\section{A uniform ratio of the two rank gaps}\label{app:signedgap} We prove \eqref{pr:gapratio} on $9/5\le p\le2$ and $1/16\le x,y\le1$. Order a distinct pair as $y0$. Differentiating $G$ gives the exact first-order equation \[ G'-a_0\tanh(v)G=a_0\frac{\sinh(\beta_0v)}{\cosh v}. \] The residual for the comparison $C$ is $E(v)=a_0\sinh(\beta_0v)+a_0C(v)\sinh v-C'(v)\cosh v$. Multiplying by the positive denominator gives the entire function \begin{align*} F(v):=\frac{(1+q_*v^2)^2E(v)}{g_2} ={}&\frac2{\beta_0}(1+q_*v^2)^2\sinh(\beta_0v)\\ &+a_0v^2(1+b_*v^2)(1+q_*v^2)\sinh v\\ &-2v(1+2b_*v^2+b_*q_*v^4)\cosh v. \end{align*} Here is a coefficient check that includes its whole tail. Write $F(v)=\sum_{j\ge0}A_jv^{2j+1}/(2j+1)!$ and $B_0=\beta_0^2$. The first two coefficients vanish. The next three are \begin{align*} A_2&=2B_0^2+80q_*B_0+240q_*^2+(132q_*-158)/7>3/2,\\ A_3&=2B_0^3+168q_*B_0^2+1680q_*^2B_0-53-24q_*>-17,\\ A_4&=2B_0^4+288q_*B_0^3+6048q_*^2B_0^2 +216q_*-3006/35>10. \end{align*} These are rational comparisons with $B_0=81/196$. For $j\ge5$, set $n_j=2j+1$. The contribution from the first line of $F$ is positive. The other two lines contribute \begin{align*} n_j\bigl\{&2a_0j-2 +(2j)(2j-1)[a_0(b_*+q_*)(2j-2)-4b_*]\\ &+b_*q_*(2j)(2j-1)(2j-2)(2j-3)[a_0(2j-4)-2]\bigr\}>0. \end{align*} Every bracket is positive already at $j=5$ and increases thereafter. Consequently, for $0v^5\left(\frac{3/2}{120}-\frac{17v^2}{5040}\right) \ge\frac{29}{5040}v^5>0. \] The integrating factor gives the desired comparison explicitly: \begin{equation}\label{sg:G} G(v)-C(v)=\cosh(v)^{a_0} \int_0^v\frac{E(t)}{\cosh(t)^{a_0+1}}\dd t\ge0. \end{equation} Cross-multiplication also gives $C(v)\ge g_2v^2/(1+k_*v^2)$, since the numerator of the difference is $b_*k_*v^4$ after canceling the positive leading factor. \subsection{A bound for the second gap with its full series} Put \[ d=a_0b_0=10/63,\quad K=(c_0^2-d^2)/2=125/882, \quad q_D=(c_0^2+d^2)/12=1325/47628, \] \[ J_0=g_2b_0^2=10/441,\quad e_0=k_*b_0^2=11/441, \quad d_2=K-J_0=5/42, \] \[ \eta_*=(Kq_D+J_0e_0)/d_2=37877/1000188, \quad q_2=1/90,\quad h_* =\eta_*-q_2>0. \] In the positive series of $\cosh(c_0v)-\cosh(dv)$, the ratio following the quadratic coefficient is $q_D$; all subsequent ratios are less than $q_2$. For the first of these later ratios the assertion is \[ \frac{c_0^4+c_0^2d^2+d^4}{30(c_0^2+d^2)}<1/90. \] For all later terms the geometric-sum identity $(1-r^{j+1})/(1-r^j)\le(j+1)/j$ gives the bound $c_0^2/[2j(2j+1)]\le c_0^2/42<1/90$ for $j\ge3$. Thus the complete series satisfies, with $w=v^2$, \[ \cosh(c_0v)-\cosh(dv) \le Kw\left(1+\frac{q_Dw}{1-q_2w}\right). \] Subtract \eqref{sg:G} at $b_0v$, using its weaker rational lower, to obtain \begin{equation}\label{sg:D} \frac{D(v)}{v^2} \le K\left(1+\frac{q_Dw}{1-q_2w}\right)-\frac{J_0}{1+e_0w} \le d_2\frac{1+h_*w}{1-q_2w}. \end{equation} The second inequality follows by expanding the first difference as $d_2+w[Kq_D/(1-q_2w)+J_0e_0/(1+e_0w)]$ and increasing the latter denominator's reciprocal to $1/(1-q_2w)$. The ratio of these two explicit bounds is \[ R_*(v)=\frac{28}{27}e^{-25v/126} \frac{(1+h_*v^2)(1+q_*v^2)}{(1-q_2v^2)(1+b_*v^2)}. \] It is log convex on $[0,\ell_0]$. The second derivatives of $\log(1+h_*v^2)$ and $-\log(1-q_2v^2)$ are positive there. For the remaining logarithmic difference the numerator has sign \[ 1-(q_*+b_*)v^2-3q_*b_*v^4>1-.654-.2724>0. \] All decimals here denote the displayed rational numbers. At zero $R_*(0)=28/27$, and at the other endpoint $R_*(\ell_0)<1019/1000$. Here is an exact bound including all tails. Write \[ L_-:=1684/1215<\ell_0(1-q_2L_+^2)(1+b_*L_-^2)>5833/5000. \end{align*} These are rational comparisons with the exact coefficients above. Set $x=(25/126)L_-=4210/15309$. The positive exponential series gives \[ e^x>\sum_{j=0}^4\frac{x^j}{j!} =\frac{216938038367782253}{164781536053084083}>\frac{2500}{1899}. \] Consequently \[ R_*(\ell_0)<\frac{28}{27}\frac{1899}{2500} \frac{1509/1000}{5833/5000} =\frac{742931}{729125}<\frac{1019}{1000}, \] with final strict margin $379/5833000$. Convexity bounds $R_*$ by its larger endpoint value, hence \begin{equation}\label{sg:chord} R_*(v)\le\lambda=28/27\qquad(0\le v\le\ell_0). \end{equation} \subsection{Restoring the physical exponent and the common mass scale} For fixed $00. \] For the second, express $g_p$ as the difference between the power mean of $1,t^{p/(p+1)}$ of order $p+1$ and their arithmetic mean. Increasing the order increases this difference. Decreasing the smaller argument also increases it, since its derivative is $[(y/M_{p+1}(1,y))^p-1]/2<0$ for $y<1$. Consequently, for $p_0\le p\le2$ and $v_0\le\ell_0$, \begin{equation}\label{sg:enclosure} d_p(t)\le d_{\rm up}\le\lambda g_{\rm low}, \qquad g_p(t)\ge g_{\rm low}>0. \end{equation} Finally let $1/16\le y