\section{Reflected targets and exact rational grids}\label{sec:reflected-comparison} This branch treats a displacement equal to the target. Unlike the fixed-shift windows in Section~\ref{sec:fixed-shift-comparison}, its rational denominators contain that target. We prove the residue-energy estimate with those denominators retained. \subsection{The exact target-grid energy} Let \(\mathcal P_y=\{p:p\le y\}\), \(y=\exp(L^{2/5})\), and \(E_0=\exp(\sqrt L)\), as in Section~\ref{sec:comparison}. The endpoint \(F\) is the centered function in \eqref{eq:gr-energy-F}. \begin{proposition}[Target-grid residue energy]\label{prop:gr-energy} Suppose that \begin{equation} q\le xL^B,\qquad \frac{Y}{qP_{\max}}\gg_A L^A \quad\text{for every fixed }A, \label{eq:gr-energy-ratio} \end{equation} where \(P_{\max}\le\exp(CL^{1/4})\) is the selected small-bank scale. For every prescribed \(D>0\), a fixed sufficiently deep density model gives \begin{equation} E_q(F):=\sum_{a\bmod q}\left| \sum_{n\equiv a\ (q)}F(n)\right|^2 \ll_D\frac{Y^2}{q}L^{-D}. \label{eq:gr-energy} \end{equation} The estimate is uniform in all valuations of \(q\). It also holds with \(F(n)\) multiplied by \(\ee{\beta n}\) when \(|\beta|Y\) is bounded by a fixed power of \(L\). \end{proposition} \begin{proof} Refine to \(Q=\lcm(q,L_0)\). Cauchy--Schwarz on the refined classes gives \(E_q(F)\le(Q/q)E_Q(F)\le L_0E_Q(F)\). For every gcd profile \(d\mid Q\), put \(r=Q/d\). Multiplicative orthogonality gives the exact identity \begin{equation} E_Q(F)=\sum_{d\mid Q}\frac1{\varphi(r)} \sum_{\chi\bmod r}\left|\sum_m F(dm)\chi(m)\right|^2. \label{eq:gr-gcdidentity} \end{equation} The character zeros impose \((m,r)=1\), so this identity retains prime powers, including primes dividing \(L_0\). We do not assume that the endpoint is prime to \(L_0\). Call a character exceptional when its primitive inducing character has conductor dividing \(L_0\). There are only a fixed number of such characters for each \(r\). In particular \(L_0\mid r\) is not required. We first identify the negligible sets in the norm actually needed. For a set \(\mathcal B\) and \(|G|\le L^C\), Cauchy--Schwarz in each residue class gives \begin{equation} E_Q(G\1_{\mathcal B}) \le(1+CY/Q)\sum_{n\asymp Y}|G(n)|^2\1_{\mathcal B}(n). \label{eq:gr-tail-energy} \end{equation} The factor is \(O(Y/Q)\) here. The following populations are exponentially small relative to \(Y\), up to fixed logarithmic factors: \begin{equation} \begin{array}{ll} n_{\mathcal P_y}>E_0:& O(YL^C e^{-cL^{1/10}}),\\ p\mid n,\ p\mid Q,\ p>y:& O(YL/y),\\ p\mid n,\ p\mid Q,\ p\text{ in the selected bank}:& O(YL/P_{\min}),\\ p^2\mid n,\ p\text{ in the selected bank}:& O(Y/P_{\min}). \end{array} \label{eq:gr-tailsets} \end{equation} For the first bound, take \(a=c/\log y\). Counting multiples of each possible smooth record gives \[ \#\{n\le CY:n_{\mathcal P_y}>E_0\} \le CY E_0^{-a}\prod_{p\le y}(1-p^{a-1})^{-1} \ll YL^C e^{-cL^{1/10}}. \] This estimate uses no property of the indicators. The other bounds follow by counting multiples, using \(P_{\min}\ge\exp(cL^{1/8})\). Thus all these sets are negligible in \eqref{eq:gr-tail-energy}. For nonexceptional characters remove only selected primes dividing \(Q\) and selected-prime squares. Do not yet impose \(n_{\mathcal P_y}\le E_0\), as that would couple the extracted prime to its cofactor. The unique selected small-bank prime gives \(n=dpm\). On a dyadic prime interval \(p\asymp P\), exact dilation invariance and the independence of \(b\) from this slot