\ge2$. \end{lemma} \begin{proof} Let $S=\sum_{i=1}^{\nu}X_i$ and $T=\sum_{i=2}^{\nu}X_i$. The first child is independent of $T$, including the random number of summands. Write \[ H=\frac1{c+S}=\frac{W_1}{1+W_1T}. \] The recursion gives $\lambda W\overset{\mathrm d}=1+H$. Consequently, $\E H/w=\lambda-1/w=d$. Since $01$ and $\nu\ge1$, the event $\nu\ge2$ has positive probability. On that event $T>0$, so the first inequality is strict for $f(t)=t^n$, $n\ge1$. Using $H\overset{\mathrm d}=\lambda W-1$ and $1-cH\overset{\mathrm d}=\lambda U$ in~\eqref{eq:chosen-child-comparison}, and multiplying by $c$, yields \[ (\lambda^n-d)a_n-(\lambda^{n+1}-d)a_{n+1}>0, \qquad n\ge1. \] Division by $h$ proves the strict decrease in~\eqref{eq:rho-monotone}; $a_0=\beta_0=1$ and $a_1=\beta_1=h/(c+h)$ prove its initial values. Finally, \[ \rho_n=\frac{\E[(\lambda U)^n]-d\E U^n}{h}\longrightarrow0, \] by bounded convergence, since $0<\lambda U<1$ almost surely. This proves the limit without a deterministic escape bound strictly smaller than one. \end{proof} \begin{lemma}[An upper bound for the speed and a lower bound for the kernel]\label{lem:harmonic-speed} For $X$ with law $\pi_\lambda$, one has $v0$. \end{lemma} \begin{proof} Let $V$ be an independent copy of $U$. Substituting $X=cU/(1-U)$ and $Y=cV/(1-V)$ into the pair kernels gives \[ \frac{XY}{A(X,Y)}=\frac{cUV}{1-\lambda UV},\qquad \frac{t(X,Y)}{A(X,Y)}= \frac{U+V-(\lambda+1)UV}{1-\lambda UV}. \] Since $\lambda UV\le1/\lambda<1$, geometric expansion and absolute convergence give \begin{equation}\label{eq:moment-speed-expansion} N=cT,\qquad Z=2Q-(\lambda+1)T,\qquad T=\sum_{n\ge1}\lambda^{n-1}a_n^2,\quad Q=\sum_{n\ge1}\lambda^{n-1}a_na_{n-1}. \end{equation} For example, the summands in $T$ are at most $\lambda^{-n-1}$ and those in $Q$ at most $\lambda^{-n}$. Thus \[ v=\frac{c}{2Q/T-(\lambda+1)},\qquad \frac QT= \frac{\sum_{n\ge1}\lambda^{n-1}a_n^2(a_{n-1}/a_n)} {\sum_{n\ge1}\lambda^{n-1}a_n^2}. \] The sequence $\beta_n$ is realized by $U_p=\lambda^{-L}$, where $\PP(L=j)=h d^{j-1}$ for $j\ge1$. Its corresponding escape coordinate is $cU_p/(1-U_p)\in(0,1]$, so its symmetric denominator is positive. Write $T_p,Q_p$ for the sums in~\eqref{eq:moment-speed-expansion} with $a_n$ replaced by $\beta_n$, and $v_p=cT_p/[2Q_p-(\lambda+1)T_p]$. The ratios \[ \frac{\beta_{n-1}}{\beta_n} =\lambda+\frac{dc}{\lambda^{n-1}-d} \] decrease with $n$. Lemma~\ref{lem:canonical-moments} implies $a_{n-1}/a_n\ge\beta_{n-1}/\beta_n$, strictly for $n\ge2$. The weights $\lambda^{n-1}a_n^2$ are the comparison weights $\lambda^{n-1}\beta_n^2$ multiplied by the decreasing sequence $\rho_n^2$. The covariance inequality for two decreasing sequences therefore gives \[ \frac QT> \frac{\sum_{n\ge1}\lambda^{n-1}\beta_n^2\rho_n^2 (\beta_{n-1}/\beta_n)} {\sum_{n\ge1}\lambda^{n-1}\beta_n^2\rho_n^2} \ge\frac{Q_p}{T_p}. \] All sums are absolutely convergent; the covariance inequality follows first for finite truncations and then by passage to the limit. Positivity of both quotient denominators gives $v0. \end{align*} To verify the second equality, subtract its summands: the difference is $h(s_n^2-s_{n+1}^2)-(s_n-s_{n+1})$. The series telescopes to $hs_1^2-s_1=0$, because $s_1=1/h$ and $s_n\to0$. Each term in the final series is nonnegative and every term with $n\ge2$ is positive. Since $d=1-h>0$, the inequality $2hR_*>Q_*$ is equivalent to $v_p1-(c+v)w=(h-v)w>0. \end{equation} Here $cz0$, and $v>0$. The polynomial $y^2+c((1+v)y-v)$ is increasing on $[0,1]$. Relative to the law weighted by $(c+y)^{-2}$, the law weighted by $A(x,y)^{-2}$ has increasing density proportional to $((c+y)/A(x,y))^2$. Its expectation of this polynomial is at least $\Psi(0)/z$. Since $\cJ1(x)\ge1/\lambda^2$, \[ \Psi(x)\ge\frac{\Psi(0)}{\lambda^2z}>0\qquad(0\le x\le1). \] \end{proof} \subsection{An identity for \texorpdfstring{$g$}{g} and a lower bound for \texorpdfstring{$D_\lambda$}{D(lambda)}} To bound $D_\lambda$ while retaining