\section{Balanced loads and finite-pool closure} \subsection{Balanced closed-walk loads and the exact closure of a finite pool} \label{sec:balanced-loads} The complete endpoint charts above retain Boolean and acyclicity constraints. We now isolate a weaker necessary test: nonnegative combinations of a fixed pool of closed-walk inequalities whose exterior incidence is invariant under real gradients. This test has an exact finite linear-programming description. Its scope is a real-gradient class on the nonstable arcs, rather than an ordinary complete-move component on the full graph. Let $k\ge2$ be even and let the alphabet have a fixed-point-free complement. In $G=B(q,k-1)$ write $S$ for the union of RC-stable PCRs, $E_0=E\setminus S$, and $N_S$ for their number. Fix an integer $T\ge k$. For $b\ge0$ define the aggregate stable budget \[ \mathcal B_b=\{c\in\R^{N_S}:c_P\ge1,\ \sum_Pc_P\le N_S+b\}. \] This relaxation has no upper bounds $c_P\le|P|$, Boolean constraints or acyclicity requirement. All projections below refer to this stated budget. For an exterior $y:E_0\to\R$, its real-gradient class is $y+\nabla f|_{E_0}$ with $f:V\to\R$ and $\nabla f(e)=f(\tail e)-f(\head e)$, as in \Cref{sec:base}. \begin{lemma}[Closed-walk counting inequality]\label{lem:walk-budget} If $M$ is an RC feedback set with $L(M)\le T-1$ and $n=\ind_M$, then every nonempty closed walk $W$, with multiplicity vector $\chi_W$, satisfies \begin{equation} \langle\chi_W,n\rangle\ge\ceil{|W|/T}.\label{eq:walk-budget} \end{equation} \end{lemma} \begin{proof} The walk meets $M$, since a nonempty residual closed walk contains a cycle. Cut its circular sequence at the selected occurrences. Each piece has one selected arc followed by at most $T-1$ residual arcs: a residual stretch cannot repeat a vertex and is bounded by $L(M)$. Thus $r$ selected occurrences give $|W|\le Tr$. Repetitions are counted with multiplicity. \end{proof} \begin{proposition}[Stable-load criterion and aggregate normal form] \label{prop:balanced-normal-form} The stable subgraph consists of vertex-disjoint directed PCR cycles. Let $F=\sum_j\lambda_j\chi_{W_j}$ with $\lambda_j\ge0$, and put $A=F|_{E_0}$. Then $A(\nabla f)=0$ for every real potential $f$ if and only if $F$ has a constant load $m_P$ on the arcs of each stable PCR $P$. For such a combination set \[ K=\sum_j\lambda_j\ceil{|W_j|/T},\qquad M_* =\max_Pm_P. \] Its exact aggregate projection is \begin{equation} A(y)\ge K-\sum_Pm_P-bM_* .\label{eq:balanced-projection} \end{equation} If $M_*>0$, adjoining nonnegative weighted stable PCR cycles makes all positive stable loads equal to $M_*$ without changing this projected inequality. Rational weights remain rational and can be cleared to integers. \end{proposition} \begin{proof} A word in a stable PCR has equal counts in every complementary letter pair, because RC is a rotation of that word. A fixed $(k-1)$-letter prefix can be completed to balanced counts by at most one letter: distinct letters add distinct signed unit vectors to its imbalance. The same argument applies to suffixes. Hence each vertex has stable indegree and outdegree at most one; rotation closure gives vertex-disjoint cycles. This argument uses necessity of balanced counts, not a converse characterization of RC stability. An arc vector annihilates all gradients exactly when its divergence is zero, as testing the potentials supported at individual vertices shows. The full vector $F$ is a circulation. Thus $A$ is a circulation exactly when $F|_S$ is one; on the disjoint stable cycles this means constant loads. Summing \eqref{eq:walk-budget} yields $A(n|_{E_0})+\sum_Pm_Pc_P\ge K$. Over $\mathcal B_b$ the maximum stable term is $\sum_Pm_P+bM_*$: place all excess on a largest-load coordinate. This proves the exact projection of this single combined row and the budget; it is not a simultaneous projection of all original walk rows. For $U=\{P:m_P>0\}$ adjoin weight $M_*-m_P$ of the PCR cycle $P$. Its length is at most $k\le T$, so its counting right side is one. The new right side is $K'=K+\sum_{P\in U}(M_*-m_P)$ and its stable maximum is $M_*(|U|+b)$; subtraction gives precisely \eqref{eq:balanced-projection}. If every load vanishes no normalization is needed. Clearing denominators proves the last assertion. \end{proof} The circulation and normalization argument