\section{Kernels and rows}\label{app:kernels} Unless a second argument is written, the functions below are functions of one variable on $[0, 1]$, a total size; $\rho$ and $M_j$ take two (a subset sum $\varsigma$ and a total $\tau$), and $XS_k$ and $\Lambda^{(j)}_k$ carry a second argument as shown. Put $s(u) = 1/u$ on $[\eta, \alpha']$ and $r(u) = 1/u$ on $(\alpha', 1/2)$, both zero elsewhere, and \[ S_j = s^{*j}/j!, \qquad R_k = r^{*k}/k!, \qquad S_0 = R_0 = \delta_0 . \] $R_{k,b}$ is the part of $R_k$ whose largest component lies in bin $b$, and $D_{j,k,b} = S_j * R_{k,b}$. ``tr'' marks multiplication by $\mathbf 1_{[1/2, 1]}$. \paragraph{Fragmentation kernels.} A large component $a \ge 1/2$ of an extended vector is fragmented as in Lemma~\ref{lem:K}. Summing Lemma~\ref{lem:K} over the fragmentations of $a$ into $j$ small and $k$ rough fragments, $j + k \ge 2$, with $B$ running over all sub-collections of the fragments ($b_s$ small and $b_r$ rough members), gives \[ \Lambda_{j,k}(a) = \sum_{b_s = 0}^{j} \sum_{b_r = 0}^{k} (-1)^{j + k + 1 + b_s + b_r} \bigl((S_{b_s} * R_{b_r})\mathbf 1_{[1/2, 1]} * S_{j - b_s} * R_{k - b_r}\bigr)(a), \qquad a \ge 1/2; \] the boundary term of Lemma~\ref{lem:K} at $a = 1$ integrates to zero. For a fragment type $y = (j_y, k_y, b_y)$ of the large component, $\Lambda_y$ is the part of $\Lambda_{j_y, k_y}$ whose largest rough fragment lies in bin $b_y$ ($\Lambda_y = \Lambda_{j_y, 0}$ when $k_y = 0$), and $c_y := c[j_y, k_y, b_y]$. These are the kernels of the constant part and of $f(1)$ below. The terms with $b_r = 0$, sub-collections made only of small fragments, are nonzero as soon as the small fragments can total at least $1/2$, and for $k = 0$ they are the whole kernel. The forced part meets only fragmentations whose small fragments total at most $\alpha' < 1/2$. There no sub-collection without a rough fragment reaches $1/2$, and the rough fragments can be resolved by number: for $k \ge 1$ rough fragments and a threshold $\vartheta > 0$, put \[ XS_k(t; \vartheta) = (-1)^k \sum_{b_r = 1}^{k} (-1)^{b_r} \bigl(R_{b_r} \mathbf 1_{[\vartheta, \infty)} * R_{k - b_r}\bigr)(t). \] With no small fragment, $\Lambda^0_k(a) = -XS_k(a; 1/2) = \Lambda_{0,k}(a)$ for $a \ge 1/2$; $\Lambda^0_1 \equiv 0$ there, since a fragmentation has at least two pieces. With $j$ small fragments of total $\tau \le \alpha'$, \[ \Lambda^{(j)}_k(a; \tau) = (-1)^{j-1} \int M_j(\varsigma, \tau)\, XS_k(a - \tau; 1/2 - \varsigma)\, d\varsigma, \qquad M_j = \rho^{*j}/j!, \] where $\rho(\varsigma, \tau) = s(\tau)[\delta(\varsigma) - \delta(\varsigma - \tau)]$ is the signed measure of one small fragment on (subset sum $\varsigma$, total $\tau$), and the convolution power is taken in both variables; equivalently $M_j(\varsigma, \tau) = \sum_{b = 0}^{j} (-1)^b S_b(\varsigma) S_{j - b}(\tau - \varsigma)$. The kernels with bins are resolved by the bin of the largest rough fragment. When every rough part exceeds $1/4$, $\Lambda^0_2 = -R_2$ and $\Lambda^0_3 = 2R_3$ on $[1/2, 1]$. \paragraph{The rows.} Consider a modulus with $j_D$ small prime factors of total $\sigma$ ($j_D = 0$, $\sigma = 0$ for $d = 1$), and put $t = 1 - \sigma$. Every vector in the row carries the sign $(-1)^{\dim}$ of its whole dimension. \begin{itemize} \item The constant part is \begin{multline*} \sum_{(j', k', b')} (-1)^{j_D + j' + k'} c[j_D + j', k', b']\, D_{j',k',b'}(t) \\ + \sum_{w} (-1)^{1 + j_D + j_w + k_w} \sum_{y} c[j_D + j_w + j_y,\ k_w + k_y,\ \max(b_w, b_y)]\, (D_w * \Lambda_y\,\mathrm{tr})(t), \end{multline*} where $w = (j_w, k_w, b_w)$ runs over the free remainders, the empty one included ($D_\emptyset = \delta_0$) except in $T_0$ (below), and $y$ over the fragment types of the large component. \item The forced part, for $h_J(\tau)$ with $J = j_D + d_j$ and $u = \tau - \sigma$, has free completions and extended completions. The free completion with $d_j = 0$ and $k$ rough parts gives the diagonal term $(-1)^{j_D + k} w_k R_k(t)$; free completions with $d_j \ge 1$ give $(-1)^{J + k} w_k S_{d_j}(u) R_k(1 - \tau)$. Extended completions, with $j_w + j_y = d_j$, $k_w + k_y = k$, $k_y \ge 1$ and sign $(-1)^{1 + j_D + j_w + k_w}$, give four kinds of terms: $(R_{k_w} * \Lambda^0_{k_y}\mathrm{tr})(t)$, $S_{j_w}(u)(R_{k_w} * \Lambda^0_{k_y}\mathrm{tr})(1 - \tau)$, $(R_{k_w} * \Lambda^{(j_y)}_{k_y}(\cdot; u)\mathrm{tr})(t)$, and $\int S_{j_w}(v - \sigma)\,(R_{k_w} * \Lambda^{(j_y)}_{k_y}(\cdot; \tau - v)\mathrm{tr})(1 - v)\, dv$, each with weight $w_k$. \item The diagonal is the sum of the terms with $d_j = 0$: $B_{j_D}(\sigma) = \sum_k w_k \bigl[(-1)^{j_D + k} R_k(t) + \sum_{k_w + k_y = k,\ k_y \ge 1} (-1)^{1 + j_D + k_w} (R_{k_w} * \Lambda^0_{k_y}\mathrm{tr})(t)\bigr]$. Lemma~\ref{lem:Beff} evaluates it for the weighted profile. \item $T_0$ is the row with $j_D = 0$. Every forced term in it has $d_j = J \ge 1$; on extended vectors these terms come with every nonempty leftover vector, including those with no rough part. The prime vector, whose single component $1$ is the large component with empty remainder, contributes $-f(1)$; this term is written separately and counted once, so in $T_0$ the sum over $w$ runs over nonempty remainders only. \item $f(1) = \sum_y c_y \Lambda_y(1) + \sum_J \int h_J(\tau)\, \Lambda^{(J)}_k(1; \tau)\, d\tau$, summed over the profile with weights $w_k$. \end{itemize} The untwisted rows, the same sums without the signs, vanish identically for any free data, since every row has $\sigma \le \alpha' < 1/2$, where (6.2) applies (\S\ref{sec:setting}); the certificates check this, to the extent stated in \S\ref{sec:certificates}, as a test of the machinery.