\documentclass[11pt,reqno]{amsart} \usepackage{amsmath,amssymb,amsthm,mathtools} \usepackage[margin=1.15in]{geometry} \usepackage{microtype} \usepackage[hidelinks]{hyperref} \usepackage[utf8]{inputenc} \emergencystretch=1.5em \numberwithin{equation}{section} \theoremstyle{plain} \newtheorem{theorem}{Theorem}[section] \newtheorem{lemma}[theorem]{Lemma} \newtheorem{proposition}[theorem]{Proposition} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{remark}[theorem]{Remark} \newtheorem{example}[theorem]{Example} \theoremstyle{remark} \newtheorem*{notation}{Notation} \DeclareMathOperator{\ord}{ord} \DeclareMathOperator{\rad}{rad} \newcommand{\bZ}{\mathbb{Z}} \newcommand{\bQ}{\mathbb{Q}} \newcommand{\bR}{\mathbb{R}} \newcommand{\bF}{\mathbb{F}} \newcommand{\cB}{\mathcal{B}} \newcommand{\cC}{\mathcal{C}} \newcommand{\cW}{\mathcal{W}} \newcommand{\cQ}{\mathcal{Q}} \newcommand{\fp}{\mathfrak{p}} \title[Wieferich co-occurrence along exponential sequences]{Wieferich co-occurrence along exponential sequences:\\a weighted tail estimate and its application to $(\star')$} \author{Chunping Zeng} \date{} \begin{document} \begin{abstract} Fix an integer $r\ge 2$ and a bound $A\ge 1$, and write $b_e=r^e-a'$ for $a'\le A$. A recent preprint (hexagon:2610.00630, Zeng) reduces the $abc$ conjecture, within a lattice-theoretic framework, to a single remaining hypothesis $(\star')$. We show that on the $k=2$ slice ($\omega(ab)=2$), the density-zero requirement on the exceptional set attached to $(\star')$ follows from one easily quoted sufficient condition $K_{r,a'}<\infty$ --- the exact threshold of the first-moment route being the strictly weaker $\Phi_r(E)=o(E)$ (\S\ref{sec:threshold}) --- where \[ K_{r,a'}=\sum_{q\in\cC_{r,a'}}s_q\log q\Big(\tfrac{1}{D_3(q)}+\tfrac{1}{e_0(q)}\Big), \] $\cC_{r,a'}$ is the set of primes passing a three-filter congruence sieve, $D_3(q)=\ord_{q^3}(r)$, and $e_0(q)$ is the least solution of $q^3\mid r^e-a'$. The condition $K_{r,a'}<\infty$ holds under $P_4+P_{6a}$, where $P_4$ is convergence of the classical Wieferich reciprocal sum and $P_{6a}$ is the finiteness of super-Wieferich primes; the remaining gap is isolated in a single weighted hypothesis (H-j) on the ``bad set'' $\{q:e_0(q)3$, \[ r^{e(q-1)}=(1+q^tw)^e\equiv1+e\,q^tw\pmod{q^3}. \] If $s_q\ge3$ then $t=3$ and $r^{q-1}\equiv1\pmod{q^3}$, so $a'^{q-1}\equiv1\pmod{q^3}$. If $s_q=2$ then $a'^{q-1}\equiv1+eq^2w\pmod{q^3}$, giving $q^2\mid a'^{q-1}-1$ and $(a'^{q-1}-1)/q^2\equiv ew\pmod q$; solving for the least representative gives $e_0\equiv\beta_qw_q^{-1}\pmod q$, and $\beta_q=0$ forces $e_0\equiv0\pmod q$, i.e.\ $e_0\ge q$. Finally $q^3\le\lvert r^{e_0}-a'\rvert\le r^{e_0}+\lvert a'\rvert$ (using $b_{e_0}\ne0$) yields $e_0\ge\frac{3\log q}{\log r}-O_{r,a'}(1)$. $\square$ \section{\texorpdfstring{Proof of Theorem~\ref{thm:main}}{Proof of Theorem 1.3}}\label{sec:mainproof} We need two lemmas. \begin{proof} Partition $[1,E]$ into dyadic blocks $(E/2^{j+1},E/2^j]$, $0\le j<\lceil\log_2E\rceil$. Inside block $j$ one has $e>E_j:=E/2^{j+1}$, hence the exceptional condition $T_s(e)\ge ce>cE_j$; Markov with the fixed threshold $cE_j$ gives at most $S(E/2^j)/(cE_j)\le(CE/2^j)/(cE/2^{j+1})=2C/c$ exceptions per block. \end{proof} \begin{lemma}[Correct counting]\label{lem:count} For $q\in\cC_{r,a'}$ and $E\ge1$, \[ N_q(E):=\#\{e\le E:q^3\mid b_e\}\ \le\ \frac{E}{D_3(q)} +\mathbf 1_{\{e_0(q)\le E\}}\ \le\ E\Big(\frac{1}{D_3(q)}+\frac{1}{e_0(q)}\Big). \] \end{lemma} \begin{proof} The solution set of $q^3\mid r^e-a'$ is the single class $e\equiv e_0\pmod{D_3}$ with $e_0\le D_3$, so $N_q(E)=\lfloor(E-e_0)/D_3\rfloor+1$ if $e_0\le E$ and $0$ otherwise; the second inequality uses $\mathbf 1_{\{e_0\le E\}}\le E/e_0$. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:main}] By Theorem~\ref{thm:two-base}, $N_q(E)>0$ forces $q\in\cW_r\cap\cW_{a'}$ with $a'\in\langle r\rangle\bmod q^3$, i.e.\ $q\in\cC_{r,a'}$; hence \[ S(E)=\sum_{q\in\cC_{r,a'}}s_q\log q\,N_q(E) \le E\sum_{q\in\cC_{r,a'}}s_q\log q\Big(\frac1{D_3(q)}+\frac1{e_0(q)}\Big) =E\,K_{r,a'}, \] For the density bound, the moving threshold is handled directly: with $c=\log r/20$, \[ \#\{e\le E: T_s(e)\ge ce\}\ \le\ \frac1c\sum_{e\le E}\frac{T_s(e)}{e},\qquad \sum_{e\le E}\frac{T_s(e)}{e} =\sum_{q\in\cC_{r,a'}}s_q\log q\sum_{\substack{e\le E\\ q^3\mid b_e}}\frac1e \ \le\ (1+\ln E)\,K_{r,a'}, \] the inner sum over the class $e\equiv e_0\pmod{D_3}$ being at most $(1+\ln E)\big(1/D_3(q)+1/e_0(q)\big)$; dividing by $cE$ gives the claim. The argument never uses $D_3\le E$: the ``$+1$'' term is absorbed by $1/e_0$. \end{proof} \begin{remark}[$K_{r,a'}<\infty$: sufficient, not necessary]\label{rem:sufficient} The condition $K_{r,a'}<\infty$ is the most quotable member of a family of sufficient conditions, not the exact requirement: the threshold of the first-moment route is the strictly weaker \textup{(M)} $\sum_{e\le E}T_s(e)/e=o(E)$ (\S\ref{sec:threshold}), which allows $\Xi_r^{(3)}=\infty$. \end{remark} \begin{proposition}[Splitting and sufficient conditions]\label{prop:split} Write $\Xi_r=\sum_{q\in\cW_r}\frac{s_q\log q}{D_3(q)}=\Xi_r^{(2)}+\Xi_r^{(3)}$. Then $\Xi_r^{(2)}$ is finite \emph{unconditionally} (Theorem~\ref{thm:xi2}), and $\Xi_r^{(3)}=\sum_{q\in\cW_r^{(3)}}\frac{s_q\log q}{d_q}\le(\log r)\,\#\cW_r^{(3)}$ by \eqref{eq:U}. Hence $P_{6a}\Rightarrow\Xi_r<\infty$; the precise squeeze around $\Xi_r^{(3)}$ and the barriers to its unconditional proof are given in \S\ref{sec:xi3}. If $a'=r^k$ then $b_e=r^k(r^{e-k}-1)$ and $T_s^{(a')}(e)=T_s^{(1)}(e-k)$; compatibility $a'\in\langle r\rangle\bmod q^3$ is automatic, and $q\in\cW_r\Rightarrow q\in\cW_{r^k}$ by $q_q(r^k)\equiv k\,q_q(r)\pmod q$, so $\cC_{r,r^k}=\cW_r$. Moreover $e_0(q)=k+D_3(q)\ge D_3(q)$, so $K_{r,r^k}\le2\Xi_r$: \emph{Case A is closed unconditionally on (H-j)}. If $a'\notin r^{\bZ}$ then $e_0\le D_3$ and $K_{r,a'}\le2J'_{r,a'}$ where $J'_{r,a'}=\sum_{q\in\cC_{r,a'}}\frac{s_q\log q}{e_0(q)}$; in particular $\Xi_{r,a'}:=\sum_{\cC_{r,a'}}\frac{s_q\log q}{D_3(q)}\le J'_{r,a'}$ is a consequence, not an independent hypothesis. \end{proposition} For $a'=1$ the count is exact with no boundary term: $V_q(e)=s_q+v_q(e/d_q)$ for $d_q\mid e$, so $\{e:V_q(e)\ge3\}=D_3(q)\bZ$ and $S(E)=\sum_{q\in\cW_r}s_q\log q\,\lfloor E/D_3(q)\rfloor\le E\Xi_r$. \section{\texorpdfstring{The bad set: proof of Theorem~\ref{thm:badset}}{The bad set: proof of Theorem 1.4}}\label{sec:badset} For the layer $s_q=2$ we have $D_3=d_qq$. By Theorem~\ref{thm:two-base}, $e_0\equiv b_q\pmod q$ with $b_q=\beta_qw_q^{-1}\bmod q\in[0,q)$; writing $e_0=b_q+j_qq$ gives $j_q\in[0,d_q)$ (since $e_0\le D_3=d_qq$). If $j_q\ge1$ then $e_0\ge q$ and $\frac{s_q\log q}{e_0}\le\frac{s_q\log q}{q}$; summing over the layer $s_q=2$ gives at most $2\Sigma_r$; the layer $s_q\ge3$ is controlled separately by $P_{6a}$, via $s_q\log q/e_0\le s_q\log q\le d_q\log r$ and the finiteness of $\cW_r^{(3)}$. If $j_q=0$ then $e_0=b_q \frac{2\log3}{3}$. For the layerwise finiteness, $e_0(q)=b$ forces $q^3\mid r^b-a'$ with $q>b$, and $q$ is a cube-divisor prime of the fixed integer $r^b-a'$; each layer is finite and enumerable. $\square$ \begin{proposition}[Closing conditions]\label{prop:close} (H-j) follows from the pair \begin{itemize} \item[\textup{(SC1)}] $\#\cB_{r,a'}(x)\ll(\log x)^A$ for some $A<1+\varepsilon$, where $\varepsilon>0$ is as in \textup{(SC2)}; \item[\textup{(SC2)}] $e_0(q)\ge(\log q)^{1+\varepsilon}$ for all but finitely many $q\in\cB_{r,a'}$; \end{itemize} \emph{with $A$ and $\varepsilon$ linked}: Abel summation gives $\sum_{q\in\cB(x)}(\log q)^{-1-\varepsilon}\ll\int^{\log x}u^{A-2-\varepsilon}\,du$, which converges if and only if $A<1+\varepsilon$. In particular the cleanest sufficient form is $\#\cB_{r,a'}(x)\ll\log x$ with any $\varepsilon>0$. Note (SC1) implies a P2-type bound for $N_r(x)$ (the bad set is a subset of $\cW_r$); it is an \emph{intersection-strengthened} version of $P_2$, not a weakening. \end{proposition} \begin{remark}[Gallagher skeleton]\label{rem:gallagher} One has $\#\cB_{r,a'}(x)=\lvert\mathcal M(x)\rvert$ in the layer $s_q=2$ (each $q$ contributes at most one $b\in[1,q)$ since $D_3=d_qq>q$). Expanding the Wieferich indicator as $\mathbf 1_{\cW_r}(q)=\frac1q\sum_{t\bmod q}e\big(\frac{t\,q_r(q)}{q}\big)$ splits $\sum_{b}\#\{\cQ_b\cap\cW_r\}$ into a main term $\sum_b\frac1q\lvert\cQ_b\rvert$ (whose random expectation is $\asymp\sum_q q^{-4}$) and a remainder of varying-modulus Fermat-quotient character sums. A mean-value hypothesis \[ \textup{(FQM-avg)}\qquad \sum_{q\le x}\Big\lvert\sum_{q'\in\cQ_b}e\Big(\frac{t\,q_r(q')}{q'}\Big)\Big\rvert \ll x^{\theta}\quad(\theta<1), \] uniform in the truncated range $b\le B_0=3\log x/\log r$. The \emph{main term} is unconditionally $O(\log x\cdot\log\log x)$: $\sum_{b\le B_0}\sum_{\substack{q\in\cQ_b\\q>b}}q^{-1} \le\sum_{b\le