\section{Uniform thresholds of order \texorpdfstring{$3p$}{3p}}\label{sec:three-p} Fix a prime $p\ge5$ and put $t=3p$. Proposition~\ref{prop:delta} gives $\delta_t(B)=1$ in every base, so $\Lambda=H_B\mathcal A_t$. We obtain a cutoff independent of $p$ without enumerating the first shell. \begin{theorem}[Order-$3p$ thresholds]\label{thm:three-p} For every prime $p\ge5$ and every integer base $B\ge9$, \begin{equation}\label{eq:three-p} K_{3p}(B)= \begin{cases} 3pB/2+2p-1,&B\text{ even},\\ 3p(B+1)/2+1,&B\text{ odd}. \end{cases} \end{equation} Moreover, \begin{equation}\label{eq:height-three-p} h_{3p}(B)=\left\lceil\frac{B^2+1}{2}\right\rceil . \end{equation} In particular $K_{3p}(10)=17p-1$. The cutoff 9 is sufficient, not asserted to be minimal. \end{theorem} \subsection{Arrays, potentials, and cycles} Choose explicitly \[ \eta_3=\zeta^{p(p^{-1}\bmod3)},\qquad \eta_p=\zeta^{3(3^{-1}\bmod p)},\qquad \zeta=\eta_3\eta_p. \] A full representative is a $3\times p$ array: cell $(i,j)$ represents $\eta_3^i\eta_p^j$ and occupies the position $r(i,j)\in[0,3p-1]$ with $r\equiv i\pmod3$, $r\equiv j\pmod p$. Deleting the last row and column, the tensor coordinates of an array $F$ are $F_{ij}-F_{i,p-1}-F_{2,j}+F_{2,p-1}$. Its kernel therefore consists exactly of arrays $u_i-w_j$. For an integer target $A$ and integer capacities $c_{ij}\ge0$, a lift to the box is equivalent to potentials with \[ 0\le A_{ij}+u_i-w_j\le c_{ij}. \] These are difference constraints on the bidirected graph $K_{3,p}$: $w_j-u_i\le A_{ij}$ and $u_i-w_j\le c_{ij}-A_{ij}$. Absence of negative-weight cycles is necessary and sufficient. For sufficiency, add a source with zero-weight edges and take shortest distances: every constraint holds and the distances are integers. A negative closed walk contains a negative simple cycle. Simple cycles here have length 2, 4, or 6. To an orientation of a $2h$-cycle associate its alternating functional $\ell$, with $h$ positive and $h$ negative cells. Its condition is \begin{equation}\label{eq:3p-cycle} \ell(A)\le\sum_{\text{positive cells}}c_{ij}. \end{equation} Both orientations must be considered; length 2 reduces to $c_{ij}\ge0$. This directly proves the integer-lift property used here. The graph and matroid context is Martin--Reiner~\cite[Theorem 1 and \S3]{MR}; the more general TU result of Beck--Hoşten~\cite[Theorem 13]{BH} is not needed for this deduction. The following elimination also applies to the signed boxes used later. \begin{lemma}[Linear-cost lifting with three rows]\label{lem:linear-lift} For integer $a\times p$ arrays $T,l,h$, $l\le h$, the system \[ l_{ij}\le T_{ij}+u_i-w_j\le h_{ij} \] is feasible if and only if the $a$-variable difference system \begin{equation}\label{eq:row-elimination} u_i-u_b\le D_{bi},\qquad D_{bi}=\min_j(h_{ij}-l_{bj}-T_{ij}+T_{bj}) \end{equation} is feasible. Integral data admit an integral lift. It can be constructed, or infeasibility detected, in $O(a^2p+a^3)$ arithmetic operations, hence $O(p)$ when $a=3$. \end{lemma} \begin{proof} For each column $j$, the admissible interval for $w_j$ is \[ \max_i(T_{ij}+u_i-h_{ij}) \ \le w_j\le\ \min_b(T_{bj}+u_b-l_{bj}). \] It is nonempty exactly when every lower endpoint is at most every upper endpoint. Rewriting those inequalities and taking the minimum over $j$ gives \eqref{eq:row-elimination}. Computing its $a^2$ weights takes $O(a^2p)$ operations. Shortest paths on the fixed $a$-vertex graph, with an added zero source, detect negative cycles or