\subsection{Subgroup sums and congruence admissibility} Throughout, \(L_0\ge2\) is even, \(G=(\Z/L_0\Z)^\times\), and \(H_1,H_2\le G\) are fixed. Define \[ \mathcal P_i=\{p\nmid L_0:p\bmod L_0\in H_i\},\quad C_i=\langle\mathcal P_i\rangle,\quad \rho_i=\frac{|H_i|}{\varphi(L_0)},\quad \kappa=2-\rho_1-\rho_2. \] An empty product is one. Fix \(U\subset H_1\), \(V\subset H_2\), and coprime positive \(L_0\)-supported integers \(A,B\). Write \[ R_{U,V}^{A,B}(N;I)=\#\{Ax+By=N:x\in C_1,\ y\in C_2, \ x\bmod L_0\in U,\ y\bmod L_0\in V,\ Ax/N\in I\}. \] Omit \(I\) when \(I=[0,1]\). Put \begin{align} c_i&=\frac{1}{\Gamma(\rho_i)} \lim_{s\downarrow1}(s-1)^{\rho_i} \prod_{p\in\mathcal P_i}(1-p^{-s})^{-1},\label{eq:marginal}\\ \Gamma_N&=\frac{L_0}{|H_1||H_2|} \#\{(u,v)\in U\times V:Au+Bv\equiv N\pmod{L_0}\},\label{eq:finite}\\ \mathcal E&=\{p\nmid L_0:p\notin\mathcal P_1, \ p\notin\mathcal P_2\},\notag\\ C_{\mathcal E}&=\prod_{p\in\mathcal E} \left(1-\frac{1}{(p-1)^2}\right),\qquad J_{\mathcal E}(N)=\prod_{\substack{p\mid N\\p\in\mathcal E}} \frac{p-1}{p-2}.\label{eq:singular} \end{align} The infinite product in \eqref{eq:singular} converges absolutely and is positive, since \(2\mid L_0\). The constants \(c_i\) are positive. For an explicit convergent expression, let \(r_i(p)>1\) denote the order of \(pH_i\) in \(G/H_i\) at a forbidden good prime. Then \begin{equation}\label{eq:marginal-product} c_i=\frac{(\varphi(L_0)/L_0)^{\rho_i}}{\Gamma(\rho_i)} \prod_{\substack{\eta\in H_i^\perp\\\eta\ne1}}L(1,\eta)^{\rho_i} \prod_{\substack{p\nmid L_0\\p\notin\mathcal P_i}} (1-p^{-r_i(p)})^{1/r_i(p)}. \end{equation} Characters here are literal modulo \(L_0\). Continuation along the positive real axis from \(s>1\) fixes the powers; conjugate factors are paired. The Euler identity proving this formula is given in Section~\ref{sec:model}. \begin{theorem}\label{thm:general} Under the analytic inputs of Section~\ref{sec:inputs}, with all these data fixed, there is an explicit \(\delta>0\) depending only on \(L_0\) such that, uniformly in \(N\) and every interval \(I\subset[0,1]\), \begin{align*} R_{U,V}^{A,B}(N;I) &=c_1c_2\Gamma_N C_{\mathcal E}|I| \frac{NJ_{\mathcal E}(N)}{AB(\log N)^\kappa}\\ &\quad+O\left(\frac{NJ_{\mathcal E}(N)} {AB(\log N)^{\kappa+\delta}}\right). \end{align*} If \(\Gamma_N=0\), the count is identically zero. For \(\Gamma_N>0\) and \(I=[0,1]\), the error is a relative \(O((\log N)^{-\delta})\). More generally it is relative \(o(1)\) when \(|I|(\log N)^\delta\to\infty\). \end{theorem} The implicit constant and threshold may depend on all fixed data. An admissible numerical choice of \(\delta\) is recorded in Section~\ref{sec:assembly}, together with the localization budgets. No numerical threshold, uniformity in a growing modulus, or power-short interval assertion is included. For noncoprime fixed coefficients, first divide by their gcd when it divides \(N\). The even-modulus convention makes the prime two a bad prime throughout the analytic theorem. Passing from an odd modulus to its even multiple requires retaining this convention; freely admitting the prime two is a separate valuation-profile operation, not a change of notation. \begin{theorem}\label{thm:c12} Assume the analytic inputs of Section~\ref{sec:inputs}. Let \(\mathcal E_{12}=\{p\nmid6:p\not\equiv1\pmod{12}\}\), and set \[ A_{12}=\frac{4\pi}{\Gamma(1/4)^2} \sqrt{\frac{\log(2+\sqrt3)}{6}} \prod_{p\in\mathcal E_{12}} \left(1-\frac{2}{p(p-1)}\right),\qquad J_{12}(N)=\prod_{\substack{p\mid N\\p\in\mathcal E_{12}}} \frac{p-1}{p-2}. \] There is an explicit \(\delta_{12}>0\) such that, for all sufficiently large \(N\equiv2\pmod{12}\), \[ R_{12}(N)=A_{12}J_{12}(N)\frac{N}{(\log N)^{3/2}} \left(1+O((\log N)^{-\delta_{12}})\right). \] All other residue classes have zero count. \end{theorem} \begin{proof}[Deduction from Theorem~\ref{thm:general}] Take \(L_0=12\), \(H_1=H_2=U=V=\{1\}\), \(A=B=1\). Then \(\rho_i=1/4\) and \(\Gamma_N=12\) on the indicated class. The four characters give the exact identity \begin{align*} F_{12}(s)^4={}&\zeta(s)L(s,\chi_{-4})L(s,\chi_{-3})L(s,\chi_{12})\, (1-2^{-2s})(1-3^{-2s}) \prod_{p\in\mathcal E_{12}}(1-p^{-2s})^2. \end{align*} Here the three nonprincipal characters displayed are primitive. Use \(L(1,\chi_{-4})=\pi/4\), \(L(1,\chi_{-3})=\pi/(3\sqrt3)\), and \[ L(1,\chi_{12})=\int_0^1\frac{1-x^4}{1+x^6}\,dx =\left[\frac{1}{2\sqrt3} \log\frac{x^2+\sqrt3x+1}{x^2-\sqrt3x+1}\right]_0^1 =\frac{\log(2+\sqrt3)}{\sqrt3}. \] Consequently \[ c_{12}=\frac{1}{\Gamma(1/4)} \left(\frac{\pi^2\log(2+\sqrt3)}{54} \prod_{p\in\mathcal E_{12}}(1-p^{-2})^2\right)^{1/4}. \] In \(12c_{12}^2C_{\mathcal E_{12}}\), combine the factors by \((1-p^{-2})(1-(p-1)^{-2})=1-2/(p(p-1))\). This is exactly \(A_{12}\). The real density integrates to one; there is no extra beta-integral factor. \end{proof} The uniformity in the target also permits a coprimality restriction. Only the primes allowed in the summands can divide their gcd. Removing them changes the leading term by the following finite product. \begin{corollary}\label{cor:coprime-c12} Assume the analytic inputs of Section~\ref{sec:inputs}. Put \[ R_{\mathrm{cop}}(N)=\#\{(a,b)\in C_{12}^2:a+b=N,\ (a,b)=1\}, \qquad P_{12}(N)=\prod_{\substack{p\mid N\\p\equiv1\ (12)}}(1-p^{-1}). \] Uniformly over sufficiently large \(N\equiv2\pmod{12}\), \[ R_{\mathrm{cop}}(N) =A_{12}J_{12}(N)P_{12}(N)\frac{N}{(\log N)^{3/2}} \left(1+O((\log N)^{-\delta_{12}/2})\right). \] Every other residue class has zero count. In particular, every sufficiently large admissible target has coprime summands. \end{corollary} \begin{proof} M\"obius inversion gives the exact identity \begin{equation}\label{eq:coprime-inversion} R_{\mathrm{cop}}(N) =\sum_{\substack{d\mid N\\d\in C_{12}}}\mu(d)R_{12}(N/d). \end{equation} Indeed, division by any common divisor preserves both prime supports, and every common divisor lies in \(C_{12}\). Moreover \(d\equiv1\pmod{12}\), so \(N/d\equiv2\pmod{12}\), and \(J_{12}(N/d)=J_{12}(N)\). Write \(L=\log N\), \(\ell=\log\log(3N)\), \(P=P_{12}(N)\), and \(\delta=\delta_{12}\). We use the elementary divisor estimates \begin{align} S_0&=\sum_{\substack{d\mid N\\d\in C_{12}}}\frac{|\mu(d)|}{d} =\prod_{\substack{p\mid N\\p\equiv1\ (12)}}(1+p^{-1}) \ll\ell,\qquad P^{-1}\ll\ell,\label{eq:coprime-products}\\ S_1&=\sum_{\substack{d\mid N\\d\in C_{12}}} \frac{|\mu(d)|\log d}{d} =S_0\sum_{\substack{p\mid N\\p\equiv1\ (12)}} \frac{\log p}{p+1} \ll\ell^3.\label{eq:coprime-logmoment} \end{align} For \eqref{eq:coprime-products}, split the prime products at \(L\). Mertens' product bound controls primes at most \(L\) by \(O(\log L)\); the remaining factors have bounded product, since \(\sum_{p\mid N,\ p>L}p^{-1}\le1/\log L\). For \eqref{eq:coprime-logmoment}, the primes at most \(L\) contribute at most \(\sum_{n\le L}(\log n)/n\ll(\log L)^2\), and the others contribute at most \(L^{-1}\sum_{p\mid N}\log p\le1\). Split \eqref{eq:coprime-inversion} at \(d=\sqrt N\). The bound \(R_{12}(M)\le M\) shows that its large-divisor part is \(O(\sqrt N\,\tau(N))\), where \(\tau\) is the divisor function. For \(d\le\sqrt N\), Theorem~\ref{thm:c12} applies uniformly, and \[ (L-\log d)^{-3/2} =L^{-3/2}+O(L^{-5/2}\log d). \] Consequently, using \(\sum_{d\mid N,\ d\in C_{12}}\mu(d)/d=P\) and extending the main sum back to all divisors, \begin{align*} R_{\mathrm{cop}}(N) ={}&A_{12}J_{12}(N)\frac{N}{L^{3/2}} \left(P+O(L^{-\delta}S_0+L^{-1}S_1 +N^{-1/2}\tau(N))\right)\\ &\quad+O(\sqrt N\,\tau(N)). \end{align*} After division by the main term, the error is \[ O\!