B_0}\lvert\cQ_b\rvert/b\le\frac{B_0\log r}{\log8}\sum_{b\le B_0}\frac1b\ll\log x\cdot\log\log x$. Hence (FQM-avg) yields the \emph{weighted} bound $\sum_{q\in\cB(x)}q^{-1}=O(\log x)$ (not the unweighted (SC1)); and the weighted bound combined with (SC2) closes (H-j), since $\sum_{\cB(x)}(\log q)^{-1-\varepsilon}\ll\cB_1(x)(\log x)^{-1-\varepsilon}+ \int^{x}\cB_1(t)\,\frac{dt}{t(\log t)^{2+\varepsilon}}$ with $\cB_1(t)=\sum_{q\in\cB(t)}q^{-1}=O(\log t)$ converges for every $\varepsilon>0$. The small-$b$ range $b\le B_0$ is covered by the trivial $O((\log x)^2)$ bound. Fixed-modulus Fermat-quotient sums are understood (Heath--Brown, Shparlinski); (FQM-avg) is the varying-modulus input, at the same depth as Katz's conjecture. \end{remark} \section{\texorpdfstring{The unconditional layer: $\Xi_r^{(2)}<\infty$}{The unconditional layer: Xi2 finite}}\label{sec:xi2} Split the governing weight by depth: \[ \Xi_r=\Xi_r^{(2)}+\Xi_r^{(3)},\qquad \Xi_r^{(2)}=\sum_{\substack{q\in\cW_r\\s_q=2}}\frac{2\log q}{d_qq},\qquad \Xi_r^{(3)}=\sum_{\substack{q\in\cW_r\\s_q\ge3}}\frac{s_q\log q}{d_q}. \] \subsection{Block decomposition}\label{sec:block} For each integer $d\ge1$ set $S_d=\{q\in\cW_r:\text{$d_q=d$, $s_q=2$, $q$ odd}\}$; the prime $q=2$ is listed separately and contributes a finite $O(\log r)$ term. \begin{lemma}[LTE identification]\label{lem:sdident} For $q\in S_d$ we have $d\mid q-1$ and \[ v_q(r^{q-1}-1)=v_q(r^d-1)+v_q\!\Big(\frac{q-1}{d}\Big)=v_q(r^d-1), \] since $0<(q-1)/dn\exp(\frac{\log n}{104\log\log n})$ --- the same; Balaji--Luca \cite{balajiluca}, Theorem~1.2, give the closest unconditional sparsity, $\#\mathcal A_{K,c,u}\cap[1,N]\ll N\exp(-\frac{\log N}{156\log\log N})$, All four inputs control $\Delta_d$ (including $\mathrm{BOOK}_d$) and yield only $\Delta_d/d=O(1)$, i.e.\ $\Phi_r(D)=O(D)$; the Chen--Ding/Graves--Murty bound $C_n'\gg a^{\varphi(n)-\varepsilon n}$ controls the squarefree part of $\Phi_n$ and gives $\mathrm{PRIM}_d\le\varepsilon d\log r$, still merely $\Phi_r(D)=O(\varepsilon D)$. None reaches $o(D)$. \item \emph{The DeKoninck--Doyon dichotomy points the other way} \cite{dekoninckdoyon}: its other branch is ``small radical $\Rightarrow$ $\cW$ infinite''. \end{enumerate} Conclusion: a family of sufficient conditions remains, of which the most easily quoted is \[ \underbrace{\Xi_r^{(2)}<\infty}_{\text{Theorem~\ref{thm:xi2}, unconditional}} \ +\ \underbrace{\Xi_r^{(3)}<\infty}_{\text{$P_{6a}$ sufficient; not necessary}} , \] and the exact threshold of the unconditional question, together with strictly weaker rungs ($\Xi_r^{(3)}(E)=o(E/\log E)$, the block-counting condition \textup{(C)}), is determined in \S\ref{sec:threshold}. \end{proposition} \begin{theorem}[Radical identity]\label{thm:radical} For every $r\ge2$ and $D\ge2$, with $\Delta_d=\log(r^d-1)-\log\rad(r^d-1)$, \begin{equation}\label{eq:radical} \sum_{d\le D}\frac{\Delta_d}{d} =\sum_{q\in\cW_r^{(3)}}\frac{(s_q-1)\log q}{d_q}\,H_{\lfloor D/d_q\rfloor} +R_2(D)+\bigl(S_{\mathrm{ord}}+o(1)\bigr)\log D, \end{equation} where $R_2(D):=\sum_{q\in\cW_r^{(2)}}\frac{\log q}{d_q}H_{\lfloor D/d_q\rfloor} \le\frac{\log r}{2}\sum_{c\le D}H_{\lfloor D/c\rfloor}=O(D)$ by Lemma~\ref{lem:mass}, and $R_2(D)=O(\log D)$ whenever $\cW_r^{(2)}$ is finite (in particular for every base examined). Hence \[ \frac23\,\Phi_r(D)\ \le\ \sum_{d\le D}\frac{\Delta_d}{d}\ \le\ \Phi_r(D)+R_2(D)+O(\log D), \] so $\sum_{d\le D}\Delta_d/d=o(D)\Rightarrow\Phi_r(D)=o(D)$ unconditionally, and if $R_2(D)=o(D)$ --- in particular if $\cW_r^{(2)}$ is finite --- then $\Phi_r(D)=o(D)\iff\sum_{d\le D}\Delta_d/d=o(D)$. \end{theorem} \begin{proof} By \eqref{eq:rad-split}, swapping the order of summation, \[ \sum_{d\le D}\frac{\mathrm{PRIM}_d}{d} =\sum_{c\le D}\frac{\Delta_c^{\Phi}}{c}H_{\lfloor D/c\rfloor} =\sum_{p\in\cW_r}\frac{(s_p-1)\log p}{d_p}H_{\lfloor D/d_p\rfloor} \] (primes with $s_p=1$ contribute $0$), and, since only primes $p\le D$ can divide a $d\le D$ with $d_p\mid d$, \[ \sum_{d\le D}\frac{\mathrm{BOOK}_d}{d} =\sum_{p\le D}\frac{\log p}{d_p}\sum_{m\le D/d_p}\frac{v_p(m)}{m} =\bigl(S_{\mathrm{ord}}+o(1)\bigr)\log D, \] using $\sum_{m\le M}v_p(m)/m=\log M/(p-1)+O_p(1)$ and the convergence of $S_{\mathrm{ord}}=\sum_p\frac{\log p}{d_p(p-1)}$ (Theorem~\ref{thm:sord}). Splitting the primitive sum by depth: the coefficients $(s_q-1)/s_q\in[2/3,1)$ on $\cW_r^{(3)}$ give the $[\frac23,1)\cdot\Phi_r(D)$ part, and the remaining $s_q=2$ part is $R_2(D)$, bounded by $O(D)$ via Lemma~\ref{lem:mass} exactly as in Theorem~\ref{thm:phicap}. \end{proof} \begin{remark}[Numerics]\label{rem:radnum} For $r=3$, $d=20$: $\Delta_{20}=6.0868=\mathrm{PRIM}\ 3.0910+\mathrm{BOOK}\ 2.9957$ ($3^{20}-1=2^4\cdot5^2\cdot11^2\cdot61\cdot1181$; the only Wieferich contribution is $\Delta_5^{\Phi}=\log11$). The bookkeeping part grows like $S_{\mathrm{ord}}\log D$ ($\mathrm{BOOK}=3.058/4.113/5.340/6.491$ at $D=10^3/3\cdot10^3/10^4/3\cdot10^4$), while the layer-specific bound $\Xi_d^{(3)}\le\frac{3}{2d}\sum_{c\mid d}\Delta_c^{\Phi,(3)}$ keeps the $s_q=2$ layer out: charging $\mathrm{PRIM}_d$ in full would add $\frac{\log11}{5}H_{\lfloor D/5\rfloor}=3.345$ at $D=3000$ --- an $O(\log D)$, not $O(D)$, term, but one that must not be charged to $\cW_r^{(3)}$. \end{remark} \subsection{\texorpdfstring{An almost-sure route: the double-zero hypothesis (D4)}{An almost-sure route: the double-zero hypothesis (D4)}}\label{sec:d4} \begin{definition}[(D4)]\label{def:d4} There is a constant $C_r>0$ such that for every odd prime $q\nmid r$, \[ \mathbb P\big(s_q\ge3\big)=\mathbb P\big(q^3\mid r^{q-1}-1\big) =\mathbb P\big(Q^{(1)}_q=0,\ Q^{(2)}_q=0\big)\le\frac{C_r}{q^2}. \] A pointwise bound for the Katz quotient at the double-zero scale. (D4) is \emph{not} a consequence of Katz's 2015 conjecture (continuous $\mathbb R/\mathbb Z$ equidistribution): attempting to place $(Q^{(1)},Q^{(2)})$ in the group-scheme framework $G=\mathbb G_m^2$, the cyclic subgroup generated by a single base lies in a proper subtorus and is likely excluded by the Zariski-density hypothesis of Katz's Lemma~4.2 (isomorphic to the $a'=r^k$ degeneration of $(\overline W_r,\overline W_{a'})$). Hence (D4) must be listed as an independent hypothesis, and it must be distinguished from (D2): (D2) is ``two bases $\times$ first quotient'' modulo $q^2$ joint, (D4) is ``one base $\times$ first and second quotients''; both give $1/q^2$, but (D2) does not imply (D4). \end{definition} \paragraph{Ensemble separation (why (D4) is not free).}\label{sec:ensembles} Let $\mathcal P=\prod_{q\ge3}(\bZ/q^3\bZ)^\times$ carry its product Haar measure $\nu=\bigotimes_q\nu_q$, and let $T_q=\mu_{q-1}(\bZ/q^3\bZ)$. Then $\nu_q(T_q)=|\mu_{q-1}|/|(\bZ/q^3)^\times|=(q-1)/(q^2(q-1))=1/q^2$ \emph{exactly}, and with the independent indicators $\mathbf 1_{T_q}$ \[ \lambda=\sum_{q\ge3}q^{-2}=P(2)-\tfrac14=0.2022475\ldots,\qquad \nu(K=0)=\prod_{q\ge3}\bigl(1-q^{-2}\bigr)=\frac{8}{\pi^2}=0.8105695\ldots , \] where $K=\#\{q:x_q\in T_q\}$; the exact law of $K$ is Poisson-binomial (the Poisson value $e^{-\lambda}=0.8168927$ is only an approximation, off by $0.78\%$), and Borel--Cantelli gives $K<\infty$ $\nu$-a.s. Thus in the \emph{Haar ensemble} the pointwise bound of Definition~\ref{def:d4} holds with $C_r=1$: (D4) is a free theorem of product measure, and Theorem~\ref{thm:d4} becomes an a.s.\ statement about a $\nu$-random element of $\mathcal P$. The primal question, however, concerns the single point \[ \iota(r)=(r\bmod q^3)_{q\ge3}\in\mathcal P,\qquad \iota:\bZ\hookrightarrow\mathcal P , \] the diagonal embedding of the integers. Its image is countable in an uncountable compact group, hence $\nu(\iota(\bZ))=0$. Every $\nu$-a.s.\ statement is therefore void on $\iota(r)$, and no product-measure argument can yield a statement about a fixed integer $r$. This resolves the apparent tension between the counting-density reading of (D4) (equivalent to $P_{6a}$, Theorem~\ref{thm:mq}) and the probabilistic reading (Definition~\ref{def:d4}, finite mean, strictly stronger): the two readings live in \emph{different ensembles}, and \[ (\mathrm{D}4)_{\rm Haar}\ \not\Longrightarrow\ (\mathrm{D}4)_{\rm diag}. \] What is missing is precisely a \emph{pointwise rigidity of $\iota(r)$ relative to $\bigotimes\nu_q$}: the assertion that a fixed integer has no systematic preference among the index-$q^2$ Teichm\"uller subgroups $T_q$ as $q$ varies. This is the exact content of specification S3 (and of the dual$\to$primal transfer of Proposition~\ref{prop:dual}); it is neither a consequence of Katz's continuous $\mathbb R/\mathbb Z$ equidistribution nor of the free Haar statement above. \begin{proposition}[Teichm\"uller anchor]\label{thm:teich} For odd $q\nmid r$, \[ q\in\cW_r^{(3)}\ \iff\ q^3\mid r^{q^2}-r\ \iff\ r\bmod q^3\in\mu_{q-1}\bigl((\bZ/q^3\bZ)^\times\bigr), \] where $(\bZ/q^3\bZ)^\times\cong C_{q-1}\times C_{q^2}$ and $|\mu_{q-1}|=q-1$: the subgroup has relative size exactly $1/q^2$. \end{proposition} \begin{proof} Write $r=\omega u$ with $\omega^{q-1}=1$ (the Teichm\"uller part) and $u\in1+q\bZ_q$ (the principal unit). Then $r^{q^2}=\omega^{q^2}u^{q^2}\equiv\omega\pmod{q^3}$, since $q^2\equiv1\pmod{q-1}$ and $(1+qx)^{q^2}\equiv1\pmod{q^3}$. Hence $r^{q^2}\equiv r=\omega u\pmod{q^3}$ iff $u\equiv1\pmod{q^3}$, i.e.