works on any finite graph whose distinguished subgraph is a union of vertex-disjoint cycles of length at most $T$. The RC de Bruijn geometry supplies this hypothesis. \begin{theorem}[Exact finite-pool closure]\label{thm:finite-pool-closure} Fix a finite pool of mixed closed walks $W_1,\ldots,W_J$, each containing a stable arc. Write $s_{ej}=\chi_{W_j}(e)$ for $e\in S$, $A_j(y)=\sum_{e\in E_0}\chi_{W_j}(e)y_e$ and $d_j=\ceil{|W_j|/T}$. Let $F_*(y)$ be the maximum of \begin{equation} \sum_j\lambda_j(d_j-A_j(y))-\sum_Pm_P\label{eq:pool-primal} \end{equation} subject to $\lambda_j\ge0$, $0\le m_P\le1$ and $\sum_js_{ej}\lambda_j=m_P$ for every $e\in P$, every stable $P$. For every $b\ge0$, the following are equivalent: \begin{enumerate} \item No nonnegative combination from this pool with gradient-invariant exterior incidence excludes $y$ by \eqref{eq:balanced-projection}. \item $F_*(y)\le b$. \item There are signed real coefficients $\gamma_e$, $e\in S$, such that \begin{align} A_j(y)+\sum_{e\in S}s_{ej}\gamma_e&\ge d_j &&(1\le j\le J),\label{eq:pool-extension}\\ \Gamma_P:=\sum_{e\in P}\gamma_e&\ge1 &&(P\text{ stable}),\notag\\ \sum_{e\in S}\gamma_e&\le N_S+b.\notag \end{align} \end{enumerate} If $y$ is RC invariant and the pool is RC closed, the lift $\gamma$ can be required to be RC invariant. Individual lift coordinates have no sign, integrality, Boolean or upper-bound restrictions. \end{theorem} \begin{proof} The primal is nonempty because zero is feasible. Every $\lambda_j\le1$: choose a stable arc in $W_j$, whose positive integer multiplicity contributes to a total load at most one. Thus the feasible set is compact and the maximum is finite and attained. Its equalities impose exactly the circulation criterion of \Cref{prop:balanced-normal-form}. A feasible objective larger than $b$ is larger than $b\max_Pm_P$ because $b\ge0$ and $m_P\le1$, and hence separates $y$. Conversely a separating combination has $M_*>0$: in a mixed pool, a nonzero nonnegative combination cannot have all stable loads zero. Divide its weights and loads by $M_*$. The normalized feasible primal point has objective larger than $b$. This proves the first equivalence. We explicitly derive the dual, using classical finite LP strong duality \cite[Chapter 5]{BoydVandenberghe2004}. Put free multipliers $\gamma_e$ on $m_P-\sum_js_{ej}\lambda_j=0$ and nonnegative $t_P$ on $m_P\le1$. The Lagrangian is \[ \sum_Pt_P+\sum_j\lambda_j\left[d_j-A_j(y)-\sum_es_{ej}\gamma_e\right] +\sum_Pm_P[-1+\Gamma_P-t_P]. \] Its supremum over nonnegative $\lambda,m$ is finite precisely when \[ \sum_es_{ej}\gamma_e\ge d_j-A_j(y),\qquad t_P\ge\Gamma_P-1,\quad t_P\ge0. \] The dual objective is $\sum_Pt_P$, and a feasible primal of finite value has an attaining dual of equal value. Formally add every stable PCR cycle as a column. This leaves the primal maximum unchanged. If its coefficient is $\eta_P$, remove it and replace $m_P$ by $m_P-\eta_P$; the remaining loads are nonnegative and at most one, and both right-side contribution and subtracted mass drop by $\eta_P$. Conversely add zero coefficients. These columns impose $\Gamma_P\ge1$ in the dual because their right sides are one. Its minimum is at most $b$ exactly when $\Gamma_P\ge1$ and $\sum_P\Gamma_P\le N_S+b$, taking $t_P=\Gamma_P-1$ for sufficiency. This is \eqref{eq:pool-extension}. Finally RC permutes the mixed rows and preserves each stable PCR set. Averaging $\gamma$ with $\gamma\circ\theta$ preserves feasibility and budget. \end{proof} The convention can also be checked directly. If an endpoint satisfies $n|_{E_0}=y+\nabla f|_{E_0}$, take $\gamma=n|_S-\nabla f|_S$. The full cochain $y+\gamma=n-\nabla f$ has the same closed-walk and PCR evaluations as $n$. It satisfies the extension whenever the endpoint satisfies its stable budget. Data using $\delta f= f(\head e)-f(\tail e)$ require $\delta=-\nabla$; transport $y=x+\delta a$ then transports the lift by $\gamma_y=\gamma_x+\delta a=\gamma_x-\nabla a$. \begin{corollary}[The MDS budget]\label{cor:mds-pool} At $b=0$, \eqref{eq:pool-extension} forces $\Gamma_P=1$ on every stable PCR. Every balanced row reduces to $A(y)\ge K-\sum_Pm_P$. \end{corollary} \begin{proof} Each of the $N_S$ masses is at least one and their total is at most $N_S$. Substitute $c_P=1$ in the combined row. \end{proof} This specialization