\left(L^{-\delta}\ell^2+L^{-1}\ell^4 +N^{-1/2}\tau(N)L^{3/2}\ell\right). \] Here \(J_{12}(N)\ge1\) and \(\tau(N)=N^{o(1)}\). The displayed error is \(O(L^{-\delta/2})\), proving the claim. \end{proof} \begin{remark}[Comparison with the quartic-norm target weight] The weight in \cite[Theorem 2]{iwaniecpomykala} is \(W_{12}(N)=\prod_{p\mid N}(1-\varepsilon_p/p)\), with \(\varepsilon_p=1\) on allowed primes and \(-1\) otherwise. On the admissible class, \[ \frac{J_{12}(N)P_{12}(N)}{W_{12}(N)} =\frac23\prod_{\substack{p\mid N\\p\equiv5,7,11\ (12)}} \frac{p(p-1)}{(p-2)(p+1)}\asymp1 \] uniformly in \(N\), since the factors are \(1+O(p^{-2})\). The factor \(2/3\) comes from \(2\mid N\); the prime three does not divide an admissible target. Thus the coprime asymptotic has the same order as the earlier conditional coprime bounds. \end{remark} We next clarify the congruence criterion without analytic inputs. Write \(C_H(U)=\{n\in C_H:n\bmod L_0\in U\}\). \begin{lemma}\label{lem:finite-images} A residue \(r\bmod M\) is represented by \(C_H(U)\) if and only if its reduction modulo \(\gcd(M,L_0)\) is that of some \(h\in U\), and every forbidden prime \(p\mid M\), \(p\nmid L_0\), satisfies \(p\nmid r\). \end{lemma} \begin{proof} Necessity follows from prime support. Choose such an \(h\), and for each allowed \(p^k\Vert M\) put \(e_p=\min(v_p(r),k)\). Set \(D=\prod p^{e_p}\), so \(D\in C_H\) and \((D,L_0)=1\). We seek a prime \(\ell\equiv hD^{-1}\pmod{L_0}\). Where \(e_p0\) is equivalent to solvability of \(Ax+By\equiv N\pmod M\) with actual \(x\in C_{H_1}(U),y\in C_{H_2}(V)\), for every \(M\ge1\). Under the analytic inputs of Section~\ref{sec:inputs}, for all sufficiently large \(N\), it is also equivalent to positive additive representability. \end{theorem} \begin{proof} Taking \(M=L_0\) proves necessity. For sufficiency choose a root counted by \(\Gamma_N\) and work modulo \(K=\lcm(M,L_0)\). At a prime dividing \(L_0\), at least one coefficient is a unit; adjust its variable to lift the equation while preserving the fixed unit states. At a prime outside \(L_0\), both coefficients are units and \(p\ge3\). Of the \(p\) solutions modulo \(p\), requiring a unit on each forbidden side removes at most two. A surviving pair lifts to every higher power. CRT and Lemma~\ref{lem:finite-images} realize the required two residues by actual semigroup values. Finally Theorem~\ref{thm:general} is positive for all sufficiently large targets with \(\Gamma_N>0\). \end{proof} \begin{remark}[A joint congruence obstruction]\label{rem:coupled} Take \(L_0=12\), \(H=\{1,11\}=\ker\chi_{12}\). Then \(H+H=\{0,2,10\}\pmod{12}\), so \(N\equiv4\pmod{12}\) is never a sum. Nevertheless \(H\) projects onto the full unit groups modulo four and three. The separate closures are \(\Z_2^\times\) and \(\Z_3^\times\); at other primes they are \(\Z_p\) or \(\Z_p^\times\), according as that prime is allowed or forbidden. Two units sum to every element of \(\Z_p\) for odd \(p\), and to every even element for \(p=2\). Thus these targets pass every separate \(p\)-adic test. The missing condition is the simultaneous residue state modulo twelve. \end{remark}