\ $r\equiv\omega\pmod{q^3}$, i.e.\ $\ord_{q^3}(r)\mid q-1$; by Theorem~\ref{thm:lift}, $\ord_{q^3}(r)=d_q\,q^{\max(0,3-s_q)}$ divides $q-1$ iff $s_q\ge3$. \end{proof} \paragraph{Layer distinction ($\bmod\ q^2$ vs.\ $\bmod\ q^3$).} The subgroup governing \textup{(D4)} and Proposition~\ref{thm:teich} is $\mu_{q-1}(\bZ/q^3\bZ)$, corresponding to $s_q\ge3$. The cluster $\{19,22,23\}$ at $q=13$ of \S\ref{sec:null} lies in $\mu_{12}(\bZ/169\bZ)$ ($\bmod\ q^2$, the $s_q\ge2$ Katz/(D2) layer): $19\in\mu(\bZ/169)$ but $19\notin\mu(\bZ/2197)$. These are different subgroups of different groups; there is no contradiction, and the census of \S\ref{sec:null} tests (D2), not (D4). In the \emph{dual} ensemble (fixed $q$, $r$ uniform in $(\bZ/q^3)^\times$) one has the exact identity $\mathbb P(r^{q-1}\equiv1\bmod q^3)=\frac{|\mu_{q-1}|}{|(\bZ/q^3)^\times|} =\frac{q-1}{q^2(q-1)}=\frac1{q^2}$; in the primal ensemble (fixed $r$, $q$ varying) there is no probability space, so the missing tool is precisely the transfer dual$\to$primal (Proposition~\ref{prop:dual}). Consequently \textup{(D4)} is the purely algebraic assertion that a fixed integer $r$ has no systematic preference among the index-$q^2$ Teichm\"uller subgroups as $q$ varies; the constant $\sum_qq^{-2}=P(2)=0.452247514\ldots$ is the reciprocal sum of the subgroup indices, not a heuristic input. The Chebotarev obstruction is unchanged: the group varies with $q$. The running example $\langle19\rangle\bmod343=\{\pm1,\pm18,\pm19\}$ of \S\ref{sec:comp} is the image of $\mu_6$. Numerically: zero counterexamples among $q<10^5\times12$ bases, and $|\mu_{q-1}\bmod q|=q-1$ for every $q\le97$ enumerated (see the ensemble separation above). Set $X_q=\frac{s_q\log q}{d_q}\mathbf1_{s_q\ge3}\ge0$, so $\Xi_r^{(3)}=\sum_qX_q$. \begin{theorem}[(D4) $\Rightarrow$ $\Xi_r^{(3)}<\infty$ and $P_{6a}$, a.s.]\label{thm:d4} Under \textup{(D4)}, \[ \mathbb E\big[\Xi_r^{(3)}\big]\le C_r(\log r)\,P(2) =0.452247514\ldots\cdot C_r\log r<\infty, \] where $P(2)=\sum_q q^{-2}$ (over all primes). Hence $\Xi_r^{(3)}<\infty$ a.s.; and by the first Borel--Cantelli lemma (no independence needed), $\#\cW_r^{(3)}<\infty$ a.s., i.e.\ $P_{6a}$/Murata's conjecture (i) holds almost surely under \textup{(D4)}. \end{theorem} \begin{proof} By (U2) (\eqref{eq:U}) the pointwise deterministic bound $X_q\le(\log r)\mathbf1_{s_q\ge3}$ holds, independent of $d_q$; hence $\mathbb E[X_q]\le(\log r)\,\mathbb P(s_q\ge3)\le C_r\log r/q^2$. Since $X_q\ge0$, Tonelli permits term-by-term expectation, and $\sum_qq^{-2}<\infty$. Finite expectation implies a.s.\ finiteness; Borel--Cantelli I needs only $\sum_q\mathbb P(s_q\ge3)\le C_rP(2)<\infty$ and yields $\#\cW_r^{(3)}<\infty$ a.s. \end{proof} \begin{remark}\label{rem:d4refine} The proof needs no conditioning on $d_q$, no $k\ge4$ layer hypothesis, no lower bound on orders, and no GRH: (U2) absorbs the $\ord$ side entirely, bypassing the ``order $\times$ Wieferich'' correlation of the (D2)/(D3) side. If (D4) is stated in conditional form one gets $\mathbb E[\Xi_r^{(3)}]\le C_r\sum_q\frac{\log q}{d_qq^2}$, whose true values are $0.0346$--$0.1470$ for $r=2,3,5,7$; the proof itself does not use this refinement. \emph{Scale.} \textup{(D4)} yields $\mathbb E[\#\cW_r^{(3)}(x)]\le C_rP(2)=O(1)$, not $\ll\log\log x$ (the latter is the $\sum_q1/q$ scale of \textup{(D1)}). Read as a \emph{counting density}, $\#\cW_r^{(3)}(x)\le C_r\sum_{q\le x}q^{-2}=O(1)$, \textup{(D4)} is \emph{equivalent} to $P_{6a}$ (Theorem~\ref{thm:mq}); it is the \emph{probabilistic} form of Definition~\ref{def:d4}, with the finite mean $\sum_q C_r/q^2$, that is strictly stronger. \emph{(K) alone is insufficient.} (D1) gives only $O(1/q)$, and $\sum_{qk$, \[ T_s^{(2)}(e)=\sum_{\substack{q\mid e-k\\q\in\cW_r^{(2)}}}2\log q \ \le\ 2\sum_{q\mid e-k}\log q\ \le\ 2\log(e-k)\ \le\ 2\log e, \] hence $\sum_{e\le E}T_s^{(2)}(e)/e\le 2\sum_{e\le E}\frac{\log e}{e} =O\big((\log E)^2\big)=o(E)$, unconditionally. Moreover \[ T_s(e)\ \le\ \sum_{q^3\mid b_e}V_q(e)\log q\ \le\ \log(r^{e-k}-1)\ <\ e\log r, \] so the exception threshold $e\log r/20$ sits a factor $20$ below the unconditional pointwise bound: density zero needs the \emph{average} to be far below the bound. Numerically ($r=2,3,5,7$; $\cW_r^{(2)}$ below $10^6$ is $\{5,11,1093,3511,20771,40487\}$): $\max_{e\le E}T_s^{(2)}(e)=$ $8.0147/8.0147/17.2122/22.0080$ at $E=10^2/10^3/10^4/10^5$, against $2\log E=9.2103/13.8155/18.4207/23.0259$ --- zero violations. \end{proposition} Thus in Case A the density-zero conclusion for the $s_q=2$ layer needs neither Theorem~\ref{thm:xi2} nor any hypothesis: Proposition~\ref{prop:caseA} suffices. Theorem~\ref{thm:xi2} remains the stronger result and the required input for Case B (where $V_q(e)\ge s_q$ fails and no pointwise bound holds). Note $\cC_{r,r^k}=\cW_r$ (Proposition~\ref{prop:split}), so Case A never meets the bad set at all: \textup{(H-j)} enters only through Case B. \subsection{\texorpdfstring{The harmonic threshold $\Phi_r(E)$}{The harmonic threshold}} For $q\in\cW_r^{(3)}$ the solution class of $q^3\mid b_e$ is $e\equiv e_0\pmod{d_q}$ (wall-free, $D_3=d_q$), and for any $a'$ the class sum satisfies \[ \sum_{\substack{e\le E\\ q^3\mid b_e}}\frac1e \ \le\ \frac{1}{e_0(q)}+\frac{H_{\lfloor E/d_q\rfloor}}{d_q}. \] Define \begin{equation}\label{eq:phiE} \Phi_r(E):=\sum_{\substack{q\in\cW_r^{(3)}\\ d_q\le E}}\frac{s_q\log q}{d_q}\, H_{\lfloor E/d_q\rfloor}. \end{equation} Then $\sum_{e\le E}\frac{T_s^{(3)}(e)}{e}\le\Phi_r(E)+O(1)$, and together with Proposition~\ref{prop:caseA} (Case A; the same bound holds for general $a'$): \[ \sum_{e\le E}\frac{T_s(e)}{e}\ \le\ (\log E)^2+\Phi_r(E)+O(\log E). \] \begin{theorem}[Unconditional linear cap]\label{thm:phicap} Let $S_d^{(3)}=\{q\in\cW_r^{(3)}:d_q=d\}$. Since $d_q=d$ forces $d\mid q-1$ (hence $q\nmid d$) and $v_q(r^d-1)=s_q$, we have $\prod_{q\in S_d^{(3)}}q^{s_q}\mid\Phi_d(r)$ (the $d$-th cyclotomic value), and $\Phi_d(r)\le(r+1)^{\varphi(d)}$; therefore \[ \operatorname{mass}_d:=\sum_{q\in S_d^{(3)}}s_q\log q\le\varphi(d)\log(r+1), \] and, with $H_m\le\log m+\gamma+o(1)$ uniformly, \[ \Phi_r(E)\ \le\ \log(r+1)\sum_{d\le E}\frac{\varphi(d)}{d} \Big(\log\frac Ed+\gamma+o(1)\Big) \ =\ \Big(\frac{6}{\pi^2}+o(1)\Big)(1+\gamma)\log(r+1)\,E . \] The constant is $C=\frac{6}{\pi^2}(1+\gamma)=0.958832\ldots$; the resulting cap $\frac{20C\log(r+1)}{\log r}$ equals $30.39, 24.20, 21.35, 20.49, 19.87$ for $r=2,3,5,7,11$ (limit $19.18$ as $r\to\infty$) --- all trivial, since $\delta(E)\le1$. The same mass bound gives the linear estimate $\Xi_r^{(3)}(E):=\sum_{d_q\le E}\frac{s_q\log q}{d_q}\le\frac{6}{\pi^2}\log(r+1)E$ ($0.668/0.843/1.089/1.264$ for $r=2,3,5,7$). \end{theorem} So the unconditional bound is \emph{linear}, while the target is $o(E)$: the required saving is a factor $\Omega(\log E)$ --- not a constant factor and not an infinite methodological gulf. It lands precisely on ``P1/P2-level counting of $\cW_r^{(3)}$'', which is the varying-modulus obstruction itself. \paragraph{Tighter budget and its saturation.