\section{Parity forces primes from \texorpdfstring{$\nu = 0.116$}{nu = 0.116}}\label{sec:converse}
This section proves Theorem~\ref{thm:B}. We work with Ford and Maynard's parameters $P = (1/2, 0, \nu)$, except in
Lemma~\ref{lem:arith715} and in the derivation of \eqref{eq:Ag-divisors}, which allow general parameters $P = (\gamma, \theta, \nu)$;
the objects below are taken for the parameters in use. We use their notation: $|z|$ and $\dim z$ for the sum and the number of
components of a vector $z$, and $z_A$ for its subvector indexed by $A \subseteq [\dim z]$ \cite[Definitions~4.1 and~4.2]{FM};
$[k] = \{1, \dots, k\}$ and $\Omega(n)$ \cite[\S5.1]{FM};
$R = R(P)$ for the fundamental region and $C(R)$ for the set of coagulations of vectors of $R$ \cite[Definitions~4.3 and~4.5]{FM};
$v(n)$ and $v(m; n)$ for the vectors of the numbers $\log p/\log n$ over the prime factors $p$ of $n$, respectively of $m$, counted
with multiplicity and in increasing order \cite[Definition~5.1]{FM}; $P^-(n)$ and $P^+(n)$ for the smallest and the largest prime
factor of $n$, with $P^-(1) = \infty$ and $P^+(1) = 1$ \cite[\S5.1]{FM}; $\mathcal P$ for the set of primes in $(x/2, x]$ and
$n \sim x$ for $x/2 < n \le x$ \cite[\S7]{FM}; $\mathcal H(P)=\{z\in C(R(P)):\dim z\ge2,\ z_i\ge\nu\text{ for all }i\}$;
$G_1$ for the set of admissible weight vectors \cite[(7.1)]{FM}; ``admissible $g$'' for a function in their class of weight
functions supported on $G_1$ \cite[Definition~7.2]{FM}; $\mathcal N = \{x/2 < n \le x : n \text{ composite},\ v(n) \in C(R)\}$ \cite[(7.12)]{FM}; and, for $g$ on $G_1$, the
ordinary-divisor weight $A_g(n)$ of \cite[(7.13)--(7.14)]{FM} and the vector convolution $V_g = 1 \star g$ \cite[Definition~7.1]{FM}.
Numbered items cited from \cite{FM} carry Ford and Maynard's numbering; Lemma~\ref{lem:arith715},
Proposition~\ref{prop:sieve-parity} and Lemma~\ref{lem:par-star} are this paper's.
On squarefree $n\in\mathcal N$ with $P^-(n)\ge n^\nu$, the ordinary-divisor
weight $A_g(n)$ agrees with $V_g(v(n))$. On non-squarefree $n$ the weights
can differ, because ordinary divisors are counted once while the vector
convolution counts subvectors by index. For example, take
$P=(1/2,0,\nu)$ with $0<\nu<1/4$ and an admissible $g$ with
$g(\emptyset)=1$ and $g((1/2))=-1/2$; its support forces it to vanish on
vectors of sum $1$. If $x/2
(p^2)^\nu$, but $A_g(p^2)=1+g((1/2))=1/2$, whereas
$V_g((1/2,1/2))=1+2g((1/2))=0$. Ford and Maynard identify the
ordinary-divisor weight with the corresponding vector convolution in
the second sentence after \cite[(7.14)]{FM}, beginning ``The above definition
of $H$ is equivalent'', and, with a roughness indicator, in
\cite[Lemma~7.18(a)]{FM} and the final sentence of the proof of part~(a).
The same counting issue occurs in the expansion over $J\subseteq[k]$
at the start of the proof of \cite[Proposition~7.22]{FM}, there for
$n\notin\mathcal P\cup\mathcal N$. For our application we need \cite[(7.15)]{FM} for the ordinary-divisor
weight, which is how it is stated; its printed proof needs
corrections, one of them at repeated prime factors, and
Lemma~\ref{lem:arith715} gives the argument in full. It follows the steps of the proofs of \cite[Lemma~7.21 and Proposition~7.22]{FM}, the latter with their parameters
$\sigma = \nu$ and $\lambda(d) = \mu(d)$.
\begin{lemma}\label{lem:arith715}
Let $(\gamma, \theta, \nu)$ satisfy Ford and Maynard's standing conditions \cite[(6.1)]{FM} with $\gamma \ge 1/2$ and
$\nu \le 1 - \gamma$, let $g$ be admissible, and let $A_g(n) = \sum_{d \mid n} G(d; n)$ be the weight of
\cite[(7.13)--(7.14)]{FM}. Let $A \ge 1$ and $\varpi \ge 1$. If $B$ is large enough in terms of
$A, \varpi, \gamma, \theta, \nu$ and $g$, and $(w_n)$ satisfies (w), (I) and (II), then
$\sum_{n \in \mathcal P \cup \mathcal N} w_n A_g(n) \ll_{g, A} x(\log x)^{-A}$.
\end{lemma}
\begin{proof}
All constants may depend on $\gamma, \theta, \nu$ and $g$, and $x$ is taken large in terms of these and of the logarithmic exponents
below. For $d \le x^\gamma$ put
\[
\Delta_d = \sup_{\text{interval } I}\, \Bigl| \sum_{\substack{dm \sim x \\ m \in I}} w_{dm} \Bigr|,
\]
so that (I) gives $\sum_{d \le x^\gamma} \tau(d)^{C} \Delta_d \le x(\log x)^{-B}$ for every $C \le B$. For a bounded function $h$ on
vectors, write $A_h(n) = \sum_{d \mid n} G_h(d; n)$, with $G_h$ defined as $G$ in \cite[(7.13)]{FM} with $h$ in place of $g$.
\emph{Boundary pieces.} Let $h$ be bounded and supported on vectors of a fixed dimension $j \ge 1$ that lie on a proper affine
hyperplane $c_1 t_1 + \dots + c_j t_j = c_0$, $(c_1, \dots, c_j) \ne 0$. A term of $A_h(n)$ comes from $d = m_1 m_2 \mid n$,
$d \le n^\gamma$, with $m_2 = p_1 \cdots p_j$, $p_i \ge n^\nu$, and $\sum_i c_i \log p_i = c_0 \log n$. If $c_0 \ne 0$, then
$m_2$ determines $n$, so for each $d \le x^\gamma$ at most $\tau(d)$ integers $n \sim x$ receive a term from $d$, and each $|w_n|$
is at most $\Delta_d$ (take $I$ to be the single cofactor). Hence $\sum_{n \sim x} |w_n A_h(n)| \ll \sum_{d \le x^\gamma} \tau(d) \Delta_d$.
