\section{The certified 141/107 component separation} Throughout the DNA atlas and weighted-component statements in this section, the parameter domain is exactly $\mathcal W$, with fixed labels and $A1$. Zero-PCR ties are fixed: the smallest RC-fixed word in a stable PCR, and the smallest word in a paired union together with its RC mate. Here is an explicit finite reduction for the root list. Let $V=\{\pm e_1,\pm e_2\}$ and $D=V+V$, a nine-element alphabet. For a lower word of seven letters write \[ x_1=v_0+v_6,\quad x_2=v_1+v_5,\quad x_3=v_2+v_4,\quad z=v_3. \] Its vector midpoint form satisfies \[ 2r=\sqrt2(x_1-x_3)-2(x_1+x_2+x_3+2z). \] Enumerating $D^3\times V$ is exhaustive and positionally reversible. For lambda use the same word, with $d_i$ the three paired differences: $\lambda=(d_1+\sqrt2d_2+d_3)/\sqrt2$. An affine scalar form $(a+c\sqrt2)\rho+b+d\sqrt2$ has, when its slope is nonzero, the unique root \[ \rho=\frac{-ab+2cd+(bc-ad)\sqrt2}{a^2-2c^2}. \] Activity must be tested on one and the same positional word; independently choosing pair sums and pair differences would be an invalid relaxation. \begin{theorem}[Complete DNA terminal atlas]\label{thm:complete-atlas} At $k=8$ there are exactly 86 distinct terminal walls in $\rho>1$: 87 open intervals and 86 point strata. They give 114 different labelled terminal sets. All 86 walls separate different adjacent terminals; at 27 walls the point is a third terminal, and at the other 59 it equals one side. Each terminal has 8230 words. The minimum number of a fixed dictionary of homogeneous linear forms whose ternary signs determine the terminal is exactly 86. \end{theorem} \begin{proof}[Exact certificates and real-parameter reduction] The reduction above enumerates every candidate; the exact comparison and activity filters produce 86 midpoint roots, including all ten lambda roots. The certificate stores each root and a realizing word. On each open interval all signs are fixed, and structural ties are fixed, so \Cref{ext:thm:two} proves constancy for every real ratio, including transcendental ratios. Rational representatives are chosen by exact dyadic bisection; point roots are evaluated in $\Q(\sqrt2)$. The producer and a separate verifier compare all membership bits; the latter reconstructs the original seed in $\Q[X]/\Phi_8$. Integer levels certify acyclicity, and explicit paths certify the reported longest lengths. PCR and RC checks are literal. Terminals are grouped by equality of bit vectors, not by assuming hash injectivity. A separating word at each wall proves necessity. A nonzero homogeneous linear form has at most one root on the normalized line, so at least one form per indispensable wall is necessary. Conversely one midpoint form at each of the 86 roots distinguishes all strata. The files and replay commands are listed in \pathfile{evidence/extended/README.md}. \end{proof} The minimum 86 concerns a fixed linear sign dictionary. It is not a lower bound for adaptive queries, arbitrary nonlinear encodings, or graph isomorphism classes. \subsection{The optimum over the weighted component union} \label{sec:weighted-component-union} The complete atlas and \Cref{prop:family-components} supply 21 ordinary components at $k=8$. These are only the components encountered by $\mathcal W$, with fixed labels and ties. The next result optimizes their complete union, rather than only the terminals of $\mathcal W$. \Cref{thm:construction-107} then separates this union from the wider RC MDS space. \begin{proposition}[Exact optima in the 21 components] \label{prop:component-optima} In each row of \Cref{tab:component-optima}, the ordinary optimum and both RC-sector optima equal the displayed value. Hence the minimum over the union of these entire components encountered by $\mathcal W$ is 141. \end{proposition} \begin{proof}[Exact certificates] For each component, a simple cycle with one selected arc gives the lower bound throughout that component. Two explicit MDS, one per RC sector, attain it. The checker verifies the clock inequalities, mate sums, gradients, Boolean membership, every PCR, RC, ranks, paths and cycle intersections. A separate sink-removal dynamic program confirms each length. The 42 sector solutions are also reconstructed from ordinary clock solutions by tight-SCC rounding. The proof does not enumerate each entire component. \end{proof} \begin{table}[ht] \centering\small \begin{tabular}{rrrrr} \toprule Component & Base row & Best weighted & Component optimum & Cycle/hits\\ \midrule 0 & c000 &161&161&162/1\\ 1 & c015 &157&157&158/1\\ 2 & c016 &153&153&154/1\\ 3 & c027 &153&153&154/1\\ 4 & c028 &149&149&150/1\\ 5 & c037 &149&149&150/1\\ 6 & c038 &149&149&150/1\\ 7 & c045 &147&143&144/1\\ 8 & c046 &149&149&150/1\\ 9 & c055 &161&157&158/1\\ 10& c056 &147&147&148/1\\ 11& c077 &163&161&162/1\\ 12& c078 &145&145&146/1\\ 13& c093 &141&141&142/1\\ 14& c094 &141&141&142/1\\ 15& c109 &155&155&156/1\\ 16& c119 &143&143&144/1\\ 17& c129 &163&161&162/1\\ 18& c130 &143&143&144/1\\ 19& c151 &155&155&156/1\\ 20& c152 &161&161&162/1\\ \bottomrule \end{tabular} \caption{DNA $k=8$. Row identifiers have prefix \texttt{k8\_}. The optimum is attained in both paired RC sectors.