} For each $d$, $\prod_{q:\,d_q=d}q^{s_q}\mid\Phi_d(r)\le(r+1)^{\varphi(d)}$, so $\operatorname{mass}_d\le\varphi(d)\log(r+1)$, a factor $1.25$--$2.95$ tighter than $\log(r^d-1)d_q$ and $q\nmid d_q$, the bookkeeping contamination requires $d\ge q\,d_q>q^2$; hence for $d\le E$ and $q>\sqrt E$ one has $q^3\mid r^d-1\iff s_q\ge3$ and $d_q\mid d$, with no contamination. \begin{remark}[Structural empty layers and the role of smoothness]\label{rem:emptylayers} Since $q\in S_d^{(3)}$ forces $d\mid q-1$ with $q$ odd, $2d\mid q-1$ for odd $d$; writing $q_{\min}(d)$ for the least prime $\equiv1\pmod d$, the divisibility $q^3\mid\Phi_d(r)\le(r+1)^{\varphi(d)}$ gives the necessary condition \[ 3\log q_{\min}(d)\ \le\ \varphi(d)\log(r+1), \] stronger than $q\ge d+1$. For $r=2$ the structurally empty $d\le299$ are $\{2,\dots,10,12,14,\allowbreak 15,18,20,24,30\}$, the first admissible being $d=11$; the witness $(r,q)=(19,7)$ has $d_q=6$ and sits on the boundary ($3\log7=5.8377$ against $\varphi(6)\log20=5.9915$). For $d\ge50$ the condition holds automatically, so it excludes only finitely many small $d$ and has no asymptotic strength. Moreover $\varphi(d)/d=\prod_{p\mid d}(1-1/p)$, so large $\varphi(d)/\log d$ occurs for \emph{rough} $d$, not smooth $d$ (pass rates $58.8\%$ versus $71.9\%$ for $r=2$, $d\in[2,50]$); summing the mass over $d$ gives $\sum_{d\le x}\varphi(d)\asymp\frac{3}{\pi^2}x^2$, a quadratic bound, as recorded before Theorem~\ref{thm:phicap}. \end{remark} \begin{theorem}[The weakest sufficient condition; none of the quoted conditions is necessary]\label{thm:weakest} Since $H_{\lfloor E/d\rfloor}\le\log E+\gamma$, \[ \Xi_r^{(3)}(E)=o\Big(\frac{E}{\log E}\Big)\ \Longrightarrow\ \Phi_r(E)=o(E)\ \Longrightarrow\ \delta(E)\to0. \] Neither $K_{r,a'}<\infty$ nor $\Xi_r^{(3)}<\infty$ is necessary for density zero. Indeed there are configurations meeting every per-block structural constraint with \[ \Xi_d=\frac{1}{(\log d)^2}\ \ (d\ge5),\qquad \Xi_r^{(3)}(E)\sim\frac{E}{(\log E)^2}\to\infty,\qquad \Phi(E)=\sum_{d\le E}\Xi_d\,H_{\lfloor E/d\rfloor}=o(E): \] \begin{center} \small \begin{tabular}{r|r|r|r|r} $E$ & $\Xi_r^{(3)}(E)$ & $\Phi_r(E)$ & $\Phi_r(E)/E$ & bound \eqref{eq:markov} on $\delta(E)$\\ \hline $10$ & $32.53$ & $69.11$ & $0.0691$ & $2.127$\\ $10^4$ & $160.1$ & $323.9$ & $0.0324$ & $0.744$\\ $10^5$ & $943.6$ & $1782$ & $0.0178$ & $0.349$\\ $10^6$ & $6245$ & $11260$ & $0.0113$ & $0.208$\\ $10^7$ & $44500$ & $77980$ & $0.0078$ & $0.142$ \end{tabular} \end{center} divergent $\Xi_r^{(3)}$ with $\delta(E)\to0$. These are logical models (satisfying all per-block constraints), not claims about the true factor sets of $r^d-1$. \end{theorem} The ladder of sufficient conditions, weakest first, all unproved unconditionally: \begin{center} \small \begin{tabular}{c|p{5.6cm}|p{6.4cm}} \# & statement & status\\ \hline \textup{(M)} & $\Phi_r(E)=o(E)$ & the exact threshold of the Markov route; equivalent to $\sum_e T_s(e)/e=o(E)$\\ \textup{(M')} & $\Xi_r^{(3)}(E)=o(E/\log E)$ & sufficient for \textup{(M)}; strictly weaker than $\Xi_r^{(3)}<\infty$\\ \textup{(C)} & $N_3(E):=\#\{d\le E:S_d^{(3)}\ne\emptyset\}=o(E/\log E)$ & sufficient for \textup{(M')} (per-block $\Xi_d^{(3)}\le\log(r+1)$); this is P1/P2-level counting of $\cW_r^{(3)}$\\ \textup{(P)} & $\Xi_r^{(3)}<\infty$ & sufficient for \textup{(M')}; the previously quoted condition --- not necessary\\ $P_{6a}$ & $\cW_r^{(3)}$ finite (Murata 1991 (i)) & the classical anchor; sufficient for \textup{(P)}\\ $(\mathrm{K}^+)$ & $(\mathrm{K})+(\mathrm{D4})$ & gives \textup{(P)} almost surely (Theorem~\ref{thm:d4}), not unconditionally \end{tabular} \end{center} For orientation: $N_3(E)\le\sum_{d\le E}\frac{\varphi(d)\log(r+1)}{3\log(d+1)} \asymp\frac{(\log r)E^2}{6\log E}$ is quadratic --- useless; a genuine sub-quadratic bound on $N_3(E)$ is itself open. In one sentence: an unconditional proof requires a sublinear bound on the deep-Wieferich mass weighted by $1/d_q$ --- a P1/P2-level statement restricted to $\cW_r^{(3)}$ and weighted by $\ord$; it is neither manifestly easier nor manifestly harder than P1, but it lives in the same no-fixed-modulus no-man's-land. Silverman's $abc$ route gives a \emph{non-Wieferich lower} bound --- the opposite direction; the other branch of \cite{dekoninckdoyon} gives $\cW$ \emph{infinite}. No unconditional tool currently produces upper bounds of this type. \begin{theorem}[Information-theoretic barrier]\label{thm:infobar} Let $q\in\cW_r$, $r^{q-1}=1+q^{s_q}u$, $q\nmid u$, and suppose $q^3\mid r^e-a'$. Raising to the $(q-1)$-st power, \[ a'^{q-1}\equiv(1+q^{s_q}u)^e\equiv1+e\,q^{s_q}u\pmod{q^3}, \] the $k\ge2$ binomial terms carrying $q^{2s_q}\ge q^4$. For $s_q=2$ the $q^2$ term survives and $e\equiv\beta_qw_q^{-1}\pmod q$ is solvable --- this is the second-Fermat-quotient formula of Theorem~\ref{thm:two-base}, the entire information content of \textup{(H-j)}/(K). For $s_q\ge3$ the term $e\,q^{s_q}u\equiv0\pmod{q^3}$: $a'^{q-1}\equiv1\pmod{q^3}$ \emph{independent of $e$}, and no congruence expansion of any finite order recovers $e$; the least solution $e_0(q)$ is determined only by the discrete-logarithm structure of $\langle r\rangle\subseteq(\bZ/q^3\bZ)^\times$ --- a varying-modulus discrete logarithm. \end{theorem} Numerically: four artificial bases $r\equiv1,57,68,124\pmod{125}$ with $r^4\equiv1 \pmod{125}$ (forced $s\ge3$ at $q=5$) give $a'^4\equiv1\pmod{125}$ identically in $e=1,\dots,29$ --- no $e$-information; while for the genuine pair $(r,q)=(3,11)$, $w_q=4$ and $a'^{10}\equiv1+100e\cdot4\pmod{1331}$ predicts $e\bmod 11$ exactly for $e=1,2,3,5,7$. Conclusion: all information about the $s\ge3$ layer lives in the $\ord$ and counting structure of $\cW_r^{(3)}$, not in any congruence expansion --- the same obstruction, which now has seven guises: \textbf{(T1)} GRH insufficiency (Theorem~\ref{thm:grh}); \textbf{(T2)} the radical channel --- withdrawn as a barrier and replaced by the identity of Theorem~\ref{thm:radical}; \textbf{(T3)} the expansion barrier above (Theorem~\ref{thm:infobar}); \textbf{(T4)} no cross-$d$ repulsion from divisibility identities --- a fixed $q$ has a unique $d_q$ with $q\equiv1\pmod{d_q}$, so distinct $d$'s carry disjoint $q$-sets, all multiples $e=md_q$ are consistent by LTE, and $\prod_{d\mid n}\Phi_d(r)=r^n-1$ with $n=\operatorname{lcm}(1..E)\sim e^E$ is too loose by $\sim33$ orders of magnitude; \textbf{(T5)} no cited theorem controls the cube-full part of $\Phi_n(a)$ (Theorem~\ref{thm:zerosaving}); \textbf{(T6)} all structural constraints together give \emph{zero} asymptotic improvement on $A_r(t)$ (Theorem~\ref{thm:zerosaving}); and \textbf{(T7)} the missing dual$\to$primal transfer of \textup{(D4)} (Proposition~\ref{prop:dual}). \begin{theorem}[Exact threshold, dyadic form]\label{thm:dyadic} Put $M_d=\sum_{q:\,d_q=d,\,s_q\ge3}s_q\log q$ and $A_r(t)=\sum_{d\le t}\frac{M_d}{d}$. Then $\Phi_r(E)=o(E)$ \emph{if and only if} $A_r(t)=o(t)$; precisely \[ \tfrac32\,A_r(E/2)\ \le\ \Phi_r(E)\ \le\ (1+\log2)\,A_r(E)+\log2\sum_{j\ge1}A_r(E/2^j), \] so $A_r(t)\le\varepsilon t$ gives $\Phi_r(E)\le(1+2\log2)\,\varepsilon E=2.386294\,\varepsilon E$. A saving $M_d=d/\log d$ is \emph{not} enough ($\Phi(E)/E\approx0.17$ at $E=10^5$); one needs $M_d/d\to0$ in Ces\`aro mean. In particular $\Xi_r^{(3)}(E)=o(E/\log E)$ of Theorem~\ref{thm:weakest} is a sufficient rung, not the threshold itself. \end{theorem} \begin{proof} Group $\Phi_r(E)=\sum_{d\le E}\frac{M_d}{d}H_{\lfloor E/d\rfloor}$ (Theorem~\ref{thm:phicap}). For $d\le E/2$ the floor is $\ge2$, so $H_{\lfloor E/d\rfloor}\ge H_2=\frac32$, giving the lower bound. For the upper bound split $[1,E]$ dyadically: on $(E/2^{j+1},E/2^j]$ one has $H_{\lfloor E/d\rfloor}\le(j+1)\log2+1$, and Abel reordering gives the stated bound. If $A_r(t)\le\varepsilon t$ for all $t\le E$, then $\sum_{j\ge1}A_r(E/2^j)\le\varepsilon E\sum_{j\ge1}2^{-j}=\varepsilon E$. \end{proof} \begin{theorem}[Container; the gap is exactly one $o(1)$ factor]\label{thm:container} Let $C_E=\prod_{d\le E}\Phi_d(r)$. Then $\prod_{q\in\cW_r^{(3)},\,d_q\le E}q^{s_q}\mid C_E$ and \[ \log C_E=\sum_{d\le E}\Bigl(\varphi(d)\log r+O\bigl(2^{\omega(d)}\log r\bigr)\Bigr) =\frac{3}{\pi^2}E^2\log r\,\bigl(1+o(1)\bigr), \] since $\sum_{d\le E}\varphi(d)=\frac{3}{\pi^2}E^2+O(E\log E)$ and $\sum_{d\le E}2^{\omega(d)}=O(E\log E)$. The volume bound gives \[ A_r(t)\ \le\ \log(r+1)\sum_{d\le t}\frac{\varphi(d)}{d} =\frac{6}{\pi^2}\log(r+1)\,t\,\bigl(1+o(1)\bigr), \] with $\sum_{d\le t}\frac{\varphi(d)}{d}\big/\bigl(\frac{6}{\pi^2}t\bigr) =1.000658, 1.000761, 1.000375, 1.000095$ at $t=10^2, 3\cdot10^2, 10^3, 3\cdot10^3$. Thus the container budget is quadratic, the first moment is linear, and what is known differs from what is needed by exactly one factor $o(1)$ per $d$ --- the $\Omega(\log E)$ saving of \S\ref{sec:threshold} in closed form. In particular the earlier $\operatorname{lcm}(1..E)\sim e^E$ argument is unnecessary: $C_E$ itself is a fixed integer of size $\exp(\Theta(E^2))$. \end{theorem} \section{Locating (H-j): Katz, Murata, and the GRH ruling}\label{sec:katz} Write $q_q(u)=k_u+q\,m_u$ with $k_u,m_u\in[0,q)$. Then $\overline{W}_u(q):= W_u(q)\bmod\bZ=\frac{q_q(u)}q\bmod\bZ$ has fractional part $k_u/q$, so $q\in\cW_u\iff k_u=0\iff\overline W_u(q)=0$; and when $q\in\cW_u$ one has $m_u(q)\equiv(u^{q-1}-1)/q^2\pmod q$, i.e.