give integral potentials in $O(a^3)$ operations. Then set $w_j=\max_i(T_{ij}+u_i-h_{ij})$ and reconstruct the array. This last step costs $O(ap)$. \end{proof} This is classical difference-constraint elimination specialized to the three-row model. It is an arithmetic operation count, not a bit bound or an empirical speed comparison. \begin{lemma}[The three-row short shell]\label{lem:3p-shell} Every $\beta\in\mathcal A_{3p}\setminus\{0\}$ with $\|\beta\|_{Z_{3p}}<2/3$ has gauge $1/2$ and a unique expression \begin{equation}\label{eq:3p-beta} \beta=\eta_3^i\sum_{j\in S}\eta_p^j,\qquad i\in\{0,1,2\},\quad\varnothing\ne S\subsetneq\Z/p\Z. \end{equation} There are $3(2^p-2)$ such elements. \end{lemma} \begin{proof} For two prime factors, \S\ref{sec:shell} says that vectors below gauge $2/3$ are precisely single-sign rectangles in the $2\times(p-1)$ basis. The three nonempty row masks are third roots, possibly with reversed sign. Absorb the sign by complementing $S$, using $\sum_{j=0}^{p-1}\eta_p^j=0$, to obtain \eqref{eq:3p-beta}. Sums over distinct nonempty proper subsets are distinct: a difference of degree at most $p-1$ can be a multiple of $\Phi_p$ only if every coefficient agrees; the two nonzero constant possibilities would require the full and empty sets. The three row factors are distinct, as their tensor masks show. The gauge and count follow from the same classification. \end{proof} Write $f_j=\mathbf1_{j\in S}$. An array for $H_B\beta$ is \begin{equation}\label{eq:3p-target} A_{qj}=(B^2+1)\mathbf1_{q=i}f_j -B\bigl(\mathbf1_{q=i+1}f_{j-1} +\mathbf1_{q=i-1}f_{j+1}\bigr). \end{equation} Indices are cyclic. For a cycle of half-length $h$ put $z=\ell(\beta)$ and $v=\ell((\zeta+\zeta^{-1})\beta)$. Then $\ell(A)=(B^2+1)z-Bv$, $|z|\le h/2$, and $|v|\le h$: root multiplication preserves $Z_{3p}$ and these elements have gauge $1/2$. A critical cycle with $z/h=1/2$ must have length 4, because for a 6-cycle $z$ is integral and $z\le1$. It compares row $i$ with $i+\sigma$, $\sigma=\pm1$, and a column $j\in S$ with $k\notin S$. Its positive cells are $(i,j),(i+\sigma,k)$, and \begin{equation}\label{eq:3p-v} z=1,\qquad v=f_{k-\sigma}-f_{j-\sigma}\in\{-1,0,1\}. \end{equation} For every nonempty proper $S$, one such cycle has $v\le0$. Otherwise, fixing $\sigma$, translation would exchange $S$ and its complement, forcing $2|S|=p$. \subsection{Exact height and odd bases} \begin{lemma}[Strict positive-inverse bound]\label{lem:strict-inverse} Let $M=(B^2+1)I-B(S+S^{-1})$ act on cyclic coefficient space, where $S$ shifts by one position. If $R_ts=H_B\beta$ and $0\le s\le u\one$, then $\|\beta\|_{Z_t}\le u/(B-1)^2$. If any coordinate of $s$ is less than $u$, the inequality is strict. \end{lemma} \begin{proof} The convergent series \[ M^{-1}=\frac1{B^2+1}\sum_{n\ge0} \left(\frac B{B^2+1}(S+S^{-1})\right)^n \] has nonnegative entries and row sum $(B-1)^{-2}$. Every entry is strictly positive: a walk between any two cycle positions contributes positively to a power. Thus $M^{-1}s\in[0,u/(B-1)^2]^t$. If any coordinate of $s$ is strictly below $u$, every upper inequality for $M^{-1}s$ is strict. Its maximum is strictly below the bound because there are finitely many coordinates. Finally $R_tM^{-1}s=\beta$. No positive margin uniform in $t$ is required. \end{proof} \begin{proof}[Proof of \eqref{eq:height-three-p}] Put $h_0=\lceil(B^2+1)/2\rceil$. For the upper bound choose $\beta=1$, or $i=0,S=\{0\}$. Critical cycles have $v=0$ or 1 and target at most $B^2+1$. Other 4-cycles have $z\le0$ and target at most $2B$. The 6-cycle targets are at most $B^2+1+3B$. Uniform capacity $h_0$ satisfies all these inequalities for $B\ge6$; for the last, $B^2-6B+1\ge0$ suffices. Integral potentials give a lift. For the lower bound, a false positive of integer height $u\le h_0$ would give, by Lemma~\ref{lem:strict-inverse}, a nonzero element with \[ \|\beta\|_{Z_{3p}}\le\frac{h_0}{(B-1)^2} \le\frac{B^2+2}{2(B-1)^2}<\frac23\qquad(B\ge9). \] The last inequality is $B^2-8B-2>0$. Lemma~\ref{lem:3p-shell} is exhaustive. The critical cycle with $v\le0$ requires $2u\ge B^2+1$, so $u\ge h_0$. \end{proof} \begin{proof}[Proof of \eqref{eq:three-p} for odd $B$] At length $3p(B+1)/2$, every capacity is $(B^2-1)/2=h_{3p}(B)-1$, so the full period is safe. At the next position choose $\beta=1$ again. The capacity of $(0,0)$ increases by $B-1$. Every critical cycle has this positive cell, giving capacity $B^2+B-2$ against target at most $B^2+1$. Other 4-cycles have capacity at least $B^2-1$ and target at most $2B$; 6-cycles have capacity at least $3(B^2-1)/2$ and target at most $B^2+1+3B$. These comparisons hold for $B\ge7$. Integral potentials and digit realization give a false positive, since $H_B\ne0$. Padding completes the proof. \end{proof} \subsection{Even bases: local obstruction and prefix cost} At length $k=3pB/2+a$, $0\le a<3p$, put $M_{qj}(a)=\mathbf1_{r(q,j)0. \] \begin{lemma}[Two local systems]\label{lem:3p-local} A nonconstant cyclic binary sequence $f$ satisfies \eqref{eq:3p-binary} if and only if it satisfies one of the following systems. \begin{description} \item[A.] No adjacent $11$ occurs; $f_j=1$ requires $M_{i,j}=1$; a pair $f_j=f_{j+1}=0$ requires $M_{i-1,j}=M_{i+1,j+1}=1$. \item[B.] No adjacent $00$ occurs; $f_j=0$ requires $M_{i-1,j}=M_{i+1,j}=1$; a pair $f_j=f_{j+1}=1$ requires $M_{i,j}=M_{i,j+1}=1$. \end{description} \end{lemma} \begin{proof} A critical cycle with $v=-1$ is impossible even with two markers. Such a cycle exists exactly when both $11$ and $00$ occur. Since $p$ is odd, a cyclic sequence cannot avoid both, so an admissible sequence belongs to exactly one system. In A, $00$ occurs. In each direction choose an unselected $k$ with unselected predecessor: then $v=0$ forces the main marker of every selected $j$ and the lateral markers precisely at the $00$ pairs. Other critical pairs have $v=1$ and are satisfied by the main marker. In B, $11$ occurs. In each direction choose a selected $j$ with selected predecessor: then $v=0$ forces both lateral markers of every unselected $k$ and the main markers at the ends of $11$. Other constraints have $v=1$ and are satisfied by the lateral marker. These arguments also prove sufficiency. \end{proof} \begin{lemma}[Exact prefix cost]\label{lem:3p-prefix} The minimum $a$ admitting a nonconstant solution to \eqref{eq:3p-binary} for some row $i$ is $2p-1$. \end{lemma} \begin{proof} Write $p=3q+r$, $r\in\{1,2\}$. At $a=2p-1$, choose system B and \[ \begin{array}{c|c|c} r&i&\text{columns with }f_j=0\\\hline 1&0&1,4,\ldots,p-3\\ 2&1&0,3,\ldots,p-2 \end{array}. \] Zeros are isolated, including across the cyclic join. Their two lateral cells occupy $j,j+p<2p-1$. Every selected main cell occupies its column's first or second occurrence, also below $2p-1$. Thus B holds. For the lower bound it suffices to exclude $a=2p-2$. The marker vector in column $j$ is \begin{equation}\label{eq:3p-marks} m_j=\begin{cases} \mathbf1_{\{j\bmod3,(j+r)\bmod3\}},&0\le j