\documentclass[12pt]{amsart} \usepackage[T1]{fontenc} \usepackage{lmodern} \usepackage{amsmath} \usepackage{amssymb} \usepackage{amsthm} \usepackage{mathtools} \usepackage{mathrsfs} \usepackage{xcolor} \usepackage{enumitem} \usepackage{needspace} \usepackage{array} \usepackage{booktabs} \usepackage[left=3cm,right=3cm,top=3cm,bottom=3.5cm]{geometry} \usepackage[expansion=false]{microtype} \usepackage{hyperref} \setlength{\emergencystretch}{3em} \allowdisplaybreaks[2] \hypersetup{colorlinks=true,linkcolor=blue,citecolor=blue,urlcolor=blue, bookmarksdepth=2,pdfauthor={Wenchang Yang, Yi Wang}, pdftitle={Smooth Isometric Embeddings of Closed Manifolds in R to the Power s(n)+5 in Large Dimension}} \setcounter{tocdepth}{1} \theoremstyle{plain} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \newtheorem{thm}[theorem]{Theorem} \newtheorem{prop}[theorem]{Proposition} \newtheorem{lem}[theorem]{Lemma} \newtheorem{cor}[theorem]{Corollary} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newtheorem{defn}[theorem]{Definition} \newtheorem{setup}[theorem]{Setup} \newtheorem{notation}[theorem]{Notation} \newtheorem{construction}[theorem]{Construction} \theoremstyle{remark} \newtheorem{remark}[theorem]{Remark} \newtheorem{rem}[theorem]{Remark} \numberwithin{equation}{section} \setlist[enumerate]{label=\textup{(\roman*)},leftmargin=*} \newcommand{\R}{\mathbb{R}} \newcommand{\C}{\mathbb{C}} \newcommand{\Z}{\mathbb{Z}} \newcommand{\N}{\mathbb{N}} \newcommand{\Ftwo}{\mathbb{F}_2} \newcommand{\eps}{\varepsilon} \newcommand{\metric}{\gamma} \newcommand{\jet}{\mathcal{J}} \newcommand{\triv}{\underline{\R}} \newcommand{\pair}[2]{\langle #1,#2\rangle} \newcommand{\ip}[2]{\left\langle #1,#2\right\rangle} \newcommand{\norm}[1]{\lVert #1\rVert} \newcommand{\abs}[1]{\lvert #1\rvert} \newcommand{\dd}{\mathrm{d}} \newcommand{\ii}{\mathrm{i}} \DeclareRobustCommand{\PaperTarget}{\ensuremath{s_n+5}} \DeclareMathOperator{\Sym}{Sym} \DeclareMathOperator{\Hom}{Hom} \DeclareMathOperator{\im}{im} \DeclareMathOperator{\Id}{Id} \DeclareMathOperator{\rank}{rank} \DeclareMathOperator{\supp}{supp} \DeclareMathOperator{\dist}{dist} \DeclareMathOperator{\sym}{sym} \DeclareMathOperator{\pr}{pr} \DeclareMathOperator{\Int}{int} \DeclareMathOperator{\diag}{diag} \DeclareMathOperator{\tr}{tr} \DeclareMathOperator{\ann}{ann} \DeclareMathOperator{\SymEnd}{SymEnd} \DeclareMathOperator{\skewp}{skew} \DeclareMathOperator{\Gr}{Gr} \DeclareMathOperator{\PD}{PD} \DeclareMathOperator{\ord}{ord} \DeclareMathOperator{\divg}{div} \renewcommand{\Re}{\operatorname{Re}} \renewcommand{\Im}{\operatorname{Im}} \title[Smooth isometric embeddings in $s_n+5$ dimensions]{Smooth Isometric Embeddings of Closed Manifolds in \texorpdfstring{$\R^{\lowercase{s_n}+5}$}{R to the power s(n)+5} in Large Dimension} \author{Wenchang Yang} \address{Tsinghua University, Beijing, China} \author{Yi Wang} \address{Tsinghua University, Beijing, China} \date{} \subjclass[2020]{53C42, 53B20, 58J60, 35F50} \keywords{global isometric embedding, closed Riemannian manifold, phase regularity, Hopf spheres, $h$-principle, differential right inverse} \begin{document} \begin{abstract} We prove that for $n\geq 10^{62}+4$, every closed smooth Riemannian $n$-manifold admits a smooth isometric embedding into $\R^{s_n+5}$, where $s_n=n(n+1)/2$. \end{abstract} \maketitle \tableofcontents \section{Introduction}\label{p00:sec:introduction} Let $(M,g)$ be a smooth Riemannian manifold of dimension $n$. An isometric embedding of $(M,g)$ into $\R^N$ is a smooth embedding $F:M\to\R^N$ satisfying \begin{equation}\label{p00:eq:isometry} \pair{\dd F_x(u)}{\dd F_x(v)}=g_x(u,v)\qquad(x\in M,\ u,v\in T_xM). \end{equation} In local coordinates this is a system of \[ s_n=\frac{n(n+1)}{2} \] equations, and $N=s_n$ is the smallest dimension in which one can expect solvability for every metric. We call $N-s_n$ the \emph{surplus}. Nash \cite{Nash56} proved that every closed Riemannian manifold embeds isometrically in a Euclidean space, and G\"unther \cite{Gunther89,Gunther91} obtained the surplus $\max\{2n,n+5\}$. In \cite{YW-28} we replaced it by the constant $28$. The main result of this paper lowers the surplus to five in large dimension. \begin{theorem}\label{thm:main} Let $n\geq 10^{62}+4$. Every closed smooth Riemannian manifold $(M^n,g)$ admits a smooth isometric embedding into $\R^{s_n+5}$. The image can be contained in any prescribed open ball, and the embedding can be chosen phase regular: at each point $x$ its second fundamental form is injective on $\Sym^2H$ for some hyperplane $H\subset T_xM$. \end{theorem} No assumption is made on the topology of $M$. The same dimension was obtained in \cite{YW-parallel} for closed stably parallelizable manifolds whose components have vanishing Euler characteristic. The lower bound on $n$ enters only through the counting estimate of Lemma~\ref{lem:relationgrowth}. \subsection*{Background} A map is \emph{free} if its first and second derivatives are linearly independent. For a free map the linearized metric operator has an algebraic right inverse, which is the mechanism of the classical proofs \cite{Nash56,GromovPDR,Gunther89}; but a free map needs $N\geq s_n+n$. We use instead the weaker condition of phase regularity. A covector $\xi$ is \emph{good} for an immersion $F$ if the second fundamental form of $F$ is injective on $\Sym^2(\ker\xi)$, and $F$ is phase regular if every point has a good covector. For phase-regular maps an exact metric correction holds as soon as $N\geq s_n+1$ (Proposition~\ref{prop:repair}), and as in \cite{YW-parallel,YW-28} the global problem reduces to the construction of a single phase-regular partial isometry on a product of Hopf spheres (Theorem~\ref{thm:model}). \subsection*{The idea of the proof} A generic bundle map $\triv^4\to TM$ has rank four outside a closed three-dimensional submanifold $P$, the \emph{core}, along which it has rank three. Near $P$ a complex structure on $\triv^4$ splits $TM=H\oplus L$ with $L$ a line, and \[ \rank\bigl(H\oplus\Sym^2H\oplus L\bigr)=(n-1)+\frac{n(n-1)}{2}+1=s_n . \] Thus a map that is free along $H$ and immersive along $L$ leaves a normal bundle $N_5$ of rank five over $P$. By Proposition~\ref{prop:characteristic}, $N_5$ splits off three trivial lines unless $n\equiv2\pmod4$, where $N_5\simeq3\mathcal D\oplus\triv^2$ for a line bundle $\mathcal D$ with possibly $w_1(\mathcal D)^3\neq0$ (for instance on $\mathbb{RP}^n$ with $n\equiv6\pmod8$). The two cases require different boundary models at the core (Sections~\ref{sec:angular} and~\ref{sec:ordinaryboundary}). Outside a tube around $P$ we have $TM=E\oplus\triv^4$ with $\rank E=m=n-4$, and we write \[ \R^{s_n+5}=\R^{s_n-152}\oplus\R^{157}. \] The first factor carries an \emph{old} map, free along $E$; the second carries an \emph{active} map on four Hopf spheres whose derivatives of order at most three are independent. Their normal spaces leave $3m+1$ directions, while the mixed metric coefficients between the two blocks give $4m$ equations. We couple the blocks by a high-frequency product term and solve the mixed equations order by order. This requires a finite-order differential right inverse for the operator \[ (\eta,u,b)\longmapsto\bigl(\pair{C_{i\alpha}}{\eta}+d_\alpha u_i+d_ib_\alpha\bigr)_{i,\alpha}, \] which we obtain from one flat normal tensor on the doubled exterior (Proposition~\ref{prop:flatmatching}) and one active map whose characteristic variety avoids the bad planes of that tensor (Theorem~\ref{thm:activeparameter}). Finally the hyperplane metric is filled across the disk bundle over $P$ (Corollary~\ref{cor:diskfill}), and the remaining terms along $L$ are absorbed by six further Hopf factors (Proposition~\ref{prop:linecompletion}). \subsection*{Organization of the paper} Sections~\ref{p01:sec:model}--\ref{sec:assembly} prove the compact-model reduction: metric correction, removal of double points, and the passage to the prescribed metric. Section~\ref{sec:foursection} studies the core and the class of $N_5$, and Section~\ref{sec:referenceproduct} constructs the reference map. Sections~\ref{sec:angular}--\ref{sec:ordinaryboundary} construct the boundary models and the flat tensor, and Sections~\ref{sec:mixedcompletion}--\ref{sec:actmodule} produce the exact metric on the exterior. Section~\ref{sec:relativefill} fills the disk, Section~\ref{sec:commoncoefficients} combines the two core cases into one set of coefficients, Section~\ref{sec:line} completes the line, and Section~\ref{sec:wholeassembly} proves Theorem~\ref{thm:main}. \section{Conventions and the compact-model reduction}\label{p01:sec:model} In this section we fix our conventions, introduce partial metrics and partial second fundamental forms along integrable subbundles, and define phase regularity. We then associate a commuting Hopf distribution with a Euclidean subbundle and state the compact-model reduction, Theorem~\ref{thm:model}. Its proof occupies Sections~\ref{sec:repair}--\ref{sec:assembly}. \subsection{Partial metrics, second fundamental forms, and phase regularity} \label{pl:sec:partial} All manifolds are smooth and second countable, and a closed manifold is compact and without boundary. We write $s_n=n(n+1)/2=\dim\Sym^2\R^n$. The Euclidean inner product $\pair{\cdot}{\cdot}$ on $\R^N$ is extended complex-bilinearly to $\C^N$ whenever complex amplitudes occur. For a smooth map $F\colon M\to\R^N$ we write $\metric(F)=F^*\pair{\cdot}{\cdot}=(\dd F)^*\dd F$; an isometric embedding of $(M,g)$ is a smooth embedding $F$ with $\metric(F)=g$. Let $D\subset TX$ be an integrable subbundle of rank $n$. A smooth map $F\colon X\to\R^N$ is a \emph{partial immersion} along $D$ if $\dd_DF=\dd F|_D$ is injective; its \emph{partial metric} is $\metric_D(F)=(\dd_DF)^*\dd_DF$, a section of $\Sym^2D^*$. Let $\mathcal N_F^D=(\dd_DF(D))^\perp$ be the partial normal bundle, of rank $N-n$, and $\Pi_F^D$ the orthogonal projection onto it. The \emph{partial second fundamental form} $B_F^D\colon\Sym^2D\to\mathcal N_F^D$ is defined on local sections $U,V$ of $D$ by \[ B_F^D(U,V)=\Pi_F^D\bigl(U(V(F))\bigr). \] It is tensorial and symmetric: the term $(Uf)\,\dd_DF(V)$ produced by the Leibniz rule lies in $\dd_DF(D)$, and so does $U(V(F))-V(U(F))=\dd_DF([U,V])$, because $D$ is integrable. Thus the projection removes the connection terms of the second derivative. In a foliation chart $(y,z)$ in which $D$ is spanned by the derivatives $\partial_i=\partial_{y^i}$ one has \begin{equation}\label{pl:eq:gauss} \partial_i\partial_jF=\Gamma_{ij}^k\,\partial_kF+B^D_{F,ij}, \end{equation} where $\Gamma_{ij}^k$ are the Christoffel symbols of $g_F=\metric_D(F)$ computed with $z$ fixed. When $D=TM$ we write $\mathcal N_F$ and $B_F$. \begin{definition}\label{pl:def:phase} A partial immersion $F$ is \emph{phase regular} along $D$ if at each point there is a hyperplane $H\subset D$ on which $B_F^D|_{\Sym^2H}$ is injective. A nonzero covector $\xi\in D^*$ is \emph{good} if $\ker\xi$ is such a hyperplane. The \emph{best phase margin} at a point is the maximum, over unit covectors $\xi$, of the least singular value of $B_F^D|_{\Sym^2\ker\xi}$. Here the norms on symmetric squares are induced by a fixed metric on $D$. \end{definition} Since $\dim\Sym^2H=s_n-n$ for a hyperplane $H$, phase regularity requires $N\geq s_n$. If $N\geq s_n+1$ and $\xi$ is good, the orthogonal complement of $B_F^D(\Sym^2\ker\xi)$ in $\mathcal N_F^D$ has dimension $N-s_n\geq1$. The best phase margin depends continuously on the point and on the leaf two-jet of $F$. Hence on a compact manifold phase regularity is equivalent to a positive lower bound for the best phase margin, and it persists under small $C^2$ perturbations. \subsection{Commuting Hopf distributions and the compact-model theorem} \label{pl:sec:hopf} On the unit sphere $S^{2h-1}\subset\C^h$ let $Y$ be the generator of the Hopf action $z\mapsto e^{2\pi\ii t}z$, $t\in\R/\Z$, so that $Y(z)=2\pi\ii z$. On $X_h=B\times(S^{2h-1})^a$ let $Y_\ell$ act as $Y$ on the $\ell$-th sphere factor and trivially on the other factors. The fields $Y_1,\dots,Y_a$ commute and are pointwise linearly independent. Let $W\subset B\times\R^a$ be a subbundle of rank $n$ with a bundle metric $g_W$. Then \[ w\longmapsto Y_w=\sum_{\ell=1}^a w_\ell Y_\ell , \qquad w\in W_b, \] defines a rank-$n$ distribution $D_W\subset TX_h$, isomorphic to $\pi_B^*W$, which we equip with the metric $\pi_B^*g_W$. All derivatives along $D_W$ keep $b\in B$ fixed. If $e_1,\dots,e_n$ is a local frame of $W$, the fields $Y_{e_1},\dots,Y_{e_n}$ commute, because their coefficient functions depend only on $b$ and are therefore annihilated by every $Y_m$. In particular $D_W$ is integrable. \begin{theorem}[Compact-model reduction]\label{thm:model} Let $n\geq2$, let $N\geq s_n+2$, and let $B$ be a closed manifold. Fix a rank-$n$ subbundle $W\subset B\times\R^a$ with a bundle metric $g_W$. Suppose that for every sufficiently large integer $h$ there is a smooth map \[ S_h\colon B\times(S^{2h-1})^a\longrightarrow\R^N \] whose partial metric on $D_W$ is $\pi_B^*g_W$ and which is phase regular along $D_W$. If $M$ is a closed $n$-manifold and $TM\simeq f^*W$ for a smooth map $f\colon M\to B$, then every smooth Riemannian metric on $M$ admits a phase-regular isometric embedding into $\R^N$ with image in any prescribed open ball. \end{theorem} \begin{remark}\label{pl:rem:quantifiers} The integer $a$, the base $B$, the subbundle $W$ and its metric are fixed before $h$ varies, and each $S_h$ is defined on the whole product $B\times(S^{2h-1})^a$. The proof uses a single value of $h$, chosen before all other parameters. The derivative bounds and the phase margin of $S_h$ may therefore depend on that value; no uniformity in $h$ is needed. \end{remark} \section{Smooth metric correction at a phase-regular map}\label{sec:repair} We prove a smooth perturbation theorem for partial metrics near a phase-regular map, with constants independent of the number of charts. Since $B_F^D$ is injective only on symmetric squares of good hyperplanes, the linearized equation has no algebraic right inverse. Along each good phase we invert the transverse block algebraically and solve the remaining equations to high order in the wavelength. Let $X$ be a compact manifold with an integrable subbundle $D$ of rank $n$. We use buffered foliation charts $(y,z)$, consisting of inner charts covering $X$ and enlarged charts containing them, in which $D$ is spanned by the $y$-derivatives. Transitions preserve the leaves, so leaf derivatives transform into combinations of leaf derivatives. Norms are maxima of coordinate norms over the enlarged charts and include derivatives in both $y$ and $z$. The symbol $\jet_D^2F$ denotes the values, first leaf derivatives and second leaf derivatives of $F$. We write \[ |u|_{j,\ell}=\sum_{b=0}^j\ell^b\|u\|_{C^b}, \qquad 0<\ell\leq1 , \] and use the same notation for tensors and for the components of $\jet_D^2F$. Then $|uv|_{j,\ell}\leq C_j|u|_{j,\ell}|v|_{j,\ell}$ and $|\partial u|_{j,\ell}\leq\ell^{-1}|u|_{j+1,\ell}$. \begin{proposition}[Uniform smooth correction]\label{prop:repair} Suppose $N\geq s_n+1$. Fix inner and enlarged chart radii, a bound on the overlap multiplicity of the enlarged charts, and bounds at every fixed order for the transitions and their inverses. Fix a compact interval of positive numbers containing the eigenvalues of the leafwise coordinate matrices of $\metric_D(F_0)$ and a positive lower bound on the best phase margin of a smooth reference partial immersion $F_0\colon X\to\R^N$. Given a sufficiently small $C^2$ tolerance $\eps_0>0$, there are a finite derivative order $m_0$ and a constant $\eta_0>0$, depending on bounds for $\|F_0\|_{C^{m_0}}$ and on the data just fixed, such that every smooth positive definite partial metric $h$ with \[ \|h-\metric_D(F_0)\|_{C^{440}}<\eta_0 \] is induced by a smooth phase-regular partial immersion $F$ with $\|F-F_0\|_{C^2}<\eps_0$. The constants $m_0$ and $\eta_0$ are independent of the number of charts. If $X=M\times P$ is a foliated product with $D=TM\oplus0$, the map $F$ is jointly smooth in the points of $M$ and the transverse parameters. \end{proposition} \subsection{Positive increments}\label{pl:sec:increments} \emph{The mode equations.} Work in an adapted chart with $y^1=\psi$, where $\dd_D\psi$ is good, and write $g_F=\metric_D(F)$. Decompose a complex amplitude as \[ Z=\dd_DF(g_F^{-1}u)+v,\qquad v\perp\dd_DF(D), \] where $u$ is a complex leaf covector. Differentiating $\pair{\partial_iF}{Z}=u_i$ and $\pair{\partial_iF}{v}=0$ with $z$ fixed and using \eqref{pl:eq:gauss}, we obtain for the linearization $(D\metric_D)_FZ=(\dd_DF)^T\dd_DZ+(\dd_DZ)^T\dd_DF$ \begin{equation}\label{eq:linear} (D\metric_D)_F Z_{ij} =\partial_i u_j+\partial_j u_i-2\Gamma_{ij}^k u_k -2\pair{B_{F,ij}^{D}}{v}. \end{equation} Conjugation by $e^{\ii\psi/\tau}$ adds $(\ii/\tau)(\delta_{i1}u_j+\delta_{j1}u_i)$. Let $\beta(v)=(\pair{B_{F,ab}^{D}}{v})_{a,b\geq2}$. For a good phase $\beta$ is onto, since its adjoint is the injective map $B_F^D|_{\Sym^2\ker\dd_D\psi}$. Put $J=\beta^*(\beta\beta^*)^{-1}$ and set \[ T(u)_{ab}=\tfrac12(\partial_a u_b+\partial_b u_a)-\Gamma_{ab}^k u_k,\qquad v=v_0+J\bigl(T(u)-\tfrac12h_H\bigr),\quad v_0\in\ker\beta , \] where $h_H=(h_{ab})_{a,b\geq2}$ is the transverse part of the desired mode amplitude $h$. This solves the transverse equations exactly. The remaining equations are equivalent to $(\ii/\tau)u+\mathcal L u=k$, where \begin{align*} (\mathcal L u)_1 &=\partial_1u_1-\Gamma_{11}^ku_k-\pair{B_{F,11}^D}{JT(u)},\\ (\mathcal L u)_a &=\partial_1u_a+\partial_a u_1-2\Gamma_{1a}^ku_k-2\pair{B_{F,1a}^D}{JT(u)},\\ k_1&=\tfrac12h_{11}+\pair{B_{F,11}^D}{v_0}-\tfrac12\pair{B_{F,11}^D}{Jh_H},\\ k_a&=h_{1a}+2\pair{B_{F,1a}^D}{v_0}-\pair{B_{F,1a}^D}{Jh_H}. \end{align*} More precisely, an error $e$ in this system contributes $2e_1$ to the $11$ entry and $e_a$ to the $1a$ entry of the metric error, and nothing to the transverse entries. For an integer $p\geq1$ we choose \[ u=\frac{\tau}{\ii}\sum_{b=0}^{p-1}\Bigl(-\frac{\tau}{\ii}\mathcal L\Bigr)^b k . \] The sum telescopes, and the error is $(-1)^{p-1}(\tau/\ii)^p\mathcal L^pk$. The coefficients of $\mathcal L$ are smooth functions of the leaf two-jet on the set where the immersion and phase margins are bounded below. In a fixed $\ell$-weighted norm each derivative costs $\ell^{-1}$ and one higher input order. The error therefore carries a factor $(\tau/\ell)^p$ relative to finitely many norms of $h$ and $v_0$, and $Z$ is bounded in each fixed slow norm by finitely many norms of $h$ and $v_0$. When $h=0$ one has moreover $Z-v_0=O(\tau/\ell)$ in the same relative sense. All formulas preserve the supports of their inputs. If a phase requires several adapted charts, we split its amplitudes by a partition of unity and add the resulting amplitudes with the same phase before forming quadratic products. \emph{Positive squares.