\ $m_r=w_q$, $m_{a'}=\beta_q$. Thus $(w_q,\beta_q)$ is the mod-$q$ higher layer of the Katz quotient at its Haar-null zeros: one order higher than $\overline W$ (the mod-$q^2$ information, equivalently $u^{q-1}-1\bmod q^3$). If $a'=r^k$ then $q_q(r^k)\equiv k\,q_q(r)\pmod q$ and the pair $(\overline W_r,\overline W_{a'})$ lies on the one-dimensional subtorus $\{(x,kx)\}$: exactly the degenerate case excluded by the Zariski-density hypothesis of the group-scheme version, matching the Case A/B dichotomy. (Numerically, the joint $\chi^2$ for $(3,9)$ and $(3,27)$ explodes to $7736$ and $3454$ against a critical value $36.4$, while independent bases pass; see \S\ref{sec:comp}.) The continuous equidistribution of Katz does not control point masses at scale $1/q$. (H-j) requires the three-layer discretized joint refinement: \begin{itemize} \item[\textup{(D1)}] \emph{Pointwise discretization}: $\#\{q\le x:\overline W_u(q)=0\} \sim\sum_{q\le x}1/q\sim\log\log x$ (Lang's fixed-residue statement \cite{lang} and the Crandall--Dilcher--Pomerance model \cite{cdp}); \item[\textup{(D2)}] \emph{Mod-$q^2$ two-base joint}: for multiplicatively independent $\langle r,a'\rangle$, the pair $(w_q,\beta_q)$ is jointly uniform on $\bF_q^2$ and $P(q\in\cW_r\cap\cW_{a'})\asymp1/q^2$; \item[\textup{(D3)}] \emph{Discrete-logarithm layer}: conditional on compatibility, $j_q$ is uniform on $[0,d_q)$, so $P(j_q=0)=1/d_q$. \end{itemize} Set (K)=(D1)+(D2)+(D3). None of (D1)--(D3) is a GRH theorem or a direct consequence of Katz's conjecture. \begin{proposition}[conditional on the random model \textup{(K)}]\label{thm:K} Assume \textup{(K)} and that $r,a'$ are multiplicatively independent (Case B). Then \[ \mathbb E\Big[\sum_{q\in\cB_{r,a'}}\frac{s_q\log q}{b_q}\Big] \ \ll\ \log r\sum_q\frac{\log q}{q^2}<\infty, \] so \textup{(H-j)} holds almost surely; together with $P_4$ and $\Xi_r^{(3)}<\infty$ one obtains $K_{r,a'}<\infty$ and the density bound of Theorem~\ref{thm:main}. \end{proposition} \begin{proof} Here (K) is used with its built-in independence: (D2) and (D3) are joint uniformities, so $(w_q,\beta_q)$ and $j_q$ are independent, and $s_q\log q$ (which depends only on $r$) is independent of $b_q$. By (D2)--(D3) $P(q\in\cB)\le P(q\in\cW_r\cap\cW_{a'})\cdot P(j_q=0)\le q^{-2}\cdot d_q^{-1} \le q^{-2}$; by (D2) conditional uniformity $b_q$ is uniform on $\bF_q^\times$ (note $w_q\ne0$ on the layer $s_q=2$), so $\mathbb E[1/b_q]=H_{q-1}/q\asymp(\log q)/q$; and $s_q\log q1$, $p^2\mid b^p-b$ and $p^2\mid c^p-c$ simultaneously for only finitely many $p$ --- exactly $\#(\cW_r\cap\cW_{a'})<\infty$, the strong form of (HW). Our (H-j) is \emph{weaker} than Murata (ii) (the intersection is allowed to be infinite; only the bad subset must be $1/b_q$-summable) and \emph{stronger} than Katz's conjecture (which is continuous equidistribution and silent on all three discretized layers). The correct description of (H-j) is: a weighted refinement interpolating between Katz 2015 and Murata 1991. \end{remark} \subsection{\texorpdfstring{The unified family $(\mathrm{K}^+)$ and the main line}{The unified family (K+) and the main line}}\label{sec:kplus} Recall $(\mathrm{K}^+)=(\mathrm{K})\cup\{\textup{(D4)}\}$ with (D4) from \S\ref{sec:d4}. \begin{theorem}[Unified conclusion]\label{thm:unified} Under $(\mathrm{K}^+)$, for every fixed $r\ge2$ and every $a'$ with $\gcd(a',r)=1$: $K_{r,a'}<\infty$ almost surely; consequently the density bound of Theorem~\ref{thm:main} holds almost surely for every $a'$. \end{theorem} \begin{proof} Case A: by Proposition~\ref{prop:split}, $K_{r,r^k}\le2\Xi_r=2(\Xi_r^{(2)}+\Xi_r^{(3)})$; $\Xi_r^{(2)}<\infty$ is the unconditional Theorem~\ref{thm:xi2}, and $\Xi_r^{(3)}<\infty$ a.s.\ is Theorem~\ref{thm:d4}. Case B: $K_{r,a'}=\Xi_{r,a'}+J'\le\Xi_r+J'$; the $\Xi_r$ part is controlled as before (Theorems~\ref{thm:xi2} and~\ref{thm:d4}); $J'$ is controlled by (D1)--(D3): on the bad set $\mathbb P(q\in\cB)\ll q^{-2}d_q^{-1}\le q^{-2}$, conditionally $\mathbb E[1/b_q]\asymp(\log q)/q$, and with (U2), \[ \mathbb E\Big[\sum_{q\in\cB_{r,a'}}\frac{s_q\log q}{b_q}\Big] \ll\log r\sum_q\frac{\log q}{q^2}<\infty, \] i.e.\ \textup{(H-j)} holds a.s., so $J'<\infty$ a.s. Finally apply Theorem~\ref{thm:main}. \end{proof} \emph{Erratum note.} An earlier draft claimed that ``(D4) does not participate in Case B''; this is inaccurate. In Case B, $\Xi_{r,a'}\le\Xi_r$ still contains the $s\ge3$ layer, which must be controlled by (D4); (K) alone controls only $J'$. \emph{Honest strength ordering.} \[ (\mathrm{K}^+)\ \Rightarrow\ P_{6a}\ \text{a.s.}\ \Rightarrow\ \Xi_r^{(3)}<\infty\ \Rightarrow\ P_4^{(3)}, \] the last three arrows by Theorem~\ref{thm:squeeze}. Read as a counting density, (D4) is \emph{equivalent} to $P_{6a}$ (Theorem~\ref{thm:mq}); its probabilistic form (Definition~\ref{def:d4}, finite mean) is strictly stronger, yielding $\Xi_r^{(3)}<\infty$ a.s. Its value lies in being same-origin with (K) and in covering the main line and the bad-set line uniformly. The final condition set of the chain may be stated either as the classical anchor $P_{6a}$ or as the unified conjecture family $(\mathrm{K}^+)$. \section{Ancillary results and the redundancy of GRH}\label{sec:ancillary} Three results decide which hypotheses are redundant in the chain; none of them fills the application-layer gap, but together they show GRH is nowhere needed. \begin{theorem}[$S_{\mathrm{ord}}<\infty$, unconditional]\label{thm:sord} $S_{\mathrm{ord}}=\sum_{p\nmid r}\frac{\log p}{d_p(p-1)}<\infty$. \end{theorem} \begin{proof} Erd\H{o}s--Murty \cite{erdosmurty} prove unconditionally $\sum_{p\le x}1/d_p\ll\sqrt{x}/(\log x)^{1+\delta}$ (Murty--Srinivasan \cite{murtysrinivasan} already gave $\ll\sqrt{x}$). With the decreasing weight $w(t)=\frac{\log t}{t-1}$ ($t\ge3$), Stieltjes partial summation gives \[ \sum_{3\le p\le x}\frac{w(p)}{d_p}=A(x)w(x)-\int_3^xA(t)w'(t)\,dt, \] with $A(x)=\sum_{p\le x}1/d_p$; the boundary term is $\ll x^{-1/2}(\log x)^{-\delta}\to0$, and after $t=e^u$ the integral is $\ll\int^{\infty}e^{-u/2}u^{-\delta}\,du<\infty$. \end{proof} The GRH version of the order bound (Kurlberg--Pomerance \cite{kurlberg}, $d_p\ge p/\varepsilon(p)$ on average) is not needed anywhere: GRH is redundant throughout the chain. \begin{theorem}[$C_r$ exact identity]\label{thm:crid} With $X_n=\log\frac{n}{\rad n}$, for $a'=1$, \[ \sum_{e\le E}X_{r^e-1} =\sum_{p\nmid r}\log p\sum_{k\ge2}\Big\lfloor\frac{E}{D_k(p)}\Big\rfloor +\mathrm{LTE}_2(E), \] so $\frac1E\sum_{e\le E}X_{r^e-1}\nearrow C_r=S_W+S_{\mathrm{ord}}+LTE_2(r)$, where $S_W=\sum_{q\in\cW_r}\frac{(s_q-1)\log q}{d_q}$. Hence \textup{(H-X)} $\iff S_W<\infty$. Numerically $C_2=0.6938$, $C_3=1.9112$, $C_5=1.9745$; for $r=3$ the decomposition is $S_W=0.479579$, $S_{\mathrm{ord}}=0.391941$, $LTE_2=1.039721$, summing to $C_3=1.911241$. \end{theorem} Indeed $X_n=\sum_p\log p\sum_{k\ge2}\mathbf1_{v_p(n)\ge k}$, and $v_p(r^e-1)\ge k\iff D_k(p)\mid e$ (the LTE tower of Theorem~\ref{thm:lift}); the prime $p=2$ contributes the finite correction $\mathrm{LTE}_2(E)$. For $p\in\cW_r$ ($s_p\ge2$) the limiting contribution is $\log p\sum_{k=2}^{s_p}1/D_k(p)=\frac{(s_p-1)\log p}{d_p}$; for $p\notin\cW_r$ it is $\frac{\log p}{d_p(p-1)}$. \begin{theorem}[Strict one-way]\label{thm:oneway} Termwise $\Xi_r\le\frac32 S_W+O(1)$ (ratios $\Xi_q/S_{\cW,q}$: $s=2$ gives $2/q$, $s=3$ gives $3/2$, $s\ge4$ gives $s/(s-1)$), so $S_W<\infty\Rightarrow\Xi_r<\infty$. The converse fails: the exact identity $S_W^{(2)}=\sum_q\frac q2\,\Xi_q^{(2)}$ shows $S_W$ is the $q$-weighted first moment of $\Xi_r$; the universal bound $\Xi_q^{(2)}\le2\log^2q/q^2$ is summable over all primes ($1.2429$), while the corresponding $S_W$ terms $\sim\log^2q/q$ diverge ($92.65$). A configuration with $d_q=q-1$ and $\cW_r$ of positive density makes $\Xi_r\approx0.818$ convergent while $S_W\sim\frac1m\log X$ diverges; since P1 is open, no theorem excludes it. \end{theorem} \begin{theorem}[Evolution of the condition set]\label{thm:evolution} \[ P_4+\Xi_r^{(3)}\ \longrightarrow\ P_4'+\Xi_r^{(3)}\ \longrightarrow\ \Xi_r^{(3)} \text{ alone}, \] where $P_4'=\sum_{q\in\cW_r}1/q<\infty$ was once supported by $d_q\ge\log q/\log r$ (\eqref{eq:U}), giving $\Xi_q^{(2)}\ll1/q$. Theorem~\ref{thm:xi2} removes $P_4$, $P_4'$ and $P_1$ from the condition set entirely (they remain open as statements about $\cW_r$). The (H-X)/$S_W$ route is a stronger-than-necessary fallback: strictly one-way (Theorem~\ref{thm:oneway}), not equivalent to the main line. \end{theorem} \medskip \begin{center} \small \begin{tabular}{p{3.4cm}|p{4.8cm}|p{2.4cm}|p{2.8cm}} route & Wieferich side & $\ord$ side & GRH\\ \hline A (main line; the $\Xi$ part of Cases A and B) & $\Xi_r^{(2)}$ unconditional $+$ $\Xi_r^{(3)}$ single condition & Hensel wall $1/D_3$ and (U2) absorb it & redundant\\ B (side line; the $J'$ part of Case B) & \textup{(H-j)} / (K) & not needed & insufficient (needs Katz pointwise strengthening) \end{tabular} \end{center} $(\mathrm{K}^+)$ collects both routes into one family: (D4) governs the deep layer of A, (D1)--(D3) govern the bad set of B. \section{Computational evidence}\label{sec:comp} \emph{Variable direction.