\section{Fractional rough models and one-variable means}\label{sec:model-preparation} We develop the rough Hankel model and its one-variable centering estimate before applying the comparison theorem. Fix an even \(L_0\), put \(G=(\Z/L_0\Z)^\times\), and take a proper subgroup \(H\le G\). Write \[ \mathcal P_H=\{p:p\nmid L_0,\ p\bmod L_0\in H\},\qquad C_H=\langle\mathcal P_H\rangle,\qquad f=\1_{C_H}, \] and set \(\rho=|H|/|G|\), \(k_H=[G:H]=1/\rho\). The outer scale is \(X\); later reflected applications take \(X=N\). Put \[ L=\log X,\quad y=\exp(L^{2/5}),\quad \mathcal P_y=\{p:p\le y\},\quad E=\exp(\sqrt L),\quad V_y=\prod_{p\le y}(1-1/p). \] Let \(n_{\mathcal P_y}\) denote the \(\mathcal P_y\)-supported part of \(n\), and \(n_{>y}=n/n_{\mathcal P_y}\). Define \[ E_+(y)=\prod_{\substack{p\le y\\p\in\mathcal P_H}}(1-1/p),\quad E_-(y)=\prod_{\substack{p\le y\\p\nmid L_0,\ p\notin\mathcal P_H}}(1-1/p), \quad \alpha_0=\varphi(L_0)/L_0. \] Then \(V_y=\alpha_0E_+(y)E_-(y)\) for large \(X\). \subsection{Fractional poles and centered rough means} Character orthogonality gives, for \(\Re s>1\), \begin{equation}\label{eq:euler-branches} F_H(s)=\prod_{p\in\mathcal P_H}(1-p^{-s})^{-1} =G_H(s)\prod_{\eta\in H^\perp}L(s,\eta)^\rho. \end{equation} At a forbidden good prime of coset order \(r>1\), the correction is \((1-p^{-rs})^{1/r}\); thus \(G_H\) converges absolutely for \(\Re s>1/2\). The literal principal character carries the bad Euler factors. The positive marginal constant is \[ c_H=\frac1{\Gamma(\rho)} \lim_{s\downarrow1}(s-1)^\rho F_H(s); \] its Euler expression is \eqref{eq:marginal-product}. Every \(\eta\in H^\perp\) equals one on the support of \(f\); every such twist therefore has the same pole branch. Define the rough series and its regular part by \[ F_R(s)=F_H(s)\prod_{\substack{p\le y\\p\in\mathcal P_H}}(1-p^{-s}) =(s-1)^{-\rho}H_R(s). \] Here \(H_R(1)=\Gamma(\rho)c_HE_+(y)\). Fixed-order Cauchy estimates on a disk of radius \(c/\log y\) give \begin{equation}\label{eq:rough-derivatives} |H_R^{(j)}(1)|/j!\ll_j H_R(1)(C\log y)^j. \end{equation} For a fixed depth \(K\), define on polynomial scales in \(X\) \begin{equation}\label{eq:hankel-model} M_{R,K}(t)=\sum_{j=0}^K\frac{H_R^{(j)}(1)}{j!} \frac{(\log t)^{\rho-j-1}}{\Gamma(\rho-j)}. \end{equation} For \(X^a\le t\le X^C\), with fixed \(00\). For definiteness fix a smooth real function \(\eta_0\), with \(0\le\eta_0\le1\), zero on \(({-}\infty,1/4]\) and one on \([1/3,\infty)\). Set \(M_{R,K}(t)=0\) for \(t\le X^{1/4}\), and multiply the displayed expression by \(\eta_0(\log t/L)\) above that point. On the transition range it has the same positive upper bound. All rough scales used in (C3) are at least \(X^{1/2}/E>X^{1/3}\), so the cutoff never changes their means. Only the polynomial bulk enters principal means. \begin{lemma}\label{lem:rough-means} For fixed \(B,A_0>0\), sufficiently large fixed \(K\) has the following property. If \(b\) is bounded and periodic of period at most \(L^B\), and \(W\) is a smooth dyadic weight at a scale \(X^{2/5}\le T\le X^2\), with fixed derivatives bounded by fixed powers of \(L\), then \begin{align*} &\sum_n W(n)b(n)f(n)\1_{(n,\prod_{p\le y}p)=1}\\ &\quad-\sum_n