\subsection{Integral formulas for moment ratios and a series lower bound} \label{subsec:marker-floor} The lower bound $F_*$ in Lemma~\ref{lem:calibration} contains the logarithmic function $g$. We bound its contribution using the moments of $U$ and the successive differences of the ratios $a_s/\beta_s$. An integral representation relates these differences at all orders. For real $s\ge0$ retain \[ a_s=\E U^s,\qquad \beta_s=\frac{h}{\lambda^s-d},\qquad \rho(s)=\frac{a_s}{\beta_s}. \] For $s\ge1$ put \begin{equation}\label{eq:marker-deficits} \begin{gathered} r_s=\frac{\beta_{s+1}}{\beta_s},\qquad \Delta_s=\rho(s)-\rho(s+1),\qquad \delta_s=a_s r_s-a_{s+1}=\beta_{s+1}\Delta_s,\\ r_0=u,\qquad \delta_0=0,\qquad k=\frac{c}{\lambda\log\lambda}. \end{gathered} \end{equation} In particular, $1/\lambda\left(1-\frac1{2s}\right)\Delta_s. \end{equation} Thus, with $e=\Delta_1=1-\rho_2$, one has $0t\}}] -d\E[(1-U)\mathbf1_{\{U>t\}}]\ge0, \qquad 01$, differentiation under this probability integral is justified locally by boundedness of $t^{s-1}|\log t|$. At $s=1$, the derivative of $a_s/\beta_s$ is finite because $U^s|\log U|$ is bounded when $s$ is bounded away from zero. Fatou's lemma applied to \[ \frac{1-t^{\varepsilon}\{1+\varepsilon(1-t)\}}{\varepsilon} \] then shows that $-\log t-(1-t)$ is $\sigma$-integrable. Hence $|\log t|$ is integrable and dominated convergence also permits differentiation from the right at $s=1$. We obtain \[ -\rho'(s)=\int_0^1 t^{s-1} \bigl[\{1+(s-1)(1-t)\}(-\log t)-(1-t)\bigr]\sigma(\dd t). \] Write $z=1-t$. Since $-\log(1-z)>z+z^2/2$ for $0\widetilde F(w)$ when $w>v$. \end{lemma} \begin{proof} Set $\Gamma=\E[br]-\E[g\cJ(y^2)]$. Apply \eqref{eq:entropy-stationarity} with the strictly positive function $r$, using $(I-\cL)r=\cJ(y^2)$. It gives \[ \Gamma=\E[\delta_g r(F_\lambda(S))]>0. \] All these expectations are finite by~\eqref{eq:entropy-integrability}. To find the dependence on $w$, symmetrize the bounded pair kernels. For $t=x+y-xy$, \[ \frac{b(x)P_c(w;y)+b(y)P_c(w;x)}2 =\frac{(1-t)(wt-xy)}\lambda. \] Since $1-t=\lambda-A$, this implies the exact correction \begin{equation}\label{eq:assigned-C-correction} \E\frac{wt-XY}{A^2}-\E[b\cJ P_c(w;\cdot)] =\frac{wZ-N}\lambda. \end{equation} At $w=v=N/Z$ the correction vanishes, giving $\widetilde F(v)=F_*$. Differentiating the affine expression with the law fixed gives \[ \widetilde F'(w)=2c\E[g\cJ(y-1)] -\E[b\cJ((1-y)(2-y))]-\frac{2A_0\Gamma}{c a_2}<0. \] Here $A_0>0$, $a_2>0$, $\Gamma>0$, $g\ge0$, and $00$ for every $n\ge1$, and the series \begin{equation}\label{eq:complete-mellin-floor} F_{\rm floor}(w)=\sum_{s\ge1}s\lambda^{s-1} \left\{\frac{(a_s-\lambda a_{s+1})J_{s-1}(w)}{\lambda c} +2\overline\gamma_{s-1}t_{s-1}(w)\right\} \end{equation} converges absolutely. Moreover, \begin{equation}\label{eq:strict-Mellin-gap} \widetilde