} Every statistic in this section is a fixed-$r$, varying-$q$ scan, and the hypotheses they test --- \textup{(D1)}, \textup{(D2)}, \textup{(D4)}, (K) --- are likewise fixed-$r$, varying-$q$ pointwise statements. A fixed-$q$ census over bases $r$ is a different, easier object: by Proposition~\ref{thm:teich} the solutions of $u^{q-1}\equiv1\pmod{q^2}$ are exactly the Teichm\"uller subgroup $\mu_{q-1}$, so a scan over $r$ hits a residue class with rate $(q-1)/q^2$ --- even the trivial family $r\equiv1\pmod{q^3}$ ($d_q=1$) is included. Such a census cannot test \textup{(D1)}/\textup{(D2)}/\textup{(D4)}; see \S\ref{sec:null} for the correction this distinction forces. All computations were performed with exact integer arithmetic (\texttt{pow} with modulus $q^3$); scripts and data are available with this note. \subsection{The bad-set scan} For $q<10^6$, twelve bases $r\le37$ and $a'<30$: the set $\cW_r$ has $23$ elements; the bad set has \emph{five} elements, all with $b_q\le4$: \[ (7,26,5,4),\ (17,19,3,2),\ (19,10,3,2),\ (37,10,3,1),\ (37,19,3,2), \] and the partial sum in (H-j) stabilizes at $6.30$. The bad-layer primes are only $3$ and $5$. A layer census for $b\le15$ finds \emph{no} cube-divisor candidates at all for $(r,a')$ equal to $(2,1)$, $(2,3)$, $(3,1)$, $(3,2)$ or $(5,2)$, matching the random expectation $\sum_q q^{-3}\approx0.17$ per integer. \subsection{First super-Wieferich sample} The scan exhibits the first observed super-Wieferich prime: $s_7(19)=3$ exactly, since $19^6-1=2^3\cdot3^3\cdot5\cdot7^3\cdot127$, and $19$ is a primitive root modulo $7$. Note $\langle19\rangle\bmod343=\{\pm1,\pm18,\pm19\}$ with $D_3=d_q=6$: on the super-Wieferich layer $e_0\in[1,d_q]$ is automatically $1$, $\gcd(r,a')=1$, $q<10^6$; a scan over all triples gives $78$ ($23$ trivial, $55$ effective), reducing to $27$ distinct ordered pairs. (The earlier ``$49=22+27$'' under-counted both layers; the trivial-pair count $22$ differs by one from the number of $\cW_r$ hits and awaits the author's reconciliation.) The expectation $27\sum_qq^{-2}\approx12.2$ of an earlier version uses the observed number of pairs as the number of candidates and is circular. More importantly, at fixed $q$ the congruence $u^{q-1}\equiv1\pmod{q^2}$ has exactly $q-1$ solutions in $(\bZ/q^2\bZ)^\times$, namely the Teichm\"uller subgroup $\mu_{q-1}$ (Proposition~\ref{thm:teich}); the correct null density for a scan \emph{over bases} is $(q-1)/q^2$, not $1/q^2$. Measured densities for $q=3,5,7,11,13,17,19$ are $0.222,0.160,0.122,0.082,0.0695,0.055,0.049$, agreeing with $(q-1)/q^2$ to three decimals; the ``factor $4$--$6$ excess'' reported in an earlier version is this missing factor $q-1\in[2,18]$, not positive correlation. Likewise the four pairs at $q=13$ are four of the six ordered pairs inside $\mu_{12}(\bZ/169\bZ)\cap[2,30]=\{19,22,23\}$, and the $q=3$ layer is the structural condition $u\equiv\pm1\pmod9=\mu_2(\bZ/9\bZ)$. All of these are consequences of the Teichm\"uller anchor, not statistical anomalies. \emph{What remains open.} This does not restore numerical support for \textup{(D2)}: its conditional-on-$k=0$ content remains untested, since below $q<3\times10^5$ only $10$ primes carry a second hit; a meaningful test needs $q\sim10^9$--$10^{10}$, conditioned on $q\in\cW_r$, and is beyond present reach. The bad-set conclusions of this section (five hits, partial sum $6.30$) are unaffected. \subsection{Katz layer tests} On the full odd-prime population $q<5\cdot10^4$ the four-dimensional statistic $(k_u,m_u)$ for independent bases passes marginal uniformity and joint independence (KS $\le0.0145$ against critical $0.0190$; joint $\chi^2\le29.9$ against $36.4$), while Case A pairs $(3,9)$, $(3,27)$ explode ($\chi^2=7736,3454$), exactly reproducing the Zariski-density restriction. The twin-Wieferich numerics of \S\ref{sec:comp} cut against \textup{(D2)} over the full observable range, so \textup{(K)} is a heuristic whose second layer is currently unsupported and whose use in Proposition~\ref{thm:K} must be read conditionally. \emph{Limitation}: below $5\cdot10^4$ the expected number of Wieferich primes is $\approx2.4$ and of twin-Wieferich $\approx0.45$, so the conditional-on-$k=0$ content of (D1), (D2) has no numerical evidence; the $s_q\ge3$ branch has the single sample $(19,7)$. \emph{Observable summary.} All deep hits ever observed have $q\le13$ and $d_q\le6$; $A_{\mathrm{obs}}(t)$ is a truncated lower bound (saturating for $t\ge6$ on the scanned range), not a proof of $A(t)=o(t)$. The statistics of this section and of \S\ref{sec:mq}--\ref{sec:dual} are the first fixed-census data for $\cW_r^{(3)}$: the $m(q)$ table tests $P_{6a}$ for \emph{all} $r\le X$ simultaneously at fixed $X$, while the pair scans test \textup{(D1)}/\textup{(D2)} pointwise at fixed $r$ --- a strictly stronger requirement (\S\ref{sec:null}, Proposition~\ref{prop:dual}). \section{\texorpdfstring{Relation to $(\star')$}{Relation to (star')}}\label{sec:star} In the framework of \cite{zeng} the remaining gap is stated in Section~7 (``The remaining gap, exactly'') as Working hypothesis~6.4: $|\log x|=o(\log c)$, equivalently $O(\log\log c)$, with $x=\det(\Lambda)^{(k-1)/k}/\Delta$ and $\Delta=N|a'|_2$; Theorem~6.5 there gives $\eta\le\frac1k+O(\frac{\log\log c}{\log c})$ and Theorem~7.1 the equivalent form $N|a'|_2=(c\theta_c/g)^{(k-1)/k}c^{o(1)}$. The expression $\frac{k-1}{k}\log c-\log\rad(ab)$ used in an earlier version of this note is a heuristic transcription, not a formula of \cite{zeng}; and the ``degeneration to $abc$ strength'' should be read as the structural ceiling of Theorem~10.1 there ($\Pi\le1/k+o(1)$, attained at $c=r^e$; full $abc$ is structurally unreachable inside the framework). Three regimes: $0.3$ (the present results, $\tfrac12-\tfrac15$), $0.5$ ($k=2$), and $(k-1)/k\to1$ ($k\to\infty$, i.e.\ $abc$). Along the line $a'=O(1)$, $b_e=r^e-a'$, one has $k_e=\omega(a'b_e)\to\infty$, so the hypothesis degenerates to $abc$-strength already for bounded $a'$; the present note therefore governs the $k=2$ slice. Theorem~\ref{thm:two-base} pins the Layer-III Pillai candidates $r^e-\prod q_i^{V_i}=a'$ to $\cW_r\cap\cW_{a'}$ with an explicit class $e\bmod q$; Theorem~\ref{thm:main} pins the Layer-III mass to $J'_{r,a'}$, the Pillai-single-solution layer ($D_3>E\ge e_0$ means ``$q$ hits exactly once in $[1,E]$''); and (H-j) shows the application-layer residual gap does not require Baker-type bounds---the Baker layer belongs to $k\ge3$ and the $V_q=2$ squarefull layer, where the first moment of the union bound diverges ($t_q=V_q-1-s_q\le-1$, no Hensel wall), and which is untouched here. The defect/order-core analysis of \cite{dw} on the Mersenne line independently corroborates this trichotomy. \begin{remark}[Notation and quantifiers in the $(\star')$ interface]\label{rem:quantifiers} The symbol $a'$ denotes three different objects in the sources combined here, and no identification between them is asserted: (i) in \cite{zeng}, Definition~3.4, $a'=h/g_h$ is a \emph{primitive normal vector} in $\bZ^k$, with $N=[\ker H:\Lambda_0]$ and $\Delta=\det(\Lambda_0)=N|a'|_2$ --- this is the $N|a'|_2$ occurring in Working hypothesis~6.4 and Theorem~7.1; (ii) in \S9 of \cite{zeng} (Theorem~9.8), $a'$ is a \emph{scalar Pillai remainder} with $|a'|\le c^{o(1)}$; (iii) throughout the present note $a'\le A$ is a \emph{fixed integer} and $b_e=r^e-a'$. In particular the line $a'=O(1)$ of the present note is a proper subfamily of Layer~III in (ii), which is why the first-moment Markov argument can produce an $e$-density-zero conclusion here without resolving that layer in general. Two uses of $k$ must likewise be separated: $k=\omega(ab)$ is a global quantity attached to a triple $(a,b,c)$ (so $k_e=\omega(a'b_e)\to\infty$ along $b_e=r^e-a'$), whereas the ``$k=2$ slice'' governed by Theorem~\ref{thm:main} is the \emph{local} per-prime valuation layer $V_q\ge3$. Working hypothesis~6.4 is a pointwise hypothesis $|\log x|=o(\log c)$ with no ``almost all'' quantifier; Theorem~6.5 is a conditional statement along a sequence of triples with $c\to\infty$; and the only density statement in \cite{zeng}, Theorem~11.7(i), is on the $c$-axis and asserts that \emph{every} coprime decomposition of a density-one set of $c$ has $q\le1+o(1)$, proved by one-variable counting. There is no quantifier position in \cite{zeng} for averaging over $a'\in[1,A]$ or discarding a zero-density set of $a'$; nor would such an average transfer to the pointwise statement of Theorem~6.5. Accordingly no averaging step is used or claimed below. \end{remark} \section*{Acknowledgments} The reductions of \S\S\ref{sec:twobase}--\ref{sec:katz} were found and machine-verified in the course of an extended collaboration on the Wieferich tail-estimation problem; the numerical protocols follow the standards of \cite{companion}. The condition set has now evolved to its minimal form. The $s_q=2$ layer is unconditionally closed (Theorem~\ref{thm:xi2}). The \emph{exact} threshold of the first-moment route is \[ \Phi_r(E)=o(E)\quad\Longleftrightarrow\quad A_r(t):=\sum_{d\le t}\frac{M_d}{d}=o(t),\qquad M_d=\sum_{\substack{q:\,d_q=d\\ s_q\ge3}}s_q\log q \] (Theorem~\ref{thm:dyadic}); this is equivalent to the statement that the cube-full part of $\Phi_d(r)$ has logarithmic mass $o(d)$ in Ces\`aro mean. $\Xi_r^{(3)}<\infty$ is a convenient sufficient rung but is \emph{not} necessary (Theorem~\ref{thm:weakest}): a configuration with $M_d=d/(\log d)^2$ has $\Xi_r^{(3)}(E)\sim E/(\log E)^2\to\infty$ while $\Phi_r(E)=o(E)$. Classically the remaining input is anchored at $P_{6a}$ (Murata 1991 (i)), probabilistically at \textup{(D4)}; GRH is redundant throughout (Theorem~\ref{thm:sord} and the Hensel-wall absorption of the $\ord$ side, Theorem~\ref{thm:d4}). \section{Postscript}\label{sec:postscript} This section records the outcome of the fifteenth--eighteenth steps of the working chain: what a new tool would have to look like, and the single-variable reformulation of the deep-layer finiteness. \subsection{Literature verdict: no existing tool applies}\label{sec:litverdict} A systematic check of the six natural candidates finds \emph{no} unconditional bound $\log\rad\Phi_n(a)\ge\varphi(n)\log a-o(n)$ --- i.e.