} Good covectors are open and dense. Indeed, in a Euclidean frame a covector $\xi$ is good exactly when some maximal minor of $B_F^D\circ\Sym^2(\abs{\xi}^2\Id-\xi\otimes\xi)$ is nonzero, because $\Sym^2(\abs{\xi}^2\Id-\xi\otimes\xi)$ maps onto $\Sym^2\ker\xi$; the sum of squares of these minors is a polynomial in $\xi$, nonzero somewhere by phase regularity. At a point, let $e_i$ be an orthonormal coframe for the reference metric. The squares of $e_i$ and of $e_i\pm e_j$ span the symmetric forms, and \[ \sum_{ir$ we use the boundedness of the smoothing at the inherited order and $(\tau/t)^b\leq(\tau/t)^r$. Derivatives of cutoffs and frames are bounded and enter only at lower orders. The same arguments apply when $X$ is a product with compact parameter spaces. Set $g_0=\metric_D(F_0)$. At stage $k\geq10$, with current scale $t$, map $G$ and normalized defect $H$, we choose \[ \delta=t^k,\qquad \ell=t^{11/10},\qquad \tau=t^{6/5}, \qquad q_0=r=40(k+1),\qquad H=\delta^{-2}(h-\metric_D(G)). \] We apply \eqref{eq:increment}--\eqref{eq:incrementerror} at $S_\ell G$ with target $S_\ell H-(\delta_+/\delta)^2g_0$, where the next reserve $\delta_+$ equals $\tau^k$ (to stay in stage $k$) or $\tau^{k+1}$ (to pass to stage $k+1$). We keep $H$ in a fixed small $C^0$ neighbourhood of $g_0$ and $W_r(t;G,H)\leq K_0$. The smoothing estimates then bound all finite input norms needed at each step in terms of $K_0$ and $k$. Let $E$ be the error in \eqref{eq:incrementerror}, put $G_+=G+U$ and $H_+=\delta_+^{-2}(h-\metric_D(G_+))$. Expanding the quadratic metrics of $G+U$ and of $S_\ell G+U$ separately gives the exact identity \begin{equation}\label{eq:true} \delta_+^2(H_+-g_0)=\delta^2(H-S_\ell H)-E -\mathcal C_D(G-S_\ell G,U), \end{equation} where $\mathcal C_D(V,U)(a,b)=\pair{\dd_D V(a)}{\dd_DU(b)}+\pair{\dd_D V(b)}{\dd_DU(a)}$. The square of $\dd_DU$ cancels between the two expansions, and the square of the smoothing error does not occur in \eqref{eq:true}. By \eqref{eq:increment} we have $|\dd_DU|_{j,\tau}\leq C_j(k)\delta$. For the smaller reserve $\delta_+=\tau^{k+1}$, the powers of $t$ of the terms in \eqref{eq:true} after division by $\delta_+^2$ are \[ \begin{array}{c|c} \text{term}&\text{power}\\ \hline \delta(\tau/\ell)^{q_0}&(13k+8)/5\\ \delta^3/\tau&(3k-18)/5\\ \delta^2(\ell/t)^r&(18k+8)/5\\ \delta^2(\tau/t)^r&(38k+28)/5\\ \delta(\ell/t)^r&(13k+8)/5\\ \delta(\tau/t)^r&(33k+28)/5 . \end{array} \] The two $(\tau/t)^r$ terms multiply $W_j(t;G,H)$. Since $k\geq10$, all powers are positive, with minimum at least $12/5$. The larger reserve multiplies every term by $\tau^2$ and only improves the estimates. In passing to $W_j(\tau;G_+,H_+)$, every inherited derivative of positive order of $G$ gains at least $\tau/t=t^{1/5}$, and the two-jet increment is $O(\delta/\tau)$. Therefore \begin{equation}\label{eq:recurrence} W_j(\tau;G_+,H_+)\leq \|\jet_D^2G\|_{C^0}+\|g_0\|_{C^0} +C_j(k)t^{1/5}(1+W_j(t;G,H)). \end{equation} We fix $K_0$ larger than four times the sum of the first two terms throughout the chosen $C^2$ neighbourhood of $F_0$. At each stage there is a positive threshold for $t$ below which the step preserves the $C^0$ neighbourhood of the defect and the bound $W_r\leq K_0$. We also require $\|U\|_{C^{\lfloor k/2\rfloor}}\leq c_1t$ for a fixed small $c_1>0$; this is possible since $\|U\|_{C^{\lfloor k/2\rfloor}}\leq C(k)t^{k+\frac65(1-\lfloor k/2\rfloor)}$ and $k+\tfrac65(1-\lfloor k/2\rfloor)>1$. We start at $k=10$ with \[ G=\sqrt{1-t^{20}}\,F_0,\qquad H=g_0+t^{-20}(h-g_0), \] so that $H$ is the normalized defect of $G$. First we choose the initial scale $t$, using the reference derivative bounds, so small that the $C^2$ displacement of $G$ and the sum of all future scales are small; since each step replaces $t$ by $t^{6/5}$, this sum is at most twice the initial scale once $t\leq2^{-5}$. Then we choose $\eta_0$, a fixed multiple of $t^{20}$, which controls the initial $C^{440}$ norm of the normalized defect. At fixed $k$, with $t$ replaced by $t^{6/5}$ at each step, \eqref{eq:recurrence} bounds any fixed higher norm after finitely many steps. Indeed, the coefficient of the inherited norm tends to zero; once it is at most $1/4$, the recurrence bounds the norm, and its limit superior is at most $K_0/4$. We wait until the next controlled order $40(k+2)$ has norm below $K_0$ and the finitely many smallness conditions of stage $k+1$ hold. Then we choose $\delta_+=\tau^{k+1}$ and pass to stage $k+1$. The estimate \eqref{eq:recurrence} at the higher order keeps its norm below $K_0$ across the transition, after decreasing the scale further within the current stage if necessary. The increment bounds keep the maps in the fixed neighbourhood where the phases are good. Each stage ends after finitely many steps, so $k\to\infty$ and, by the bound $\|U\|_{C^{\lfloor k/2\rfloor}}\leq c_1t$, the increments sum absolutely in each fixed smooth norm. The limit $F$ induces $h$, because the bounded normalized defect is multiplied by $\delta^2\to0$. The initial choices depend on finitely many reference derivatives. It remains to obtain uniformity. The bounded reference two-jets with the given phase and metric margins form a compact set. We choose a finite collection of positive-square and projected-normal constructions valid over this set. A fixed subdivision of each enlarged chart makes the chosen data valid on each piece, by the uniform reference $C^3$ bound. We colour the intersection graph of the resulting supports and assign the hierarchy of weights to the pairs consisting of a colour and a square index. Distinct supports of the same colour are disjoint. Hence the number of required weights and the number of local terms meeting a chart are bounded independently of the size of the atlas. The good differentials of the finitely many phase types yield uniformly buffered adapted coordinates, obtained by shrinking the chart and inverting a leaf derivative that is bounded away from zero. All estimates above then have uniform constants at each fixed order, including the finitely many starting thresholds at $k=10$. This proves Proposition~\ref{prop:repair}. \section{Removal of double points with the metric fixed}\label{sec:separation} The output of Proposition~\ref{prop:repair} need not be injective. We now remove double points without changing the metric or the target dimension. The tool is a smooth finite-parameter family of exact isometric immersions driven by prescribed oscillating normal perturbations. Let $F_0$ be an immersion with normal projection $\Pi_0$. For an immersion $F$ close to $F_0$ in $C^1$, the restriction $P_F=\Pi_0|_{\mathcal N_F}$ is an isomorphism onto $\mathcal N_{F_0}$, and its polar part $\Phi_F=(P_FP_F^*)^{-1/2}P_F$ is an isometry $\mathcal N_F\to\mathcal N_{F_0}$. We call $\widehat B_F=\Phi_F\circ B_F$ the \emph{identified second fundamental form}. In local coordinates $\widehat B_F(\partial_\alpha,\partial_\beta)=\mathcal Q_F(\partial_\alpha\partial_\beta F)$, where $\mathcal Q_F$ is a smooth function of $\dd F$ with $\mathcal Q_{F_0}=\Pi_0$. For $\nu\in\mathcal N_{F_0}$ and a covector $\xi$ we write $\nu\otimes\xi^{\otimes2}$ for the map $Q\mapsto Q(\xi,\xi)\nu$. \begin{lemma}[Exact parameter families]\label{lem:probes} Let $F_0\colon M\to\R^N$ be a phase-regular immersion of a closed manifold, with $N\geq s_n+1$, and let $g_0=\metric(F_0)$. Take finitely many coordinate phases $\psi_l$ whose differentials are good throughout their charts, smooth normal fields $\nu_l$ perpendicular to $B_{F_0}(\Sym^2\ker\dd\psi_l)$ on these charts, and smooth cutoffs $\chi_l$ compactly supported in the charts. There are positive constants $w_l$ and a ball of parameters $u=(u_{l,1},u_{l,2})$ about zero with the following property. For each sufficiently small $t>0$ there is a family $F_{t,u}$, jointly smooth in $(x,u)$, with $\metric(F_{t,u})=g_0$ for all $u$ and $F_{t,u}\to F_0$ in $C^2$ uniformly in $u$ as $t\to0$. Put \[ \tau=t^{6/5},\quad \sigma=t^{44/5},\quad b=t^{56/5},\quad (c_{l,1},c_{l,2})= \bigl(\cos(w_l\psi_l/\tau),\sin(w_l\psi_l/\tau)\bigr), \] and let $\widehat B_{t,u}$ be the identified second fundamental form of $F_{t,u}$. Then \begin{align*} \partial_{u_{l,i}}\widehat B_{t,u} &=\sigma\,\chi_l\nu_l\otimes\dd\psi_l^{\otimes2}\,c_{l,i} +\sigma\,O(\eta(t)+|u|),\\ \partial_{u_{l,i}} F_{t,u} &=-b\, w_l^{-2}\chi_l\nu_l c_{l,i}+b\,O(\eta(t)+|u|), \end{align*} where the remainders are uniform in $C^0$ and $\eta(t)\to0$. \end{lemma} \begin{proof} We include the prescribed phases in the hierarchy of weights used for \eqref{eq:squares}, so that all doubled phases and all sums and differences of distinct phases with intersecting supports are good. Set $\delta=t^{10}$, $\ell=t^{11/10}$ and $G_0=\sqrt{1-\delta^2}F_0$. Constant scaling preserves the normal spaces and the kernels of the transverse pairings. To the homogeneous modes we add the free amplitudes \[ -\delta\tau\chi_l w_l^{-2}\nu_l (u_{l,1}-\ii u_{l,2}) \] with the weighted phases $w_l\psi_l$. Since the mode operator is complex linear, the resulting amplitude is $\delta\tau(u_{l,1}-\ii u_{l,2})Z_l^{(0)}$, where $Z_l^{(0)}$ is the output for $-w_l^{-2}\chi_l\nu_l$ and differs from it by $O(\theta)$, $\theta=\tau/\ell=t^{1/10}$. The leading quadratic mean of these modes, divided by $\delta^2$, is the tensor \[ H_{\mathrm p}(u)=\sum_l \frac{\chi_l^2|\nu_l|^2(u_{l,1}^2+u_{l,2}^2)}{2 w_l^2} \,\dd\psi_l^{\otimes2}. \] We use the target $H_{\mathrm{tar}}=(1-(\tau^{10}/\delta)^2)g_0=(1-t^4)g_0$ and begin the positive-square formulas for the background at $A_0=H_{\mathrm{tar}}-H_{\mathrm p}(u)$. For a fixed small parameter ball this stays in the positivity neighbourhood. The zero-phase error $T_0(A,u)$ involves only each label paired with its own conjugate, so it is the sum of a term depending only on $A$ and a term quadratic in $u$. In the finite recursion \[ A_{j+1}=H_{\mathrm{tar}}-H_{\mathrm p}(u)-T_0(A_j,u) \] all background functions are therefore even in $u$. Their first parameter derivatives are $O(|u|)$ in the required slow norms, since their second parameter derivatives are uniformly bounded by the finite formulas. We take accuracy $q_0=440$ in this first increment and cancel the nonzero quadratic modes as in \eqref{eq:incrementerror}. The target is $(\delta^2-\tau^{20})g_0$, and the bounds \eqref{eq:increment}--\eqref{eq:incrementerror} hold in every fixed weighted joint norm in $(x,u)$. Since $\tau^{20}=t^{24}$ and $\theta^{440}=t^{44}$, the first normalized defect $\tau^{-20}(g_0-\metric(G_1))$ at scale $\tau$ differs from $g_0$ by $O(t^{30}+t^{24/5})$ in these norms. The first map $G_1$ is within $O(\sigma)$ of $F_0$ in $C^2$ and within $O(\delta)$ in $C^1$. The leading terms of its first parameter derivatives are those in the second line of the lemma, because $\delta\tau=b$. Two spatial derivatives falling on the phase give the leading term of the first line, because $b\tau^{-2}=\delta/\tau=\sigma$; terms in which fewer derivatives fall on the phase are $O(\delta)=\sigma\,O(\tau)$. The homogeneous solver contributes a relative error $O(\tau/\ell)$, the quadratic correction a relative error $O(\delta/\tau)$, and the background parameter derivatives a relative error $O(|u|)$. This gives the displayed estimates before the exact correction, with ordinary second derivatives in place of the second fundamental form. For the correction, we extend the parameters to a compact torus by a smooth map into the small ball that is the identity on a smaller ball. We run the iteration of Proposition~\ref{prop:repair} on the product, with the partial metric along $M$, starting at scale $\tau$ and stage $10$, with scale and waiting choices common to all parameters. The foliated estimates give smooth convergence jointly in the parameters and the points of $M$. To retain the leading responses we impose additional finite waits on entering a stage $k>10$: for $j=1,2,3$ we require $C_j(k)t_i^{k-10}\leq1$, where $t_i$ is the scale at the step and $C_j(k)$ is the constant in \eqref{eq:increment}. Each wait is finite because $k-10>0$. The $C^j$ norm of a later increment is then at most a fixed constant times $t_i^{10+(6/5)(1-j)}$ for $j=1,2,3$. Since the first scale is $\tau$, the sums of these norms are $O(t^{12})$, $O(t^{264/25})$ and $O(t^{228/25})$ respectively. The first is $o(b)$, the second $o(\delta)$ and the third $o(\sigma)$. Hence one exact family has both parameter estimates, including the second-derivative estimate. Its first jets and their first parameter derivatives differ from those of $F_0$ by $O(\delta)$. Orthogonal projection and polar identification are smooth functions of the first jets on a uniform immersion neighbourhood, and differentiating either costs $O(\delta)=o(\sigma)$ in the estimate of the second fundamental form. This proves the lemma, with $\eta(t)=t^{1/10}$. \end{proof} We use the following parameter-avoidance principle. \begin{lemma}\label{pl:lem:avoidance} Let $\Sigma$ be a manifold of dimension $m$, let $\mathcal U\subset\R^P$ be open, and let $\Phi\colon\Sigma\times\mathcal U\to\R^k$ be smooth. Let $S\subset\Sigma\times\mathcal U$ and suppose that $\dd\Phi$ has rank at least $r>m$ at every zero of $\Phi$ in $S$. Then the set of $u\in\mathcal U$ for which $\Phi(\cdot,u)$ has a zero in $S$ has measure zero. The same holds for countably many such equations. \end{lemma} \begin{proof} Near a zero we choose $r$ components of $\Phi$ with independent differentials. By the implicit function theorem their common zero set is a submanifold of dimension $m+P-r0$ choose a basis $Q_i$ of $K$ and set $a(\xi)=(Q_i(\xi,\xi))_{i=1}^d$. Choose $\xi_0\in\Omega$ with $a(\xi_0)\ne0$. Since $\dd(Q_i(\xi,\xi))=2Q_i\xi$ and the vectors $Q_i\xi$ are independent, $J=\dd a_{\xi_0}$ is onto. For $P\in\mathcal R$ the differential of $p(\xi)=P(\xi,\xi)$ lies in the span of those of the $a_i$; in local submersion coordinates it follows that $p=f\circ a$ near $\xi_0$. The differentials $\dd p_{\xi_0}$, $P\in\mathcal R$, span a $d$-dimensional space, since they include those of the $Q_i$. On the kernel of $P\mapsto\dd p_{\xi_0}$ we have $\dd f_{a(\xi_0)}=0$ and $\dd^2p=J^* C J$ with $C=\dd^2f_{a(\xi_0)}$. Homogeneity gives $CJ\xi_0=0$, because $\dd^2p(\xi_0,\cdot)=\dd p_{\xi_0}=0$, and $J\xi_0=2a(\xi_0)\ne0$. The assignment $P\mapsto C$ on this kernel is injective, since the Hessian determines a quadratic form. The symmetric forms on $\R^d$ annihilating a fixed nonzero vector form a space of dimension $d(d-1)/2$. Thus $\dim\mathcal R-d\leq d(d-1)/2$, and the span has dimension at least \[ \mu-\frac{d(d-1)}2 \ \geq\ (q+2)-\frac{(n-1)(n-2)}2=n+1 .\qedhere \] \end{proof} Finally, if $L(\alpha)=\sum_l(\alpha_{l,1}\cos\theta_l+\alpha_{l,2}\sin\theta_l)v_l$ for vectors $v_l$ in an inner product space, then $LL^*w=\sum_l\pair{w}{v_l}v_l$ for all angles $\theta_l$. Thus a cosine--sine pair of columns has the same Gram operator as the single column $v_l$, and the nonzero singular values of $L$ do not depend on the angles. \begin{proposition}[Same-metric approximation by embeddings]\label{prop:embed} Let $M$ be closed, let $N\geq s_n+2$, and let $F\colon M\to\R^N$ be a smooth phase-regular immersion. For every $\eps>0$ there is a smooth phase-regular embedding $G$ with $\metric(G)=\metric(F)$ and $\norm{G-F}_{C^2}<\eps$. \end{proposition} \begin{proof} Since phase regularity persists under small $C^2$ perturbations, we may assume that every map within $\eps$ of $F$ in $C^2$ is a phase-regular immersion. Write $q=s_n-n$ and $\mu=N-n\geq q+2$. By phase regularity, $\rank B_F\geq q$ everywhere. \emph{Removal of the minimal-rank locus.} At a point where $\rank B_F=q$, the kernel $K$ of $B_F$ has dimension $n$, and $U=(\im B_F)^\perp\subset\mathcal N_F$ has dimension $\mu-q\geq2$. In frames adapted to the splittings of $\Sym^2TM$ and $\mathcal N_F$ defined by $K$ and $\im B_F$ at this point, $\widehat B$ has a block matrix whose $(1,1)$ block is invertible nearby. Near the point the rank-$q$ equation is the vanishing of its Schur complement in $\Hom(K,U)$, whose differential with respect to $\widehat B$ is the $K$-to-$U$ block. For every good $\xi$ at the point, $B_F(\Sym^2\ker\xi)=\im B_F$, so $U$ is the space of free normal directions for the phase $\xi$. Such a perturbation changes the Schur complement to leading order by $Q\mapsto Q(\xi,\xi)\nu$ with $\nu\in U$. The quadratic evaluations at good $\xi$ span $K^*$, since a quadratic form vanishing on an open set is zero. Hence finitely many such perturbations span $\Hom(K,U)$. We extend each of them by a coordinate-linear phase and a smooth projected normal field, and choose a cutoff equal to one near the point. The rank-$q$ locus is compact, so a finite list suffices. By Lemma~\ref{lem:probes} and the cosine--sine identity, the Schur equations have, for a small parameter ball and small $t$, parameter derivative of rank $n(\mu-q)\geq2n>n$ near this locus. Outside the chosen neighbourhoods the $(q+1)$-st singular value of $B_F$ has a positive minimum. Lemma~\ref{pl:lem:avoidance} gives an immersion $F_1$ with the same metric, arbitrarily close to $F$ in $C^2$, with $\rank B_{F_1}\geq q+1$ everywhere. \emph{Removal of double points.