\section{Recipes and certified formulas}\label{sec:certified} This section combines the reductions proved above with finite exact identities. Section~\ref{sec:repro} identifies the required files and checkers. The solvers used to discover the certificates are not premises. \subsection{The totally unimodular regime}\label{subsec:tu} By Beck--Hoşten~\cite[Theorem 13 and the decomposition in \S2]{BH}, the CRT root matrix $R_t$ is totally unimodular when $t$ has at most two distinct odd prime factors. We use this property of the matrix, rather than a property of a different convex body. \begin{corollary} Under this hypothesis, the least eventual coefficient period of $K_t(B)$ divides $\rho_t$. \end{corollary} \begin{proof} Every nonzero minor has absolute value 1, so $L_t=1$ in Theorem~\ref{thm:effective}. \end{proof} There is a smaller recipe for computing the intercepts. Let $r_i$ be the columns of $R_t$. Let $\mathcal N_t$ contain both primitive normals to each span of $d-1$ independent columns, with repetitions removed. These are the facet directions of the zonotope. Total unimodularity and the cofactor construction give $u\cdot r_i\in\{0,1,-1\}$. Put \[ h(u)=\sum_{i0$, the integer fibre at length $tL+a$ is nonempty precisely when \begin{equation}\label{eq:facet-test} u\cdot H_B\beta\le(B-1)(Lh(u)+h_a(u)) \quad\text{for every }u\in\mathcal N_t . \end{equation} Indeed these inequalities describe the sum of segments of capacities $(B-1)(L+\mathbf1_{i0\}}. \] The reciprocal term therefore matters when the affine part is integral. The intercept recipe is \begin{equation}\label{eq:tu-intercept} b_{t,j}=\min_{\beta\in\mathcal S,\ 0\le a0\}}-\mu j\right)+a\right\}. \end{equation} For an effective cutoff, require for each winning pair $(c,d)$ \[ \begin{aligned} (\mu-r_u)B&>|c-c_u|+|d-d_u|&&(r_u<\mu),\\ (B-1)(c-c_u)&>|d-d_u|&&(r_u=\mu,\ c_u|c|+|d|.& \end{aligned} \] If $c_u=c$, the inequality $d\ge d_u$ resolves the tie. Also require $|d|/(B-1)<\varepsilon_j$, where $\varepsilon_j$ is the distance of the nonintegral number $\mu j+c$ to the nearest adjacent integer, or 1 if it is integral. Taking the maximum of these rational bounds and $C_{\rm red}$ proves the recipe without fitting sampled bases. \subsection{Orders fifteen and thirty} \begin{corollary}\label{thm:15-30} The formulas of Theorems~\ref{thm:three-p} and \ref{thm:six-p} at $p=5$ give the exact thresholds for order 15 when $B\ge9$, and for order 30 when $B\equiv j\pmod4$ and $B\ge(16,13,10,19)_j$. The least eventual coefficient periods are 2 and 4, respectively. In particular $K_{15}(10)=84$ and $K_{30}(10)=91$. \end{corollary} \begin{proof} Apply Theorems~\ref{thm:three-p} and \ref{thm:six-p} with $p=5$, including their period assertions. \end{proof} The facet recipe provides a redundant check for these two orders. Its computational details are recorded in \path{docs/COMPUTATIONS.md} in the companion; its sufficient cutoffs do not restrict this corollary. \subsection{Orders one hundred five and two hundred ten} \begin{theorem}[Thresholds beyond total unimodularity]\label{thm:105-210} For $B=4n+j\ge44$, $0\le j<4$, \begin{equation}\label{eq:K105} K_{105}(B)=105n+(74,86,106,127)_j. \end{equation} For $B=8n+j\ge65$, $0\le j<8$, \begin{equation}\label{eq:K210} K_{210}(B)=210n+(71,91,127,144,161,186,197,211)_j. \end{equation} The least eventual coefficient periods are 4 and 8. The cutoffs are sufficient; minimality of the cutoffs is not claimed. \end{theorem} Use the $2\times4\times6$ basis of monomials $\eta_3^i\eta_5^j\eta_7^k$, where $\eta_3=\zeta_{105}^{70}$, $\eta_5=\zeta_{105}^{21}$ and $\eta_7=\zeta_{105}^{15}$. Put $R=R_{105}$, and let $T$ represent $\zeta_{105}+\zeta_{105}^{-1}$. For order 210 take $\xi=-\zeta_{105}$. The targets are, respectively, \[ ((B^2+1)I-BT)\beta,\qquad ((B^2+1)I+BT)\beta. \] \begin{proof}[Reduction and certification] For order 105, a failure of length at most $105n+c_j-1$ has gauge at most \[ \frac{n+q_j}{4n+j-1},\qquad(q_0,q_1,q_2,q_3)=(1,1,1,2). \] The inequality $7(n+q_j)<2(4n+j-1)$ holds for $n\ge(10,8,6,11)_j$. Corollary~\ref{cor:105-shell} therefore reduces every earlier failure to one of the 5670 signed rectangles. For order 210 the same argument gives \[ \|\beta\|_{Z_{210}}\le\frac{n+1}{8n+j-1}<\frac17 \quad\Longleftrightarrow\quad n\ge9-j. \] This list is again exhaustive. The reduction excludes length $K-1$; it does not classify the fibre at $K$. Put $(t,b,\epsilon)=(105,4,-1)$ or $(210,8,1)$, and $B=bn+j$. For any integer normal $u$, define \[ h=\sum_r\max((R_t^Tu)_r,0),\quad h_a=\sum_{r1$. The difference concerns representation in the box, not a new ideal index. \end{proof} \subsection{No real--integer gap at the certified thresholds} Define $K_t^{\mathrm{fib},\R}(B)$ by allowing real folded sums within the digit capacities, while requiring their image to remain a nonzero point of the same discrete lattice $\Lambda$. The lattice variable is not relaxed. \begin{corollary}\label{cor:real-fibres} \[ K_{105}^{\mathrm{fib},\R}(B)=K_{105}(B)\quad(B\ge44),\qquad K_{210}^{\mathrm{fib},\R}(B)=K_{210}(B)\quad(B\ge65). \] \end{corollary} \begin{proof} The maximum-principle reduction also holds for real sums, while $\beta=H_B^{-1}R_ts$ remains in the discrete lattice. Thus the shell to exclude at length $K-1$ is unchanged. All lower-bound separators are supporting inequalities valid on the whole real box. No real lift exists at $K-1$, whereas the upper certificates give integer lifts at $K$. \end{proof} This does not make all fibres integral. The non-TU example in Section~\ref{sec:shell} uses a different box and does not imply a gap at these thresholds.