\section{The lower sieve and the split factors}\label{sec:sieve} Fix \(\kappa,\xi>0\) with \(4\kappa+\frac52\xi<\varepsilon/2\). We use a divisor level \(D=n^\xi\) and prime bounds \(Z=n^\eta\) and \(Y=n^{\lambda_0}\), with \(0<\lambda_0<\eta\) chosen below. Consider triples of pairwise coprime squarefree integers \(d,u,v\), supported on split primes at most \(Y\), satisfying \begin{equation}\label{eq:triples} d\mid n,\qquad (uv,n)=1,\qquad duv\le n^\kappa . \end{equation} On the shell \(m=n/d\) use the prescriptions of Lemma~\ref{lem:joint} at \(u,v\), and require \(p\nmid A_1A_2\) at every other prime \(5\le p\le Z\), including the primes of \(d\). Since \(d\equiv1\bmod6\), the target \(m\) belongs to the same family. The weight on the shell is \(W(t)=\phi((dA_1-x)/H)\), so after rescaling the norm values by \(d\) the original window is enforced. \subsection{Finite inclusion-exclusion} We give the sieve weights together with the level they require. Partition the sieving primes into blocks \(Z^{2^{-j-1}}
0\). In fact the formulas in Section~\ref{sec:local} lie uniformly below one and are \(O(1/p)\). Prime harmonic Mertens gives a bounded sum of \(1/p\) in each block. Omitting primes, including the primes of \(uv\), preserves these bounds. We also use the progression version of Mertens modulo \(3\), \(\sum_{p\le T,\ p\equiv a\ (3)}1/p=\tfrac12\log\log T+O(1)\) for \(a=1,2\); see \cite[Lemma~2.4]{LM}. In block \(j\) truncate inclusion-exclusion at consecutive orders \(k_j=2\lceil(K+j)/2\rceil\) and \(k_j-1\), where \(K\) is a fixed large integer. Denote the resulting polynomials in the bad-prime indicators by \(U_j,L_j\). Bonferroni gives \(0\le\mathbf1_j\le U_j\) and \(L_j\le\mathbf1_j\), where \(\mathbf1_j\) indicates avoidance in the block. Telescoping \(\prod_j U_j-\prod_j\mathbf1_j\) gives the lower polynomial \begin{equation}\label{eq:lower-poly} \prod_j\mathbf1_j\ge \prod_j U_j-\sum_j(U_j-L_j)\prod_{i\ne j}U_i . \end{equation} Under independent bad events with probabilities \(g_m(p)\), \[ \mathbf E U_j\ge v_j,\qquad \mathbf E(U_j-L_j)\le \frac{C^{k_j}}{k_j!}. \] Choose \(K\) so that the sum over \(j\) of \(C^{k_j}/(c k_j!)\) is at most \(1/2\). The expectation of the lower polynomial is at least \(\frac12\prod_j v_j\). Also \(0\le \mathbf E U_j-v_j\le\mathbf E(U_j-L_j)\), so the expectation of the upper polynomial \(\prod_j U_j\) is \(O(\prod_j v_j)\). We will use this upper estimate later with integer-sieve probabilities. Put \(S=\sum_{j\ge0}k_j2^{-j}<\infty\) and \(\eta=\xi/S\). Every divisor in either polynomial is at most \(Z^S=D\). The upper coefficients have absolute value at most one, and the lower coefficients at most one plus the number of nonempty blocks, which is \(O(\log\log(3n))\). Apply \eqref{eq:joint} to the lower polynomial with the nonnegative shell weights. For each triple the main contribution is at least \begin{equation}\label{eq:lower-main} \tfrac12 X_m V_m(Z)\prod_{p\mid uv}a_p,\qquad V_m(Z)=\prod_{5\le p\le Z}(1-g_m(p)). \end{equation} Here \(a_p=t_p/(2B_p)\) accounts for replacing the ordinary sieve factor at a prime of \(uv\) by the exact-valuation factor. Summing the remainders over all divisors uses \(\sum_{\ell\le D}\ell^{3/2}\ll D^{5/2}\). There are at most \(n^{\kappa+o(1)}\) triples because a fixed product \(duv\) has at most \(\tau(duv)^2\) decompositions. The total absolute remainder, for any \(\zeta>0\), is \[ O_{\phi,\zeta}\!