} At every point of $F_1$, Lemma~\ref{pl:lem:freespan} provides $n+1$ independent free normal vectors. There is a fixed neighbourhood of the diagonal in $M\times M$ on which every immersion sufficiently $C^1$-close to $F_1$ is injective on pairs of distinct points. To see this, we project, in each of finitely many small convex coordinate balls, to a fixed tangent plane and compare the derivative with an invertible constant matrix; integration along segments gives a uniform lower bound for the displacement on each ball. The remaining pairs form a compact set disjoint from the diagonal. At a coincidence $(x,y)$ in this set we choose $n+1$ independent free normal vectors at $x$, with phases and cutoffs supported away from $y$. A finite list covers the coincidence set. In the equation $F_{t,u}(x)-F_{t,u}(y)=0$ the parameter derivatives divided by $b$ in Lemma~\ref{lem:probes} produce these $n+1$ normal directions, uniformly in the phase angles. Together with the $n$ independent tangent directions $\dd F_1(T_xM)$, which are orthogonal to them, they give rank at least $2n+1>2n=\dim(M\times M)$. This inequality persists for a small parameter ball and small $t$. Away from the chosen neighbourhoods the value difference has a positive minimum on the compact set of pairs. Lemma~\ref{pl:lem:avoidance} removes every coincidence. Taking both perturbations small gives the required $C^2$ bound, and by compactness the injective immersion is an embedding. \end{proof} \section{From a compact model to the prescribed metric}\label{sec:assembly} We now prove Theorem~\ref{thm:model}. We wind $M$ at a high frequency $T$ into the sphere product, pull back a single model map $S_h$, and correct the resulting metric by Proposition~\ref{prop:repair}. \begin{lemma}\label{p03:lem:factorization} Let $(M,g)$, $f$ and $W$ be as in Theorem~\ref{thm:model}. There are a smooth immersion $\phi\colon M\to\R^d$ and smooth bundle maps $L\colon TM\to M\times\R^a$ and $R\colon M\times\R^d\to M\times\R^a$ such that, for every $x\in M$, $L_x$ is an isometry of $(T_xM,g_x)$ onto $(W_{f(x)},g_W)$ and $R_x\,\dd\phi_x=L_x$. \end{lemma} \begin{proof} Take a finite coordinate cover and compactly supported functions $\beta_i\geq0$, at least one positive at every point. The coordinates of $\phi$ are the $\beta_i$ together with their products by each of the corresponding local coordinates. At a point with $\beta_i>0$, the derivative of $\phi$ determines all components of a tangent vector, so $\phi$ is an immersion. We normalize a bundle isomorphism $TM\to f^*W$ by the inverse positive square root of the change of metric, and obtain isometries $L_x\colon T_xM\to W_{f(x)}$ for the prescribed $g$. Using the inclusion $W\subset B\times\R^a$ we regard $L_x$ as $\R^a$-valued. Let $A_x=\dd\phi_x$, let $A_x^+$ be the smooth left inverse obtained from orthogonal projection onto $\im A_x$, and set $R_x=L_xA_x^+$. Then $R_xA_x=L_x$. \end{proof} \begin{lemma}\label{p03:lem:hopf-pullback} In the setting of Theorem~\ref{thm:model}, let $\phi$, $L$ and $R$ be as in Lemma~\ref{p03:lem:factorization}. There is an integer $h$ for which $S_h$ exists, and there are $K_0,T_0\geq1$, with the following property. For $K\geq K_0$ and $T\geq\max(K,T_0)$ there is a smooth map $z=(z_1,\dots,z_a)\colon M\to(S^{2h-1})^a$ such that \[ F_{K,T}(x)=S_h\bigl(f(x),z_1(x),\ldots,z_a(x)\bigr) \] satisfies the following, in the charts $x=\exp_p(y/T)$, $\abs{y}<4$, centred at a $T^{-1}$-net: \begin{enumerate} \item the inner charts $\abs{y}<2$ cover $M$, the overlap multiplicity is bounded, the transitions and the coefficients of $T^2g$ are bounded at every fixed order, and $T^2g$ is uniformly positive, all uniformly in $K$ and $T$; \item $\abs{F_{K,T}}\leq\sup\abs{S_h}$, and every fixed-order derivative of $F_{K,T}$ in these charts is bounded uniformly in $K$ and $T$; \item for a constant $C$ independent of $K$ and $T$, \begin{equation}\label{pl:eq:scaled-defect} \|\metric(F_{K,T})-T^2g\|_{C^{440}_{\mathrm{scaled}}}\leq C(K^{-1}+K/T), \end{equation} and the best phase margin of $F_{K,T}$ in these charts has a positive lower bound independent of $K$ and $T$ once $K^{-1}+K/T$ is small. \end{enumerate} \end{lemma} \begin{proof} Take a maximal $K^{-1}$-separated net $\{x_i\}$ and cover $M$ by the balls of radius $2K^{-1}$ about its points. Choose a squared partition of unity $\sum_i\chi_i^2=1$ with $\supp\chi_i$ in these balls and $\norm{\nabla^j\chi_i}_\infty\leq C_jK^j$; such functions are obtained by normalizing the sum of squares of rescaled bumps. Packing of disjoint smaller balls bounds the overlap multiplicity independently of $K$. We colour the intersection graph of the closed supports with a fixed number $h$ of colours, increasing $h$ if necessary so that $S_h$ is available. At most one summand of a given colour is nonzero at any point. This map $S_h$ is selected before the large parameters $K$ and $T$. For $T\geq K$ and $1\leq\ell\leq a$ we define \[ z_{\ell,c}(x)=\sum_{i\text{ of colour }c} \chi_i(x)\exp\bigl(2\pi\ii T(R_{x_i}\phi(x))_\ell\bigr), \qquad c=1,\ldots,h . \] Disjointness within each colour gives $\abs{z_\ell(x)}^2=\sum_i\chi_i(x)^2=1$, so $z_\ell$ takes values in $S^{2h-1}$ and $\abs{F_{K,T}}\leq\sup\abs{S_h}$. On $\supp\chi_i$ we have $R_{x_i}\dd\phi_x-L_x=O(K^{-1})$. Hence, for a local vector field $v$ of bounded length, \[ T^{-1}v(z_{\ell,c})=2\pi\ii\,(L_xv)_\ell\, z_{\ell,c}+O(K^{-1}+K/T). \] More precisely, the error is a sum with coefficients $T^{-1}v\chi_i$ or $2\pi\ii\chi_i(R_{x_i}\dd\phi_x-L_x)v$. In charts scaled by $T^{-1}$ their derivatives at each fixed order satisfy the same bound: a derivative falling on a cutoff costs $K/T$, the derivatives of positive order of the exponent are bounded, and derivatives falling on slow data gain $T^{-1}$. Since $2\pi\ii z_\ell=Y_\ell(z_\ell)$, the scaled first derivatives of $F_{K,T}$ are, up to $O(K^{-1}+K/T)$, the Hopf derivatives $Y_{L_xv}S_h$. Differentiating once more in the scaled frame gives the same error bound for the second derivative columns, whose leading terms are the successive Hopf derivatives $Y_{L_xu}Y_{L_xv}S_h$. The terms in which $L_x$, the source frame or $f(x)$ are differentiated carry a factor $T^{-1}$. These estimates hold at each fixed scaled derivative order, since the number of overlapping supports is bounded. Since $S_h$ induces $g_W$ on $D_W$ and $L_x$ is an isometry, the leading Gram matrix of the first columns with respect to a $g$-orthonormal frame is the identity. The leading normal projection and second fundamental form are exactly those of $S_h$ along $D_W$ at the image point, composed with $L_x$. We use the charts $x=\exp_p(y/T)$ with centres $p$ in a maximal $T^{-1}$-separated net. Their scaled transitions $y\mapsto T\exp_q^{-1}(\exp_p(y/T))$ have uniformly bounded derivatives at every fixed order, by smoothness on a compact manifold and rescaling, and packing again bounds the overlap multiplicity. In these charts $T^2g$ equals $\Id+O(T^{-1})$ with uniform derivative bounds. This proves~(i). The previous estimates give~(ii) and~\eqref{pl:eq:scaled-defect}. The best phase margin of $S_h$ has a positive minimum on the compact domain of the model, and the normal projector depends smoothly on the first columns. Hence the phase margin of $F_{K,T}$ is bounded below for $K^{-1}+K/T$ small. \end{proof} \begin{proof}[Proof of Theorem~\ref{thm:model}] Apply Lemmas~\ref{p03:lem:factorization} and~\ref{p03:lem:hopf-pullback} with $K=T^{1/2}$, so that $K^{-1}+K/T=2T^{-1/2}$. In the scaled charts all the data required by Proposition~\ref{prop:repair}, namely chart radii, overlap multiplicity, transition bounds, ellipticity, reference derivative bounds and phase margin, are uniform in $T$. We apply Proposition~\ref{prop:repair} with a fixed tolerance $\eps_0\leq1$ in the scaled charts. Its threshold $\eta_0$ does not depend on the number of charts, hence not on $T$. By \eqref{pl:eq:scaled-defect}, for $T$ large the proposition yields a phase-regular map $F_T^*$ with $\metric(F_T^*)=T^2g$ and $\abs{F_T^*-F_{T^{1/2},T}}\leq1$. The map $F_T^*/T$ induces $g$, is phase regular, since scaling by a constant preserves tangent and normal spaces, and satisfies \[ \sup_M\abs{F_T^*/T}\leq\frac{\sup\abs{S_h}+1}{T}. \] We choose $T$ so that its image lies in the ball of half the prescribed radius centred at the origin. We then apply Proposition~\ref{prop:embed} with a $C^0$ displacement smaller than the remaining radius, and translate to the prescribed centre. \end{proof} \section{The four-section construction at dimension \texorpdfstring{\(s_n+5\)}{s(n)+5}} \label{sec:foursection} In this section we describe the geometry that a compact model for a general tangent bundle \(TM\) has to accommodate. It accounts for the normal ranks used later, and for the difference between a construction of formal columns and the partial-metric hypothesis of Theorem~\ref{thm:model}. \subsection{The three-dimensional singular set}\label{sec:fourgeometry} Let \(M\) be closed with \(n\geq5\). A generic bundle map \(\Psi:\triv^4\to TM\) has rank at least three everywhere, and it drops rank exactly on its rank-three locus \(P\), a closed embedded three-manifold (possibly empty) along which \(\Psi\) is transverse to the rank stratification. Indeed, in a matrix chart in which a \(j\)-by-\(j\) minor is invertible, block elimination writes the matrix as an invertible lower triangular factor, times the block diagonal matrix formed by the minor and its Schur complement, times an invertible upper triangular factor. Hence rank \(j\) is equivalent to the vanishing of the \((n-j)\)-by-\((4-j)\) Schur complement, and the rank-\(j\) stratum has codimension \((n-j)(4-j)\). For \(j\leq2\) this is at least \(2(n-2)>n\), while for \(j=3\) it is \(n-3\). Transversality follows from local matrix perturbations, cut off on a finite cover, and the parametric transversality theorem. On \(P\) we write \(K=\ker\Psi|_P\) for the kernel line and \(Q=(\im\Psi|_P)^\perp\) for the cokernel, of rank \(n-3\). The normal derivative of the Schur complement identifies \(\nu P\) with \(\Hom(K,Q)\). On a sufficiently small tube around \(P\) we use the Schur complement itself as the normal coordinate \(v\in\Hom(K,Q)\). In the splittings \(\triv^4=K^\perp\oplus K\) and \(TM=\im(\Psi|_P)\oplus Q\), transported radially from \(P\), block elimination factors \(\Psi\) into the diagonal blocks given by the rank-three isomorphism and by \(v\), with invertible triangular factors on the two sides. Multiplying the off-diagonal entries of these factors by a radial cutoff switches them off on a smaller tube, keeps both factors invertible and creates no new rank changes. Let \(A_0\) be the invertible block at the zero section and \(U_0=A_0(A_0^*A_0)^{-1/2}\) its polar isometry. After multiplication by \(U_0^{-1}\), the block equals \((A_0^*A_0)^{1/2}\) at the zero section, so its symmetric part is positive definite on a smaller tube, and convex interpolation with the identity preserves invertibility. Altogether we obtain a prepared form \[ \Psi_{(p,v)}=\begin{pmatrix} U_0(p) & 0\\ 0 & v\end{pmatrix}: K_p^\perp\oplus K_p\longrightarrow\im(\Psi_p)\oplus Q_p \] on a neighbourhood of \(P\), with \(U_0:K^\perp\to\im(\Psi|_P)\) a bundle isometry over \(P\). Fix an orthogonal complex structure \(J\) of \(\triv^4\) and set \(L=JK\) and \(D_2=(K\oplus L)^\perp\). Write \[ A=K\oplus D_2=L^\perp,\qquad H=Q\oplus D_2,\qquad m=n-4 . \] Via \(U_0\), the pulled-back tangent bundle is \(H\oplus L\) on the tube. Let \(\Sigma=S(\Hom(K,Q))\) be the unit normal sphere bundle, of dimension \(n-1=m+3\). A point \(v\in\Sigma\) is an isometric embedding of the line \(K_p\) into \(Q_p\), and we put \(E_v=Q\ominus vK\), a bundle of rank \(m\) over \(\Sigma\). The link identification \[ E_v\oplus A\longrightarrow H,\qquad (e,k,b)\longmapsto(e+v(k),b), \] uses \(v\) on the summand \(K\) of \(A\). It is a metric identification and requires no choice of a unit vector in \(K\), so it is defined even when \(K\) is nonorientable. For a Euclidean bundle \(U\) write \(J_2(U)=U\oplus\Sym^2U\), and put \(j_k=\rank J_2(\R^k)=k(k+3)/2\). The link identification splits \(J_2(H)\) orthogonally into \(J_2(E_v)\) and the small summand \[ A\oplus(E_v\otimes A)\oplus\Sym^2A, \] of rank \(3m+9\). The same small summand is the orthogonal complement of \(v(K)\otimes A\) in the bundle \[ \mathcal S=A\oplus(Q\otimes A)\oplus\Sym^2A \] over \(P\), of rank \(3m+12\). The second description allows us to compare the full columns over the whole link. We shall use the following standard facts. \begin{lemma}\label{pl:rankavoid} Let \(X\) be a compact manifold of dimension \(d\), possibly with boundary, and let \(U_1,U_2\) be Euclidean bundles over \(X\) of ranks \(k\leq\ell\), with rank gap \(g=\ell-k\). \begin{enumerate} \item If \(g+1>d\), every isometric monomorphism \(U_1\to U_2\) given near a closed subset \(X_0\subset X\) extends to an isometric monomorphism over \(X\) which agrees with it near \(X_0\). In particular, if \(g>d\), any two isometric monomorphisms \(U_1\to U_2\) are homotopic through isometric monomorphisms, with stationary endpoints. \item Let \(Z_t:U_1\to U_2\), \(t\in[0,1]\), be a smooth path of isometric monomorphisms. The skew endomorphism \[ \Omega_t=\dot Z_tZ_t^*-Z_t\dot Z_t^*-Z_t(Z_t^*\dot Z_t)Z_t^* \] of \(U_2\) satisfies \(\Omega_tZ_t=\dot Z_t\). Its linear flow \(O_t\), with \(O_0=\Id\), is orthogonal, satisfies \(O_tZ_0=Z_t\), maps \((\im Z_0)^\perp\) onto \((\im Z_t)^\perp\), and is the identity wherever \(Z_t\) is independent of \(t\). Similarly, for a smooth path of orthogonal projectors \(\Pi_t\), the skew generator \([\dot\Pi_t,\Pi_t]\) has an orthogonal flow \(V_t\) with \(\Pi_tV_t=V_t\Pi_0\). \end{enumerate} \end{lemma} \begin{proof} In \(\Hom(\R^k,\R^\ell)\) the matrices of rank \(j\dim\Sigma\) and \(j_m>\dim\Sigma\), there is an isometric monomorphism \(K\otimes A\to\mathcal S\) over \(P\), and its pullback to \(\Sigma\) and the moving inclusion \(k\otimes a\mapsto v(k)\otimes a\) are homotopic through isometric monomorphisms. Indeed, rank failure for an inclusion with rank gap \(R\) has codimension \(R+1\), which exceeds the dimension of the cylinder \(\Sigma\times[0,1]\); one extends a path of bundle maps from the endpoints, uses relative transversality and applies polar normalization (Lemma~\ref{pl:rankavoid}(i)). If \(Z_t\) is the resulting isometric path, the skew matrix \[ \dot Z_t Z_t^* - Z_t\dot Z_t^* - Z_t(Z_t^*\dot Z_t) Z_t^* \] takes \(Z_t\) to \(\dot Z_t\). Its linear orthogonal flow \(R_t\) transports the complement of the moving inclusion, that is, the small summand above, to a fixed small bundle \(S_0\) pulled back from \(P\). The two inclusions of \(S_0\) into \(J_2(H)\) obtained in this way, namely an inclusion \(i_0\) over \(P\) and the small summand composed with \(R_1^{-1}\), can be compared by the same procedure, now with rank gap \(j_m\). The resulting orthogonal flow \(T_t\) of \(J_2(H)\) over \(\Sigma\) gives an isometry \(T_1\) of \(J_2(E_v)\) over the whole link onto a bundle \(J_0\) pulled back from \(P\), and it can be used to clutch the old column bundle on the exterior to one on the disk tube. The two rank inequalities hold for \(n\geq7\): one has \(R-\dim\Sigma=2m+6\) and \(j_m-\dim\Sigma=(m-2)(m+3)/2\). The corresponding extension of formal columns is as follows. We write \(J_2(H)=J_0\oplus i_0(S_0)\) for the fixed splitting over \(P\) obtained after the second transport. Let \(I_0:J_0\to W_0\) and \(I_{\mathcal S}:\mathcal S\to W_0\) be isometric inclusions, over \(P\), with orthogonal ranges in a target bundle \(W_0\). On the link we prescribe the inclusion \[ x+y\longmapsto I_0(T_1x)+I_{\mathcal S}(y), \qquad x\in J_2(E_v),\quad y\in A\oplus(E_v\otimes A)\oplus\Sym^2 A . \] First apply \(R_t\) in the second summand, keeping the first one unchanged. At \(t=1\) the result equals \(I_*T_1\), where \[ I_*(x_0+i_0s)=I_0(x_0)+I_{\mathcal S}(s),\qquad x_0\in J_0,\ s\in S_0, \] is independent of \(v\). Then follow \(I_*T_t\) backwards to \(I_*\). Placing these two paths on radial annuli with stationary endpoints, and using \(I_*\) near the zero section, we obtain a smooth monomorphism on the disk bundle which coincides with the prescribed one on a boundary collar. This is a statement about formal column data only. The coefficient inclusion itself can also be extended across the tube, without changing the number of original Hopf factors. On the boundary collar, polar normalization of the prepared map sends \(D_2\oplus L\) identically to itself and \(v(K)\) to \(K\) by \(v^*\). Accordingly, if \(I_Q:Q\to\triv^a\) is an isometric inclusion over \(P\), the link inclusion sends \(E_v\) through \(I_Q\), \(vK\) through \(v^*:vK\to K\), and \(D_2\oplus L\) identically into \(\triv^4\). For \(q\in Q\), \(b\in D_2\) and \(\ell\in L\), the map \[ (q,b,\ell)\longmapsto \bigl(I_Q(\pi_{E_v}q+\cos(\theta)\,\pi_{vK}q),\ \sin(\theta)\,v^*q+b+\ell\bigr) \] is isometric for \(0\leq\theta\leq\pi/2\), since \(\abs{v^*q}=\abs{\pi_{vK}q}\). Choose \(\theta\) as a function of the radius, equal to \(\pi/2\) on a boundary collar, where the map is the link inclusion above, and equal to zero near the zero section, where the map becomes \((q,b,\ell)\mapsto(I_Q(q),b+\ell)\) and is independent of \(v\). This gives a coefficient inclusion of \(TM\) across the whole tube. \subsection{The dimension count}\label{p03:subsec:dimcount} In the four-direction construction the old and active Euclidean blocks have dimensions \(s_n-152\) and \(157\). The full old columns \(J_2(E)\) use \(j_m\) dimensions, and the four active first columns and ten symmetric second columns use \(4+10=14\). Thus, if the corresponding partial maps have these ranks, the remaining dimensions in the two blocks are \[ \begin{array}{c|r} \text{old block complement} &s_n-152-j_m=3m-142\\ \text{active block complement}&157-14=143\\ \hline \text{total complement} &3m+1 . \end{array} \] Here we used \(s_n=s_m+4m+10\), and the count involves \((s_n-152)+157=s_n+5\) coordinates. On the core, the bundle \(J_2(H)\oplus L\) has rank \(j_{n-1}+1=s_n\), so any isometric column inclusion of it into \(\R^{s_n+5}\) has a complement of rank five. For any such inclusion over the zero section, two perpendicular copies of \(L\) can be chosen in this complement: rank failure in \(\Hom(2L,\R^5)\) has codimension at least four, which exceeds \(\dim P\). Such a choice extends along any given homotopy of column inclusions on the disk bundle and a product collar, by orthogonal transport of their complements. For instance, if \(\Pi_t\) denotes the projector onto the complement, one may use the skew generator \([\dot\Pi_t,\Pi_t]\) of Lemma~\ref{pl:rankavoid}(ii). \subsection{The characteristic-class condition}\label{p03:subsec:charclass} Write \[ r=n-3,\qquad a=w_1(Q),\qquad b=w_2(Q),\qquad c=w_3(Q),\qquad t=w_1(L) \] on \(P\). Let \(N_5\) be the complement of an isometric monomorphism \(J_2(H)\oplus L\to P\times\R^{s_n+5}\) over \(P\). Such a monomorphism exists, and is unique up to isometric homotopy, by Lemma~\ref{pl:rankavoid}, since the rank gap is five and \(\dim P=3\). \begin{proposition}\label{prop:characteristic} On every component of \(P\), \[ w_3(N_5)= \begin{cases} 0, & n \text{ odd or } n\equiv0\pmod4,\\ w_1(\det\Hom(K,Q))^3, & n\equiv2\pmod4. \end{cases} \] Three linearly independent sections of \(N_5\) exist on a component precisely when this class vanishes there. If \(n\equiv2\pmod4\), put \(\mathcal D=\det\Hom(K,Q)\); then \(N_5\simeq 3\mathcal D\oplus\triv^2\). \end{proposition} \begin{proof} All characteristic classes are taken over \(\Ftwo\), and we work on one component of \(P\). Since \(J\) identifies \(K\) with \(L\), we have \(H\oplus2L\simeq Q\oplus\triv^4\). Expanding \(\Sym^2\) of both sides, and using \(\Sym^2(2L)\simeq\triv^3\) and \([H\otimes L]=[Q\otimes L]+4[L]-2[\triv]\), we obtain in stable bundle notation \[ [\Sym^2 H]=[\Sym^2Q]+4[Q]-2[Q\otimes L]-8[L]+11[\triv]. \] The first columns have bundle \(H\oplus L\), with \([H\oplus L]=[Q]+4[\triv]-[L]\). Since \([J_2(H)\oplus L]+[N_5]\) is trivial, it follows that \[ w(N_5)=w(\Sym^2Q)^{-1}w(Q)^{-5}w(Q\otimes L)^2(1+t)^9 . \] Put \(r_2=r\bmod 2\). Since \(\dim P=3\), we have \(w(Q)^4=(1+t)^8=1\) and \(w(Q\otimes L)^2=1+(a+r_2t)^2\). Hence, before multiplying by the first factor, the terms of degrees one through three are \[ a+t,\qquad b+at+r_2t^2,\qquad c+bt+r_2at^2+r_2t^3 . \] The positive-degree terms of \(w(\Sym^2Q)\) through degree three are \[ (r-1)a,\qquad \tbinom{r-1}{2}a^2+r_2 b,\qquad \tbinom{r-1}{3}a^3+r_2(ab+c). \] These formulas follow by expanding \(\prod_{i3\), rank avoidance gives a section \(v_0:K\to Q\) of the unit link bundle over \(P\). The argument of Section~\ref{sec:fourgeometry}, with \(v_0\) as the fixed inclusion, gives an orthogonal homotopy \(T_t\) of \(J_2(H)\) over the link, with \(T_0=\Id\), carrying \(J_2(E_v)\) at