remains a necessary test on the exterior. Pure $E_0$ walks must be checked separately, and feasibility of a signed lift does not imply a Boolean stable completion, a clock or a full endpoint. Adding mixed walks can strictly strengthen the test, as the following single example shows. \begin{proposition}[An exterior passing one pool and excluded by new returns] \label{prop:core-pool-strictness} There is an RC exterior $Y$ of 8102 nonstable DNA8 arcs, one per nonstable PCR, whose residual core $(V,E_0\setminus Y)$ is a DAG of length exactly 106. At $T=107$ a specified pool of 1804 mixed walks admits an RC signed rational lift of total stable mass 128 for $y=\ind_Y$. Thus this entire balanced-pool closure passes at both $b=0$ and $b=1$. Nevertheless every RC feedback set $M$ with $L(M)\le106$ and \[ \ind_M|_{E_0}=\ind_Y+\nabla f|_{E_0},\qquad f:V\to\R, \] has $|M|\ge8235$. Attainment of 8235 in this class is not asserted. \end{proposition} \begin{proof}[Finite certificate reduction] Encode $A,C,G,T$ by $0,1,2,3$. A length-eight word has ID $e<65536$, tail $\floor{e/4}$ and head $e\bmod16384$; RC reverses the digits and replaces $d$ by $3-d$. Rotation partitions the arcs into 8230 PCRs. Exactly 128 PCRs are stable, containing 988 arcs. The other 8102 PCRs form 4051 RC pairs. All paths in this proof are relative to \pathfile{evidence/balanced_loads/}. The literal witness and exact payload are \pathfile{results/joint_core106.words.txt} and \pathfile{results/joint_core106.json.gz}. The companion verifier \pathfile{verify_case.py} uses the declared shared exact kernel to reconstruct the carrier and check all selections, RC, nonstable PCR counts, and a rank equal to the longest incoming core distance. Every residual core arc increases this rank, whose maximum is 106; a listed 106-arc residual path proves equality. The core omits every stable arc freely and is not a complete feedback-set witness. The mixed pool is reconstructed from the literal sources and declared alphabet images, joined walks and their RC images, deduplicating by arc multiplicity. Its 1804 walks compress to 902 mixed rows on 622 stable RC arc orbits. The exported rational lift has denominators one or two. Exact row evaluation checks all mixed lowers, each of the 128 PCR masses $\Gamma_P=1$, and total mass 128. By \Cref{thm:finite-pool-closure}, no balanced combination of this pool and the stable PCR cycles separates $Y$. For the stronger bound, forty further walks supply a return $Q_e$ entirely in $E_0$ from $\head e$ to $\tail e$ for every arc $e$ of five stable PCRs. Each $eQ_e$ is closed. Put \[ K_P=\sum_{e\in P}\ceil{(1+|Q_e|)/107},\qquad A_P(z)=\sum_{e\in P}\sum_{a\in Q_e}z(a). \] Their stable restriction is the incidence of the PCR cycle, so their external incidence is a circulation. Hence $A_P(\nabla f)=0$ and \eqref{eq:walk-budget} gives $c_P\ge K_P-A_P(\ind_Y)$. The checked closed-walk lengths range from 53 to 111, with the following values: \begin{center} \begin{tabular}{lrrr}\toprule PCR representative & $K_P$ & $A_P(\ind_Y)$ & lower on $c_P$\\\midrule \texttt{ATCCCGGG} & 10 & 8 & 2\\ \texttt{ATCGCGCG} & 9 & 7 & 2\\ \texttt{CCCGGGCG} & 9 & 7 & 2\\ \texttt{CCGCGCGG} & 9 & 7 & 2\\ \texttt{CCGCGGCG} & 9 & 7 & 2\\\bottomrule \end{tabular} \end{center} The verifier checks adjacency, closure, $E_0$ support, complete PCR coverage, ceilings, multiplicities, circulation and evaluation. The remaining 123 stable PCRs require one hit each, so stable cost is at least 133. A gradient telescopes on each nonstable PCR, preserving its selected count one. Nonstable cost is therefore 8102, proving $|M|\ge8102+133=8235$ for arbitrary real $f$. \end{proof} Summing the five return rows gives $A(y)+\sum_{P\in U}c_P\ge46$. For an RC endpoint of cost at most 8231, nonstable surplus is even and must vanish. Its stable budget is at most 129; after the other 123 mandatory hits, $\sum_{P\in U}c_P\le6$. Thus $A(y)\ge40$, or $A(y)\ge41$ in the MDS branch. On $Y$ the value is 36. These are weighted occurrence thresholds, not counts of distinct newly selected words. The example proves strictness relative to the stated pool, without asserting independence from every retained historical row. It leaves $8230\le\tau_{\RC}(106)\le8232$ unchanged and gives no global lower bound or attained optimum in the new real-gradient class.