If $c_0 = 0$, let $\mathcal D$ be the set of $m_2 \le x^\gamma$ whose $j$ prime factors are at least $P_0 = (x/2)^\nu$ and satisfy
$\sum_i c_i \log p_i = 0$. Choosing $i$ with $c_i \ne 0$, the other $j - 1$ primes determine $p_i$, so
$\sum_{m_2 \in \mathcal D} 1/m_2 \ll P_0^{-1} (\log x)^{j-1}$. With $S_d = \#\{m_2 \mid d : m_2 \in \mathcal D\}$ and
$\rho(n) = \sum_{d \mid n,\, d \le x^\gamma} S_d$ we have $|A_h(n)| \ll \rho(n)$,
\[
\sum_{n \sim x} \rho(n) \le x \sum_{d \le x^\gamma} \frac{S_d}{d} \ll x^{1 - \nu} (\log x)^{j}, \qquad
\Bigl| \sum_{n \sim x} \rho(n) w_n \Bigr| \le \sum_{d \le x^\gamma} \tau(d) \Delta_d,
\]
and, since $w_n \ge -x^{\nu/10}$ by (w), $\sum_{n \sim x} |w_n| \rho(n) \le |\sum_n \rho(n) w_n| + 2x^{\nu/10} \sum_n \rho(n)$. In both cases
$\sum_{n \sim x} |w_n A_h(n)| \ll x(\log x)^{-B} + x^{1 - 9\nu/10} (\log x)^{j}$, and the bound holds for the sum over any set of $n$.
\emph{Reduction to Lipschitz pieces.} Write the restriction of $g$ to ordered vectors as the finite sum of functions $g_a$ on polytopes $Q_a$ given by its definition, and
work with each summand separately. In ambient dimension $j \ge 1$, intersect $Q_a$ with $\{t_i \ge \nu,\ \sum_i t_i \le \gamma\}$, and
call the intersection $P_a$. This does not change any contributing term of $A_{g_a}$. If $P_a$ is full-dimensional, the bounded first
derivatives of $g_a$ on its interior give a bounded Lipschitz function there, which extends continuously to its closure. Restrict this
extension to $P_a$. Its difference from $g_a \mathbf 1_{P_a}$ is supported on the boundary of $P_a$, which is contained in a finite union of proper affine
hyperplanes. If $P_a$ is not full-dimensional, including when it is a point in $\mathbb R^j$ with $j \ge 1$, the whole piece is supported
on a proper affine hyperplane. The boundary estimate treats all these differences in absolute value, over any set of integers. Keep
the value $g(\emptyset)$ separately. By linearity it therefore suffices to handle a bounded Lipschitz function $\tilde g$ on a convex
polytope $\mathcal G$, multiplied by its indicator, and the term at $\emptyset$.
\emph{The full sum.} By linearity fix one Lipschitz piece and fix $m_1, m_2$ with $m_1 m_2 \le x^\gamma$. Write $m_2 = q_1 \cdots q_j$
with prime factors counted with multiplicity; for $m_2 = 1$ this is the empty list. The conditions $m_1 m_2 \le n^\gamma$,
$P^+(m_1) < n^\nu \le P^-(m_2)$ and $v(m_2; n) \in \mathcal G$ restrict $n \sim x$ to an interval, since
\[
v(m_2; n) = (\log q_1, \dots, \log q_j)/\log n
\]
moves along a ray. On this interval the Lipschitz bound gives $|\tilde g(v(m_2; t)) - \tilde g(v(m_2; t'))| \ll |t - t'|/(x \log x)$.
Its total variation, including the bounded endpoint jumps when extended by zero, is therefore $O(1)$. Partial summation within the
interval and (I) give
\[
\sum_{x/2 < n \le x} w_n A_{\tilde g \mathbf 1_{\mathcal G}}(n) \ll \sum_{d \le x^\gamma} \tau(d) \Delta_d \le x(\log x)^{-B}
\]
for each reduced Lipschitz piece $(\tilde g, \mathcal G)$, and likewise for the term at $\emptyset$. This is the Type I argument of \cite[Lemma~7.21]{FM}. It remains to bound their sum over
$n \notin \mathcal P \cup \mathcal N$.
\emph{The complementary sum: factorisation and polytopes.} Write uniquely
\[
n = a p_1 \cdots p_k, \qquad P^+(a) < n^\nu \le p_1 \le \dots \le p_k, \qquad 0 \le k \le 1/\nu.
\]
This is the factorisation of \cite[(7.13)]{FM}. Let $\mathcal U_1, \dots, \mathcal U_{J_0}$ be the complement polytopes
constructed below, in positive ambient dimension, for small vectors with positive coordinates less than $\nu$, including the empty vector. For $n \sim x$ with $\Omega(n) \ge 2$, the
condition $v(n) \notin C(R)$ is equivalent to $k = 0$ or $v(p_1 \cdots p_k; n) \in \bigcup_j \mathcal U_j$. For $k = 0$, all coordinates
of $v(n)$, and hence of every fragmentation of $v(n)$, are less than $\nu$; no such fragmentation lies in $R$
\cite[Lemma~5.6(i)]{FM}, so $v(n) \notin C(R)$. For $k \ge 1$, use the form verified below with the
small vector $v(a; n)$. In particular, if $t \in \mathcal U_j$ and $q$ is any vector, possibly empty, of positive coordinates less
than $\nu$ with $|q| = 1 - |t|$, then $(t, q) \notin C(R)$.
We use \cite[Lemma~5.8]{FM} only for small vectors $y$ whose coordinates are positive and less than $\sigma = \nu$, possibly
$y = \emptyset$, and we check this form directly, following its proof. As there, $(y, x) \in C(R)$ if and only if $x$ is a
coagulation of a vector $z$ of bounded dimension with positive coordinates and $(y, z) \in R$. If $y \ne \emptyset$, then for each
$J \subseteq [\dim z]$ the sums $|z_J| + y_1 + \dots + y_i$ $(0 \le i \le \dim y)$ increase from $|z_J|$ to $|z_J| + |y|$ in steps
less than $\nu$; using $y_1 < \nu$ when $J = \emptyset$, and $\theta + \nu < 1$ when $J = [\dim z]$, it follows that $(y, z) \in R$
if and only if $|y| + |z| = 1$, $z \in (0, 1 - \gamma)^{\dim z}$, and $|z_J| > \theta + \nu$ or $|z_J| < \theta - |y|$ for every
$J \subseteq [\dim z]$, a condition on $|y|$ alone. If $y = \emptyset$, the same holds with $J$ restricted to
$\emptyset \ne J \ne [\dim z]$, since the empty and the full subsum are not proper. With $\mathcal Q_\ell(\mathcal J)$ as in the proof
of \cite[Lemma~5.8]{FM}, define $\mathcal Q^0_\ell(\mathcal J)$ in the same way using only the proper nonempty subsets $J$ of
$[\ell]$. For $1 \le k \le \lfloor 1/\nu \rfloor$ and $k \le \ell \le k + \lfloor 2/(1-\gamma) \rfloor$, fix an ordered partition
$\pi = (I_1, \dots, I_k)$ of $[\ell]$ into nonempty parts and write $T_\pi(z) = (|z_{I_1}|, \dots, |z_{I_k}|)$. Each of the sets
\[
T_\pi(\mathcal Q_\ell(\mathcal J)) \cap [\nu, 1]^k \cap \{|x| < 1\}, \qquad
T_\pi(\mathcal Q^0_\ell(\mathcal J)) \cap [\nu, 1]^k \cap \{|x| = 1\}
\]
is a convex polytope with boundedly many constraints, each strict or not: $T_\pi$ is linear, and eliminating the
coordinates of $z$ (Fourier--Motzkin elimination, which applies to strict and non-strict inequalities alike) describes the image
by boundedly many such constraints, as at the end of the proof of \cite[Lemma~5.8]{FM}. Take their union over the finitely many choices of $\ell$, $\mathcal J$ and $\pi$.