its endpoint onto \(J_0=J_2(E_{v_0})\), pulled back from \(P\). In detail, we first homotope the three-column inclusion \(vK\otimes A\) in \(A\oplus(Q\otimes A)\oplus\Sym^2 A\) to the one for \(v_0\); its complement is transported to the fixed small summand \(A\oplus(E_{v_0}\otimes A)\oplus\Sym^2A\) of rank \(3m+9\). The two inclusions of that summand in \(J_2(H)\) are then homotopic, since the rank gap is \(j_m\), and extending this homotopy by orthogonal transport gives \(T_t\). Both rank gaps exceed the dimension \(m+4\) of the link cylinder. Choose an isometric old monomorphism \(I_{\rm o}^P:J_0\to\triv^{N_{\rm o}}\) over \(P\). The rank gap, over \(P\) and on the double, is \(3m-142\). We extend the composite \(I_{\rm o}^P\circ T_1\) on \(J_2(E_v)\), taken radially constant on a collar, to an isometric formal old monomorphism \[ I_{\rm o}:J_2(E)\longrightarrow\triv^{N_{\rm o}} \] on the double. Independently, we prescribe a constant isometric inclusion of the fourteen full active columns \(J_2(A_4)\) into the \(157\)-dimensional target block, with values in \(\R^{143}\oplus\{0\}\); its complement is a fixed \(143\)-plane containing the fourteen unused axes \(\{0\}\oplus\R^{14}\). Over the link, since \(A\oplus L=A_4\), we have the orthogonal decomposition \[ J_2(H)\oplus L=J_2(E_v)\oplus(A_4\oplus\Sym^2A)\oplus(E_v\otimes A) \] into summands of ranks \(j_m\), \(10\) and \(3m\). On the core, we extend the inclusion \(I_{\rm o}^P\) of \(J_0\) to an isometric inclusion of \(J_2(H)\oplus L\) into \(\triv^N\): the complement of \(J_0\) has rank \(3m+10\), the complement of the old columns in \(\R^N\) has rank \(3m+15\), so the remaining rank gap is five, and the base has dimension three. On the link we precompose this inclusion with \(T_1\oplus\Id_L\), so that its old columns agree with the slow old inclusion \(I_{\rm o}\) specified above. In the orthogonal complement of the old columns, which has rank \(3m+15\), its ten active columns \(A_4\oplus\Sym^2 A\) can be moved by an isometric homotopy to the corresponding columns in the fixed active block. The rank gap here is \(3m+5\), larger than \(m+4\), so this relative homotopy of the ten columns follows by rank avoidance on the link cylinder, and we extend it by orthogonal transport of the complement. We have thus obtained the slow augmented class over the link, and it extends over the disk: on a radial collar we run these two homotopies backwards to the original inclusion over \(P\). At the attaching end, the comparison keeps the old columns and the ten active columns at their prescribed placements. We denote by \(J_{\Sigma}\) the resulting mixed reference inclusion of \(E_v\otimes A\) into the complement of the old columns and the ten active base columns. Its normal bundle is the pullback of the rank-five bundle \(N_5\) of Proposition~\ref{prop:characteristic}. The slow retained complement \(W\) on the double is the sum of the complement of the formal old columns and the fixed \(143\)-plane in the active block. It has rank \((3m-142)+143=3m+1\), and on the collar it is pulled back from \(P\), since there the old columns \(I_{\rm o}(J_2(E_v))=I_{\rm o}^P(J_0)\) are independent of \(v\). Adding the four cofactor axes, namely the images of the columns \((L\otimes A)\oplus\Sym^2L\) of the fixed active inclusion, gives the complement of the old columns and the ten selected active columns. Restricted to the marked section \(v_0\), this complement is identified with \((E_{v_0}\otimes A)\oplus N_5\) by the class of the full injection. This gives the two bundle identifications for the ordinary and the exceptional moment models used in Sections~\ref{sec:ordinaryboundary} and \ref{sec:angular}. On the original Hopf product of any fixed size at least two, we realize the formal old inclusion on the closed double by Proposition~\ref{prop:relativefill}, with \(K=\emptyset\) (the double is closed) and the unit old metric; this applies since \(N_{\rm o}\geq j_m+5\). We obtain an exact unit old block \(F_{\rm o}\) together with a homotopy of its full columns to the formal old ones. After polar normalization of the columns, orthogonal complement transport along this homotopy gives a smooth isometry of the slow old normal complement onto the complement of the full old columns of \(F_{\rm o}\). In the active block we use the map \(B\) of Section~\ref{sec:physicalactive}. The homotopy of its full active columns that gives the normal identification can be chosen to fix the fourteen unused axes throughout. For the rank-nine bank over a core we use the map \(\eta_\kappa\) of that section, with a positive constant \(\kappa\) on an annulus around the link and with a slow cutoff to zero before the end of a larger collar. We transport the normal complement from \(\eta_0=\nu\) to \(\eta_\kappa\) within the complement of the full active columns of \(B\). This modifies the active normal isometry by an orthogonal path which is the identity off the collar. In particular it realizes the bank \(\eta_\kappa\) without changing the old part of the normal isometry. With these choices, \[ F_0=(F_{\rm o},B):\text{original Hopf product over the double} \longrightarrow \R^{N_{\rm o}}\oplus\R^{157} \] is the unit free product reference of Section~\ref{sec:metricproduction}. The old part of its full normal isometry \(I_x\) is independent of the active variables, and the active part is independent of the old variables. On every annulus used below, these frames and the slow augmented class determine the same data. \section{The active map on an exceptional link} \label{sec:angular} Throughout this section we assume \[ n\equiv 2\pmod 4,\qquad r=n-3\geq 10^{62}. \] In this residue class Proposition~\ref{prop:characteristic} gives \(N_5\simeq 3\mathcal D\oplus\triv^2\), and three independent normal sections need not exist over \(P\). We therefore work directly over the unit sphere bundle \(\Sigma=S(Q\otimes L)\), which has dimension \(r+2\). Write \(A_4=\triv^4=L\oplus A\) with \(A=L^\perp\) of rank three. An isometric identification of \(\R^4\) with the quaternions \(\mathbb H\) trivializes \(U=L\otimes A\): for a unit vector \(\ell\in L\) the vectors \(\ell\mathbf i,\ell\mathbf j,\ell\mathbf k\) form an orthonormal basis of \(A=\ell^\perp\), and the sections \[ e_1=\ell\otimes\ell\mathbf i,\qquad e_2=\ell\otimes\ell\mathbf j,\qquad e_3=\ell\otimes\ell\mathbf k \] of \(U\) do not depend on the sign of \(\ell\). Via the metric contraction \(L\otimes L\simeq\triv\) we also identify \(A\simeq L\otimes U\). Put \(V=A_4\oplus\R z\), with the inclusion \(I:A_4\hookrightarrow V\), and define over \(\Sigma\) \[ \mathcal H_v=E_v\oplus L\mathcal D,\qquad L\mathcal D=L\otimes\mathcal D . \] The line \(L\mathcal D\) belongs to the old generating bundle \(\mathcal H\), which has rank \(r\). By rank avoidance choose two orthogonal copies of \(L\) in \(Q\), so that \(Q=D\oplus L\oplus L\) with \(\rank D=r-2\). To compare the old full columns we use the section of \(\Sigma\) given by the first copy of \(L\). Along this section the transported bundle \(J_0\) of Section~\ref{sec:referenceproduct} restricts to \(J_2(Q-L)\), and the four first and ten second active columns have bundle \[ A_4\oplus\Sym^2 A\oplus (L\otimes A)\oplus\Sym^2 L . \] The last two summands are trivial, of total rank four. Hence the retained complement \(W\) of the old full columns and the ten active columns \(A_4\oplus\Sym^2A\) has stable class \[ W+\triv^4=\bigl((Q-L)\otimes A\bigr)+N_5 . \] Since \(L\otimes A\simeq\triv^3\) and \(N_5\simeq3\mathcal D\oplus\triv^2\), stable cancellation gives \[ W\simeq(D\otimes A)\oplus3\mathcal{D}\oplus\triv,\qquad w=\rank W=3r-2. \] For the cancellation we regard both sides as orthogonal complements of two isometric inclusions of the common trivial summand \(\triv^4\) into one bundle of rank \(w+4\); by the rank stratification argument these inclusions are homotopic through isometric inclusions, and we transport their complements along the homotopy. This applies because the rank gap \(w\) exceeds the dimension of the three-dimensional base and of its homotopy cylinder. The bundle \(W\) over \(\Sigma\) is pulled back from \(P\). \subsection{The angular seed and its four cofactor rows} The splitting \(Q=D\oplus L\oplus L\) and the contraction \(L\otimes L\simeq\triv\) give \(Q\otimes L\simeq(D\otimes L)\oplus\triv\oplus\triv\), and we write accordingly \[ v=(v_D,p,q),\qquad \abs{v_D}^2+p^2+q^2=1 . \] Using \(A\simeq L\otimes U\), a tensor in \(\mathcal H_v\otimes A\) has coordinates \((Z,x,y,\zeta)\) with \[ Z\in(D\otimes L)\otimes U,\qquad x,y\in U,\qquad \zeta\in (L\mathcal D)\otimes A\simeq\mathcal D\otimes U=3\mathcal D, \] subject to the condition that \((Z,x,y)\) be orthogonal to the image of \(v\), namely \begin{equation}\label{eq:angularconstraint} v_D^*Z+px+qy=0 , \end{equation} where \(v_D^*\) contracts the factor \(D\otimes L\) with \(v_D\). In the basis \((e_1,e_2,e_3)\) of \(U\) let \[ S_0=\begin{pmatrix}0&1&0\\1&0&1\\0&1&0\end{pmatrix},\qquad P_0=\begin{pmatrix}0&0&1\\0&0&0\\1&0&0\end{pmatrix}; \] thus \(S_0\) is tridiagonal with zero diagonal, and \(P_0\) interchanges \(e_1,e_3\) and kills \(e_2\). Identifying \(W\) with \(((D\otimes L)\otimes U)\oplus3\mathcal D\oplus\triv\), we define the retained seed \(T_A\) and the four base cofactor rows \(C_{\mathrm{base}}\), with values in \(U\oplus\triv\simeq\triv^4\), by \[ T_A(Z,x,y,\zeta)=(Z+v_D\otimes P_0y,\ \zeta,\ y_2),\qquad C_{\mathrm{base}}(Z,x,y,\zeta)=(x+S_0y,\ y_3). \] \begin{lemma}\label{pl:seed} The augmented map \((T_A,C_{\mathrm{base}})\) is injective on \(\mathcal H_v\otimes A\) for every \(v\in\Sigma\). Its restriction to \(E_v\otimes A\) has normal bundle in \(W\oplus\triv^4\) isomorphic to \(\triv^2\oplus3\mathcal D\simeq\pi_\Sigma^*N_5\). \end{lemma} \begin{proof} If \((T_A,C_{\mathrm{base}})(Z,x,y,\zeta)=0\), then \(\zeta=0\) and \(y_2=y_3=0\), and since \(S_0e_1=e_2\), \(P_0e_1=e_3\), \[ y=d_1e_1,\qquad x=-d_1e_2,\qquad Z=-d_1v_D\otimes e_3,\qquad \zeta=0 . \] Then \eqref{eq:angularconstraint} gives \[ d_1\bigl(-\abs{v_D}^2e_3-pe_2+qe_1\bigr)=0 . \] The vector \(\mathbf w(v)=-\abs{v_D}^2e_3-pe_2+qe_1\) satisfies \(\abs{\mathbf w(v)}^2=(\abs{v_D}^2-\tfrac12)^2+\tfrac34>0\) for every unit \(v\), so \(d_1=0\). For the normal bundle we note that \(\zeta=0\) on \(E_v\otimes A\), and we temporarily omit both \eqref{eq:angularconstraint} and \(\zeta\). The map of \((Z,x,y)\) to the other retained outputs and the cofactor outputs goes from a space of dimension \(3r\) to one of dimension \(3r-1\), and by the computation above its kernel is precisely the line spanned by \((-v_D\otimes e_3,-e_2,e_1)\); hence it is surjective. The constraint map \((Z,x,y)\mapsto v_D^*Z+px+qy\) is onto \(U\), has kernel \(E_v\otimes A\), and sends that line onto \(\R\mathbf w(v)\). Imposing \eqref{eq:angularconstraint} therefore leaves the cokernel \(U/\R\mathbf w(v)\). The unit vector \(\mathbf w/\abs{\mathbf w}\) has nonpositive \(e_3\)-component, so its normalized straight-line homotopy to \(-e_3\) is defined, and the quotient is trivial of rank two. The unused summand \(3\mathcal D\) of \(W\) accounts for the other three normals, and \(\triv^2\oplus3\mathcal D\simeq\pi_\Sigma^*N_5\) by Proposition~\ref{prop:characteristic}. \end{proof} This normal identification allows us to move the seed into the prescribed formal class. After the independent base columns are identified, let \(J_{\rm ref}=J_\Sigma\) be the prescribed mixed injection of \(E_v\otimes A\) into \(W\oplus\triv^4\); its normal bundle is \(N_5\). We polar-normalize both injections. By Lemma~\ref{pl:seed} and Proposition~\ref{prop:characteristic} their normal bundles are isometric, so there is an orthogonal bundle automorphism of \(W\oplus\triv^4\) carrying the normalized seed injection to \(J_{\rm ref}\). The restriction of this automorphism to the four extra axes and the original inclusion of these axes are homotopic through isometric four-frames, since the rank-failure codimension in a path is \[ w+1=3r-1>\dim\Sigma+1 . \] We extend this homotopy by orthogonal transport. At its endpoint the resulting automorphism fixes the four axes, and so acts by an orthogonal change of \(W\) alone. Applying this change to the retained seed places its augmented mixed injection in the prescribed class. Along the comparison the four extra axes may move, but at the retained endpoint they are unchanged; the positive polar normalizations are themselves homotopies through injections. If the retained tensor is presented in moment form, the same change conjugates the selfadjoint moment operator and postcomposes the injector. The rows \(C_{\mathrm{base}}\) arise by contracting \(A\) against the reader \[ Q\longrightarrow\mathcal B=L\otimes\bigl(\SymEnd(U)\oplus U^*\bigr),\qquad (h_D,\ell_1,\ell_2)\longmapsto\bigl(\ell_1\otimes\Id_U+\ell_2\otimes S_0,\ \ell_2\otimes e_3^*\bigr), \] extended by zero on \(L\mathcal D\); here \(\SymEnd(U)\) denotes the symmetric endomorphisms of \(U\), and \(\mathcal B\simeq9L\). By rank avoidance we split a copy of \(9L\) off \(D\) and identify it with \(\mathcal B\). Adding to the reader above a small positive multiple of the orthogonal projection onto these coordinates, we obtain a reader \(R_{\mathrm{amb}}:Q\to\mathcal B\) which is onto. Let \(R_v=(R_{v,\mathrm{sym}},R_{v,*})\) be its restriction to \(\mathcal H_v\), extended by zero on the generating line \(L\mathcal D\). Since \(E_v\) is a hyperplane in \(Q\), \(R_v\) has rank nine except on the sphere \(\Sigma_8\) of \(\im R_{\mathrm{amb}}^*\) in \(Q\otimes L\), where it has rank eight. This angular stratum has codimension \((r-1)-8=r-9\) in \(\Sigma\). Define the new cofactor rows by \[ C_0(h\otimes d)=\bigl(R_{v,\mathrm{sym}}(h)d,\ R_{v,*}(h)d\bigr),\qquad h\in\mathcal H_v,\ d\in A, \] using the contraction with \(L\); the formula does not depend on the choice of a local unit of \(L\). If the added multiple is small, the compact margin of the starting augmented injection persists under the change from \(C_{\mathrm{base}}\) to \(C_0\). \begin{lemma}\label{pl:cofactoronto} For every \(v\in\Sigma\), including \(v\in\Sigma_8\), and every complex two-plane \(B\subset A_\C\), the restriction \(C_0|_{\mathcal H_{v,\C}\otimes B}\) is onto \(\C^4\). \end{lemma} \begin{proof} Identify \(\SymEnd(U_\C)\) with \(\Sym^2U_\C\) by the bilinear trace pairing. A dual vector \((u,s)\ne0\) killing the image would force the map \(d\mapsto(u\odot d,\,sd)\), \(d\in B\), to take values in the annihilator of \(\im R_v\), which has dimension at most one. But this map is injective on the two-plane \(B\): if \(s\ne0\) this is clear, and if \(s=0\) then \(u\ne0\) and multiplication by \(u\) is injective in the symmetric algebra. This contradiction proves surjectivity. \end{proof} The retained seed extends to the five labels as a moment tensor. Diagonalize \(P_0\) in \(U\), keeping \(e_2\) as one eigenvector; this splits \(W\) orthogonally into three target blocks, and on each block the seed uses only the corresponding single \(L\)-valued reader on \(A\). Assign distinct real scalar nodes to the three blocks, and interpolate the three readers, extended by zero to the two other labels. Thus the map on each block is a fixed old injector composed with the evaluation, at the node of the block, of one common \(L\)-valued polynomial system of degree at most seven. It agrees with \(T_A\) on \(A\). A slow orthogonal change of the target carries this presentation along by conjugation. In each block we split off trivial lines, by rank avoidance over \(P\), until a summand of rank at most three remains. Let \(X\) be the sum of these residual summands; it has rank at most nine. We assign distinct simple real nodes to the trivial lines, arbitrarily close to the original nodes, and let the normalized rows on \(X\) vary freely in \(\Hom(\mathcal H\otimes V,X)\). The simple rows keep the evaluation form, and each old injector covector varies freely. We call these data the \emph{simple moment and residual-row model}. By compactness, the augmented injection at \(I:A_4\hookrightarrow V\) persists under sufficiently small changes of the nodes, the polynomial system, the injector covectors and the rows on \(X\). \subsection{Polynomial labels} The five \(L\)-valued label polynomials can be chosen homogeneous of degree \(70\), with no common projective zero and with linearly independent fifteen pair products. To the initial polynomials, of degree at most seven, we add \[ \eps\, t^{40+b_i}\beta(t),\qquad (b_i)=(0,1,3,9,27),\qquad \beta(t)=\sum_{j=0}^3 t^j\ell_j , \] where \(\ell_j\) are the projections to \(L\) of the four ambient coordinate vectors. The polynomial \(\beta\) is nonzero in every fibre. In a linear relation among the pair products, the terms that involve an initial polynomial have degree at most \(77\). The remaining terms have degree at least \(80\) and equal \(\eps^2t^{80}\beta^2\) times a linear combination of the powers \(t^{b_i+b_j}\), \(i\leq j\), which are distinct. This proves independence. We homogenize and perturb the coefficients over \(P\). A common real projective zero imposes five independent conditions, and one condition is returned by its location; a nonreal zero imposes ten real conditions and returns its two real coordinates, since \(1\) and \(t\) interpolate an arbitrary complex value with real coefficients in an affine chart. Likewise, proportional evaluation rows at two distinct simple nodes, or the failure of the rows at five such nodes to span at least four dimensions, imposes codimension at least four. This follows from surjectivity of polynomial interpolation at up to five fixed real nodes together with the codimension of the determinantal strata. Since \(P\) has dimension three, finite local parameter sprays over \(P\) avoid all these incidences. The perturbations preserve the positive compact margin at \(I\). Write \(\mathcal P\) for the resulting five-dimensional space of scalar polynomials over \(\C\), in a local unit of \(L\). \begin{lemma}\label{pl:multiplier} Let \(f,g\in\mathcal P\) be linearly independent. If \(u\) is a rational function with \(uf,ug\in\mathcal P\), then \(u\) is constant. The same holds after a subsystem of \(\mathcal P\) is divided by a common scalar factor. \end{lemma} \begin{proof} The identity \(f(ug)=g(uf)\) holds in \(\Sym^2\mathcal P\), by independence of the pair products. Comparing coefficients in a basis of \(\mathcal P\) extending \(f,g\) gives \(uf=\alpha f\) and \(ug=\alpha g\) for a constant \(\alpha\). After division by a common factor \(h\), multiplication by \(h^2\) reduces the claim to the previous case. \end{proof} \subsection{A path of active derivatives} \begin{theorem}\label{thm:angularpath} There is a normalized tensor \(T:\mathcal H\otimes V\to W\) in the simple moment and residual-row model with the following property. Let \(Z_{\mathrm a}\) be any fixed product of four original Hopf spheres. On \(\Sigma\times Z_{\mathrm a}\) there are a smooth map \(a_{\mathrm a}:\Sigma\times Z_{\mathrm a}\to V\) and a smooth path \(D_s:A_4\to V\), \(s\in[0,1]\), of linear maps of rank four, from \(D_0=I\) to the active derivative \(D_1=\dd_{\mathrm a}a_{\mathrm a}\), such that throughout the path \[ T(\Id\otimes D_s)|_{\ker C_0}\quad\text{is injective},\qquad \ker C_0\subset\mathcal H\otimes A. \] The four rows \(C_0\) remain fixed, and the injection has a positive uniform margin on the compact path. \end{theorem} \begin{proof} We use the complexified incidence spaces with fixed simple nodes and with the polynomial system on its open locus. The simple injector rows and all residual rows on \(X\) vary independently. In what follows every generic preparation is chosen within the starting compact injection margin at \(I\). \emph{Step 1: small kernel tensors.