\section{Technical supplement: local exclusion domains} \subsection{What the finite exclusions do and do not prove} \label{sec:scoped-exclusions} A PCR group below is either a stable PCR or a pair of nonstable PCRs interchanged by RC. A chart site is a lower-vertex pair $\{u,\theta u\}$ in \Cref{thm:bounded-chart}; it is not a PCR group or a legal move. Changing these definitions changes a neighbourhood and can invalidate a purported exclusion. \begin{proposition}[Exact local exclusions] \label{prop:local-exclusions} The following finite statements hold for the exact packaged bases. \begin{enumerate} \item In the 99-PCR region generated by the displayed critical cycle of $M_{121}$ and its RC image, with the exterior fixed, the minimum RC MDS length is 121. \item Changing at most four RC-PCR groups of $M_{111}$ cannot produce an RC MDS of length at most 110. \item Changing at most three RC-PCR groups of $M_{109}$ and then applying arbitrary integral retiming cannot attain length at most 108. The minimum in that family is 109. Retiming the base before measuring the radius is not included. \item In the chart based at the canonical $M_{107}$, allowing at most three sites, arbitrary integral site coefficients, all 128 stable bits and an arbitrary integral gradient gives minimum RC MDS length 107. \item In each of eight specified regions around $S$ of \Cref{prop:8232}, with respectively $1,13,139,64,128,256,512,1024$ nonstable groups free and all stable choices unrestricted, the minimum cardinality at path bound 106 is 8232. These are eight specified regions, not all regions of those sizes. \end{enumerate} \end{proposition} \begin{proof}[Exact finite proof objects] For the first statement the verifier reconstructs 51 RC-compatible option domains with 382 initial options and checks 53 walk cuts, 202 option eliminations and a terminal contradiction. No numerical solver status enters this implication chain. For the four-group radius, 2486 path or cycle cuts support an exhaustive branching certificate with 533,138 nodes. Each node has its declared domain and cut bound recomputed, so complete branch coverage supplies the exclusion. A two-group change repairs the old critical cycles but creates a seven-cycle, showing why PCR and critical-cycle conditions alone are insufficient. For the three-group radius around $M_{109}$, the 1069 cuts are all closed walks; their weights are unchanged by subsequent retiming. The 45,845-node proof thus excludes the full retiming closure of the stated radius. A separate acyclic three-group change hits each of the 6,931,456 old critical cycles three times but has component optimum 111. The exhaustive old-cycle count uses a tight-clock DAG and is independently replayed. For three chart sites, globally checked support conditions reduce the $\binom{8192}{3}=91,592,417,280$ triples to 24,592 necessary ones. Lifting a rational certificate to an omitted site requires both vanishing of its combined coefficients and the divisibility conditions used in row normalization. The verifier checks these conditions before using a lifted support clause. It reconstructs the complete support recursion, then checks every surviving triple by rational Farkas combinations, integer strips, modular equalities or rounded valid inequalities. There are no unresolved triples. The 1323 closed walks and all rational coefficients are included; the checker uses integer and rational arithmetic only. For the eight regions, deleting all 988 stable-PCR arcs is a valid relaxation. The verifier checks 844 sequential eliminations of nonstable options using residual cycles or 107-arc paths. An option is removed only using already proved residual arcs or fixed exterior arcs. Every region ends with an empty domain if all its nonstable PCRs are singly hit. Cost at most 8231 has this property; the base $S$ attains 8232 in each region. \end{proof} The finite cut relaxation for the four-group radius has exact fractional minimum $1/2$, certified by a feasible rational point and matching dual multipliers. The corresponding integer cut system requires at least five changed groups, hence its gap is at least ten. This concerns that particular relaxation. Adding more valid inequalities or changing the objective changes the statement; it is not a universal integrality-gap theorem for the global problem. \subsubsection{A boundary clause needs its assumptions} Each regional elimination proof supplies a set $B$ of exterior RC literals and a set $\mathcal T$ of nonstable PCR representatives whose one-selection property the proof used. Writing $c_P=|M\cap P|$, its valid extension to any RC feedback set at the target length is \begin{equation} \sum_{b\in B}\ind_M(b)+\sum_{P\in\mathcal T}(c_P-1)\ge1. \label{eq:regional-surplus} \end{equation} Indeed, if the nonnegative integral left side were zero, all boundary arcs would remain residual and all used PCRs would be singly hit, so the identical inductive contradiction would apply. At cost at most 8231 the nonstable surplus vanishes, and the bare boundary clause is valid. At higher cost it need not vanish. This surplus is different from the stable-PCR surplus in \Cref{cor:surplus-two-arm}. The shortest supplied boundary clause has 21 RC literals and uses 12 group assumptions and eight cycles. Both $S_A,S_B$ satisfy all eight bare clauses while moving the double-PCR surplus to a different nonstable group. Thus these clauses are useful necessary constraints, not global refutations. The companion also checks 214 globally valid closed-walk cuts at length 106, of which 187 require one hit and 27 require two. Their validity does not imply their sufficiency. \subsubsection{Reproducibility and scope} The companion retains the literal inputs and exact proof objects for these local exclusions. Its separate base-four checks recompute the positive endpoints and sharp parity example; small-instance tests exercise the conditional stable-completion and lifted two-arm implementations. Such checks support implementations; the uniform results depend on their proofs. Each exclusion is tied to the stated base, exterior, neighbourhood or affine family. None certifies the missing global optimum.