contributions from all powers of $\cL$ in the series for $\mathcal R$, we use the function \begin{equation}\label{eq:entropy-function} g(x)=\frac{x(c+x)\log(1+c/x)}{\lambda c\log\lambda}-\frac{x}{c} \quad(x>0),\qquad g(0)=0. \end{equation} The formula defines $g$ for all positive $x$, since it will also be applied to a sum of child escapes. On $[0,1]$ it is continuous, nonnegative and bounded. \begin{lemma}[An identity for $g$ under the escape recursion]\label{lem:entropy-response} For a finite positive child list $x_1,\ldots,x_n$, with $S=\sum_i x_i$, \begin{equation}\label{eq:entropy-pair} \sum_i g(x_i)-g(S) =\frac1{\lambda c\log\lambda}\int_0^c(c-a) \sum_{i0. \] For example, $c+F_\lambda(s)=\lambda(c+s)/(\lambda+s)$ and $1+c/F_\lambda(s)=\lambda(1+c/s)$ separate the logarithm into the two required terms. Thus \[ g(F_\lambda(S)) =b(F_\lambda(S))+F_\lambda'(S)\sum_i g(X_i)-\delta_g. \] Multiply by $f(F_\lambda(S))$, take expectations, and use the fixed-point equation and conditional exchangeability. The middle term becomes $\E[g(X)(\cL f)(X)]$, proving~\eqref{eq:entropy-stationarity}. These expectations are justified for unbounded $\nu$. If $M=\max_{[0,1]}g$, then $g(S)\ge-S/c$ and $S\le\nu$, so \begin{equation}\label{eq:entropy-integrability} 0\le\delta_g\le\frac{M+1/c}{\lambda}\nu. \end{equation} The exchangeability term is at most $M\|f\|_\infty\nu/\lambda$ in absolute value. Both bounds are integrable. No derivative of $g$ at zero is used as a test in Lemma~\ref{lem:full-response}. Apply~\eqref{eq:entropy-stationarity} with $\mathcal R f$ in place of $f$. For nonnegative continuous $f$, positivity of the resolvent gives \[ \eta(f)=\E[g(X)f(X)]+\E[\delta_g(\mathcal Rf)(F_\lambda(S))] \ge\E[g(X)f(X)]. \] The Neumann series separately gives $\mathcal R f\ge f$, hence $\eta\ge b\pi_\lambda$. These inequalities hold as inequalities of finite Borel measures, by continuous approximation on the compact interval. Splitting a Borel set into its intersections with $\{g\ge b\}$ and $\{g0, \qquad H_* =\Psi-\alpha\cJ(y^2),\\ F_* =2\alpha\E[br]-C+2\E[gH_*]. \end{gathered} \end{equation} Then $H_*\ge0$ on $[0,1]$ and \begin{equation}\label{eq:floor-remainder} D_\lambda=F_*+2\E[\delta_g(\mathcal RH_*)(F_\lambda(S))]\ge F_*. \end{equation} If $\alpha=1$, then $D_\lambda>0$. In the remaining case $0<\alpha<1$, \begin{equation}\label{eq:hard-calibration} \begin{gathered} A_0=1-(\lambda+1)u+\lambda a_2>0,\qquad \alpha=\frac{u-(v/c)A_0}{a_2}>\lambda^2u,\\ u<\lambda^{-2},\qquad h<\frac1{\lambda+1}. \end{gathered} \end{equation} In particular, $(\lambda+1)h\ge1$ implies $\alpha=1$ and hence $D_\lambda>0$. \end{lemma} \begin{proof} The polynomial $(1-\alpha)y^2+c((1+v)y-v)$ is increasing on $[0,1]$. Its inverse-square weighted expectation at zero is nonnegative by the choice of $\alpha$. The increasing-density comparison used in Lemma~\ref{lem:harmonic-speed} therefore gives $H_*\ge0$ everywhere. Since $(I-\cL)r=\cJ(y^2)$, we have \[ \mathcal R\Psi=\alpha r+\mathcal RH_*. \] Apply~\eqref{eq:entropy-stationarity} with $f=\mathcal RH_*$ and insert this decomposition into~\eqref{eq:full-sign}. This proves~\eqref{eq:floor-remainder}. If $\alpha=1$, then $\cJ h_v(0)\ge0$, and the same increasing-density comparison gives $\cJ h_v(x)\ge0$ for all $x$. Identity~\eqref{eq:positive-K} remains valid without assuming the strict weighted-mean inequality. Its second term is strictly positive because $00$. Formula~\eqref{eq:floor-remainder} and $gH_*\ge0$ prove $D_\lambda>0$. Suppose $\alpha<1$. The identities \[ \cJ(y^2)(0)=a_2,\qquad \Psi(0)=u-\frac vc\{1-(\lambda+1)u+\lambda a_2\} \] give the formula for $\alpha$. Also $A_0=\E[(1-U)(1-\lambda U)]>0$. Since $v \frac{u-uA_0/(1-u)}{a_2} =\frac{\lambda u(u-a_2)}{(1-u)a_2}. \] Lemma~\ref{lem:canonical-moments} gives $a_2<\beta_2=u/(\lambda+1-\lambda u)$. Hence $(u-a_2)/[(1-u)a_2]>\lambda$, proving $\alpha>\lambda^2u$. Because $\alpha<1$, this implies $u<\lambda^{-2}$, and $h=cu/(1-u)