make its remaining coefficient \(\gamma_d(m)\) independent of \(p\), with \(|\gamma_d(m)|\le L^C\). It includes the condition that \(m\) has no other selected-bank prime, and every other bank weight. Its containing length is \(M\asymp Y/(dP)\). Mellin inversion of the product cutoff separates the character sum as an integral of \begin{equation} A_\chi(t)B_{d,\chi}(t),\qquad A_\chi(t)=\sum_{p\asymp P,\ p\in\mathcal P_g}\chi(p)p^{it}, \quad B_{d,\chi}(t)=\sum_{m\asymp M}\gamma_d(m)\chi(m)m^{it}. \label{eq:gr-primefactorization} \end{equation} The omitted Mellin scaling has modulus one. The transform has polynomial-logarithmic weighted \(L^1\) norm, including weight \(1+|t|\). Condition \eqref{eq:gr-energy-ratio} gives \[ M/r\asymp Y/(QP)\gg1, \] independently of \(d\). Elementary character orthogonality therefore yields \begin{equation} \frac1{\varphi(r)}\sum_{\chi\bmod r}|B_{d,\chi}(t)|^2 \le(1+CM/r)\sum_{m\asymp M}|\gamma_d(m)|^2 \ll \frac{M^2}{r}L^C. \label{eq:gr-cofactorenergy} \end{equation} Apply Lemma~\ref{lem:gr-primes} to \(A_\chi\) and Minkowski to the integral in \eqref{eq:gr-primefactorization}. Its nonexceptional energy at profile \(d\) is \begin{equation} \ll \frac{Y^2}{Qd}L^C\exp(-cL^{1/8}). \label{eq:gr-nonexceptional} \end{equation} The prime dyadic decomposition costs only a fixed power of \(L\). Summing \(d\) is harmless because \(\sum_{d\mid Q}1/d\le Q/\varphi(Q)\ll\log\log(3Q)\). There is no number-of-characters or unweighted divisor-count loss. For exceptional characters use the further restriction \(n_{\mathcal P_y}=A\le E_0\), and remove \(Q\)-prime divisors above \(y\). The errors were already controlled in energy. Now the gcd profile is determined by the smooth record: \(\gcd(A,Q)=d\), \(n=Av\), with \(v\) \(\mathcal P_y\)-rough. If \(\chi^*\) is the primitive character inducing an exceptional \(\chi\bmod r\), then on the retained profile \[ \chi(n/d)=\chi^*(A/d)\chi^*(v). \] Indeed \(\operatorname{cond}\chi^*\mid r\) and \((A/d,r)=1\). One must use \(\chi^*\) here: a literal character modulo \(L_0\) could vanish on a retained bad-prime prefix. Every other list weight depends only on \(A\), is bounded by \(L^C\), and the remaining periodic test in \(v\) has fixed period \(L_0\). The one-variable hypothesis must center this full periodic test; matching only a scalar principal mean would not suffice. There is still the condition that \(v\) avoid primes \(p\mid Q\) above \(y\). Remove it before applying the fixed-period centering lemma, extending the same bounded periodic test to those new points. At each retained record, multiples of such \(p\) number at most \(CY/(Ap)\) on the containing interval \(0y}}\frac1p \le \frac{Y}{d}L^CV_y^{-1}\frac{\log Q}{y\log y} \ll_D (Y/d)L^{-D}. \label{eq:gr-avoidance} \end{equation} We used the exact harmonic identity \begin{equation} \sum_{\substack{A\ \mathcal P_y\text{-smooth}\\d\mid A}}\frac1A =\frac1{dV_y} \label{eq:gr-record-harmonic} \end{equation} when \(d\) is \(\mathcal P_y\)-smooth; otherwise this restricted profile is empty. After this extension, hypothesis \textup{(C3)} applies record by record at length \(Y/A\). It permits arbitrary bounded fixed-period tests, arbitrary choices of retained records, and the stated smooth phases. Choose its logarithmic accuracy after the list and cutoff costs. Using \eqref{eq:gr-record-harmonic} gives, for every prescribed \(D_0\), \begin{equation} \left|\sum_mF_{\rm good}(dm)\chi(m)\right| \ll (Y/d)L^{-D_0}. \label{eq:gr-exceptional} \end{equation} Restoring