Its intersection with $\{|x| = 1 - |y|\}$ is exactly the set of $x$ with $x_i \ge \nu$ and $(y, x) \in C(R)$: the first family
applies when $y \ne \emptyset$, and the second when $y = \emptyset$. For each $k$, subdivide the complement of this union in
$[\nu, 1]^k \cap \{|x| \le 1\}$ into disjoint convex polytopes. These are the $\mathcal U_j$. The number of pieces and of their
defining constraints is bounded in terms of the fixed parameters.
We split the sum according to the polytope, to the set
$\mathcal W = \{i : p_i = p_{i+1}\}$, and to the term of $A_g$: if $J$ runs over the subsets of $[k]$ that meet each maximal
block of equal primes in an initial segment, then $J \mapsto \prod_{j \in J} p_j$ is a bijection onto the divisors of
$p_1 \cdots p_k$, and
\[
A_g(n) = \sum_{u \mid a} \sum_{J} \mathbf 1\bigl(u \textstyle\prod_{j \in J} p_j \le n^\gamma\bigr)\, \mu(u)\,
g\bigl(v(\textstyle\prod_{j \in J} p_j; n)\bigr).
\]
There are $O(1)$ choices of $k$, $\mathcal W$, $J$ and a piece $(\tilde g, \mathcal G)$; the term at $\emptyset$ is the piece
$\mathcal G = \{\emptyset\}$, $\tilde g \equiv g(\emptyset)$, which occurs only with $J = \emptyset$, and a piece in $\mathbb R^j$ with
$j \ge 1$ occurs only with $|J| = j$. For $k \ge 1$ and each such choice,
let $\mathcal T'$ be the intersection of the selected complement polytope with $\{\nu \le t_1 \le \dots \le t_k\}$, with $t_i = t_{i+1}$ $(i \in \mathcal W)$ and
$t_i < t_{i+1}$ $(i \notin \mathcal W)$, and with $\{t : t_J \in \mathcal G\}$. It is a convex polytope with boundedly many
constraints, each strict or not, and it keeps the property just stated. For each $u \mid a$ write $v = a/u$, so that $a = uv$. For $k \ge 2$ the condition $\Omega(n) \ge 2$
is automatic; for $k \le 1$ it is $uv > 1$, which we keep. For $k = 0$ there is one empty equality pattern, $J = \emptyset$, and the
factor $g(\emptyset)$ is constant; no $\mathcal T'$ condition is imposed.
In the reductions below, fixed bounds for the separate coefficient functions are absorbed into the implied constants, so that these
functions may be taken $1$-bounded. When variables are grouped to apply \cite[Lemmas~7.9 and~7.17]{FM}, the coefficient arrays are
bounded by $|w_n| \tau(n)^C$ for a fixed $C$, and are set to zero off $n \sim x$. By (w) and the divisor bound,
\[
|w_n| \tau(n)^C \le x(\log x)^\varpi \tau(n)^C < x^2
\]
for sufficiently large $x$. Thus their coefficient hypotheses hold.
\emph{Prime decomposition ($k \ge 1$).} We apply the box and prime decomposition argument of \cite[Lemma~7.14]{FM} to the tuples
with $P^+(uv) < n^\nu \le p_1$, taking the first logarithmic box interval for $u$ and $v$ closed at zero, to include unit factors.
The argument needs only the following: for fixed $(u, v)$ the
conditions define a convex region in the coordinates $\log p_i/\log x$; $p_i \ge n^\nu > (x/2)^\nu$ gives the lower bound used
there; within each box the Lipschitz bound for $\tilde g$ applies, since $\mathcal T' \subseteq \{t : t_J \in \mathcal G\}$, and, as
$n > x/2$, each $\log p_i/\log n$ varies over a box by at most a constant times its side, also with the first interval closed at zero;
the factorisation of $n$ is unique, so the error of the box approximation is
$O((\log x)^{-D} \sum_n |w_n| \tau(n))$ with a fixed $D \ge A + \varpi + 2$, taking $B$ sufficiently large after $D$ is chosen; and
\cite[Lemma~7.12]{FM} applies to these regions, including the faces given by $\mathcal W$; with the first box interval closed at zero, the box condition for $u$ or $v$ is a
one-sided bound $u \le x^c$ or $v \le x^c$, which \cite[Lemma~7.9]{FM} removes as in the proof of \cite[Lemma~7.13]{FM}. From here on $\ell = 6\lceil 1/(1-\gamma)\rceil$, as in \cite[Lemma~7.12]{FM}.
In applying the prime decomposition here and in \S\ref{ssec:large-nu}, we make the following cutoff explicit. In the proof of \cite[Lemma~7.12]{FM} (in the notation of that proof), retain
$f_{h,j}^{r_h-1}\le x$ when separating the identity $d_{h,j}=f_{h,j}^{r_h-1}$.
This cutoff is automatic before separation, since $d_{h,j}\mid n\le x$, and is absorbed into the bounded coefficient of $f_{h,j}$ afterwards.
On its support the compared quantities $d_{h,j}$ and $f_{h,j}^{r_h-1}\pm\tfrac12$ lie in $[1/2,x+1/2]$, and their distance is at least $1/2$.
Off its support, use the constant functions $1/2$ and $3/2$ in place of $f_{h,j}^{r_h-1}-1/2$ and $f_{h,j}^{r_h-1}+1/2$, respectively; the coefficient there is zero.
The coefficient arrays there are sums over the remaining variables, so they are at most $\max_{n\le x}\tau_K(n)\sum_n|w_n|\tau(n) x^{\nu/3}\ (i < s), \qquad v_j > x^{\nu/3}\ (j < r), \qquad u_s, v_r \le x^{\nu/3}.
\]
Let $\mathcal I$ be the indicator of the coarse conditions
\[
u_i > x^{\nu/3}\ (i < s), \qquad v_j > x^{\nu/3}\ (j < r), \qquad u_s, v_r \le x^{\nu/3}, \qquad m_h > (x/2)^\nu\ (h \le k).
\]
The last conditions follow from $\tilde y_h \ge \nu$. Retain $u_i < n^\nu$ for $i < s$ and $v_j < n^\nu$ for $j < r$ as separate summation
conditions. The corresponding bounds for $u_s, v_r$ follow from their coarse upper bounds and $n > x/2$, for large $x$.
\emph{A Type II factor.} Form $z$ by taking $\log b/\log n$ for every entry $b$ of the indexed list
\[
(u_1, \dots, u_s, v_1, \dots, v_r, (d_{h,j})_{h,j}),
\]
omitting entries equal to $1$. Equal entries at different indices are retained.
The coordinates of $z$ are positive, sum to $1$, and are less than $1 - \gamma$, since $u_i, v_j < n^\nu \le n^{1-\gamma}$ and
$d_{h,j} < n/x^\gamma \le n^{1-\gamma}$. Let $q$ be the vector of $\log u_i/\log n$ and $\log v_j/\log n$ for those $u_i, v_j > 1$.