} We exclude rank-one kernel tensors and pairs of distinct projective rank-two kernel tensors. A projective rank-one tensor has \(r+3\) parameters. Its scalar entry is a nonzero polynomial of degree \(70\), so it vanishes at no more than \(70\) simple nodes, and the kernel condition imposes at least \(w-70\) equations. A projective rank-two tensor has \(2r+5\) parameters. Two distinct such tensors are linearly independent over the rational functions, by Lemma~\ref{pl:multiplier}. Hence their wedge has a nonzero minor of degree at most \(140\), which bounds the sum of their rank deficits over the simple nodes; on the free rows over \(X\), their constant independence gives two conditions on each row. Thus, after the projective tensor parameters are returned, the codimensions are at least \[ w-70-(r+3),\qquad 2w-140-2(2r+5). \] Both exceed \(r+2=\dim\Sigma\). \emph{Step 2: the pencil on a principal slice.} Fix a principal covector on \(A_4\) which is nonzero on \(A\), and let \(B\subset A\) be its kernel in \(A\), the frozen two-plane. Write \(F:B\to V\) for the frozen injection, \(F_i=F(b_i)\) in a basis of \(B\), and put \(K_B=\ker(C_0|_{\mathcal H\otimes B})\). By Lemma~\ref{pl:cofactoronto}, \(\dim K_B=2r-4\). At an allowed jet, \(T(\Id\otimes F)\) embeds \(K_B\) in \(W\). Choose \(e\in A\) on which the principal covector equals one and a linear lift \(k_e(h)\in\mathcal H\otimes B\) with \(C_0(k_e(h))=C_0(h\otimes e)\). The remaining quotient of \(\ker C_0\) is then tested by the affine pencil \[ h\longmapsto T(h,D(e))-T(\Id\otimes F)k_e(h)\pmod{T(\Id\otimes F)K_B}. \] It maps an \(r\)-dimensional space to a space of dimension \(w-(2r-4)=r+2\). Pass to the highest homogeneous part in \(z_*=D(e)\). For a label outside \(\mathbb P\langle F_1,F_2\rangle\), a rank defect of the homogeneous pencil is witnessed by a tensor \begin{equation}\label{eq:witness} q_*=X_1\otimes F_1+X_2\otimes F_2+h_*\otimes z_*,\qquad h_*\ne0 , \end{equation} in \(\ker T\), with \(C_0(X_1\otimes b_1+X_2\otimes b_2)=0\). By Step 1 there is at most one projective rank-two kernel tensor, and it contributes at most a projective two-plane of labels: its row plane must meet the frozen plane in a line and determines the three-plane containing both, and if it equals the frozen plane it contributes no label away from \(\langle F_1,F_2\rangle\). Consequently a hypersurface of rank defects of the homogeneous pencil requires an irreducible family of rank-three witnesses whose marked moving labels have image of dimension at least three. \emph{Step 3: rank-three families.} For a generic member of such a family the vector polynomial of old vectors has a fixed common scalar divisor \(g\), up to scale. It divides the greatest common divisor of the two frozen polynomials, so there are finitely many possibilities. If \(g\) is nonconstant, the subsystem divisible by \(g\) has dimension exactly four: dominance of the moving label forces at least four, and five would contradict the absence of common projective zeros. We divide this subsystem by \(g\). This deletes at most one simple node, since two distinct evaluation hyperplanes cannot both contain the four-system. Let \(q_0\in\{0,1\}\) be the number of deleted nodes, with \(q_0=0\) in the undivided case. After the division, generic vector polynomials of the family are primitive, that is, they have no common scalar zero. Take \(\ell\) generic members of the family that are linearly independent over the rational functions, where \(\ell\) is the smaller of the rational dimension of the family and \(\lfloor13r/20\rfloor+1\). Let \(d\) be the dimension of their combined old span and \(\delta=3\ell-d\). Let \(R_{\mathrm{rel}}\) be the \(\delta\)-plane of linear relations among the \(3\ell\) separate old columns of the tensors \eqref{eq:witness}. At each projective parameter the \(\ell\) evaluations in the separate blocks span an \(\ell\)-plane; let \(\nu\leq\delta\) be the dimension of the span of its intersections with \(R_{\mathrm{rel}}\) as the parameter varies. Let \(d_1,d_2,d_3\) be the rank deficiencies of the arrays of first, second and third old columns, and put \(j=\max(\nu,d_1,d_2,d_3)\leq\delta\). On the span of dimension \(\nu\), the blockwise wedge with the separate evaluations is generically injective by rational independence. A nonzero \(\nu\)-minor has degree at most \(70\nu\), and its order of vanishing at a node is at least the rank deficiency there, by local row elimination. Hence the total number of equations lost at the distinct simple nondeleted nodes is at most \(140\nu\). On the free rows over \(X\) there are exactly \((\rank X)\ell\) equations, by constant independence. Together the retained rows impose at least \((w-q_0)\ell-140\nu\) equations; the deleted node has no effect on the free rows. For the count of source parameters we require \(R_{\mathrm{rel}}\) to contain the span of \(\nu\) independent lines in the \(\ell\)-planes of separate evaluations. Each line costs at most \(\ell\) parameters (one for the node and \(\ell-1\) for ratios), while the containment imposes \(d\nu\) conditions. Alternatively, we mark a \(d_i\)-dimensional subspace in the fixed \(\ell\)-space of one block slot and require the relation plane to contain it; this saves \(d_i(d-\ell+d_i)\) parameters. The surjective combined column matrix with fixed old span projects submersively to its relation Grassmannian, so we may use the largest of these savings, which is at least \(j(d-\ell)\). We temporarily treat the nine-dimensional reader as independently variable. The four cofactor equations on each member of the tuple impose at least \(4\ell-4(d_1+d_2)\) equations on the reader: we use two diagonal symmetric coordinates for \(X_1\) and \(X_2\), one off-diagonal coordinate involving the third basis direction for \(X_1\), and one scalar coordinate for \(X_1\); in a complex basis we use dual coordinates for the symmetric form. These variations are independent of the retained row variables. For each projective tensor with fixed combined old \(d\)-space, the tensor itself takes \(3d-1\) parameters, the extension of the frozen two-plane to a three-plane takes two, and its marked label takes two. The combined \(d\)-space takes \(d(r-d)\) parameters. In the divided four-system we save one extension parameter per tensor, which covers \(q_0\leq1\). The fibre of generic tuples in the same witness family returns at least \(3\ell\) parameters. Thus the codimension at fixed \(F\), and hence also jointly with \(F\) as a free parameter, is at least \begin{equation}\label{eq:jointcount} \delta(r-d)+2\ell+j(d-\ell-148). \end{equation} For a whole rational basis we have \(j\geq1\). Indeed, if \(\nu=0\), the evaluated vectors are independent everywhere, so the rational coefficients of a polynomial of the same degree in their span are regular at every point of \(\mathbb P^1\), hence constant. If moreover \(d_3=0\), the moving old columns are independent. Modulo the frozen pair, a generic combination with all coefficients nonzero then has rank equal to the dimension of the span of the moving labels, and for a generic basis tuple this span has dimension at least two by dominance; this contradicts the form \eqref{eq:witness}. A whole basis of length one cannot occur: by Lemma~\ref{pl:multiplier} it gives a single projective tensor, whose marked image has dimension at most two. For \(\ell=\lfloor13r/20\rfloor+1\) the bound \eqref{eq:jointcount} exceeds \(r+300\): if its last bracket is negative we use \(j\leq\delta\) and obtain \(2\ell+\delta(r-\ell-148)\), and otherwise the bound is at least \(2\ell\). For a smaller whole basis of length greater than \(300\), a nonnegative bracket gives at least \(r+\ell-148>r+150\), since \(\delta,j\geq1\); a negative bracket gives the same strict bound, by \(\delta>2\ell-148\) and \(\ell\leq13r/20+1\). For length at most \(300\) and \(\delta\geq2\) we use \(d\leq900\) and \(j\leq\delta\) to obtain a bound greater than \(r+150\). \emph{Step 4: a whole basis with one relation.} It remains to treat a whole rational basis of length \(T_0\), \(2\leq T_0\leq300\), with \(\delta=1\), so that \(d=3T_0-1\). Project each separate three-block away from the projection of the one-dimensional relation. This projection factors through the combined old space and leaves at least two independent scalar entries of the original polynomial. In the projection of another member of the family, its rational coefficient in that block multiplies both entries, so it is constant by Lemma~\ref{pl:multiplier}. The family therefore lies in the constant span of the chosen basis. Its unique old relation involves only the moving columns, since independence of these columns would lead to the contradiction with the label quotient obtained above. Write \(b_i\) for the coefficients of the relation, \(g_i\) for the moving labels in the quotient by the frozen pair, and \(g_*\) for the rank-one row direction of a generic combination with coefficients \(c_i\ne0\). Every covector \(\alpha\) annihilating \(g_*\) has \((c_i\alpha(g_i))_i\) proportional to \((b_i)_i\). Hence the \(g_i\) span exactly two dimensions, by dominance and because the evaluation of functionals restricted to the annihilator of \(g_*\) has rank one. Generically \(g_*\) differs projectively from each \(g_i\). Consequently all \(b_i\ne0\), and the projective combination is determined by \([c_i]=[b_i/\alpha(g_i)]\) for a nonzero quotient covector \(\alpha\) killing \(g_*\) in their two-dimensional span. The unmarked tensor family is thus one-dimensional and uses one four-system over the frozen pair. For a generic unmarked tensor, the marked lifts sweep a dense subset of its projective two-plane, since the marked image has dimension at least three. Fix a canonical lift by a linear complement to the frozen pair. Changing the lift by the two frozen labels changes the first two old columns by multiples of the moving old column. Hence, for a generic unmarked tensor, the reader annihilates its moving column tensored with both frozen directions, in addition to the four equations on the canonical frozen columns. The cost for the reader is \[ 7(T_0-1)+4T_0 : \] the two symmetric matrix columns have five independent entries and the scalar reader has two, and the canonical frozen columns are independent modulo the span of the moving columns. At a simple nondeleted node at most one equation is lost, and a loss forces both frozen polynomials to vanish there, because every coefficient of the unique relation is nonzero. Let \(d_0\) be the number of such nodes. Then \(d_0\leq4\), since the rows at five nodes span at least four dimensions in the original five-system. If \(d_0\geq3\), requiring \(F\) to vanish under two of these evaluations imposes four independent additional conditions, because the evaluations are projectively distinct. We combine the retained cost \(wT_0-d_0-q_0T_0\), the reader cost above and the source saving \(d-T_0\). Canonical unmarked tensors take \(3dT_0\) parameters and their fibre returns \(T_0\); the common old span takes \(d(r-d)\); and the four-system returns two parameters in the primitive case and none in the divided case. Substituting \(d=3T_0-1\) gives the codimensions \[ r+9T_0-9-d_0,\qquad r+(9-q_0)T_0-7-d_0 , \] respectively. Both exceed \(r+6\) when \(d_0\leq2\); when \(d_0\geq3\) we add the four extra conditions. \emph{Step 5: the fixed reader and the principal slices.} Apply a small isomorphism \(B_{\mathrm o}\) of the source to \(T\). In the incidence this replaces the effective reader by \(R_vB_{\mathrm o}^{-1}\), and it preserves both types of retained row parameters. Where \(R_v\) has rank nine, this variation is onto all readers. On the rank-eight stratum it varies freely into the given eight-plane, and restricting readers to this plane loses at most \(r\) equations in the long counts. For the count with one relation we freeze the combined old span; the reader equations use at most its \(900\) directions and so lose at most \(900\) equations. Adding the codimension \(r-9\) of the rank-eight stratum, we obtain a bound of at least \(r+6\) in the combined space in both cases, even if \(72\) frame parameters for the eight-plane are returned. In the original variables the reader, and hence \(C_0\), is fixed. The incidence counts involve finitely many tuple lengths and rank charts. The constraints and the changes of scalar system are constructible, since the polynomial degrees are bounded. We take the loci where the fibres have dimension at least \(3\ell\), or at least \(T_0\) in the canonical unmarked count, and then their closures; the bounds above apply to them without change. Take four principal covectors on \(A_4=L\otimes\R^4\), represented locally by \((1,x_i,x_i^2,x_i^3)\) for four distinct real numbers \(x_i\), and identify \(A\) with the first three coordinates. Any three of them restrict to independent covectors on \(A\). Let \(B_i\) be their frozen two-planes and \(B_{ij}=B_i\cap B_j\). Denote by \(Z_i\) the closed exceptional locus for the map on \(B_i\). Project it to the map on \(B_{ij}\) and take the closure of the locus where the fibre has dimension at least four, in both orders; the union \(Z_{ij}\) has codimension at least \(r+5\), since the space of affine extensions has dimension five. Forgetting that line map and taking the closures where the fibres have dimension at least four for the three pairs of any triple, we obtain loci of codimension at least \(r+4\). Small perturbations of the coefficient data and of an initial formal derivative \(D_0\) close to \(I\) avoid all these loci over \(\Sigma\), by stratified parametric transversality. The source sprays span the normalized row coordinates and the coordinates of the source isomorphism, and the sprays of the initial derivative span the coordinates of the frozen maps. The angular variable is the sphere variable itself, including the rank-eight stratum; after pullback, polynomial variations over \(P\) span their coefficients at each angular point. The dimension is at most \(r+2\), below each of the codimension bounds. The strict injection is retained along the straight path from \(I\) to \(D_0\). \emph{Step 6: the \(h\)-principle.} In the open relation we require rank four, injectivity on \(\ker C_0\), and avoidance of \(Z_i\) and \(Z_{ij}\) by the corresponding restrictions. In principal slice \(i\) the homogeneous pencil has no hypersurface of rank defects once its frozen data are fixed. Hence the affine pencil has a rank-defect set of complex codimension at least two: if its maximal minors had a common nonconstant factor, the highest homogeneous part of that factor would divide every maximal minor of top degree \(r\), and these are precisely the minors of the homogeneous pencil. Failure of rank four itself has codimension two in the five free coordinates of the slice, since the other three active columns are independent. The varying exclusions \(Z_j,Z_{jk}\) have dimension at most three in that slice, by the fibre bounds for pairs and triples: for \(Z_j\) the map on \(B_{ij}\) is fixed, and for \(Z_{jk}\) with \(i\) distinct from \(j,k\) the covector \(i\) is nonzero on \(B_{jk}\), so the bound for triples applies. The real trace of each excluded set therefore has codimension at least two. The complement of all excluded sets in the slice is connected and has full convex hull: a generic two-segment path joins any two points, and generic opposite displacements exhibit every point as a midpoint. Thus each principal slice is ample. In Hopf flow coordinates, the four chosen covectors together with transverse coordinate covectors form integrable principal frames; one can use the functions \(\lambda_i(s)\cdot\theta\) and the transverse coordinates, where \(\theta\) denotes the flow coordinates and \(s\) the transverse variables. The transverse slices do not change the active test. We apply the open principal-frame \(h\)-principle \cite[Theorem 3.11]{MA} to a formal first jet with active part \(D_0\). It gives a homotopy within the relation from this formal jet to a holonomic one. Appending it to the straight path from \(I\) and using stationary collars at the endpoints proves the theorem. The positive margin follows by compactness. \end{proof} \subsection{Removing the semidefinite target weight} \label{sec:angularweight} For a basis \(v_1,\dots,v_5\) of \(V\) define the alternating Gram tensor \[ \Omega(T)_{ij}(e,f)=\pair{T_i e}{T_j f}-\pair{T_j e}{T_i f}, \qquad i0\), let \(L_{\mathrm{orb}}\) be its mean over the orbit and put \(q=v/L_{\mathrm{orb}}-1\), which has zero mean on each orbit. For the Hopf action \(\rho_t\) the function \[ b(z)=\int_0^1 t\,q(\rho_t z)\,\dd t \] satisfies \(Yb=q\), by integration by parts. The variable-time map \(z\mapsto\rho_{b(z)}z\) preserves each orbit, has derivative \(1+Yb=v/L_{\mathrm{orb}}>0\) along it, and is homotopic to the identity through \(\rho_{sb(z)}z\); it is therefore a diffeomorphism. We compose the initial map with its inverse and divide by \(L_{\mathrm{orb}}\); the resulting speed is one. The product \(b_0\) of four such maps has partial Gram operator the identity on the four generators. We place it in \(\R^{15}\), using eight coordinates and padding with zeros. Choose \(64\) covectors \(\lambda_\nu\) on \(\R^{15}\) which distinguish every nonzero quadratic and every nonzero cubic form on every real four-plane, that is, no such form vanishes at all the restrictions \(\lambda_\nu|_E\). For cubics the projective marker space of pairs \((E,[p])\) has dimension \(4(15-4)+(20-1)=63\), and for quadratics the corresponding dimension is \(4(15-4)+(10-1)=53\). For a fixed nonzero form each equation \(p(\lambda_\nu)=0\) costs at least one dimension. Consequently the projected exceptional sets in \((\R^{15})^{64}\) have dimension at most \(63+64\cdot14=959<960\), and we choose a tuple in their complement. For a small constant \(\eps_0>0\) put \[ A_0=\Bigl(\Id-\eps_0^2\sum_\nu\lambda_\nu\lambda_\nu^*\Bigr)^{1/2}, \qquad \Phi(x)=\Bigl(A_0x,\, \bigl(\eps_0\cos(\lambda_\nu\cdot x),\,\eps_0\sin(\lambda_\nu\cdot x)\bigr)_{\nu=1}^{64}\Bigr). \] The positive square root is invertible and \((\dd\Phi)^*\dd\Phi=\Id\). Thus \(B_{143}=\Phi\circ b_0\) has unit partial metric in \(\R^{143}\). The four first, ten second and twenty third active columns of \(B_{143}\) are linearly independent. In a linear relation among these columns, the \(A_0\) block first kills the corresponding linear combination of the first, second and third columns of \(b_0\). The tangential component of the circle at frequency \(\lambda_\nu\) then gives the homogeneous cubic in \(\lambda_\nu\cdot Y_ib_0\) with the coefficients of the third columns, and by the choice of frequencies all these coefficients vanish. The radial components then give the analogous quadratic, so the coefficients of the second columns vanish as well. Finally the four first columns of \(b_0\) are independent. This argument uses only the injectivity of the first array of \(b_0\), and so it applies to any formal first, second and third arrays with this property. The four first columns lie in four fixed disjoint planes, and so they are homotopic through independent columns to four constant unit vectors: lift each into an unused independent constant coordinate and then rotate the endpoint. At the same time we interpolate the formal value and the second and third arrays to constants. Through the chain rule this gives a homotopy of the injection of all \(34\) columns, and in particular of the first \(14\) columns, to a constant injection. Orthogonal transport gives a normal identification along this homotopy. Put \(B=(B_{143},0)\) in \(\R^{157}\), with fourteen unused coordinates. Let \(N\) be the orthogonal complement of all first and second active columns of \(B\), and define \[ T_3=\pi_N D_{\mathrm a}^3 B,\qquad R_3=T_3(T_3^*T_3)^{-1}. \] By the independence just proved, \(R_3\) is a smooth right inverse of \(T_3^*\), with zero components in the unused block. The rank-nine bank bundle \(\mathcal B=L\otimes(\SymEnd(U)\oplus U^*)\) admits an isometric inclusion \(\nu\) into the unused \(\R^{14}\) over \(P\), by rank avoidance. For \(\xi=\ell\otimes(M,\rho)\in\mathcal B\) in a local unit \(\ell\) of \(L\), we prescribe a symmetric cubic form \(\mathcal K(\xi)\) on \(A_4=A\oplus L\) by its values on the four types of triples: \[ \begin{aligned} \mathcal K(\xi)|_{A^3}&=0,\\ \mathcal K(\xi)(a,b,\ell)&=\pair{Mu}{v}, &&a=\ell\otimes u,\quad b=\ell\otimes v,\\ \mathcal K(\xi)(a,\ell,\ell)&=\rho(u),\\ \mathcal K(\xi)(\ell,\ell,\ell)&=0 . \end{aligned} \] These formulas use \(A=L\otimes U\) and are compatible with a change of sign of the unit \(\ell\). For sufficiently small \(\kappa>0\) put \[ \eta_\kappa=-\kappa R_3\mathcal K +\nu\Bigl(\Id-\kappa^2(R_3\mathcal K)^*R_3\mathcal K\Bigr)^{1/2}. \] Its two terms lie in orthogonal factors of the target, so \(\eta_\kappa\) is an isometric bank in \(N\), and \(\eta_\kappa^*D_{\mathrm a}^3B=-\kappa\mathcal K^*\). Differentiating the orthogonality of \(\eta_\kappa\) to the first and second columns along \(a\in A\) gives \[ \pair{D_a\eta_\kappa(\xi)}{D_i B}=0,\qquad \pair{D_a\eta_\kappa(\xi)}{D^2_{\mathrm a}B(b,c)}=\kappa\,\mathcal K(\xi)(a,b,c). \] Thus \(D_a\eta_\kappa\) is orthogonal to the four first columns and to the six second columns \(D^2_{\mathrm a}B(b,c)\), \(b,c\in A\). On the remaining four directions of the span of the active full columns we use the coordinates given by the pairings with the columns in \(L\otimes A\) and \(\Sym^2L\), after projecting off the ten selected columns; this is an isomorphism because all fourteen columns are independent. In these coordinates the derivative of the bank has precisely the four rows \(\kappa(Mu,\rho(u))\). The other component, normal to all full columns, is the derivative of \(\eta_\kappa\) itself and is retained. At \(\kappa=0\) the bank is the fixed slow inclusion \(\nu\) into the fourteen unused axes. The smooth path in \(\kappa\) is a path of isometric banks in \(N\), so its orthogonal transport matches the final bank to this slow inclusion. Together with the normal identification of \(B_{143}\), this identifies the \(143\)-dimensional active normal bundle, including the specified bank, with a fixed block. We compose its slow bank input with \(\kappa^{-1}R_v\), in the notation of the exceptional link. The cofactor response of the reference map is then \(C_0\) in the coordinates above. This realizes the bank hypothesis of the preceding subsection in \(157\) active coordinates. The construction does not depend on a trivialization of the line \(L\). \section{Flat matching on the doubled exterior} \label{sec:spectralmatching} Let \(m=n-4\) and let \(B\) be the doubled slow exterior; it is a closed manifold with \(\dim B\leq m+4\). Put \[ k=\lceil1000\sqrt m\,\rceil,\qquad R=m+k,\qquad w=3m+1 . \] The generating bundle is \(\mathcal E=E\oplus\triv^k\), of rank \(R\), while the retained target \(W\) keeps rank \(w\); note that \(w=3R+1-3k\). On an exceptional collar the generating line \(L\mathcal D\) embeds in \(\triv^k\) over the three-dimensional core \(P\) by rank avoidance, since \(k>4\). The angular data of Section~\ref{sec:angular} therefore extend to \(\mathcal E\) on the collar, by zero on the orthogonal complement of \(L\mathcal D\) in \(\triv^k\). In this section we extend these data to an exactly flat tensor on all of \(B\), and we arrange at the same time that its old planes satisfy the following goodness conditions, both for generic full planes and for all reduced planes. \begin{definition}\label{pl:def:goodplanes} Let \(G\colon\mathcal E\otimes V\to W\) be a five-label tensor at a point of \(B\), and let \(S\subset\mathcal E_\C\) be an old plane. A normal two-plane \(N\subset V_\C^*\) is \emph{bad} for \(S\) if the restriction of \(G_\C\) to \(S\otimes\ann N\) is not injective; here \(\ann N\subset V_\C\) is three-dimensional. A \emph{star} is the three-dimensional family of all normal two-planes containing a fixed normal line, its \emph{center}. We call \(S\) \emph{good} if the ordinary skew map of \(G\) on \(S\) is onto and the complex variety of bad normal two-planes for \(S\) has dimension at most four and contains no star. \emph{Reduced old planes} are the complex \((m-1)\)-planes in \(E_\C\oplus\ell_{\rm aux}\), where \(\ell_{\rm aux}\subset\C^k\) is any auxiliary line. \end{definition} \subsection{Source estimates for polynomial moments} \label{sec:polynomialrelations} Throughout this subsection all dimensions and incidences are over \(\C\). Let \(P_d\) be a basepoint-free five-dimensional linear subsystem of \(H^0(\mathbb P^1,\mathcal O(d))\) with independent pair products, that is, the fifteen products of pairs of elements of a basis are linearly independent in \(H^0(\mathbb P^1,\mathcal O(2d))\). We also use four-dimensional subsystems, after division by their common factor; the pair-product property persists for a subsystem and after the division. We write \(q\leq70\) for the degree after any such division. Consider an irreducible projective family of rank-three old tensors with labels in such a system, satisfying one of the following two conditions on its normal image. In five labels, the unique annihilator two-planes of its members form a subvariety of \(\Gr_2(V_\C^*)\) of dimension at least five. In four labels, the annihilator lines of its members are dense in \(\mathbb P^3\). In both cases we assume that the general member is \emph{primitive} as a vector polynomial, meaning that its scalar coordinates have no common zero on \(\mathbb P^1\). Let \(T\) be the rank of the family over the field of rational functions on \(\mathbb P^1\). Choose \(T\) generic members \(u_1,\dots,u_T\) that are linearly independent over this field, and write the three old columns of \(u_i\) in the \(i\)-th summand of \[ \mathcal V=(\C^3)^{\oplus T}. \] Let \(h\) be the dimension of the combined old span of \(u_1,\dots,u_T\). The column map from \(\mathcal V\) onto this span has a kernel \(R_{\rm rel}\), the \emph{relation plane}, of dimension \[ \delta=3T-h . \] All old coefficients of members of the family lie in the combined span: a member is a rational combination of \(u_1,\dots,u_T\), and its projection to the constant quotient by the span vanishes. \begin{lemma}\label{lem:relationgrowth} For such a tuple \((u_1,\dots,u_T)\), \[ \delta\geq 2,\qquad 211\,\delta\geq T^{2/3}. \] More precisely, there are integers \(d_0,e_0\geq0\) with \(T^{2/3}\leq211(d_0+e_0)\leq211\,\delta\) such that, at fixed scalar row spaces of the \(u_i\), the source incidence in \(\Gr_\delta(\mathcal V)\) admits the following codimension savings: \begin{enumerate} \item if \(d_0\geq100\), a saving of at least \(100(h-T-7070)\); \item if \(d_0<100\) and \(e_0\geq100\), a saving of at least \[ d_0(h-T-69d_0-70)+100(h-T-70d_0-69). \] \end{enumerate} These savings are alternatives to the saving obtained from the span of evaluated relations alone. \end{lemma} \begin{proof} \emph{Step 1: a function-space inequality.} Let \(X\) be an irreducible variety, let \(U_b\) be a \(b\)-dimensional space of rational functions on \(X\), and let \(x_1,x_2,x_3\) be algebraically independent rational functions on \(X\). We claim that \begin{equation}\label{pl:eq:loomiswhitney} \dim(U_b+x_1U_b+x_2U_b+x_3U_b)\geq b+b^{2/3}. \end{equation} Choose a smooth point \(p_0\) at which all elements of \(U_b\) and the \(x_i\) are regular after clearing a common denominator, and at which \(z_i=x_i-x_i(p_0)\), \(i=1,2,3\), are part of a system of local parameters; this is possible since the \(x_i\) are algebraically independent. By elimination in power series, choose a basis of \(U_b\) whose leading exponents, for a fixed monomial order, are distinct; they form a set \(A\) with \(\abs{A}=b\). Let \(\pi_i\) be the projection dropping the \(i\)-th coordinate. On each line parallel to the \(i\)-th axis that meets \(A\), multiply the basis element with the topmost exponent on that line by \(z_i\); the leading exponent of the product is the successor of that point, which lies in \((A+e_i)\setminus A\). Functions with distinct leading exponents are independent, so \[ \dim(U_b+x_iU_b)=\dim(U_b+z_iU_b)\geq b+\abs{\pi_iA}. \] The discrete projection inequality \(\abs{A}^2\leq\prod_{i=1}^3\abs{\pi_iA}\) gives \(\max_i\abs{\pi_iA}\geq b^{2/3}\), which proves \eqref{pl:eq:loomiswhitney}. For \(A\subset\Z^3\) the projection inequality follows by applying Cauchy--Schwarz to the sum over \(A\) of the product of the three indicator functions of the pairwise projections. In general apply it on each slice with fixed remaining coordinates and use \(\sum_j\sqrt{a_jb_jc_j}\leq\sqrt{\sum_ja_j\sum_jb_j\sum_jc_j}\). We apply \eqref{pl:eq:loomiswhitney} to a constant basis \(w_1,\dots,w_b\) of any family with the specified normal image, writing its rank-three members again in \(b\) separate three-blocks; the kernel of the resulting column map is the \emph{constant-basis relation plane}. On an affine chart of the family write a member as \(\sum_jc_jw_j\), with linearly independent coefficient functions \(c_1,\dots,c_b\). In five labels choose a chart of \(\Gr_2(V_\C^*)\) in which the normal plane has basis rows \((1,0,x_1,x_2,x_3)\) and \((0,1,y_1,y_2,y_3)\). Since the normal image has dimension at least five, one of the two triples, say \((x_1,x_2,x_3)\), is algebraically independent. The contraction of a member by either row vanishes, so the separate contractions \((c_j\,\alpha(w_j))_j\) by a row \(\alpha\) lie in the constant-basis relation plane. Before the separate five-to-three coefficient projections are applied, the contractions by the first row have coordinates \((c_j,0,x_1c_j,x_2c_j,x_3c_j)\) and span at least \(b+b^{2/3}\) dimensions by \eqref{pl:eq:loomiswhitney}. The contractions by the second row contribute at least \(b\) further dimensions, seen in the second normal slot of each block. The separate projections lose at most \(2b\) dimensions. In four labels the single varying normal is \((1,x_1,x_2,x_3)\), with \(x_1,x_2,x_3\) algebraically independent because the normal image is dense; the same argument uses only this row and loses at most \(b\) under the four-to-three projections. Consequently, in either case, \begin{equation}\label{pl:eq:constantbasis} \dim(\text{constant-basis relation plane})\geq b^{2/3}. \end{equation} \emph{Step 2: torsion on \(\mathbb P^1\).} Return to the rational basis. The primitive columns \(u_i\) give subbundles \(\mathcal O(-q)\to\C^3\otimes\mathcal O\) on \(\mathbb P^1\). Let \(Q_{\rm pol}\) be their combined quotient; the map given blockwise by wedge product with \(u_i\) embeds \(Q_{\rm pol}\) as a subbundle of a sum of copies of \(\mathcal O(q)\). The map \[ R_{\rm rel}\otimes\mathcal O\longrightarrow Q_{\rm pol} \] is generically injective by rational independence, so its cokernel has a torsion part of finite length \(L_{\rm pol}\). At each point the torsion classes are represented by the principal parts of meromorphic \(R_{\rm rel}\)-valued sections with regular image in \(Q_{\rm pol}\), modulo regular relation sections. Let \(D_{\rm pol}\subset R_{\rm rel}\) be the constant span of all coefficients of these principal parts, and put \(d_0=\dim D_{\rm pol}\). By the principal-part description, the torsion length of the map from \(D_{\rm pol}\otimes\mathcal O\) is the same number \(L_{\rm pol}\). A nonzero \(d_0\)-minor of this map, composed with the wedge embedding, is a nonzero section of \(\mathcal O(qd_0)\); its zero divisor bounds the torsion, so \[ L_{\rm pol}\leq qd_0\leq70\,d_0 . \] Let \(K'\) be a constant subspace with \(D_{\rm pol}\subset K'\subset R_{\rm rel}\). The projected basis columns form a subsheaf of \((\mathcal V/K')\otimes\mathcal O\) isomorphic to \(\mathcal O(-q)^T\), and its saturation has quotient of length \(L_{\rm pol}\) over this subsheaf. To see this, lift a meromorphic combination of the columns that is regular modulo \(K'\); its principal part is a meromorphic \(K'\)-section with regular image in \(Q_{\rm pol}\). Conversely, every such relation principal part cancels the poles of a meromorphic combination of the columns, and holomorphic coefficients give back the original subsheaf. The natural maps between the saturated bundles for different \(K'\) are therefore generic isomorphisms between bundles of the same rank \(T\) and degree \(-qT+L_{\rm pol}\); their determinant has no zeros, so they are isomorphisms. After twisting by \(\mathcal O(q)\) the space of sections has dimension \(T+L_{\rm pol}\), by the torsion exact sequence and \(H^1(\mathbb P^1,\mathcal O)=0\). In particular, taking \(K'=D_{\rm pol}\), every member of the rational family lifts to this space of sections. In \(R_{\rm rel}/D_{\rm pol}\) take the span of the contractions of all lifted tensors by the annihilators of their three-dimensional scalar row spaces, using the full degree-\(q\) scalar space. All these contractions project to zero in \(\mathcal V/R_{\rm rel}\). Let \(e_0\) be the dimension of their span; thus \(d_0+e_0\leq\delta\). After passing to the quotient by this span, the lifted rank-three family has exactly its original scalar row spaces, including membership in the designated label subsystem. It contains the projected basis tensors in its closure, so its old span has dimension \(3T-d_0-e_0\), and its constant dimension \(b\) satisfies \(T\leq b\leq T+L_{\rm pol}\). The estimate \eqref{pl:eq:constantbasis} for this family, together with \(b-T\leq L_{\rm pol}\leq70d_0\), gives \[ T^{2/3}\leq b^{2/3}\leq3(b-T)+d_0+e_0\leq211(d_0+e_0)\leq211\,\delta . \] Suppose \(\delta\leq1\). Project each initial three-block away from the projection of \(R_{\rm rel}\) into it. The resulting projection is defined on the old span and retains at least two independent scalar entries of that basis member. For any member of the family, its rational coefficient in that block multiplies both entries into the label subsystem. Such a multiplier is constant: if \(a,b\) are independent and \(fa,fb\) belong to the subsystem, then \(a(fb)=b(fa)\) is an identity in its symmetric square, so by the pair-product property \(fa=ca\) and \(fb=cb\) for a constant \(c\). Hence the constant dimension of the original family is \(T\), and \eqref{pl:eq:constantbasis} gives \(\delta\geq T^{2/3}\). Since \(T\geq1\), this forces \(\delta=T=1\); but a single projective tensor has a single point as normal image, which is incompatible with either condition on the normal image. This proves \(\delta\geq2\). \emph{Step 3: the source savings.} We count the constrained relation planes at fixed basis scalar row spaces. The surjective combined coefficient matrices onto a fixed \(h\)-space map to \(\Gr_\delta(\mathcal V)\), by taking the kernel, with equidimensional fibres. Requiring a given \(j\)-plane to be contained in the relation plane costs \(jh\) conditions. If \(d_0\geq100\), list the principal-part coefficients obtained from local Smith generators chain by chain, from each highest pole order down to the lowest, and stop at the hundredth independent one. The segments already processed and their shorter shifts give independent torsion classes for the current coefficient span: a truncated segment is represented by multiplying its full Smith generator by a power of a local parameter, and the shorter shifts form the basis of its cyclic submodule. Their independence holds in the full torsion module, hence in the module for the current span. Thus, by the same wedge-minor bound, at most \(qj\) positions have been processed when the span has dimension \(j=100\). If \(d_0<100\), the argument applies to the full list with \(j=d_0\). The root locations cost at most \(qj\) parameters. With the preceding negative coefficients fixed, the next one is a known translate of a vector in the \(T\)-dimensional span of the separate block values at the root. Indeed, a meromorphic principal part mapping to a regular section of the quotient can be written locally as a combination of the separate columns plus a regular section, and its new coefficients enter through these independent block values. A step that increases the span therefore costs at most \(T\) parameters, and a step that does not increase it costs at most \(j\). The possible \(j\)-planes thus form a family of dimension at most \(jT+(q-1)j^2+qj\), and requiring their containment in the relation plane saves at least \[ j\bigl(h-T-(q-1)j-q\bigr)\geq j(h-T-69j-70). \] For \(j=100\) this gives the first saving. In the second case, after \(D_{\rm pol}\) has been chosen (with \(j=d_0\)), the projected contractions of any lifted section are parametrized projectively by at most \(T+L_{\rm pol}-1+q\leq T+70d_0+69\) parameters: a projective lifted section and a scalar functional. Mark \(100\) independent such lines in the quotient by \(D_{\rm pol}\) and require their containment in the relation plane as well. With \(q\leq70\) this gives the second saving. These counts use only finite data: degrees, root multiplicities, finite arrays of Laurent coefficients, and finite rank charts. For a fixed proposed span the space of possible lifts in degree \(q\) is defined by the identical vanishing of the \((T+1)\)-minors formed with the basis columns, which is linear in the extra column. All dimension estimates therefore apply to constructible source incidences, before or after projectivizing the members of the tuple. \end{proof} \subsection{Transversality and global flat matching} \begin{proposition}\label{prop:flatmatching} Suppose \(m\geq10^{62}\). Given the angular preparation of Section~\ref{sec:angular} on finitely many disjoint collars, there is a smooth tensor \(G\colon\mathcal E\otimes V\to W\) over \(B\) with \(\Omega(G)=0\) which retains the marked path tests on the prescribed angular source in an inner band of each collar, and such that at every slow point: \begin{enumerate} \item a nonempty Zariski-open set of full complex \(m\)-planes in \(\mathcal E_\C\) is good; \item every reduced old plane is good. \end{enumerate} \end{proposition} \begin{proof} \emph{The outer polynomial system.