\ no theorem giving $M_d=o(d)$: \begin{itemize} \item Stewart \cite{stewart13,stewartyu} controls the \emph{greatest prime factor}, $P(\Phi_n(\alpha,\beta))>n\exp(\frac{\log n}{104\log\log n})$; a large single-power prime factor is compatible with the remaining mass being concentrated in high powers of small primes. \item Bugeaud--Corvaja--Zannier \cite{bugeaudcorvajazannier} gives $\log\gcd(a^n-1,b^n-1)=o(n)$: the \emph{common} factor of two sequences, not the repeated-prime mass of one. \item Balaji--Luca \cite{balajiluca} gives $\#\mathcal A_{K,c,u}\cap[1,N]\ll N\exp(-\frac{\log N}{156\log\log N})$ for the $Kn$-smooth part, for \emph{most} $n$; smooth parts include first powers and do not exclude high powers. \item Zsigmondy / Bilu--Hanrot--Voutier \cite{biluhanrotvoutier} guarantees \emph{one} primitive divisor, not that all primitive divisors occur once. The paraphrase ``$p\nmid n$ implies $p$ occurs at most once in $\Phi_n(r)$'' is \emph{false}: $v_{1093}(\Phi_{364}(2))=2$ with $1093\nmid364$, and $\Phi_5(3)=121=11^2$. Indeed \emph{all} deep Wieferich primes satisfy $q\nmid d_q$ (since $d_q\mid q-1$), so high exponents occur precisely at primitive divisors. \item Murty--S\'eguin \cite{ms} gives conditional greatest-prime-factor bounds and ``super-Wieferich finite $\Rightarrow$ non-Wieferich infinite''; no upper bound on $W_r^{(3)}(x)$ better than $O(x^2)$. \item Fellini--Murty \cite{fellinimurty}, Silverman \cite{silverman}, Graves--Murty \cite{gravesmurty} and Ding \cite{gravesmurty} give \emph{lower} bounds for non-Wieferich primes or $abc$-conditional results; Ding's $C_n'\gg a^{\varphi(n)-\varepsilon n}$ yields $\log\rad\Phi_n(a)\ge(1-\varepsilon)\varphi(n)\log a$, which controls $M_d$ only with $\varepsilon\to0$ \emph{and} $abc$. \item W.~Xie, \emph{Primary components and large prime factors of cyclotomic values}, arXiv:\allowbreak 2611.00075v1 (math.RA, 36~pp., preprint, not peer-reviewed). This is the closest same-object reference: Theorem~3.2 gives an unconditional lower bound for the primary part, but Theorem~3.6 takes $\log\rad\Phi_r(a)\ge q^{1-\gamma}$ as a \emph{hypothesis} and its counting input is $\vartheta(Y;r,1)$ with a \emph{fixed} conductor and an average over bases. It therefore supplies no unconditional $\log\rad\ge\varphi(n)\log a-o(n)$ and does not transfer to the fixed-base, varying-modulus side; the verdict above is unchanged. \emph{Citation caveat}: the GenPDF header prints ``arXiv:2611.00075v1'' followed by ``[math.RA] 16 Oct 2026''; the ``2611'' identifier and the October date are incompatible at the metadata level. We cite the v1 PDF as deposited and do not assert a submission date. \end{itemize} The unconditional side has no lower bound of strength $\log\rad\Phi_n(a)\ge\varphi(n)\log a-o(n)$; the strongest unconditional statements are $\log\rad=O(\log n)$ type. \textup{(D4)} has no standard name: Katz \cite{katz} is a continuous $\mathbb R/\mathbb Z$ equidistribution conjecture, Gras \cite{gras} is a $1/p$ heuristic, and Berger \cite{berger}'s formal-group rigidity treats fixed formal groups. If (D4) is read as a counting density it is already $P_{6a}$; a probabilistic reading must specify its probability space (Proposition~\ref{prop:dual}). \begin{remark}[Transcription check]\label{rem:transcription} Three paraphrases in the commissioned verification report were wrong and are corrected here; the body of this note is unaffected: (i) $v_q(r^{q-1}-1)=s_q$, not $s_q+v_q(q-1)$, since $q\nmid(q-1)/d_q$ (Lemma~\ref{lem:sdident}); (ii) the translation $r\bmod q^3\in\mu_{q-1}$ of Proposition~\ref{thm:teich} stands ($q\le97$ enumeration, zero counterexamples); (iii) the ``primitive divisors occur at most once'' paraphrase is false (counterexamples above). \end{remark} \subsection{The power of structural constraints: zero}\label{sec:zeropower} \begin{theorem}[Zero saving]\label{thm:zerosaving} The small-$q$ layer $q\le\sqrt t$ is unconditionally $o(t)$, and for $q>\sqrt t$ one has $d_q>c_r\log t$ with $c_r=3/(2\log r)$ and $q\nmid d_q$ (no BOOK contamination, cf.\ the stratification trap above). Nevertheless \emph{every} volume-type upper bound obtained from this split is strictly worse than the trivial volume bound: \[ C(t)-V(t)=\sqrt t\,\Bigl[\log r-\frac{6}{\pi^2}\log(r+1)\Bigr]+O(\log t)>0 \qquad\text{for every }r\ge2 \] (coefficient $0.0253$ at $r=2$, $0.2558$ at $r=3$, $1.7995$ at $r=100$, since $\log r>\frac{6}{\pi^2}\log(r+1)$ for all $r\ge2$). Hence all known structural constraints together save \emph{zero} linear mass: the small-$d$ structural emptiness of Remark~\ref{rem:emptylayers} excludes only finitely many $d$ and has no asymptotic force. Moreover, a saturated pseudo-configuration satisfying \emph{all} per-$d$ constraints $3\log q\le\min(\varphi(d)\log(r+1),d\log r)$, $q>\sqrt t$, $q\equiv1\pmod d$ still gives $A(t)/t=0.552/0.775/1.051/1.234$ for $r=2,3,5,7$ at $t=10^3$, stabilising at $0.561/0.783/1.058/1.243$ at $t=10^4$. Thus block/volume/partial-summation methods are exhausted (logical pseudo-configurations, not claims about real factor sets). \end{theorem} \subsection{The ladder has two independent weakenings}\label{sec:incomparable} \begin{theorem}[$A_r(t)=o(t)$ and $P_4^{(3)}$ are incomparable]\label{thm:incomparable} Model M1 ($d_q=q-1$, $s_q=3$, all primes) gives, by Mertens' first theorem $\sum_{p\le x}\frac{\log p}{p}=\log x+O(1)$, $A(t)\approx3\log t=o(t)$ while $P_4^{(3)}\approx\log t\to\infty$; model M2 (one saturated deep prime per $d$ with $\varphi(d)/d\ge\log r/\log(r+1)$) gives $P_4^{(3)}<\infty$ while $A(t)=\delta_r t$ with $\delta_r=0.417/0.289/0.214/0.184/0.129/0.095$ for $r=2,3,5,7,19,100$. Both satisfy the tight structural necessary condition. Hence the ladder \[ P_{6a}\Rightarrow(\mathrm{D4})_{\mathrm{count}} \Rightarrow\Xi_r^{(3)}<\infty \Rightarrow\{A=o(t)\}\wedge\{P_4^{(3)}\} \] has two independent weakenings, and $\Xi_r^{(3)}<\infty$ is strictly stronger than either. \end{theorem} \subsection{\texorpdfstring{The single-variable translation of $P_{6a}$}{The single-variable translation of P6a}}\label{sec:mq} \begin{theorem}[$P_{6a}\iff m(q)\to\infty$]\label{thm:mq} Let $m(q)=\min\bigl(\mu_{q-1}(\bZ/q^3\bZ)\setminus\{1\}\bigr)$ be the least nontrivial Teichm\"uller representative modulo $q^3$. Then \[ P_{6a}\ \Longleftrightarrow\ m(q)\to\infty . \] \end{theorem} \begin{proof} If $q>r^{1/3}$ then $rr^{1/3}\}\subseteq\{q:m(q)\le r\}$; if $m(q)\to\infty$ the right side is finite, and the finitely many primes $q\le r^{1/3}$ add nothing: $\cW_r^{(3)}$ is finite. Conversely, if $m(q)\not\to\infty$, there is $X$ with infinitely many $q$ having $m(q)\le X$; choosing for each such $q$ an element of $\mu_{q-1}\setminus\{1\}$ in $[2,X]$, the pigeonhole principle gives a fixed $r\in\{2,\dots,X\}$ lying in $\mu_{q-1}\bmod q^3$ for infinitely many $q$, all but finitely many with $q>r^{1/3}$; by Proposition~\ref{thm:teich} again, $\cW_r^{(3)}$ is infinite. \end{proof} Numerically, among the $1228$ odd primes $q<10^4$ the median of $m(q)/q^2$ is $0.7222$ (so $m(q)\asymp q^2$, the exact source of the $1/q^2$ heuristic, since $q-1$ elements are spread over $[0,q^3)$). The ten smallest $(q,m(q))$ are $(7,18)$,\allowbreak $(3,26)$,\allowbreak $(23,42)$,\allowbreak $(5,57)$,\allowbreak $(113,68)$,\allowbreak $(11,124)$,\allowbreak $(17,158)$,\allowbreak $(13,239)$,\allowbreak $(47,295)$,\allowbreak $(19,333)$. For each fixed $X\le10^3$ the set $\{q:m(q)\le X\}$ saturates already by $Q=300$ ($3/5/12/15$ primes for $X=50/100/500/1000$, largest such $q=113$), and for $X=10^4$ by $Q=3000$ ($40$ primes, largest $q=1483$). This is the first \emph{fixed-$X$, all-$r\le X$} numerical evidence for $P_{6a}$; all previous scans were fixed-$r$, varying-$q$. Consequently the counting-density reading of (D4) is \emph{equivalent} to $P_{6a}$, while the probabilistic form with finite mean $C_r/q^2$ is strictly stronger (yielding $\Xi_r^{(3)}<\infty$ a.s.). \subsection{Dual exactness and the trivial family}\label{sec:dual} \begin{proposition}[Dual ensemble]\label{prop:dual} In the dual ensemble (fixed $q$, $r$ uniform in $(\bZ/q^3)^\times$) one has the exact identity $\mathbb P(s_q\ge3)=1/q^2$, verified for $q=3,5,7,11,13,17,19$ ($|\mu_{q-1}|=q-1$ throughout). In short intervals $r\ll q^3$ the expectation is correct but the variance is large (at $R=100$ the counts for $q=11,13,17,19$ are all $0$ against expectations $0.82,0.59,0.34,0.27$). In the primal ensemble (fixed $r$, $q$ varying) there is \emph{no} probability space; the missing tool is exactly the transfer dual$\to$primal (cf.