Its coordinates are less than $\nu$ and $|q| = 1 - |\tilde y|$. If $k \ge 1$ and $z \in R$, joining the $d$-coordinates belonging to each
$m_h$ gives $(\tilde y, q) \in C(R)$, contrary to the property of $\mathcal T'$. Hence, for $k \ge 1$, $z \notin R$, so it has a nonempty
proper subsum in $[\theta, \theta + \nu]$. If $k = 0$, every
coordinate of $z$ is less than $\nu$; adding coordinates until the sum reaches $\theta$ (or taking one coordinate if $\theta = 0$)
gives a positive subsum in $[\theta, \theta + \nu]$, proper since $\theta + \nu < 1$. In both cases the subsum lies in
\[
\mathcal K_n^* = \Bigl( \theta \frac{\log(x/2)}{\log n},\ (\theta + \nu) \frac{\log x}{\log n} \Bigr],
\]
because $n > x/2$ when $\theta > 0$, the subsum is positive when $\theta = 0$, and $n \le x$. Label all blocks, units included, as
$n_1, \dots, n_K$, where $n_i = u_i$ for $1 \le i \le s$, $n_{s+j} = v_j$ for $1 \le j \le r$, followed by the $d_{h,j}$
in lexicographic order; thus $K = k\ell + r + s$. For nonempty $E \subseteq [K]$, $\sum_{e \in E} \log n_e/\log n \in \mathcal K_n^*$ is
equivalent to $(x/2)^\theta < \prod_{e \in E} n_e \le x^{\theta + \nu}$, the range of (II), and this excludes products of units.
By inclusion--exclusion over the nonempty subsets of $[K]$ it suffices to treat sums with these conditions for every $E$ in a fixed
nonempty collection $\mathcal E$. We remove each condition $d_{h,j} < n/x^\gamma$, written as
$n/d_{h,j} > \lfloor x^\gamma \rfloor + 1/2$, by \cite[Lemma~7.9]{FM}; the twists it introduces depend on one block each and are
absorbed into the coefficients.
\emph{Removing the polytope conditions.} We treat $uv = 1$ and $uv > 1$ separately. If $uv = 1$, then every $u_i$ and $v_j$ is $1$,
these unit conditions are absorbed into the coefficients, $\mathcal I$ forces $r = s = 1$, and $k \ge 2$, since for $k \le 1$ we
kept $uv > 1$. On both branches we keep every constraint of $\mathcal T'$, the cutoff $u \prod_{h \in J} m_h \le n^\gamma$, and the strict
bounds $u_i < n^\nu$ $(i < s)$ and $v_j < n^\nu$ $(j < r)$ as separate conditions to be removed by \cite[Lemma~7.11]{FM}. Its designated blocks are the sets of variables whose products are the $m_h$, and the single
variables $u_i$ $(i < s)$ and $v_j$ $(j < r)$; on $\mathcal I$ each has product at least $x^{\nu/3}$, as that lemma requires with the
$\nu$ of (w). At least one exists, since $k = 0$ and $r = s = 1$ would give $n \le x^{2\nu/3}$. Each condition is a monomial
inequality $n_1^{c_1} \cdots n_K^{c_K} \le 1$ or $< 1$ with equal nonzero exponents on some designated block, as follows. For
$i < s$, the condition $u_i < n^\nu$ is $n_i^{1-\nu} \prod_{e \ne i} n_e^{-\nu} < 1$, with the designated singleton $\{i\}$; the same
argument applies to each $v_j$ with $j < r$ using its own index. The
cutoff has exponent $1 - \gamma$ on the variables of $u$ and of $m_h$ $(h \in J)$ and $-\gamma$ on the others, and every block is
wholly of one kind. A linear constraint $\sum_h e_h \tilde y_h \le e_0$ (or $< e_0$) of $\mathcal T'$ reads
$(uv)^{-e_0} \prod_h m_h^{e_h - e_0} \le 1$ (or $< 1$); if $e_h \ne e_0$ for some $h$, the block of $m_h$ serves. If $e_h = e_0$ for
every $h$, the constraint is a condition on $|\tilde y|$ alone, and $|\tilde y| = 1$ when $uv = 1$, $|\tilde y| < 1$ when $uv > 1$, so on each branch
it either always holds, and is dropped, or never holds, and the summation piece is empty. The constraints from $\mathcal W$ have
$e_0 = 0$ and nonzero exponents on the blocks of $m_i$ and $m_{i+1}$; an equality is written as two weak inequalities. After a
bounded positive rescaling the exponents lie in $[-D', -1] \cup \{0\} \cup [1, D']$ for a fixed $D'$. We apply \cite[Lemma~7.11]{FM} with $\phi_{\mathbf n}$
equal to the product of the coefficients, $\mathcal I$, the branch condition and the conditions indexed by $\mathcal E$; both sums
it produces have the same form, and its error term is $O(x^{1 - \nu/20})$. The block conditions
$\prod_{e \in \mathcal M} n_e \ge x^{\nu/3}$, one for each designated block $\mathcal M$, that remain in those sums are implied by $\mathcal I$, which is part of
$\phi_{\mathbf n}$, and are dropped.
\emph{Type II.} Every condition in $\mathcal I$ now compares an integer product $\Pi$ with a fixed threshold $\Xi$. Write a strict lower
condition as $\Pi > \lfloor \Xi \rfloor + 1/2$ and a weak upper condition as $\Pi \le \lfloor \Xi \rfloor + 1/2$. Apply \cite[Lemma~7.9]{FM}
to these conditions and, on the second branch, to $uv > 1$, written as $u_1 \cdots u_s v_1 \cdots v_r > 3/2$. The unit indicators on the
first branch remain in the separate coefficients. If $[K] \in \mathcal E$, the whole summation piece is empty for large $x$, since
$n > x/2 > x^{\theta + \nu}$, and is discarded. Otherwise every member of $\mathcal E$ is proper. Choose one $E \in \mathcal E$ and remove
the interval conditions indexed by the other members of $\mathcal E$ using the same half-integer thresholds. The gaps are at least
$1/2$ and both compared quantities lie in $[1/2, 2x]$, as \cite[Lemma~7.9]{FM} requires, and all twists factor into
one-variable coefficients. The remaining sum is
\[
\sum_{\substack{n' n'' \sim x \\ (x/2)^\theta < n' \le x^{\theta + \nu}}} Y_1(n') Y_2(n'') w_{n' n''},
\]
where $Y_1$ and $Y_2$ sum the coefficients over the factorisations of $n'$ into the blocks of $E$ and of $n''$ into the others, so
that $|Y_1(n')| \le \tau(n')^K$ and $|Y_2(n'')| \le \tau(n'')^K$. By (II) it is at most $x(\log x)^{-B}$ once $B \ge K$. All losses
were fixed powers of $\log x$ and power savings, so for $B$ large in terms of $A$ and $\varpi$ the complementary sum is $O(x(\log x)^{-A})$.
Subtracting the complementary sum from the full sum, and restoring the boundary pieces by their absolute-value estimate, proves the
lemma for the original $g$.