} We first work in normalized coordinates, allowing a semidefinite target weight as in Section~\ref{sec:angularweight}. After the residual-row collar we have a pure twisted moment, described by a slow selfadjoint operator \(K\) of \(W\), an injector on \(\mathcal E\otimes L\), and the \(L\)-valued five polynomials. We extend these data a little further across a product collar. On the untwisted side we choose a plain five-polynomial system, for example the span of \[ 1+\eps,\quad t+\eps t,\quad t^2+\eps t^3,\quad t^3+\eps t^9,\quad t^4+\eps t^{27} \] for a small generic \(\eps>0\), homogenized to degree \(27\). It is basepoint-free, since the first polynomial vanishes only at \(t=\infty\), where the last one does not. It spans five labels, and its pair products are independent: the highest power of \(\eps\) in a suitable maximal minor comes from the pair products of the monomials \(1,t,t^3,t^9,t^{27}\), which are pairwise distinct because the fifteen sums of two exponents from \(\{0,1,3,9,27\}\) are distinct. The functional calculus of \(K\) realizes either moment by evaluating its polynomial row on the corresponding eigenspace. \medskip\noindent\emph{The spectral bridge.} Extend \(K\) to a smooth selfadjoint endomorphism of \(W\) over \(B\). Perturb it away from the prescribed simple-moment collar so that all eigenvalue multiplicities are at most \(M=\lfloor2\sqrt R\rfloor\). By the dimension count of orthogonal orbits, a repeated eigenvalue of prescribed multiplicity \(j\), with the eigenvalue free, has real codimension \(j(j+1)/2-1\). For \(j>M\) this exceeds \(2R-1\), hence \(\dim B\). The residual eigenblocks on the prepared inner band have rank at most three, so this is a relative transversality condition. By compactness there is a window size such that every window of four times that size contains at most \(M\) eigenvalues, counted with multiplicity; otherwise a limit of windows of vanishing size would give an eigenvalue of multiplicity larger than \(M\). On the transition collar sweep a threshold from below the spectrum to above it, with room at both ends. Let \(u\) be the signed spectral displacement from the threshold divided by the window size. We use a weight \(g(u)\) which vanishes for \(\abs u\leq2\), is positive for \(\abs u>2\), and equals one for \(\abs u\geq3\). The twisted moment is used on the positive side, multiplied by a cutoff supported in \(u>1/2\) and equal to one for \(u\geq3/2\); the plain moment is used with the reflected cutoff. In \(\abs u<1.65\) we allow general normalized rows, by adding an arbitrary tensor multiplied by a scalar function which is positive there and vanishes for \(\abs u\geq1.75\). Every normalized row is then either freely variable or an independent injector row evaluated against exactly one curve. Where the weight is nonzero the weighted tensor is a pure moment, so its alternating curvature vanishes: \(\Omega(gT)=0\). The spectral formulas are smooth through multiple eigenvalues. For instance, write the cutoffs by Fourier inversion, after a smooth compactly supported modification outside a bounded interval containing the spectra. The exponential \(e^{\ii\xi K}\) and its derivatives of fixed order have polynomial bounds in the Fourier parameter \(\xi\), which permits differentiation under the rapidly decaying integral. At the two ends of the collar the construction agrees with the respective pure moments. \medskip\noindent\emph{Ordinaryness of the live rows.} We first check that the rows with nonzero weight are ordinary, that is, that their skew map is onto. For a nonzero skew stress let \(u_0\) be the smaller of its generic contraction ranks on the two evaluation curves that occur. The spanning-curve argument of Section~\ref{sec:angularweight} bounds its coefficient rank by \(3u_0-2\), so the projective stress parameters number at most \((3u_0-2)(R+10)-1\). A nonzero \(u_0\)-minor on each curve has degree at most \(70u_0\), so rank is lost at no more than that many nodes per curve. The zero-weight window costs at most \(M\) further rows. Even if \(140\) additional whole eigenblocks are deleted, the equations cost at least \[ u_0(w-141M)-140u_0^2M . \] Since \(u_0\leq4\), \(k\leq1001\sqrt R\) and \(M\leq2\sqrt R\), the net codimension exceeds \(2R-20000\sqrt R-100>m+4\). Thus ordinaryness holds after the small simultaneous preparations described below. Near the angular band the same construction includes the residual-row bridge of Section~\ref{sec:angularweight}. At most nine residual rows are lost there. Their normalized variables are free while their weight vanishes, and their coefficient cutoff can be taken positive until the single-curve term has become nonzero. All counts in that band use one curve. \medskip\noindent\emph{Low-rank exclusions and a kernel cap.} The following dimension counts give the bound on the bad family for generic full old planes. We exclude rank-one kernel tensors and pairs of distinct projective rank-two kernel tensors in the ambient \(R\)-space. On either curve, rational proportionality of two rank-two tensors would force constant proportionality, by independence of pair products. The bounds on root losses, with two curves and multiplicities at most \(M\), give the codimensions \[ w-140M-(R+3),\qquad 2w-280M-2(2R+5), \] where \(R+3\) and \(2R+5\) are the dimensions of the projective rank-one and rank-two tensors. Free rows impose the full one or two equations. Both codimensions exceed \(m+4\). When these exclusions hold, a single rank-two kernel tensor with two-dimensional label image \(V_2\) makes \(N\) bad only if \(N\subset\ann V_2\); these \(N\) form a projective plane, which has dimension two and contains no star. We also cap the kernel on each three-label plane: its dimension is at most \(a-1\), where \[ a=\lfloor7000\sqrt R\,\rfloor+1 . \] At each node where the three-plane does not vanish identically, an element with zero evaluation has coefficient rank at most two. Hence in an \(a\)-dimensional kernel of tensors the evaluation kernel has dimension at most one, and it is zero except at the roots of the single possible rank-two ray. On each curve there are at most \(70\) such nodes and at most \(70\) basepoint nodes of the three-plane. The corresponding loss over both groups is at most \(140M(a+1)\). Marking the kernel plane and its three labels costs \(a(3R-a)+6\). Consequently the forbidden incidence has codimension at least \[ a(a+1-3k)-140M(a+1)-6>m+4 . \] A freely variable normalized block incurs no loss in this count or in the other kernel-tuple counts. \medskip\noindent\emph{General full old planes.} Suppose that a generic \(m\)-plane in the old source has a bad family of dimension at least five. Then there is an irreducible projective rank-three kernel family \(\Gamma\) in the ambient space, of dimension at least \(3k+5\), whose normal image has dimension at least five. Indeed, the normal two-plane of a rank-three tensor is its unique annihilator. Marking an \(m\)-plane containing its three-dimensional old span takes \(k(m-3)\) parameters, while the \(m\)-planes vary in a family of dimension \(km\). The single rank-two ray and the excluded rank-one tensors cannot account for a five-dimensional normal image. Thus a component that gives at least five dimensions in the generic marked \(m\)-plane has the asserted dimension before marking. For each label curve, the vector polynomials in \(\Gamma\) are generically primitive: a common root would put their normal line through a point of that evaluation curve, and the lines in \(\mathbb P^4\) meeting a curve form a family of dimension at most four. Choose \[ \ell=\min\bigl(T_1,T_2,\lfloor R-300\sqrt R\,\rfloor\bigr) \] generic members, independent over the rational function field on both curves, where \(T_1,T_2\) are the two rational span dimensions; each condition is open on the same irreducible family of tuples. With one curve, use just its dimension. Let \(d\) be the combined old dimension, put \(\delta=3\ell-d\), and let \(\nu\) be the larger of the dimensions of the spans of evaluated relations on the two curves, as in the proof of Theorem~\ref{thm:angularpath}. The relation plane itself is the same for both curves, since both scalar systems are identifications of the same label space. The wedge-minor argument bounds the total spectral loss by \(140M\nu\). The free rows give every equation for each constant-independent tuple. We use the larger relation-span saving \(\nu(d-\ell)\), once. Each rank-three flag in the common old \(d\)-space takes \(3d+5\) parameters, including its projective three-label space, and the common old span takes \(d(R-d)\). Returning the \((3k+5)\ell\) parameters of the fibre of tuples in \(\Gamma\) and using \(w=3R+1-3k\), we obtain the net codimension \[ w\ell-140M\nu-\bigl(d(R-d)+(3d+5)\ell-\nu(d-\ell)-(3k+5)\ell\bigr) =\delta(R-d)+\ell+\nu(d-\ell-140M). \] If \(\ell\) is the truncation \(\lfloor R-300\sqrt R\rfloor\), this is at least \(\ell>m+4\): when the bracket is negative, use \(\nu\leq\delta\) and \(R-\ell-140M>0\). If instead \(\ell\) is a shorter whole basis on one curve, then \(\nu,\delta\geq1\). Otherwise all evaluated vectors would be independent on that curve, and regularity on \(\mathbb P^1\) would force all rational coefficients of any member of the family to be constant. The contraction-drop locus of this constant basis in the dual \(\mathbb P^4\) would then be closed and disjoint from the evaluation curve, and so of dimension at most two by the projective dimension theorem. Every normal line of the family lies in that locus, and such a family of lines has dimension at most two, since two distinct marked points determine a line. This contradicts the dimension of the normal image. Using \(\delta,\nu\geq1\), a nonnegative bracket gives a bound of at least \(R-140M>m+4\); for a negative bracket the same follows from \(\nu\leq\delta\). Thus the projected bad incidences have codimension greater than the slow dimension. At a tensor outside them there is, by the dimension of projective fibres, a Zariski-open set of full old planes with the claimed property. It contains real planes, because the real Grassmannian is Zariski-dense in the complex one. \medskip\noindent\emph{All reduced old planes.} Fix \(\ell_{\rm aux}\) and work in the old \((m+1)\)-space \(E_\C\oplus\ell_{\rm aux}\). The restrictions of the injector and of the arbitrary rows to it are freely variable. Write \(r_0=m+1\) and \(X=\sqrt R\); then \(w=3r_0-2\). If some reduced plane has a bad family of dimension five, the low-rank exclusions leave an irreducible rank-three family with normal image of that dimension, generically primitive on both curves as before. Its members lie in the reduced plane, so all combined old spans counted below have \(h\leq r_0-2\), and no separate mark for the reduced plane is needed. For \(\ell\) entries that are rationally independent on both curves we use the same relation-plane saving and spectral loss. The fibre of tuples has dimension at least \(5\ell\). The resulting count in the \(r_0\)-space is at least \[ B_{\rm red}=\delta(r_0-h)-2\ell+\nu(h-\ell-140M),\qquad \delta=3\ell-h . \] Let \(\ell\) be the smaller rational rank or \(\lfloor4r_0/5\rfloor\), whichever is less. In the truncated case with nonnegative last bracket, \(r_0-h\geq2\) and \(\delta\geq3\ell-r_0+2\) give \(B_{\rm red}\geq4\ell-2r_0\geq6r_0/5-4\). With negative bracket, \(\nu\leq\delta\) and \(\delta\geq3\ell-r_0\) give \(B_{\rm red}\geq\delta(r_0-\ell-140M)-2\ell>2r_0\). For a shorter whole basis we apply Lemma~\ref{lem:relationgrowth} on the controlling curve. If \(h\leq0.9r_0\), then \(B_{\rm red}\geq\delta r_0/10-2\ell\), using \(\nu\leq\delta\) when the bracket is negative; this is greater than \(3r_0\) for \(\delta\geq50\). For \(\delta<50\) the lemma gives \(\ell<2\cdot10^6\) and \(h<6\cdot10^6\), and \(\delta\geq2\) yields the sharper inequality \[ B_{\rm red}\geq2(r_0-6\cdot10^6)-4\cdot10^6-7000M>1.1r_0 . \] If \(h>0.9r_0\), then \(\ell>0.3r_0\) and \(h-\ell>0.1r_0\). For \(\nu\geq50\) the bound exceeds \(-1.6r_0+50(0.1r_0-140M)>1.1r_0\). For \(\nu<50\), the lemma gives \(d_0+e_0>201\), so that one of its two source savings is greater than \(9r_0\). Using that saving in place of \(\nu(h-\ell)\) in \(B_{\rm red}\) gives a bound greater than \(1.1r_0\); this saving uses only the controlling curve. All these bounds exceed \(m+4+k-1\). Letting the auxiliary line vary costs \(k-1\) parameters, so the bound holds simultaneously for all reduced planes. The resulting condition is open, by properness of the projective flag space of auxiliary lines and reduced planes. \medskip\noindent\emph{Regular tuples for stars.} At the weighted tensor prepared so far, which we call the \emph{seed}, consider a rank-three kernel family dominating the star through some center. Its old coefficients may lie anywhere in \(\mathcal E_\C\), while all its labels lie in the four-dimensional annihilator of the center. On each curve divide these labels by their common factor; the removed nodes carry in total \(q_\star\leq140M\) rows. After the division a generic tensor of the family is primitive on both curves, since a common root would force its normal point in \(\mathbb P^3\) to lie on the divided evaluation curve. Rank-one tensors are already absent, and a single rank-two tensor cannot dominate a star. Let \(b\) be the constant span dimension of the rank-three family, and put \(c=w-3R=1-3k\). Suppose \(b\leq R/16\). Choose a whole rational basis of length \(\ell\leq b\) on the curve of smaller rank, independent on both curves. Outside the common basepoint blocks the loss of evaluation equations is at most \(140M\ell\), by the zero divisor of a nonzero \(\ell\)-minor on each curve. A generic projective flag of a rank-three tensor in four labels takes \(3h+2\) parameters in the combined \(h\)-space. Returning \(3\ell\) in the fibre of tuples and allowing four parameters for the shared center, we obtain a net count of at least \[ \delta(R-h)+(c+1-q_\star-140M)\ell-4 . \] By Lemma~\ref{lem:relationgrowth}, \(\delta\geq\max(2,\ell^{2/3}/211)\), and \(h\leq3\ell\leq3R/16\). The negative linear term has magnitude at most \(3700X\ell\), which is less than \(0.01R\delta\). Hence this count exceeds \(m+4\). Suppose \(R/16m+4\). Therefore, outside a set of codimension greater than \(m+4\), every covering rank-three family at the prepared seed has \(b>3R/4\). Let \(l_*=\lfloor13R/20\rfloor+1\). Since \(b-a>l_*\), each covering family has \(l_*\) constant-independent members whose evaluations are independent at every moment node that is not a basepoint; the relevant nonvanishing conditions are finitely many open conditions on the irreducible power of the family. On the zero-weight blocks and on the basepoint blocks we first prescribe arbitrary outputs of the tuple by an unrestricted variation of the normalized tensor, and compensate its linearized curvature on the remaining live blocks by ordinaryness there. This uses the deletion of at most \(140\) whole basepoint blocks besides the zero-weight window, which is within the ordinaryness count above. On each remaining moment block we freely vary its injector to prescribe the remaining outputs; this variation preserves the weighted curvature to first order. Near the angular residual-row bridge the live residual blocks, if used, have pure moment form by construction, so the same argument applies. This proves that the \(l_*\) evaluations are jointly surjective on the tangent space of the normalized weighted-flat relation, with the weight held fixed. We call such a tuple \emph{regular}. \medskip\noindent\emph{Star avoidance.} Star avoidance is a condition in the unrestricted weighted-flat relation. The moment estimates by themselves give only regular tuples, and a star may still be present, so we pass from the prepared seed as follows. Near the seed the derivative of the weighted curvature is onto. Regularity is the open full-rank condition on the stacked vertical evaluation and curvature matrices; it is an algebraic condition with the weight matrix among the parameters. Every nearby star failure admits at least \(3l_*\) dimensions of regular tuples with its fixed center and source. Otherwise the constructible locus of failures with smaller fibres would specialize to the seed along an algebraic trait, the source and center marks being compact and projective. Take the projective closure over the trait of a generic covering component of the family of kernel flags, after a finite extension if needed. This closure is flat over the trait, since its local rings have no torsion over the discrete valuation ring. By properness, a component of the special fibre still covers the whole specialized star, because the generic star is covered and its closure is the star family. At the seed such a component has general rank three, by the low-rank exclusions, and it contains a regular tuple, by the seed calculation. The \(l_*\)-fold fibre product of the flat closure is flat, so its generic fibre is dense, and the open regular minor lifts to the power of the same generic component. An open subset there has dimension at least \(3l_*\), a contradiction. For a marked old \(s\)-plane and a fixed center there are \(3s+2\) parameters per projective flag, and each entry of a regular tuple imposes \(w\) independent equations in the weighted-flat relation. Returning \(3l_*\) fibre dimensions gives a codimension of at least \(l_*(w-3s+1)\) before the marks are forgotten. For \(s=m-1\) we have \(w-3s+1=5\), and allowing all reduced planes and the center costs \((k-1)+2(m-1)+4\); the resulting codimension is at least \[ 5l_*-(k-1)-2(m-1)-4>m+4 . \] For \(s=m\), use first the dimension bound in coefficient--plane pairs, allowing four parameters for the center. If all full planes contain a star, the projection to the coefficients returns the whole Grassmann fibre, of dimension \(km\), and the resulting codimension is at least \(2l_*-4>m+4\). For a fixed tensor the full planes containing a star form a closed set, by properness of the incidence of normal flags. Thus, after simultaneous avoidance, there are nonempty open sets of full planes without a star and with the dimension bound, and their intersection is nonempty. \medskip\noindent\emph{Stratification and transversality.