\subsection{Subgroup applications}\label{sec:model} Use the separate models of Section~\ref{sec:model-preparation} for the two subgroup restrictions, with outer scale \(X=N\). Thus \(L=\log N\), \(y=\exp(L^{2/5})\), and \(E=\exp(\sqrt L)\). We evaluate their correlation and verify the positive record envelope. \subsubsection{The model correlation} Use separate constructions for the two subgroups with common \(\mathcal P_y\). \begin{proposition}\label{prop:model-correlation} For a smooth \(0\le\Phi\le1\) supported away from zero and one, \begin{align*} &\sum_{Ax+By=N}\Phi(Ax/N)\mathcal D_U(x)\mathcal D_V(y)\\ &\quad=c_1c_2\Gamma_N C_{\mathcal E}J_{\mathcal E}(N) \frac{N}{AB L^\kappa} \left(\int_0^1\Phi(t)\,dt+O(L^{-1/2})\right)+O(N^{1/2}). \end{align*} The statement is uniform in target valuations. Endpoint strips of relative length \(\epsilon\) have upper bound \(O(\epsilon NJ_{\mathcal E}(N)L^{-\kappa}+N^{1/2})\). For every interval \(I\subset[0,1]\), the unsmoothed model correlation has the same leading coefficient multiplied by \(|I|\), with absolute error \(O(NJ_{\mathcal E}(N)L^{-\kappa-\delta_0})\), where \(\delta_0=1/(8192\varphi(L_0))\). \end{proposition} \begin{proof} Discard records larger than \(E\) using \eqref{eq:record-tail}. The rough logarithms on the rest are \(L+O(\sqrt L)\), giving \[ \frac{M_{R,K}(n_{>y})}{V_y} =\frac{c_H L^{\rho-1}}{\alpha_0E_-(y)}(1+O_K(L^{-1/2})). \] Count the remaining small-prime sieve set in the fixed residues. The fundamental lemma applies at level \(N^{\epsilon_0}\), for small fixed \(\epsilon_0>0\), with \(u=\log(N^{\epsilon_0})/\log y\asymp L^{3/5}\). Its relative error is \(e^{-cu\log u}\), and CRT gives endpoint error \(O(\nu(d))\) at each squarefree sieving modulus \(d\). Summing those errors costs \(N^{\epsilon_0}L^{O(1)}\). This is the standard fixed-dimensional fundamental-lemma application \cite[Theorem 11.7]{richert}; target-root mergers decrease its dimension and preserve uniformity. The positive integer line is \((x,y)=(x_0+Bt,y_0-At)\), with parameter length \(N/(AB)\). Modulo \(L_0\), this parameter bijects onto all roots: at each prime at least one coefficient is a unit. If \(R\) roots have the prescribed states, their raw probability is \(R/L_0\). Multiplying the normalized marginal projections gives \[ \frac{R}{L_0}\frac{L_0c_1}{|H_1|}\frac{L_0c_2}{|H_2|} =c_1c_2\Gamma_N. \] A prime forbidden on only one side has relative factor one. A prime forbidden on both sides has two excluded roots, merging exactly when \(p\mid N\), and relative factor \[ \frac{1-(2-\1_{p\mid N})/p}{(1-1/p)^2}. \] Their product up to \(y\) is \(C_{\mathcal E}(y)J_{\mathcal E,y}(N)\). The ordinary tail is \(O(1/y)\), and the target-prime tail is \(O(L/(y\log y))\). Partial summation inserts \(\Phi\), costing its total variation. The same upper sieve on an endpoint strip gives its length times the sparse density, with a power-smaller endpoint error. To obtain the uniform interval assertion, put \(\epsilon=L^{-\delta_0}\) and remove strips of this width near zero and one. On the rest the rough logarithms are \(L+O(\sqrt L+\delta_0\log L)\), giving the same relative \(O(L^{-1/2})\) replacement. The fundamental-lemma count in any remaining interval has the same CRT endpoint error; partial summation costs the total variation, bounded independently of the interval. Restoring the strips costs \(O(\epsilon)\) of the sparse scale. All other tails are superlogarithmic or power-smaller, proving the claim. \end{proof} \subsubsection{Verification of the comparison hypotheses} The indicators lie between zero and one, and the model bound after \eqref{eq:corrected-model} gives \(\mathcal D_U\le1\) eventually for each fixed depth. Thus (C1) holds, with a threshold allowed to depend on that depth. The full-unit-group exact model also lies between zero and one. Choose banks of primes \(p\equiv1\pmod{L_0}\), with small banks in \([\exp(L^{1/8}),\exp(L^{1/4})]\) and big banks in \([\exp(L^{1/3}),\exp(L^{2/5})]\). Delete primes dividing \(N\). Every bank prime is a unit at \(A,B\), belongs to both allowed sets, and preserves the fixed states. Hence \begin{equation}\label{eq:exact-invariance} f_U(pn)=f_U(n),\quad \mathcal D_U(pn)=\mathcal D_U(n),\quad \xi_U(pn)=\xi_U(n) \end{equation} exactly, including higher powers. The deletion changes reciprocal mass by at most \(O(L/(P_{\min}\log P_{\min}))\), which is smaller than every fixed inverse logarithmic power. This