} \label{tab:component-optima} \end{table} Thus no sequence of complete moves and inverses starting at a terminal of $\mathcal W$ can go below 141, even if RC is abandoned in the intermediate states. Changing component is necessary. The corresponding small regression optima are 17 at $k=4$ and $63,65,57$ for the three $k=6$ components. \subsection{Sharp parity cost and a shorter DNA construction} \label{sec:construction-107} The atlas of $\mathcal W$ is a starting family within the space of MDS. Other labelled antisymmetric weight maps are not classified here. We now give finite witnesses outside it. All lengths in this and the following sections count residual arcs in $B(q,k-1)$. \begin{proposition}[The one-unit sector cost is attained] \label{prop:sharp-parity} For DNA at $k=8$ there is an RC MDS $M_*$ whose ordinary component optimum is 124, while the optima in $\mathcal C_0(M_*)$ and $\mathcal C_1(M_*)$ are respectively 125 and 124. \end{proposition} \begin{proof}[Proof with finite witnesses] The directory \pathfile{evidence/advances/r1/witnesses/} supplies \pathfile{Mstar125.words.txt} and \pathfile{M124_sector1.words.txt}, each an RC set of 8230 words. Direct DAG reconstruction gives lengths 125 and 124. Their difference is the integer gradient stored in \pathfile{r1/certificates/retiming_Mstar125_to_M124.json}, whose mate sum is odd, so they are in the indicated sectors of one component. The parity certificate supplies a residual path $P:A^7\leadsto T^7$ of length 71 and a return path $Q$ of length 54 containing exactly one selected arc of $M_*$. Their concatenation is a simple cycle of length 125 with one selected arc. It gives the ordinary lower bound 124. At period $T=125$, the two clock inequalities force \[ H(T^7)-H(A^7)=71. \] Sector zero requires $H(T^7)+H(A^7)=124$, hence $2H(A^7)=53$, impossible for an integer clock. The length-125 endpoint attains the next possible value, whereas the other endpoint attains 124. The independent verifier checks all arc differences, both paths, the cycle and both DAGs. \end{proof} This resolves sharpness for a prescribed sector. It does not show that the union of the two RC sectors ever has a worse optimum than the ordinary component. \begin{theorem}[Certified DNA upper bound 107] \label{thm:construction-107} There is an RC DNA MDS $M_{107}$ of cardinality 8230 with $L(M_{107})=107$. Its ordinary component has exact optimum 107. Consequently the global RC MDS optimum at $k=8$ is at most 107. The upper-graph residual has length 106; a longest avoiding DNA word has 114 letters, and every DNA word of length 115 contains a selected eight-letter window. \end{theorem} \begin{proof}[Exact certificates] The canonical witness is \pathfile{r5/witnesses/M107.words.txt}. Its sorted ASCII lines, including the final newline, have SHA-256 \begin{center}\ttfamily\small 4e0c44d87ea9782f261cf3ed2c0e918784fa404bbf0f57041b494509ff7d5edc. \end{center} The checker enumerates all 65,536 arcs and 16,384 vertices, verifies RC and one selection in each of the 8230 PCRs, and checks a rank in $[0,107]$ increasing along each residual arc. A residual path of 107 arcs attains the bound. A simple cycle of 108 arcs has exactly one selected arc, proving the component lower bound by \Cref{thm:unit-retiming}. Both path and cycle are included in \pathfile{r5/certificates/M107.json}. The length conversions follow from overlapping consecutive eight-letter words. \end{proof} Earlier component-optimal endpoints and their one-hit cycles are retained in the companion; the first explicit escape is described in \Cref{sec:construction-131}. A four-witness panel at length 107 certifies two ordinary components with both parity sectors represented in each. The complete stable fibre in \Cref{prop:stable-fibre} supplies further components. Neither finite panel exhausts the global component space. The 108-cycle forces any improvement to $L\le106$ to meet that cycle at least twice, hence to leave this ordinary component. Changing the intersection with one old critical cycle is only a necessary condition: \Cref{prop:local-exclusions} includes examples where all old critical cycles are repaired but the result is cyclic or belongs to a component with a worse optimum.