the avoidance condition costs \eqref{eq:gr-avoidance}. In \eqref{eq:gr-gcdidentity} the total exceptional contribution is at most \begin{equation} \frac{Y^2}{Q}L^{-2D_0} \sum_{d\mid Q}\frac1d\,\frac{Q/d}{\varphi(Q/d)} \ll \frac{Y^2}{Q}L^{-2D_0+C}. \label{eq:gr-exceptionalenergy} \end{equation} Choose \(D_0\) large, add the negligible sets, and use the refinement inequality to prove \eqref{eq:gr-energy}. Multiplying the cutoff by \(\ee{\beta Yu}\) changes only its fixed logarithmic derivative bounds when \(|\beta|Y\le L^B\), proving the last assertion. \end{proof} This proof uses only boundedness, exact prime invariance, and the recordwise periodic means in \textup{(C1)--(C3)}. In particular no subgroup support, pole order, positivity of an individual character branch, or unit-support condition at the fixed modulus enters the energy estimate. \subsection{From target-grid energy to comparison cancellation} Specialize the common Fourier reduction of Section~\ref{sec:prime-estimates} to the reflected family \(k=T=x\). Use the product scales \(U_i\) and \(Y=xU_1U_2\) from that reduction. Choose an outer cutoff \(B_{\rm out}\), after the finite comparison family has been fixed. Dirichlet approximation with denominator ceiling \(U_1U_2/L^{B_{\rm out}-1}\) makes the multiplier smaller than any requested inverse logarithmic power unless \(T\theta\) lies within \(L^{B_{\rm out}}/(U_1U_2)\) of a rational of denominator \(r\le L^{B_{\rm out}}\). The corresponding physical frequencies are covered by grids of denominator \(q=Tr\), with widths \(O(L^{B_{\rm out}}/(TU_1U_2))=O(L^{B_{\rm out}}/Y)\). Moreover \[ \frac{Y}{qP_{\max}} \gg \exp(cL^{1/3}-CL^{1/4})L^{-B_{\rm out}}, \] so \eqref{eq:gr-energy-ratio} holds uniformly in every gcd profile. Discrete additive Parseval gives the exact normalization \begin{equation} \sum_{j\bmod q}|\widehat F(j/q+\beta)|^2 =qE_q(F\ee{\beta n}). \label{eq:gr-gridparseval} \end{equation} Integrating over width \(L^{B_{\rm out}}/Y\) and applying Proposition~\ref{prop:gr-energy} gives \(YL^{-D}\) for every desired \(D\), after demanding a larger residue-energy saving. The polynomial-logarithmic family of \(r\)'s is paid at the same step. In particular the \(T\) grid points do not create a factor \(T\): they are already included in \eqref{eq:gr-gridparseval}. Cauchy--Schwarz and Parseval on the other endpoint bound \eqref{eq:gr-fouriercomparison} on these grids by an arbitrary inverse power of \(L\). Off the grids, use \eqref{eq:gr-multiplier} and the endpoint norms. This proves that every comparison in \eqref{eq:gr-comparison} is negligible to the requested logarithmic accuracy. Finally \(\Psi(t)=t\Phi(t)\) makes \[ \frac{\Psi(|w|/x)}{|w|} =\frac{\Phi(|w|/x)}x. \] Thus the harmonic raw correlation is \(1/x\) times the ordinary banked correlation on the reflected line. The comparison at scale \(Y\) has exactly the corresponding factor \(1/Y\). Paying the finite comparison coefficients therefore proves the centered banked replacement required in the assembly. Fixed bad-supported coefficients in \(Aa+Bb=T\) cause no new analytic condition. Embed, for example, \(\widetilde F(n)=\1_{A\mid n}F(n/A)\). Its exact residue energy is \[ E_q(\widetilde F)=E_{q/(q,A)}(F), \] and its core length is divided by the fixed \(A\). All energy ratios change only by fixed factors. Bank primes are units at \(A,B\), so dilation invariance and shared-label removal persist. This embedding also avoids treating the physical summands as units modulo \(L_0\) when they are not.