verifies (C2). For (C3), fix an \(\mathcal P_y\)-smooth record \(r\le E\). If \(f(r)=0\), both the indicator and model on that record vanish. Otherwise \(r\bmod L_0\in H\), and insert the residual state \(U/r\) into the periodic test of Lemma~\ref{lem:rough-means}. Its smooth scale is \(Y/r\), so the error is \((Y/r)L^{-D}\), as required. An additional periodic test of period at most \(L^B\) increases the period by only a fixed factor. For every required accuracy, the Hankel remainder works for every sufficiently large fixed depth; constants and thresholds, rather than earlier graph exponents, may depend on that depth. This proves (C3) with all surviving character branches retained. It remains to check the positive envelope (C4). Write \(\mu_{\mathcal B}\) for the law \eqref{eq:bank-law}. \begin{lemma}\label{lem:envelope} For a nonnegative record function \(F(r,s)\le L^B\), and either indicator or model on each side, its weighted correlation on the positive line restricted by \(Ax\in J\subset[0,N]\) is at most \[ C\frac{|J|J_{\mathcal E}(N)}{AB L^\kappa}\E_{\mu_{\mathcal B}}F +NL^{O(1)}e^{-cL^{1/10}}+O(N^{1/2}). \] The constant may depend on the fixed coefficients. \end{lemma} \begin{proof} Discard records \(r,s>E\) using \eqref{eq:record-tail} and bounded weights. Their total contribution is \(NL^{O(1)}e^{-cL^{1/10}}\). There are at most \(E^2\) retained record pairs. A shared bank prime makes a stratum empty, since every bank prime avoids \(N\). For coprime \(r,s\), their divisibilities select one progression of length \(|J|/(ABrs)\). At \(p^e\Vert r\), write \(x=r\ell\). Exact valuation excludes \(\ell\equiv0\pmod p\), while \(y\) is automatically a unit because \(p\nmid N\). The same argument applies to \(s\). At an unused bank prime there are two distinct excluded roots. Thus the complete bank factor, including the progression density, is \[ \frac1{rs}\prod_{\substack{p\in\mathcal B\\p\nmid rs}}(1-2/p) \prod_{p\mid rs}(1-1/p) =\mu_{\mathcal B}(r,s). \] All forbidden-prime slopes remain units. Apply the upper Selberg sieve with \(z=N^{1/10}\) and primes up to \(w=z^{1/16}\), which contains every bank for sufficiently large \(N\). The nonbank sieve denominators are bounded below, up to fixed positive constants, by \[ \begin{array}{c|c} \text{weights}&\text{sieve denominator}\\ \hline f_1,f_2&(\log w)^\kappa/J_{\mathcal E}(N),\\ \mathcal D_1,f_2& (\log y)^{1-\rho_1}(\log w)^{1-\rho_2}/J_{\mathcal E}(N),\\ \mathcal D_1,\mathcal D_2& (\log y)^\kappa/J_{\mathcal E}(N). \end{array} \] The middle row also applies with the sides interchanged. A model excludes its forbidden primes up to \(y\); an indicator excludes them up to \(w\). A prime forbidden on only one side enters only that side's denominator. Target-root mergers cost at most the displayed factor \(J_{\mathcal E}(N)\). Each model bound contributes \(L^{-3(1-\rho_i)/5}\), so all three rows give the same sparse factor \(L^{-\kappa}J_{\mathcal E}(N)\). Consequently every retained stratum, for \(w_i\in\{f_i,\mathcal D_i\}\), has the single uniform estimate \[ \sum_{\substack{Ax+By=N,\ Ax\in J\\ x_{\mathcal B}=r,\ y_{\mathcal B}=s}} w_1(x)w_2(y) \le C_K\frac{|J|}{AB}\mu_{\mathcal B}(r,s) \frac{J_{\mathcal E}(N)}{L^\kappa} +O_K(z^{4+\epsilon}), \] where \(\epsilon>0\) is arbitrarily small and fixed. The remainder follows from the squared-weight CRT root counts. Multiplying by \(F(r,s)\le L^B\) and summing at most \(E^2\) retained records leaves \[ L^BE^2z^{4+\epsilon}=O(N^{1/2}) \] after fixing, for example, \(\epsilon=1/2\). Summing the main terms gives \(\mathbb E_{\mu_{\mathcal B}}F\). The endpoint exponent and sieve level are independent of the final model depth; constants and thresholds may depend on it. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:general}] The preceding checks verify (C1)--(C4), with \(\mathcal S_N=NJ_{\mathcal E}(N)/(AB L^\kappa)\). This scale is between \(NL^{-C}\) and \(NL^C\) for a fixed \(C\), by \(J_{\mathcal E}\ge1\) and Mertens' upper product bound. Theorem~\ref{thm:model-comparison} replaces both indicators by their models with error \(O(\mathcal S_NL^{-\delta_0})\), uniformly in the interval. Proposition~\ref{prop:model-correlation} evaluates that model count with exactly the coefficient in the theorem. If \(\Gamma_N=0\), the residue restrictions give zero identically. Otherwise its finitely many positive values have a fixed positive minimum, yielding the relative assertion. \end{proof}