\end{proof}
We use Lemma~\ref{lem:arith715} for $A_g$ and, separately for the vector weight $V_g$, condition (b.2) as in the proof of \cite[Lemma~7.20]{FM};
Proposition~\ref{prop:sieve-parity} imposes an arithmetic sign condition on non-squarefree $n \in \mathcal N$ with $P^-(n) \ge n^\nu$ outside an
exceptional set with negligible absolute $w$-mass.
For any parameters $P = (\gamma, \theta, \nu)$, let $g$ be admissible and $n \in \mathcal N$. As in the proof of \cite[Lemma~7.18(a)]{FM},
the cutoff $d \le n^\gamma$ in $A_g(n)$ can be dropped. Indeed, let $d = m_1 m_2 \mid n$ with $P^+(m_1) < n^\nu \le P^-(m_2)$ and
$g(v(m_2; n)) \ne 0$, so that $v(m_2; n) \in G_1$. Since $m_1 m_2 \mid n$, the vector $(v(m_2; n), v(m_1; n))$ is a subvector of a
rearrangement of $v(n) \in C(R)$, and $C(R)$ is closed under rearrangement, because the parts of a coagulation may be listed in any
order; the coordinates of $v(m_1; n)$ are less than $\nu$. So the definitions of $\psi$ and $G_1$ in \cite[(7.1)]{FM} give
$|v(m_1; n)| \le \psi(v(m_2; n)) \le \gamma - |v(m_2; n)|$, that is, $d \le n^\gamma$. The divisors of $n$ are the products
$m_1 m_2$ of a divisor $m_1$ of $n$ with $P^+(m_1) < n^\nu$ and a divisor $m_2$ of $n$ with $P^-(m_2) \ge n^\nu$, each obtained
once, so $A_g(n) = \sum_{m_1} \mu(m_1) \sum_{m_2} g(v(m_2; n))$. The first sum is $\mathbf 1_{P^-(n) \ge n^\nu}$, and when it is $1$
every divisor of $n$ is an $m_2$. Hence
\[
A_g(n) = \mathbf 1_{P^-(n) \ge n^\nu} \sum_{m \mid n} g(v(m; n)) \qquad (n \in \mathcal N), \tag{$\ast$}\label{eq:Ag-divisors}
\]
the sum running over the ordinary divisors $m$ of $n$. In particular, $A_g(n) = 0$ for $n \in \mathcal N$ with $P^-(n) < n^\nu$, and
$A_g(n) = V_g(v(n))$ for squarefree $n \in \mathcal N$ with $P^-(n) \ge n^\nu$, whose divisors correspond to the subvectors of $v(n)$.
\subsection{A sieve bound with one parity term}
\begin{proposition}\label{prop:sieve-parity}
Let $P = (1/2, 0, \nu)$ with $R(P)$ nonempty, and let $g$ be admissible, extended by zero outside $G_1$, with $g(\emptyset) = 1$, and
such that its restriction to ordered vectors is a finite sum of functions each supported on a convex polytope and Lipschitz continuous
on it. For $c < 1$ put $\Phi(x) = V_g(x) + c\,(-1)^{\dim x}$. Let
$\mathcal N_r = \{n \in \mathcal N : P^-(n) \ge n^\nu\}$. Assume, in addition to Ford and Maynard's hypotheses, parity balance on the primes
and $\mathcal N_r$: $\bigl|\sum_{n \in \mathcal P \cup \mathcal N_r} \mu(n) w_n\bigr| \le x(\log x)^{-B}$. Suppose that $\Phi \le 0$ on
$\mathcal H(P)$, and that $A_g(n) \le 0$ on non-squarefree $n \in \mathcal N_r$ apart from an exceptional set $E_x$ such that
$\sum_{n \in E_x} |w_n| \le 2x(\log x)^{-B}$ for every pair in Ford and Maynard's class at exponent $B$ and every large $x$. Then
$C^-(1/2, 0, \nu)$, taken over the class with the parity condition on $\mathcal P \cup \mathcal N_r$ adjoined, satisfies
\[
C^- \;\ge\; 1 + \frac{1}{1 - c} \sum_{k \ge 2} \int_{\mathcal H(P) \cap \{x_1 \le \dots \le x_k\}} \frac{\Phi(x)}{x_1 \cdots x_k}\, dx .
\]
\end{proposition}
\begin{proof}
Put $F(n) = A_g(n) + c\,\mu(n)\mathbf 1_{P^-(n) \ge n^\nu}$. A vector of $R(P)$ has components in $(0, 1/2)$ and sum $1$, so it has at
least three components, each a proper subsum and hence greater than $\nu$; thus $\nu < 1/3$, and in particular
$\nu \le 1/2 = 1 - \gamma$, so Lemma~\ref{lem:arith715} applies. Together with the parity hypothesis it gives
$\sum_{\mathcal P \cup \mathcal N} w_n F(n) \ll_A x/\log^A x$ once $B \ge A$ is large enough. On primes $F = 1 - c$; on squarefree
$n \in \mathcal N_r$, $F = \Phi(v(n)) \le 0$, by \eqref{eq:Ag-divisors} and the hypothesis on $\Phi$; on non-squarefree
$n \in \mathcal N_r$, $F = A_g(n)$, which is $\le 0$ off $E_x$; and $F = 0$ on $\mathcal N \setminus \mathcal N_r$, where $A_g = 0$
by \eqref{eq:Ag-divisors}. On $\mathcal N_r$ we have $\Omega(n) \le 1/\nu$, so $|A_g(n)| \le \|g\|_\infty 2^{1/\nu}$ by
\eqref{eq:Ag-divisors}, and $F$ and $\Phi(v(n))$ are bounded there. Since $a_n \ge 0$, and the terms from $E_x$ are
$O(x(\log x)^{-B})$ by the hypothesis on $E_x$,
$(1 - c)\sum_p w_p \ge \sum_{n \in \mathcal N_r \setminus E_x} b_n F(n) - O_A(x/\log^A x)$.
The last sum is $(\sum_p b_p)$ times the sum of integrals in the statement, with error $O_g((\sum_p b_p)/B)$. Here $\mathcal N_r$ is the set of $n \sim x$ with
$P^-(n) \ge n^\nu$ and $v(n) \in \mathcal H(P)$, and $\mathcal H(P) \cap \mathbb R^k$ is the set of $x$ with $|x| = 1$ and every
$x_i \in (\nu, 1 - 2\nu) \cup (2\nu, 1 - \nu)$ \cite[Lemma~8.4]{FM}. So its ordered vectors form a finite disjoint union of convex
polytopes inside $C(R) \cap \{\nu \le x_1 \le \dots \le x_k\}$, where Ford and Maynard's condition (b.2) applies. For
$J \subseteq [k]$ and one of the functions of the hypothesis on $g$, Lipschitz on a convex polytope $\mathcal G$, the ordered $x$ of
one of these polytopes with $x_J \in \mathcal G$ form a convex set, on which that function of $x_J$ is Lipschitz. Applying (b.2) on
each such set, as in the proof of \cite[Lemma~7.20]{FM}, and summing over $J$ and the functions gives the part from
$V_g(v(n)) = \sum_{J \subseteq [k]} g(v(n)_J)$, with error $O_g((\sum_p b_p)/B)$; (b.2) with $f = 1$ on the same polytopes gives the
part from $c\,(-1)^k$. The faces with repeated coordinates are convex sets of projection measure zero, so (b.2) with $f = 1$, and
$b_n \ge 0$, bound their $b$-mass by $O_g((\sum_p b_p)/B)$. As $F$ and $\Phi(v(n))$ are bounded on $\mathcal N_r$, replacing $F(n)$
by $\Phi(v(n))$ at the non-squarefree $n \in \mathcal N_r$, including those in $E_x$, changes the sum by $O_g((\sum_p b_p)/B)$.