W(n)b(n)\1_{(n,\prod_{p\le y}p)=1} \left(\sum_{\eta\in H^\perp}\eta(n)\right) \frac{M_{R,K}(n)}{V_y}\ll T L^{-A_0}. \end{align*} The same holds with a factor \(n^{iu}\), \(|u|\le L^B\). \end{lemma} \begin{proof} Let \(q\le L^B\) be the period of \(b\). Refine it to \(Q=\lcm(q,L_0)\) and expand on unit residues by characters. All primes dividing \(Q\) are at most \(y\). For \(\psi\bmod Q\), the twisted rough series is \[ \prod_{\eta\in H^\perp}L(s,\psi\eta)^\rho\,G_\psi(s) \prod_{\substack{p\le y\\p\in\mathcal P_H}}(1-\psi(p)p^{-s}). \] If \(\psi\) is inflated from \(H^\perp\), this equals \(F_R\); otherwise all its character factors are nonprincipal. Use smooth Mellin inversion at height \(\exp(L^{1/10})\), moving to \(1-c/\log y\). The classical zero-free region is wider there at these heights and polylogarithmic conductors. Siegel's theorem places a possible exceptional real zero outside the strip for large \(X\), after choosing its fixed exponent depending on \(B\). The finite Euler product costs \(L^{O(1)}\); the displacement gives \(\exp(-cL^{3/5})\), dominating polynomial \(L\)-function growth at that height. Smooth tails are smaller than every fixed logarithmic power. The threshold can be ineffective. Hankel integration gives \eqref{eq:hankel-model} for each principal branch. Its relative Taylor error is \(O_K(L^{-3(K+1)/5})\), by \eqref{eq:rough-derivatives}. Choose \(K\) after the derivative and character-count costs. For the unweighted rough population the series \(\zeta(s)\prod_{p\le y}(1-p^{-s})\) has principal density \(V_y\), and only the trivial character contributes. The projection \(\sum_{\eta\in H^\perp}\eta\) therefore reproduces every principal branch of the first sum with its exact coefficient. The factor \(n^{iu}\) changes only the derivative budget and choice of \(K\). \end{proof} The contour argument uses the fractional pole and the absolute convergence and fixed-order derivative bounds of its Euler correction. It therefore also applies when that correction is inverted and remains absolutely convergent for \(\Re s>1/2\). The quadratic-indicator application uses precisely this variant, with exponent \(1/2\) and its two principal character branches. For \(U\subset H\), set \begin{equation}\label{eq:corrected-model} \mathcal D_U(n)=k_H\1_{n\bmod L_0\in U}f(n_{\mathcal P_y}) \frac{M_{R,K}(n_{>y})}{V_y},\qquad \xi_U(n)=f(n)\1_{n\bmod L_0\in U}-\mathcal D_U(n). \end{equation} Then \(\mathcal D_U\ge0\), \(\mathcal D_U\ll L^{-3(1-\rho)/5}\) and \(|\xi_U|\ll1\). Conditioning on an exact record \(n_{\mathcal P_y}=r\le E\) leaves residual state \(U/r\), so Lemma~\ref{lem:rough-means} centers its principal mean. Fixed divisor-growth record coefficients cost only powers of \(L\), since \(\sum_{r\ \mathcal P_y\text{-supported}}d_j(r)/r\ll L^{O_j(1)}\). Rankin's bound with exponent \(c/\log y\) gives \begin{equation}\label{eq:record-tail} \#\{n\le T:n_{\mathcal P_y}>E\}\ll TL^{O(1)}e^{-cL^{1/10}} \quad(X^a\le T\le X^C). \end{equation} Indeed the Euler moment is \(L^{O(1)}\), whereas \(E^{-c/\log y}\) gives the exponential saving. The same argument permits fixed divisor weights. For \(H=G\), use the exact model \(\mathcal D_U=\1_{n\bmod L_0\in U}\) and \(\xi_U=0\).