\subsection{A fixed-geometry value on the joint cone} We now bound the signed power-law value, then restore its clipped tail. The estimates use the rectangle earned in Lemma~\ref{mid:jointstrip}, \[ 2u-1\leq m\leq u,\qquad0\leq d=m-r\leq4\delta/5. \] In this subsection $q=q(m,m,m-d)$ and $H=H(m,m-d;m)$. The signed mesh has no dependence on $d$: \[ S(u,m)=u(1+pa-Z)+m\sum_{j\geq0}G_u(u+jm). \] Use the fixed constants in \eqref{mid:valuemultipliers} and put $\Phi(m,d)=\Phi_0(m,m-d)=S+\lambda_*q+\mu_*H$; thus $d$ is a gap coordinate, not the fourth mass $r$. The complete enclosures give $\lambda_*/\mu_*>8/3$ and $\lambda_* -\mu_*>.1573$. No pure-root equation is assumed in the following value estimates. \begin{lemma}[Convexity at zero gap]\label{mid:zerogapconvexity} At fixed $p,u$, the function $m\mapsto\Phi(m,0)$ is strictly convex on $2u-1\leq m\leq u$. \end{lemma} \begin{proof} Let $h(s)=s/(1-e^{-s})$. Then $h'=1-C$ and $h''=-C'>0$. At $d=0$ the exact adjacent second derivative gives \[ q_{mm}=\int\rho_p(s)a_0(us)s h''(ms)\dd s \geq m^{-p-1}K_p,\qquad K_p=\int\rho_p(s)a_0(s)s h''(s)\dd s. \] The function $e^sh''(s)$ increases, as its derivative has positive numerator $e^{2s}+(4-4s)e^s-2s-5$. Gamma order consequently gives $K_p/p\geq K_1=6[\zeta(3)-\zeta(4)]>7/10$. For the last inequality retain $1/16+2/81+3/256+4/625$ and the integral tail $1/81$ in $\zeta(3)-\zeta(4)$. The diagonal identity \[ H(m,m;m)=\int\rho_p(s)b_0(ms)e^{-us}[u+\beta(1;s)]\dd s -(Z-1)m^{1-p} \] has convex first term, hence $H_{mm}>-p(p-1)(Z-1)m^{-p-1}$. The common upper bound \eqref{pf:bose} gives $(p-1)(Z-1)<132/125$. Therefore the source curvature is positive, since $(\lambda_*/\mu_*)K_p>(8/3)(7/10)p>(132/125)p$. The second derivative of $S$ has integrand $\rho_p e^{-us}g(s)s h''(ms)$. After moving $e^{-ms}$ into the signed weight its remaining multiplier $se^{ms}h''(ms)$ is positive increasing. Apply signed comparison to the positive seed $G_u(u+m)$: $u+m\leq2$, $G_{1,7/8}(2)=37/1920>0$, and raising $u$ or $p$ preserves the seed. Thus the second derivative of $S$ is positive too. \end{proof} \subsection{Bounding the clipping error} For the actual and every interpolated kernel write \[ c=1/2-a,\quad b=1/6-a/2,\quad A=c(p+1),\quad B=b(p+1)(p+2),\quad C_2=\frac{A+\sqrt{A^2-4\delta B}}{2\delta}. \] The square root is real. In fact the number $C_1=A/\delta-(p+2)/2>2$ lies strictly between the two roots: \[ \delta C_1^2-AC_1+B =(p+2)[3\delta(p+2)-(p+1)]/12<0. \] To explain why these roots control the tail, put \[ h_a(s)=1-a-\left[\frac{1-e^{-s}}s-ae^{-s}\right]. \] This is positive increasing, and on the whole positive ray \[ h_a(s)>\frac{cs}{1+s/2},\qquad h_a(s)>cs-bs^2. \] For the first inequality multiplication gives $2s[(1+s/2)h_a-cs]=(1+as)[s-2+(s+2)e^{-s}]>0$. For the second, use the positive remainder $R_3(s)=1-e^{-s}-s+s^2/2$ and $R_3(ts)>t^3R_3(s)$ for $00$, since $a<1/4$. Under the law with density $w_\theta(s)se^{-xs}/U(x)$, we have $G_u(x)/U(x)=\E h_a-\delta$. Its likelihood ratio against Gamma$(p+1,x)$ is increasing. Under that Gamma law, Cauchy--Schwarz gives $\E[s/(1+s/2)]\geq (\E s)^2/\E[s+s^2/2]$, so \[ \frac{G_u(x)}{U(x)}> \max\left\{\frac{A}{x+(p+2)/2},\ \frac A x-\frac B{x^2}\right\}-\delta. \] The first bound gives positivity up to $C_1$ and the second extends it to $C_2$. For $x\ge C_2$, factor the negative of the second bound as $\delta(1-C_2/x)(1-C_-/x)$, where $C_->0$ is the smaller root. Dropping the last factor and using $U(x)\leq px^{-p-1}$ gives \[ [-G_u(x)]_+\leq p\delta x^{-p-2}(x-C_2)_+. \] The majorant is unimodal. Comparing samples to the cells toward its maximum bounds the mesh sum by its integral plus one sample. The integral is $\delta C_2^{-p}/(p+1)$. At the maximizing point $x=C_2(p+2)/(p+1)$ the factor $[(p+1)/(p+2)]^{p+1}$ is at most $4/9<1/2$, since $v\log(1+1/v)$ increases for $v>0$. Thus for mesh step at most one the two contributions give the clipping bound \begin{equation}\label{mid:clipprice} \mathcal C(p,\delta) =\delta C_2^{-p}\left[\frac1{p+1} +\frac{p}{2(p+2)C_2}\right]. \end{equation} A