Finally, by (b.1) the error $O_A(x/\log^A x)$ is $o(\sum_p b_p)$ for $A > \varpi$; then use Ford and Maynard's quantifiers with $B$
arbitrarily large, as in the proof of \cite[Theorem~7.3]{FM}.
\end{proof}
\begin{lemma}\label{lem:par-star}
Let $P = (1/2, 0, \nu)$ with $0 < \nu < 1/3$. If $B$ is large enough in terms of $A$, $\varpi$ and $\nu$, then (w), (I) and (II) imply
$\sum_{n \notin \mathcal P \cup \mathcal N} \mu(n) w_n \ll x(\log x)^{-A}$. Hence global parity balance $(\mathrm{Par})$ at a large enough
logarithmic exponent implies parity balance on $\mathcal P \cup \mathcal N$ at any prescribed exponent.
\end{lemma}
\begin{proof}
Only squarefree integers contribute to the sum. First consider those
with $P^-(n)\le n^\nu$, retaining the boundary because the components
of $C(R)$ are strictly greater than $\nu$. Writing $n=pm$ with $p$ the
least prime factor, their contribution is
\[
-\sum_{\substack{pm\sim x,\ p\le x^\nu\ {\rm prime}\\
P^-(m)>p,\ p^{(1-\nu)/\nu}\le m}}\mu(m)w_{pm}.
\]
Apply \cite[Lemma~7.10]{FM}, with designated block $\{m\}$ (on the sum $m > x^{1-\nu}/2 \ge x^{\nu/3}$), to the
monomial inequality $p^{(1-\nu)/\nu} m^{-1} \le 1$ and
\cite[Lemma~7.9]{FM} to $P^-(m)>p$, using the gap between an integer and
$p+1/2$ (as $m>1$ on the sum, the value $P^-(1)=\infty$ does not enter). The resulting separated sums are covered by (II).
Both parts of (w) supply the error bounds in those lemmas; the bounded
losses of logarithmic powers are absorbed by choosing $B$ sufficiently
large. The other squarefree composite $n$ have $P^-(n)>n^\nu$, so every
component of $v(n)$ lies in $(\nu,1-\nu)$. For $\nu\le1/4$ every such $n$ lies
in $\mathcal N$ by \cite[Lemma~8.4]{FM} (at $\nu=1/4$ a component $1/2$ cannot
occur in a squarefree composite), so this exhausts the contributing complement
of $\mathcal P\cup\mathcal N$.
If $1/4<\nu<1/3$, such an $n$ has two or three prime factors, since $4\nu>1$.
With three, each component is less than $1-2\nu$, as the other two exceed $\nu$,
so $n\in\mathcal N$; with two, $v(n)=(\beta,1-\beta)$ with
$\beta<1/2<2\nu$, which lies in $\mathcal H(P)$ if and only if
$\beta<1-2\nu$ \cite[Lemma~8.4]{FM}. So the remaining integers are semiprimes
$qr$, with $qq$ and $m\le q^{2\nu/(1-2\nu)}$ are interval
conditions, so (I) bounds the sum over all such $m$. Subtract its
composite cofactors, written uniquely as $m=pr$ with
$p=P^-(m)$, $r\ge2$ and $P^-(r)\ge p$; here $p\le m^{1/2}\le (qm)^\nu\le x^\nu$ (as $r\ge p$), since $m\le q^{2\nu/(1-2\nu)}$ says
$m\le (qm)^{2\nu}$, and we keep $p\le x^\nu$ as a condition on $p$ alone.
\cite[Lemmas~7.10--7.11]{FM}, with designated block $\{q\}$ ($q \ge (x/2)^{1-2\nu} > x^{\nu/3}$; in the second condition the exponent $-2\nu/(1-2\nu)$ of $q$ is at most $-1$,
since $\nu > 1/4$), separate $q \nu,\ y_1 + y_2 \le 1/2\}$: for $\nu < y < 1/2$ and for such $(y_1, y_2)$, the vectors $(y, (1 - y)/2, (1 - y)/2)$,
$(1/4, 1/4, 1/4, 1/4)$, which has the coagulation $(1/2, 1/2)$, and $(y_1, y_2, s, s)$ with $s = (1 - y_1 - y_2)/2$ lie
in $R$, since their components lie in $(\nu, 1/2)$ and sum to $1$, so that no proper subsum lies in $[0, \nu]$. For $1/4 \le \nu < 1/3$,
the composite vectors of $C(R)$ are the triples with all components in $(\nu, 1 - 2\nu)$ and the pairs $(\beta, 1 - \beta)$ with
$\nu < \beta < 1 - 2\nu$ \cite[Lemma~8.4]{FM}; a two-component subvector of such a triple has sum greater than $2\nu \ge 1/2$, and
the larger component of such a pair exceeds $2\nu \ge 1/2$, so $G_1 = \{\emptyset\} \cup \{(y) : \nu < y < 1 - 2\nu\}$. In each case $g$, on ordered vectors, is constant on finitely many convex
polytopes, so it is admissible and satisfies the Lipschitz hypothesis of Proposition~\ref{prop:sieve-parity}. Then
$\Phi = 1 - N_1/2 - N_2/2 - (-1)^k/2$ on a $k$-dimensional vector, where $N_1$ and $N_2$ count its subvectors of dimension $1$ and $2$ with sum
at most $1/2$, and $N_1 \ge k - 1$. Hence:
\begin{center}
\begin{tabular}{@{}ll@{}}
\toprule
vector & $\Phi$ \\
\midrule
$k = 2$, unequal components & $0$ \\
$E = (1/2, 1/2)$, in $\mathcal H$ when $\nu < 1/4$ & $-1/2$ \\
$k = 3$, no component equal to $1/2$ & $0$ \\
$k = 3$, one component equal to $1/2$ & $-1/2$ \\
$k \ge 4$ & $\le 1 - 3/2 + 1/2 = 0$ \\
\bottomrule
\end{tabular}
\end{center}
So $\Phi \le 0$ on $\mathcal H(P)$. On non-squarefree $n \in \mathcal N_r$ with $\nu < 1/4$, $A_g(n) = 1 - (D_1(n) + D_2(n))/2$, where $D_1$
counts the distinct prime divisors up to $\sqrt n$ and $D_2$ the divisors $d \le \sqrt n$ with $\Omega(d) = 2$, squares included; this is
$\le 0$ except at prime squares and cubes, where it equals $1/2$. Indeed, if $\Omega(n) \ge 4$, the two least prime factors $p \le p'$ of
$n$, counted with multiplicity, satisfy $p \le n^{1/4}$ and $pp' \le n^{1/2}$, so $D_1(n), D_2(n) \ge 1$; if $n = p^2 q$ with primes
$p \ne q$, then $D_1(n) = 2$ unless $p \le n^{1/4}$, and in that case $p$ and $p^2$ are counted. Prime squares and cubes form the exceptional set: (I), applied with the modulus $p$ and the one-point
cofactor interval $\{p^{r-1}\}$, gives $\sum_{x/2 < p^r \le x} |w_{p^r}| \le x(\log x)^{-B}$ for $r = 2, 3$. (For $\nu \ge 1/4$, the only non-squarefree members of $\mathcal N$ are $p^2 q$ with $p \ne q$, where $A_g = 0$, and $p^3$, where
$A_g = 1/2$; see the description of $C(R)$ and $G_1$ above, and the proof of part~(b) below for these values of $A_g$, which also
hold at $\nu = 1/4$.)