\ the ensemble separation of \S\ref{sec:ensembles}). (D4) is therefore not a Katz/Gras theorem and has no standard name. \end{proposition} \begin{theorem}[The trivial family is thin]\label{thm:trivialfam} $d_q=d$ implies $q^3\mid r^d-1\le r^d$, hence $q\le r^{d/3}$; in particular $d_q=1$ forces $q\le(r-1)^{1/3}$, so the trivial family $r\equiv1\pmod{q^3}$ contributes only finitely many $q$ for each fixed $r$. A sweep over $q<5\cdot10^6$ ($348512$ odd primes) and $24$ fixed bases (including the $r=1+q^3$ families $28,126,344,1332,2198$ and large bases $10^4,10^5$), $8.36\cdot10^6$ congruence tests, yields $15$ deep hits, \emph{all} with $q\in\{3,5,7,11,13\}$, $d_q\le6$, $s_q\in\{3,4,5\}$; the bound $q\le r^{d/3}$ has zero violations and the $r=1+q^3$ families give $d_q=1$ exactly. The observed $A_{\mathrm{obs}}(t)$ saturates for $t\ge6$ and is a lower bound truncated at $q\sqrt E$); \item[\textbf{S6}] bypass finite-order congruence expansions (Theorem~\ref{thm:infobar}) or handle the varying-modulus discrete logarithm. \item[\textbf{S7}] \emph{No $r$-uniform output}. By the CRT family of Remark~\ref{rem:perbase}, any bound uniform in $r$ is false; the constant $C_r$ of \textup{(D4)} must depend on $r$, and the tool must produce a statement for each fixed base. \end{enumerate} Self-checks: the tool must not imply $P_4$; must give $\Phi=o(E)$ rather than $\Xi^{(3)}<\infty$; must control the $1/d_q$-weighted mass, not $N_3(E)$ alone. \textbf{T1--T6}: GRH-only, gcd/divisibility-identity-only, greatest-prime-factor/\allowbreak single-power/\allowbreak smooth-part/\allowbreak gcd-only, and structure-constraint/\allowbreak block/\allowbreak volume-only mechanisms all fail (see \S\ref{sec:threshold} and \S\ref{sec:litverdict}--\ref{sec:zeropower}). The \emph{minimal} tool is a single deterministic statement, $m(q)\to\infty$ (Theorem~\ref{thm:mq}), or its quantitative form $\#\{q\le x:m(q)\le X\}=O_X(1)$. \begin{thebibliography}{99} \bibitem{zeng} C.~Zeng, \emph{A lattice-theoretic approach to the $abc$ conjecture}: hexagon:2610.00630v1; Zenodo 22995922 (v1, 2026-09-27); Zenodo 23745797 (v2.0, 2026-10-28, revised). Not peer-reviewed; the v1 abstract (``min--max barrier'') was withdrawn by errata E12/E30 after Theorem~4.5. Hexagon is a deposition platform, not arXiv; no arXiv record 2610.00630 exists. The results of this note are proved independently and use only the problem setting. \bibitem{katz} N.~M. Katz, \emph{Wieferich past and future}, Contemp.\ Math.\ 632 (2015), 253--270, Conjecture 3.1. DOI 10.1090/conm/632/12632. \bibitem{murata} L.~Murata, \emph{The distribution of primes satisfying the condition $a^{p-1}\equiv1\pmod{p^2}$}, Acta Arith. \textbf{58}(2) (1991), 103--111, Conjectures (i), (ii). \bibitem{lang} S.~Lang, \emph{Old and new conjectured Diophantine inequalities}, Bull.\ AMS \textbf{23} (1990), 37--75, p.~42. \bibitem{cdp} R.~Crandall, K.~Dilcher, C.~Pomerance, \emph{A search for Wieferich and Wilson primes}, Math.\ Comp.\ \textbf{66} (1997), 433--449. \bibitem{silverman} J.~H. Silverman, \emph{Wieferich's criterion and the $abc$-conjecture}, J.\ Number Theory \textbf{30} (1988), 226--237. \bibitem{dk} D.~Dorais, D.~Klyve, \emph{A Wieferich prime search up to $6.7\times10^{15}$}, J.\ Integer Seq.\ \textbf{14} (2011). \bibitem{ms} M.~R. Murty, F.~S\'eguin, \emph{Prime divisors of sparse values of cyclotomic polynomials and Wieferich primes}, J.\ Number Theory \textbf{201} (2019), 1--22. \bibitem{hooley} C.~Hooley, \emph{On Artin's conjecture}, J.\ Reine Angew.\ Math.\ \textbf{225} (1967), 209--220. \bibitem{heathbrown} D.~R. Heath--Brown, \emph{The least square-free number in an arithmetic progression}, and Fermat-quotient estimates via Burgess bounds (see Shparlinski's survey, arXiv:1104.3909, for the Ihara-sum improvement). \bibitem{acvz} A.~Alexandru, C.~Cobeli, M.~V\^aj\^aitu, A.~Zaharescu, \emph{Fast covering of $\mathbb{Z}/p\mathbb{Z}$ by Fermat quotients}, J.\ Number Theory \textbf{161} (2022). \bibitem{dillon} R.~Dillon, \emph{The Wieferich locus of an algebraic unit}, Zenodo 21456694 (order-lift dichotomy). \bibitem{chua} K.~Chua, \emph{A note on $p$-adic valuations}, arXiv:2112.04173. \bibitem{cgl} A.~Caruso, Q.~Gazda, G.~Lucas, \emph{On $\phi$-Wieferich places of Drinfeld modules}, arXiv:2412.11588 (Thm.~C gives the family average and explicitly not the fixed-module infinitude). \bibitem{dw} W.~Laniewski, \emph{Defect and order-core analysis of $abc$ data}, arXiv:2609.00039 (Mersenne-line density $47/210$). \bibitem{kurlberg} P.~Kurlberg, C.~Pomerance, \emph{On the periods of the linear congruential and power generators}, Acta Arith. \textbf{119}(2) (2005), 149--169 (unconditional $\ord_q(r)>q^{1/2+\varepsilon(q)}$ on a positive density set; GRH $>q^{1-\varepsilon(q)}$). \bibitem{companion} Companion verification report and scripts for this note (machine-assisted proofs of Theorems~\ref{thm:lift}--\ref{thm:badset} and Proposition~\ref{thm:K} on all tested samples; available from the authors). \bibitem{erdosmurty} P.~Erd\H{o}s, M.~R. Murty, \emph{On the order of $a$ (mod $p$)}, CRM Proc. Lecture Notes \textbf{19} (1999), 87--97. \bibitem{murtysrinivasan} M.~R. Murty, S.~Srinivasan, \emph{Some remarks on Artin's conjecture}, Canad. Math. Bull. \textbf{30}(1) (1987), 78--88. \bibitem{fellinimurty} N.~Fellini, M.~R. Murty, \emph{Wieferich primes in number fields and the conjectures of Ankeny--Artin--Chowla and Mordell}, J.~Number Theory \textbf{278} (2026), 109883 (published version of arXiv:2508.08472v2; the authors note explicitly that no unconditional result in this direction is known). \bibitem{dekoninckdoyon} J.-M. DeKoninck, N.~Doyon, \emph{On the set of Wieferich primes and of its complement}, Ann. Univ. Sci. Budapest. Sect. Comput. \textbf{27} (2007), 3--13 (notation: their $\cW_r$ denotes the depth-$r$ layer for the fixed base $2$, i.e.\ $W_2^{(r)}$; not the base-$r$ set $\cW_r$ used here). \bibitem{gravesmurty} Y.~Ding, \emph{Non-Wieferich primes under the $abc$ conjecture}, C.~R.~Math. Acad. Sci. Paris \textbf{357} (2019), 483--486 (the intermediate bound $C_n'\gg a^{\varphi(n)-\varepsilon n}$ on the squarefree part of $\Phi_n(a)$). See also H.~K. Graves, M.~R. Murty, \emph{The $abc$ conjecture and non-Wieferich primes in arithmetic progressions}, J.~Number Theory \textbf{133} (2013), 1809--1813. \bibitem{stewartyu} C.~L. Stewart, K.~Yu, \emph{On the $abc$ conjecture, II}, Duke Math.~J. \textbf{108}(1) (2001), 169--181 (the $\kappa R^{1/3}(\log R)^3$ bound). \bibitem{stewart13} C.~L. Stewart, \emph{On divisors of Lucas and Lehmer numbers}, Acta Math. \textbf{211} (2013), 291--314. Cf. also Acta Arith. \textbf{26} (1975), 427--433; Proc. London Math. Soc. (3) \textbf{35} (1977), 425--447. \bibitem{balajiluca} N.~Balaji, F.~Luca, \emph{Terms of Lucas sequences having a large smooth divisor}, Canad. Math. Bull. \textbf{66} (2023), 225--231 (arXiv:2203.12364, Theorem~1.2). \bibitem{granvillesound} A.~Granville, K.~Soundararajan, \emph{A binary additive problem of Erd\H{o}s and the order of $2$ (mod $p^2$)}, Ramanujan J. \textbf{2} (1998), 283--298. \bibitem{moreesurvey} P.~Moree, \emph{Artin's primitive root conjecture --- a survey}, Integers \textbf{12A} (2012), \#A13. \bibitem{bugeaudcorvajazannier} Y.~Bugeaud, P.~Corvaja, U.~Zannier, \emph{An upper bound for the greatest common divisor of $a^n-1$ and $b^n-1$}, Math.~Z. \textbf{243} (2003), 79--84. \bibitem{biluhanrotvoutier} Y.~Bilu, G.~Hanrot, P.~M. Voutier, \emph{Existence of primitive divisors of Lucas and Lehmer numbers}, J.~Reine Angew. Math. \textbf{539} (2001), 75--122. \bibitem{gras} G.~Gras, \emph{Probabilistic study of Fermat quotients}, Comment. Math. \textbf{54} (2016), 115--140 (heuristic, not a pointwise theorem). \bibitem{berger} L.~Berger, \emph{Rigidity and unlikely intersections for formal groups}, arXiv:1811.05824v2 (framework only; not a pointwise theorem for fixed $r$/varying $q$). \bibitem{xie} W.~Xie, \emph{Primary components and large prime factors of cyclotomic values}, arXiv:2611.00075v1 (math.RA; v1 PDF header dated 16 Oct 2026, identifier/date incompatible; not peer-reviewed). \end{thebibliography} \end{document}