\left( n^{1/2+4\kappa+(5/2)\xi+\zeta}\sqrt{M/H}\right) =o_\varepsilon(H/\log n). \] The equality uses \(\sqrt{nM/H}\le Hn^{-\varepsilon/2}\) and the strict choice of \(\kappa,\xi\), with \(\zeta\) small enough. All of these estimates are uniform in \(x,H\). \subsection{The logarithm gained from small factors}\label{sec:cofactor-log} The target-independent product \[ V_*(Z)= \prod_{\substack{5\le p\le Z\\ p\ {\rm split}}}B_p \prod_{\substack{5\le p\le Z\\ p\ {\rm inert}}}b_p, \qquad b_p=1-\frac2{p(p-1)} \] satisfies \(V_*(Z)\asymp(\log Z)^{-2}\), since \(B_p=1-4/p+O(p^{-2})\) and the inert factors have a convergent positive product. To compare this with \(V_m(Z)\), include the shell factor \(R_o(m)/m\) from \eqref{eq:odd-shell}. For \(p^e\Vert m\) split and \(p\le Z\), \[ \frac{\sigma(p^e)}{p^e}(1-g_m(p)) =\frac{(p-1)^2}{p(p+1)}=B_p E_p,\qquad E_p=\frac{p-1}{p-2}. \] For \(p^e\Vert m\) inert and \(p\le Z\) the combined factor is \(1+1/p=b_p E_p\), by \eqref{eq:inert}, including \(e=1\). The inert primes dividing \(m\) and \(n\) are the same. For the inert primes dividing \(m\) beyond \(Z\), the shell factor is at least \(1+1/p=b_pE_p\), and the product of the extra factors \(b_p\) is bounded below. All split enhancements are at least one. Using \(h m=(H/n)m=H/d\), we get \begin{equation}\label{eq:target-gain} X_m V_m(Z)\gg_\phi \frac H d J(n)V_*(Z) \prod_{\substack{p\le Y,\ p\ {\rm split}\\ p\mid m}}E_p . \end{equation} Temporarily drop the product bound in \eqref{eq:triples}. The sum of the cofactor weights in \eqref{eq:lower-main} and \eqref{eq:target-gain} factors over split \(p\le Y\). For \(p\nmid n\) the three options of absence, inclusion in \(u\), or inclusion in \(v\) have weights \(1,a_p,a_p\). If \(p^e\Vert n\), the two options of absence or inclusion in \(d\) have weights \(E_p,E'_p/p\), where \(E'_p=1\) for \(e=1\) and \(E'_p=E_p\) for \(e\ge2\). The compensation at target divisors is the identity \[ E_p+\frac{E'_p}{p}\ge E_p+\frac1p=1+2a_p . \] Thus the full cofactor sum is at least \(\prod_{p\le Y,\ p\ {\rm split}}(1+2a_p)\asymp\log Y\), since \(a_p=1/p+O(p^{-2})\). Normalize the local weights to independent probabilities. At \(p\nmid n\) the probability of \(p\mid duv\) is \(2a_p/(1+2a_p)\ll p^{-1}\), and at \(p\mid n\) it is at most \(1/p\). Hence \(\mathbf E\log(duv)\le C'\log Y\), uniformly in \(n\), by \(\sum_{p\le Y}(\log p)/p\ll\log Y\). Take \(0<\lambda_0<\min(\eta,\kappa/(3C'))\). Markov's inequality then retains at least half the weight on \(duv\le n^\kappa\). Inserting \eqref{eq:target-gain} into \eqref{eq:lower-main} and subtracting the aggregate remainder proves that the total retained weight is \begin{equation}\label{eq:weight} \gg_\varepsilon H J(n)\frac{\log Y}{(\log Z)^2} \gg_\varepsilon \frac{HJ(n)}{\log n}. \end{equation}