In dimensions $4, 5, 6, 7, 8$, $\Phi$ is bounded below by $-3$, $-6$, $-10$, $-25/2$ and $-35/2$. In dimensions $5$ to $8$ this holds everywhere,
since $N_1 \le k$ and $N_2 \le \binom k2$; in dimension $4$ it holds off the hyperplanes $x_i + x_j = 1/2$, which have projection
measure zero (on them it can fail: $\Phi(1/4, 1/4, 1/4, 1/4) = -9/2$). In dimensions $2$ and $3$, $\Phi$
vanishes almost everywhere. Every vector of $\mathcal H(P) \cap \mathbb R^k$ has its components in $(\nu, 1 - \nu)$ \cite[Lemma~8.4]{FM}, and integration over
ordered vectors is $1/k!$ times integration over all orderings, so the integral of $1/(x_1 \cdots x_k)$ over the ordered vectors of
$\mathcal H(P) \cap \mathbb R^k$ is at most $I_k(\nu)$.
For $P = (1/2, 0, \nu)$, $\mathcal N_r = \mathcal N$: a vector of $R(1/2, 0, \nu)$ has no proper subsum in $[0, \nu]$, so every
component of a vector of $C(R)$ with at least two components exceeds $\nu$, and hence $P^-(n) > n^\nu$ for every $n \in \mathcal N$.
Ford and Maynard's class shrinks as $B$ grows. Given $\varpi$ and a constant below the right-hand side of
Proposition~\ref{prop:sieve-parity}, let $B'$ be an exponent that the proposition supplies, and take $B \ge B' + 1$ so large in
terms of $B'$, $\varpi$ and $\nu$ that Lemma~\ref{lem:par-star} applies with $A = B' + 1$. Since $\mathcal N_r = \mathcal N$, a pair
of the class at exponent $B$ that satisfies $(\mathrm{Par})$ then has
$\bigl|\sum_{n \in \mathcal P \cup \mathcal N_r} \mu(n) w_n\bigr| \le x(\log x)^{-B} + O(x(\log x)^{-B'-1}) \le x(\log x)^{-B'}$
for large $x$, so it lies in the class of Proposition~\ref{prop:sieve-parity} at exponent $B'$; the bound on $E_x$ holds at every
exponent, by (I) as above. Hence, for $1/9 \le \nu < 1/3$,
\[
C^-(1/2, 0, \nu \mid \mathrm{Par}) \ge 1 - 2I_4(\nu) - 4I_5(\nu) - \tfrac{20}{3} I_6(\nu) - \tfrac{25}{3} I_7(\nu) - \tfrac{35}{3} I_8(\nu),
\]
with $I_k$ as in Theorem~\ref{thm:B} ($I_k = 0$ once $k\nu \ge 1$, so dimensions $9$ and up do not occur, and $I_8 = 0$ for $\nu \ge 1/8$).
\subsection{The numbers}
Each $I_k(\nu)$ is non-increasing in $\nu$: as $\nu$ increases the domain of integration shrinks, and the integrand is positive and does not depend on
$\nu$. A separately written program, from the audit of this theorem, encloses the
$I_k(\nu)$ in exact integer arithmetic: $[0, 1]$ is cut into $N$ cells of length $1/N$, the density $\mathbf 1_{\nu < x < 1 - \nu}/x$ is
bounded almost everywhere on each cell from below and above by integers at scale $10^{18}$, the convolutions are formed as exact integer polynomial
products and divided by $k$ with outward rounding, the results are enclosed cell by cell with the support below $k\nu$ removed, and
$I_k(\nu)$ is enclosed at $1$. The program and its outputs are in the supplement's \texttt{thmB/} directory.
With $N = 16000$ at $\nu = 3/20$ and $N = 9800$ at $\nu = 1/7$ this gives
\begin{gather*}
C^-(1/2, 0, 3/20 \mid \mathrm{Par}) \ge \tfrac{132606075869165383}{187500000000000000} > 0.7072324046, \\
C^-(1/2, 0, 1/7 \mid \mathrm{Par}) \ge \tfrac{232429204005956881}{375000000000000000} > 0.6198112106,
\end{gather*}
and with $N = 8000$ at $\nu = 1/8$ it gives the lower bound
$\tfrac{144641353317676797}{500000000000000000} > 0.2892827066$ there, with $I_7(1/8) \le 9.15 \cdot 10^{-7}$. Below $\nu = 1/8$ the term
$I_8$ enters. A copy of the program that adds it to the bound, with its enclosure arithmetic unchanged, gives with $N = 32000$ at
$\nu = 0.116$ the lower bound $\tfrac{26309423794885631}{750000000000000000} > 0.0350792317$ there, with $I_8(0.116) \le 8.49 \cdot 10^{-10}$; with the term $I_8$ set to zero, the copy reproduces the stored exact outputs at $3/20$, $1/7$ and
$1/8$ byte for byte. By monotonicity the bound is positive on the whole range $0.116 \le \nu < 1/3$, at least $0.0350$ throughout, at
least $0.2892$ for $\nu \ge 1/8$, at least $0.6198$ for $\nu \ge 1/7$ and at least $0.7072$ for
$\nu \ge 3/20$. A floating-point integral recurrence, computed in the same audit, gives the underlying values $0.70749$ at $3/20$ and
$0.62035$ at $1/7$; the same recurrence with the term $I_8$, run with the copy after the audit, gives $0.03551$ at $0.116$. This proves
Theorem~\ref{thm:B}(a). \qed
\subsection{Proof of Theorem~\ref{thm:B}(b)}
Let $1/4 < \nu < 1/3$. As in the proof of part~(a), the composite vectors of $C(R)$ are the triples with all components in $(\nu, 1 - 2\nu)$, and the pairs $(\beta, 1 - \beta)$ with $\nu < \beta < 1 - 2\nu$ \cite[Lemma~8.4]{FM}; moreover $\psi \equiv 0$ and
$G_1 = \{\emptyset\} \cup \{(y) : \nu < y < 1 - 2\nu\}$. Take $g(\emptyset) = 1$, $g((y)) = -1/2$ for the one-component vectors of $G_1$,
and $c = -1/2$. On squarefree $n \in \mathcal P \cup \mathcal N$ we have $A_g(n) - \mu(n)/2 = 1 - N_1(n)/2 - \mu(n)/2$, where $N_1(n)$
counts the prime factors $p$ of $n$ with $(\log p/\log n) \in G_1$; for a prime, a pair and a triple, $(N_1, \mu) = (0, -1), (1, 1), (3, -1)$, so
this is $\tfrac32 \mathbf 1_{\mathcal P}(n)$. The only non-squarefree members of $\mathcal N$ are $p^2 q$ with $p \ne q$ and $p^3$. Their
ordinary divisors $m$ with $v(m; n) \in G_1$ are $1, p, q$ and $1, p$ respectively, so $A_g(p^2 q) = 0$ and $A_g(p^3) = 1/2$, and $\mu = 0$ at both.