uniform upper bound is $\mathcal C<\delta T_*(p)$, where \begin{equation}\label{mid:clipT} \alpha_*(p)=\frac{p+27}{16},\qquad T_* =\alpha_*^{-p}(p+1)^{-p-1} \left[1+\frac{p}{2\alpha_*(p+2)}\right]. \end{equation} Indeed $C_2>\alpha_*(p)(p+1)$: substitute this test point into the quadratic. Its value, divided by $p+1$, increases with $\delta\leq1/8$; at $1/8$ it is \[ [\alpha_*^2/8-(211/800)\alpha_*](p+1)+(233/4800)(p+2). \] This is convex in $p$ and negative at both endpoints, with values $-.011875$ and less than $-.01084$. The function $T_*$ is convex. Writing $T_*=fh$ and $L=-f'/f$ gives $L=\log[\alpha_*(p)(p+1)]+1+p/(p+27)$, $23.4-1.8-.6>0$. Its endpoints satisfy $T_*(1)=23/147$ and $T_*(9/5)<.0224$, the latter by a fifth-power comparison. \subsection{One value variable and two endpoints} Appendix~\ref{app:middlecone} proves, throughout $2u-1\leq m\leq u$, $0\leq d\leq9(u-m)/10$, \[ \Phi(m,d)-\Phi(m,0) \leq dm^{-p}[C_*(p)-\kappa_p(u-m)/m],\qquad\kappa_p>0. \] With $C_+=[C_*]_+$, an upper on this cone is therefore \[ \Psi(m)=\Phi(m,0)+\frac9{10}(u-m)m^{-p}C_+. \] It is convex by Lemma~\ref{mid:zerogapconvexity} and $[(u-m)m^{-p}]''=pm^{-p-2}[(p+1)u-(p-1)m]>0$. The clipping error bound is independent of $m$, so only two values remain. At $m=u$, Lemma~\ref{mid:upperzerogap} and \eqref{mid:unitchord} give $[\Psi(u)+\mathcal C]/\delta<-.0097-.54(p_{\max}-p)-.0043\delta<0$. At $m=1-2\delta$, retain the diagonal curvature of Lemma~\ref{mid:diagonalcurvature}. Also $m^{-p}C_+$ is convex in $\delta$, and $\mathcal C/\delta^2$ increases: with $v=\delta C_2$, \[ v'=\frac{A'v-(B+\delta B')}{2v-A}<0, \quad A'v<.125,\quad B+\delta B'\geq.28, \] and \eqref{mid:clipprice} gives $\mathcal C/\delta^2=\delta^{p-1}v^{-p} [1/(p+1)+p\delta/(2(p+2)v)]$. The curvature lemma gives convexity of $\Phi(1-2\delta,0)/\delta$. The clipping quotient satisfies $\mathcal C/\delta\le(\delta/\delta_0)\mathcal C(p,\delta_0)/\delta_0$ for $\delta_0=1/8$. Combining this linear bound with the convex diagonal terms gives the chord from the unit limit to the upper bound at $\delta_0$. The unit bound \eqref{mid:unitchord} with $k=1$ and $c\leq4/5$ already pays $(4/5)C_+$, choosing $c=4/5$ or $0$ according to its sign. The extra tenth costs at most $.0069(p_{\max}-p)$, since $C_*(p_{\max})=0$ and $C_*'>-.069$. At the far endpoint, Lemma~\ref{mid:farvalue} already pays a coefficient $1/10$ of $(3/4)^{-p}C_+$. The required coefficient is $(9/10)\delta_0=9/80$, leaving $1/80$. Thus the upper is \[ \mathcal F_0(p)+\frac1{80}(3/4)^{-p}C_+. \] Lemma~\ref{mid:farvalue} gives $\mathcal F_0<-.00052-.009(p_{\max}-p)$, while the additional term is less than $.00154(p_{\max}-p)$. Both endpoints of the chord are negative. Consequently $\Psi(m)+\mathcal C<0$ on the full $m$ interval. Every feasible point supplied by Lemmas~\ref{mid:jointcap} and \ref{mid:outerfeasible} lies in the rectangle and cone. At that point, $H_1=0$, $q_1\geq0$ and $\lambda_*>0$, so \[ \mathcal L_1(m)+\mathcal C \leq\Phi_1(m,r)+\mathcal C <\Phi_0(m,r)+\mathcal C \leq\Psi(m)+\mathcal C<0. \] The strict inequality is the fixed-geometry comparison of Lemma~\ref{mid:jointreturn}. The clipping error bound controls the negative atoms discarded from $\mathcal L_1$, proving $\mathcal B_u+H_{J_u}(u,m)<0$. \subsection{Returning to the two physical geometries} For the inner geometry, use $m=M$ from Lemma~\ref{mid:jointcap}. The negative value just proved contradicts \eqref{mid:innermesh}. For the outer geometry, use $m=m_R$ from Lemma~\ref{mid:outerfeasible}. On the active region $G=G_u(u)>0$, $J_u(u)=G$ and the complete clipped tail decreases with its step. Since $m_R0$. For $u\leq7/8$, the direct adjacent source bound and Lemma~\ref{mid:adjacentvalue} contradict the comparison functional. For $u\geq7/8$, the near-unit inner cap or affine outer cap applies according as $v\leq u$ or $v\geq u$; the common feasible-point estimate then gives the contradiction. Thus unequal masses are impossible. Strict angular convexity then makes the equal-mass configuration regular. \end{proof}