F(w)-F_{\rm floor}(w) =2\sum_{s\ge2}s\lambda^{s-1} \left\{k\beta_s[-\rho'(s)] -k\left(1-\frac1{2s}\right)\frac{\delta_s}{r_s}\right\} t_{s-1}(w)>0. \end{equation} In particular, $D_\lambda\ge F_*>F_{\rm floor}(v)$. \end{lemma} \begin{proof} The choice of $a(w)$ gives $t_0(w)=0$. The quadratic $Q_{a(w)}$ is negative at zero, strictly concave, and has value $(\lambda-a(w))/\lambda^2>0$ at $1/\lambda$. It therefore has exactly one zero $z_0$ in $(0,1/\lambda)$, with the sign of $z-z_0$ on that interval. For $n\ge1$, \[ t_n(w)=\E[(U^n-z_0^n)Q_{a(w)}(U)]>0. \] The integrand is nonnegative and vanishes only at $U=z_0$. The tree starts with a chain of single-child vertices of any given length with positive probability. Along such a chain of length $j$, the transform $U$ is multiplied by $\lambda^{-j}$. Thus the essential infimum of $U$ is zero, although $U>0$ almost surely. In particular, its law cannot be concentrated at $z_0>0$. Write $V=Y/(c+Y)$ for an independent copy of $U$. Direct substitution gives, with $a=a(w)$, \begin{align*} \frac{b(X)Y}{A^2} &=\frac{UV(1-\lambda U)(1-V)}{\lambda(1-\lambda UV)^2},\\ \frac{b(X)P_c(w;Y)}{A^2} &=\frac{U(1-\lambda U)(1-\lambda V) \{2w-(c+(\lambda+1)w)V\}} {\lambda c(1-\lambda UV)^2},\\ \frac{(1-a)Y^2+c((1+w)Y-w)}{A^2} &=\frac{(1-U)^2Q_a(V)}{(1-\lambda UV)^2}. \end{align*} Since $\lambda UV\le1/\lambda<1$, expand the common denominator in its uniformly convergent series. With \[ \gamma_{s-1}=\E[g(X)(1-U)^2U^{s-1}] =k\E[U^s\log(1/U)]-a_s+a_{s+1}, \] the first two kernels produce $(a_s-\lambda a_{s+1})J_{s-1}(w)/(\lambda c)$, and the third produces $2\gamma_{s-1}t_{s-1}(w)$. Their sum is exactly $\widetilde F(w)$. In particular, the first factor is $a_s-\lambda a_{s+1}$ for every index $s\ge1$, including the first. Differentiate $a_s=\beta_s\rho(s)$ with respect to the moment order $s$. Since $-k\beta_s'/\beta_s=1-r_{s-1}$, one obtains \[ \gamma_{s-1}=a_s(r_s-r_{s-1})-\delta_s +k\beta_s[-\rho'(s)]. \] Use Lemma~\ref{lem:canonical-marker}, $\beta_s\Delta_s=\delta_s/r_s$, and $t_{s-1}(w)>0$ for $s\ge2$. The term $s=1$ has $t_0(w)=0$. This proves the equality and strict inequality in~\eqref{eq:strict-Mellin-gap}. For completeness, $a_s\le\lambda^{-s}$ and $\delta_s\le\lambda^{-s-1}$. The ratios $r_s$ are bounded away from zero at fixed $\lambda,h$, and $r_s-r_{s-1}$ is bounded. The integral defining $\gamma_{s-1}$ is bounded by $\|g\|_\infty\lambda^{-s+1}$; its identity in the preceding display therefore bounds $k\beta_s[-\rho'(s)]$ by a constant times $\lambda^{-s}$. Also $t_{s-1}(w)$ and $J_{s-1}(w)$ have bounds of the same geometric order. Every expansion and regrouping above is dominated by a constant times $\sum_{s\ge1}s\lambda^{-s}$. No unweighted logarithmic moment or inverse escape moment is used. \end{proof}