Lemma~\ref{lem:arith715}, together with $(\mathrm{Par})$ and Lemma~\ref{lem:par-star},
gives $\sum_{n \in \mathcal P \cup \mathcal N} w_n (A_g(n) - \mu(n)/2) \ll_A x(\log x)^{-A}$, that is,
\[
\tfrac32 \sum_{x/2 < p \le x} w_p + \tfrac12 \sum_{x/2 < p^3 \le x} w_{p^3} \ll_A x(\log x)^{-A}.
\]
The sum over cubes is $O(x(\log x)^{-B})$ by (I), applied with intervals that isolate the cofactor $p^2$, as in part~(a). Hence
$\sum_{x/2 < p \le x} (a_p - b_p) \ll_A x(\log x)^{-A}$ once $B$ is large enough, which is the first assertion of part~(b). Since $\sum_p b_p \ge x(\log x)^{-\varpi}$ by (b.1), taking
$A > \varpi$ gives $C^-(1/2, 0, \nu \mid \mathrm{Par}) \ge 1$. The pair $a_n = b_n = 1$ on $(x/2, x]$ has $w = 0$, so it satisfies
(w), (I), (II) and $(\mathrm{Par})$, and (b.1) and (b.2) hold for it by \cite[Lemma~4.6]{FM} with $q = 1$, $y = x/2$
(\S\ref{sec:setting}); as $\sum_p a_p = \sum_p b_p$ for it, $C^-(1/2, 0, \nu \mid \mathrm{Par}) \le 1$. \qed
\subsection{The remaining range}\label{ssec:large-nu}
Let $1/3 \le \nu \le 1$. We show directly that (w), (I) and (II), with $\gamma = 1/2$ and $\theta = 0$, imply
$\sum_{x/2 < p \le x} w_p \ll_A x(\log x)^{-A}$ when $B$ is sufficiently large. For $x/2 \le t \le x$, put
\[
\Pi(t) = \sum_{x/2 < p \le t} w_p \frac{\log p}{\log x}.
\]
Apply \cite[Lemma~7.12]{FM} with $k = 1$ (its $\ell = 6\lceil 1/(1-\gamma)\rceil$ is $12$, as $\gamma = 1/2$), with
$\psi_{u,v} = \mathbf 1_{u = v = 1}$ and $\mathcal T_{1,1} = [\log(x/2)/\log x, \log t/\log x]$; the other regions may be empty, and
the sum of that lemma is then $\Pi(t)$. Its hypotheses are (w) and (I), with $B$
sufficiently large, and the indicated region is convex. Its implied constant depends only on $k$ and the fixed parameters, not on the region, so it gives, uniformly in $t$,
\[
|\Pi(t)| \ll (\log x)^{12} \sup_{|\beta_i| \le 1} |S_{\boldsymbol\beta}(t)| + x(\log x)^{-B+1}, \qquad
S_{\boldsymbol\beta}(t) = \sum_{\substack{x/2 < n = d_1 \cdots d_{12} \le t\\ d_i < n/\sqrt x\ (1 \le i \le 12)}} w_n \prod_{i=1}^{12} \beta_i(d_i).
\]
Every tuple in this sum has at least three factors greater than $1$. Indeed, the factors greater than $1$ have product $n > 1$ and are
each less than $n/\sqrt x < n$, so there are at least two of them; if there were exactly two, we would have
$n < (n/\sqrt x)^2 = n \cdot n/x \le n$, a contradiction. The least factor greater than $1$ is therefore at
most $n^{1/3} \le x^\nu$. Partition the tuples by the first index $j$ for which $1 < d_j \le x^\nu$. For $i < j$, the conditions
$d_i = 1$ or $d_i > x^\nu$ depend on one variable only and are absorbed into its coefficient.
For each $j$, remove the conditions $d_i < n/\sqrt x$ and $n \le t$ using \cite[Lemma~7.9]{FM}. Write them as
\[
\frac{n}{d_i} > \lfloor \sqrt x \rfloor + \frac12, \qquad n < \lfloor t \rfloor + \frac12.
\]
The compared quantities are positive, are bounded by $2x$, and have gaps at least $1/2$. For the condition on $d_i$, the two variables of that
lemma are $n/d_i$, the product of the other eleven factors, and $d_i$, with the threshold as a constant function; for the last
condition one may group the variables by their total product and use a dummy second variable equal to $1$. Constant unimodular
factors of the twists are taken out of the absolute value. The grouped coefficient arrays have absolute value
at most $|w_n| \tau(n)^C < x^2$ for a fixed $C$ and large $x$, by (w) and the divisor bound. Thus all the hypotheses of
\cite[Lemma~7.9]{FM} hold. Its twists, $(n/d_i)^{it_i}$ and $n^{it_0}$, factor into one-variable coefficients of unchanged absolute
value. After these thirteen separations, at a cost $O((\log x)^{13})$ and a negligible error, the sums have the form
\[
\sum_{\substack{x/2 < mr \le x\\ 1 < m \le x^\nu}} \xi(m) \kappa(r) w_{mr}, \qquad |\xi(m)| \le 1, \quad |\kappa(r)| \le \tau(r)^{12},
\]
where $m = d_j$ and $\kappa(r)$ sums over the ordered factorisations of $r$ into the other eleven factors, so that
$|\kappa(r)|$ is at most their number $\tau_{11}(r) \le \tau(r)^{10}$ \cite[Lemma~7.7]{FM}. Condition (II) bounds each
sum by $x(\log x)^{-B}$ for $B \ge 12$. Consequently $\sup_t |\Pi(t)| \ll_A x(\log x)^{-A}$ once $B \ge A + 26$ is also large enough
for the preceding lemmas. Partial summation with the bounded function $\log x/\log t$, whose total variation on $[x/2, x]$ is
bounded, gives $\sum_{x/2 < p \le x} w_p \ll_A x(\log x)^{-A}$. Taking $A > \varpi$ and using (b.1), we get
$C^-(1/2, 0, \nu) \ge 1$ for $1/3 \le \nu \le 1$, and adjoining $(\Hoo_\alpha)$ can only increase $C^-$ (\S\ref{sec:setting}). For
$0.116 \le \nu < 1/3$, part~(a) applies, since $(\Hoo_\alpha)$ implies $(\mathrm{Par})$. This proves the final assertion of
Theorem~\ref{thm:B}. \qed