} The counts used in preparing the weighted tensor are finite constructible incidences on its spectral strata. On a stratum of fixed eigenvalue multiplicities we use separate local orthonormal frames of the eigenspaces and the eigenvalues as coordinates; the row-type strata also record the inequalities between these eigenvalues and the threshold. On a row with nonzero moment coefficient we absorb that scalar into the free injector, and on a freely variable row we absorb its moment part into the independent map of that row. These are invertible changes of parameters on the strata, and the counts above are the algebraic counts in the resulting injector and coefficient variables. We first choose the selfadjoint section transverse to the spectral and row-type strata, which is possible by local sprays of symmetric matrices relative to the pure collars. The same local sprays of the injector and free-row variables, applied to a stratification of the constructible incidences and their closures, then give the stated avoidances. Here we use that the real trace of a complex constructible set of complex codimension \(c\) lies in a finite union of submanifolds of real codimension at least \(c\): near a smooth point, the real and imaginary parts of \(c\) local defining functions with independent differentials contain \(c\) functions with independent real differentials. On the pure inner band the counts use only the single twisted curve and the freely variable injector and residual rows. All these perturbations can be taken small there, so the compact marked path tests are preserved. In the two-curve region the twisted data are used only inside the product collar; the line \(L\) is not extended over \(B\). For the subsequent unrestricted perturbation we use local implicit charts of the normalized weighted-flat relation to spray its whole vertical tangent space. On each spectral and row-type stratum the weight has constant rank and inertia, and smooth local congruences identify it with a fixed weight of that rank for the dimension counts, without changing the kernel tests. The smooth spectral section may first be chosen transverse to these strata, as above. The real traces of the closures of the complex constructible failure sets then have codimension greater than the dimension of each slow stratum, and relative stratified transversality gives one arbitrarily small perturbation avoiding all of them. The implicit charts are smooth also where the rank of the weight changes, since the vertical curvature derivative is onto there. Thus the perturbations can be made on the whole product collar and the double, preserving the dimension bounds and the strict angular comparisons. On the inner band one may first prepare the single-curve data on the link and pull them back radially; openness provides a buffer for all the relative perturbations. \medskip\noindent\emph{Removal of the weight.} On the resulting normalized solution \(T_*\) of the weighted equation \(\Omega(gT_*)=0\), ordinaryness of the live rows gives a surjective vertical curvature derivative, by the same half-pairing formula as in Section~\ref{sec:angularweight}. Replace the weight by \(g_\eta=(g+\eta)/(1+\eta)\) with a single small constant \(\eta>0\) on the compact doubled exterior, and solve the curvature equation \(\Omega(g_\eta T)=0\) on the smooth normal complement given by the right inverse, obtaining \(T_\eta\to T_*\); put \(G=g_\eta T_\eta\). By openness this retains the simultaneous tests on full and reduced planes. Since \(g_\eta\) is positive, row scaling by it preserves each kernel test, and ordinaryness transfers by the transpose on variation columns. The marked angular path persists in its inner band throughout the small normalized correction and the invertible scaling, as in Section~\ref{sec:angularweight}. This proves the proposition. \end{proof} \section{Boundary data when the rank-five normal splits} \label{sec:ordinaryboundary} In this section we treat the components of the three-dimensional core \(P\) on which the rank-five normal bundle \(N_5\) of the formal disk columns splits as \[ N_5\simeq N_2\oplus\triv^3 ,\qquad \rank N_2=2 . \] By Proposition~\ref{prop:characteristic} this holds on every component of \(P\) when \(n\) is odd or \(n\equiv0\pmod 4\), and on the components where \(w_1(\det\Hom(K,Q))^3=0\) when \(n\equiv2\pmod4\). We use the core \(P\), the kernel line \(K\), the cokernel \(Q\) of rank \(m+1\), the line \(L=JK\) and its orthogonal complement \(A=L^\perp\subset\triv^4\) from the four-section construction of Section~\ref{sec:fourgeometry}. Choose an isometric monomorphism \(v_0\colon K\to Q\) over \(P\), that is, a section of the unit link bundle, and identify its image with \(L\). Put \(D=Q\ominus L\), a bundle of rank \(m=n-4\). The old and active reference blocks of Section~\ref{sec:referenceproduct} give a slow bundle \(W\) over \(P\) of rank \(w=3m+1\). Along the marked section of the link the complete columns decompose orthogonally as \[ J_2(D)\oplus(A_4\oplus\Sym^2 A)\oplus(D\otimes A). \] The four active columns omitted from the first two summands span \(L\otimes A\oplus\Sym^2L\simeq\triv^4\), so the orthogonal complement of the first two summands is \(W\oplus\triv^4\), while the complement of all columns is \(N_5\simeq N_2\oplus\triv^3\). Hence \((D\otimes A)\oplus N_2\oplus\triv^3\simeq W\oplus\triv^4\), and cancelling three trivial lines over \(P\) gives \[ W\oplus\triv\simeq(D\otimes A)\oplus N_2 . \] Here and below, stable cancellation over \(P\) of bundles of rank greater than three is implemented by a homotopy of a common unit-line inclusion followed by orthogonal transport of its complement. \subsection{A moment with one scalar row} Let \(p_b(t)\), \(b\in V=\R^5\), be the fixed untwisted degree-\(27\) label system introduced in the proof of Proposition~\ref{prop:flatmatching} (Section~\ref{sec:spectralmatching}). It is basepoint-free and has linearly independent pair products. Its evaluation curve \(\mathbb P^1\to\mathbb P^4\) is injective: on the affine part the ratio of the first two coordinates recovers \(t\), and the point at infinity has a distinct image. \begin{proposition}\label{pl:splitseed} Over \(P\) there are a unit section \(e_*\) of \(D\), a covector field \(\ell_*\) on \(A\), and a pure moment \(G_D(h\otimes b)=p_{gb}(B_*)J_D(h)\) on \(D\), with \(B_*\) a smooth selfadjoint endomorphism of \(W\), \(g\) a smooth automorphism of \(V\) arbitrarily close to the identity and \(J_D\colon D\to W\) an injector, such that, with \[ \sigma(h,a)=\pair{h}{e_*}\,\ell_*(a),\qquad h\in D,\quad a\in A, \] the map \((G_D|_{D\otimes A},\sigma)\colon D\otimes A\to W\oplus\triv\) is injective with normal bundle isomorphic to \(N_2\). Moreover, on \(Q=D\oplus L\) adjoin the three rows \(h\otimes a\mapsto(\pi_Lh)\otimes a\) and use \(G_D\) on \(\pi_Dh\). At the included derivative \(I\colon A_4\hookrightarrow V\) the resulting map \(Q\otimes A\to W\oplus(L\otimes A)\oplus\triv\) is injective with normal \(N_2\), and its restriction to \(E_v\otimes A\) over the link is injective with normal \(N_5\). Both statements persist under sufficiently small variations of \(G_D\) with \(\sigma\) fixed. \end{proposition} \begin{proof} \emph{The spectrum of \(B_*\).} Split off a unit section \(e_*\), so that \(D=D'\oplus\R e_*\). Take a generic smooth real selfadjoint endomorphism \(B_*\) of \(W\) and a smooth label automorphism \(g\) close to the identity. A triple eigenvalue has codimension five and two distinct double eigenvalues have codimension four; since \(\dim P=3\), the eigenvalues of \(B_*\) have multiplicity at most two, with at most one double eigenvalue in each fiber. Write \(x(t)\in\R^3\) for the three evaluations of \(p_{gb}(t)\) on a local basis of \(A\). At a simple eigenvalue the condition \(x(t)=0\) imposes three conditions on the label three-plane \(g(A)\). The evaluations at two distinct eigenvalues are independent, so simultaneous vanishing costs six conditions, and vanishing at a double eigenvalue costs at least five, including the codimension of the multiplicity. Hence only simple eigenlines can be \emph{inactive} on \(A\), that is, have \(x(t)=0\); there is at most one in each fiber, and the points of \(P\) carrying one form a finite transverse set \(P_{\rm inact}\). \emph{Injectivity on \(D'\otimes A\).} For a generic injector \(J'\colon D'\to W\) the moment \(G'(h'\otimes a)=p_{ga}(B_*)J'(h')\) is injective on \(D'\otimes A\). Indeed, a projective rank-three kernel witness has \(3(m-1)-1\) parameters and imposes one equation on each spectral row except the inactive one, that is, at least \(3m\) equations; this leaves codimension at least four. A witness of rank one or two is a nonzero vector polynomial of degree at most \(27\), which vanishes on at most \(54\) spectral rows; its projective parameter count is at most \(2m-1\), so the codimension is at least \(m-52>3\). Let \(C_4\) be the rank-four orthogonal complement of \(G'(D'\otimes A)\) in \(W\). At a point of \(P_{\rm inact}\) the inactive eigenline \(U_0\) is orthogonal to \(G'(D'\otimes A)\), hence \(U_0\subset C_4\). \emph{The projected evaluation.} We claim that the map \[ \Psi\colon W\to\Hom(A,C_4),\qquad \Psi(u)(a)=\pi_{C_4}\,p_{ga}(B_*)u, \] is onto off \(P_{\rm inact}\), and onto \(\Hom(A,C_4\cap U_0^\perp)\) at each point of \(P_{\rm inact}\). A cokernel witness is a triple of vectors in \(C_4\); in an eigenbasis of \(B_*\) it has entries \(y_i\in\R^3\) at the spectral rows \(t_i\), and it satisfies \(x(t_i)\cdot y_i=0\). Thus the matrices \(x(t_i)y_i^{\mathsf t}\) are trace-free. Let \(S\) be their span, of dimension \(b\leq8\). Orthogonality of the witness to \(G'(D'\otimes A)\) imposes \((m-1)b\) linear conditions on \(J'\), and marking \(S\) costs at most \(16\) parameters. With \(S\) fixed, let \(d\leq2\) be the generic dimension of the space of \(y\) with \(x(t)\cdot y=0\) and \(x(t)y^{\mathsf t}\in S\). A nonzero maximal minor of these constraints has degree at most \(81\), so the witness dimension is at most \(wd+486\). We bound \(b\) in terms of \(d\). \begin{itemize} \item If \(d=2\), every trace-free matrix orthogonal to \(S\) preserves every evaluation line \(\R x(t)\). Such a matrix is scalar: a finite union of its eigenspaces cannot contain the spanning polynomial curve unless one eigenspace contains the curve. Hence it vanishes and \(b=8\). \item If \(d=1\), a rational generator \(y(t)\) cannot have constant projective direction, since \(x(t)\) spans \(\R^3\) and \(x(t)\cdot y(t)=0\). The entries of \(y(t)\) therefore span at least two rational functions, and those of \(x(t)\) span three. Their pair products span at least four: choose bases with distinct vanishing orders at a regular point and use that the sumset of two increasing integer lists of lengths \(a_1,a_2\) has at least \(a_1+a_2-1\) elements. Hence \(b\geq4\). \item If \(d=0\) and \(b>0\), we use \(b\geq1\). \end{itemize} In each case \((m-1)b-wd-502>3\). These fiberwise incidence counts give simultaneous real transversality over \(P\) for a generic \(J'\). The remaining witnesses, with \(b=0\), have \(y_i=0\) whenever \(x(t_i)\ne0\); they are exactly the triples supported on an inactive row. This proves the claim; no global eigenbasis is needed. \emph{The last column.} Since \(W=(D'\otimes A)\oplus C_4\), the identity \(W\oplus\triv\simeq(D\otimes A)\oplus N_2\) shows that \(C_4\oplus\triv\) and \(A\oplus N_2\) are stably isomorphic; as these bundles have rank five over the three-dimensional base \(P\), they are isometric. Choose an isometric inclusion \(i=(i_{C_4},i_{\triv})\colon A\to C_4\oplus\triv\) with normal \(N_2\), and put \(\ell_*=i_{\triv}\). At each point of \(P_{\rm inact}\) a local orthogonal isotopy moves \(i\) so that \(i_{C_4}(A)\subset C_4\cap U_0^\perp\) at that point; the isotopy can be supported in a small ball, since the special orthogonal group of a five-dimensional space acts transitively on its orthonormal three-frames. In that ball choose a local section \(u\) of \(W\) with \(\Psi(u)=i_{C_4}\) at the center, and blend \(i_{C_4}\) to \(\Psi(u)\) in a smaller ball while keeping \(i_{\triv}=\ell_*\). The modified map remains injective by closeness to its value at the center. Away from the smaller balls solve \(\Psi(u)=i_{C_4}\) using the surjectivity of \(\Psi\), and patch the local solutions by a partition of unity; since the equation is linear in \(u\), the patched section still satisfies it. Put \(J_D|_{D'}=J'\) and \(J_D(e_*)=u\). After eliminating the component of the last three columns lying in \(G'(D'\otimes A)\), the last three columns together with the scalar row give the inclusion \((i_{C_4},\ell_*)\). Hence \((G_D|_{D\otimes A},\sigma)\) is injective with normal \(N_2\). \emph{Extension to \(Q\) and to the link.} Adjoining the rows \((\pi_Lh)\otimes a\) adds the trivial three-plane \(L\otimes A\) to both source and target, so the map on \(Q\otimes A\) is injective with normal \(N_2\). Over the link, \(Q=E_v\oplus vK\), and restriction to \(E_v\otimes A\) removes the three dimensions \(vK\otimes A\simeq L\otimes A\), a trivial three-plane. The normal of the restricted inclusion is therefore \(N_2\oplus\triv^3\simeq N_5\). Both statements are open conditions over the compact base and persist under small variations of \(G_D\) with \(\sigma\) fixed. \end{proof} \subsection{Simultaneous rank and incidence preparation} \label{sec:ordinaryselection} Embed \(K\) in the trivial bundle \(\triv^k\) over \(P\) by rank avoidance, and let \(K^\perp\) be its orthogonal complement. Over the link, the restriction of \(\mathcal E=E\oplus\triv^k\) is identified isometrically with the pull-back of the slow bundle \(Q\oplus K^\perp\) of rank \(R\), by sending the auxiliary copy of \(K\) to \(vK\), that is, \((e_v,k_0,k_\perp)\mapsto(e_v+v(k_0),k_\perp)\). Extend the moment \(G_D\) of Proposition~\ref{pl:splitseed} by zero on \(L\oplus K^\perp\), and vary its injector and spectrum on this rank-\(R\) bundle. \begin{lemma}\label{pl:ordinaryprep} Arbitrarily close to the moment of Proposition~\ref{pl:splitseed} there is a prepared pure moment, together with an arbitrarily short path in the ordinary flat locus over \(P\) to a flat tensor \(G\) on \((Q\oplus K^\perp)\otimes V\), such that: \begin{enumerate} \item the full and reduced goodness of Proposition~\ref{prop:flatmatching} hold at every point of the link, with the same tensors over \(P\); \item \(G(\cdot,z)\) is injective on \(Q\oplus K^\perp\); \item on \(D\), the bad normal family \(\mathcal B\) has dimension at most four, contains no full star, and contains no \(\Gr_2(\ann b)\) for a projective label \(b\). \end{enumerate} These properties persist under subsequent small variations in the ordinary flat locus, and the preparation retains the augmented injection of Proposition~\ref{pl:splitseed} and its normal class. \end{lemma} \begin{proof} We carry out the pure-moment preparation of Proposition~\ref{prop:flatmatching} over \(P\). For reduced sources we mark both the angular point \(v\) and the auxiliary line; the complex angular quadric adds \(m\) parameters and the auxiliary line \(k-1\). The seed bound \(1.1(m+1)\) for reduced incidences exceeds \(m+k-1+3\), and the reduced star bound becomes \[ 5l_*-(k-1)-2(m-1)-4-m>3 . \] The generic full-plane condition is invariant under the orthogonal change of source at \(v\). Thus the prepared seed, and its subsequent unrestricted flat perturbation, satisfy the exterior tests at every point of the link. This gives~(i). Part~(ii) follows from the rank-one exclusion, and is an open condition. For~(iii), avoidance of full stars on the fixed bundle \(D\) has codimension \(2l_*-4>3\) in the same unrestricted flat perturbation as in Proposition~\ref{prop:flatmatching}. The other two conditions are imposed at the pure moment seed, by the following additional counts. For a bad family of dimension at least five, the rank-three witnesses are generically primitive on the evaluation curve, because projective lines meeting that curve form a family of dimension at most four. If the bad family contains all normals of three-label planes through one label \(b\), a common root of a rank-three witness must be one of the finitely many roots of \(p_{gb}\), and the three-planes through \(b\) in such an evaluation hyperplane form only a two-dimensional family; so the witnesses are generically primitive in this case as well. The possible single rank-two ray cannot dominate either condition. We now use the relation-plane count of Proposition~\ref{prop:flatmatching}, with source of dimension \(m\). For \(\ell\) rationally independent witnesses let \(d\) be their common old source dimension, put \(\delta=3\ell-d\), and let \(\nu\) be the dimension of the span of evaluation relations. The two codimensions are at least \[ \delta(m-d)+\ell+\nu(d-\ell-140M),\qquad \delta(m-d)+2\ell-4+\nu(d-\ell-140M). \] In the first count the parameter count per flag is \(3d+5\) and the tuple fiber returns \(5\ell\). In the second these numbers are \(3d+3\) and \(4\ell\), with four parameters for the marked label. The scalar degree and the source relation fee are as before. Take \(\ell\) to be the smaller of the rational rank and \(\lfloor m/2\rfloor+1\). In the truncated case both bounds exceed three, using \(\nu\leq\delta\) when the bracket is negative. If \(\ell\) is the rational rank, then \(\nu,\delta\geq1\). Indeed, if the evaluations of the basis were independent everywhere, all rational coefficients of the relations would be regular on the projective line, hence constant. Their contraction-drop locus would then be disjoint from the evaluation curve, so of dimension at most two, and so would its family of contained lines; this contradicts either prescribed normal image. With \(\nu,\delta\geq1\) both bounds again exceed three, since \(m-\ell-140M>0\). Both avoidances are open by properness of the incidences. Finally, we keep the prepared moment together with the short flat path to the resulting tensor. Such a path in the ordinary part of the curvature equation is obtained in its smooth local implicit charts, using the smoothly varying right-inverse complement over the compact base. All preparations can be taken arbitrarily small, so they retain the augmented injection and the source normal class. \end{proof} \subsection{The active map on a split-normal core} \label{sec:ordinaryactivity} All projective incidences in this subsection are over \(\C\); a normal two-plane in \(V_\C^*\) is viewed as a line in \(\mathbb P(V_\C^*)\simeq\mathbb P^4\). For the bad normal family \(\mathcal B\) of \(G\) on \(D\) put \[ \mathcal E_{\rm bad}=\{N:\dim\{M\in\mathcal B:M\cap N\ne0\}\geq3\}, \qquad \mathcal C_0=\mathcal B\cup\mathcal E_{\rm bad}. \] For a closed family \(\mathcal C'\) of normal planes let \(\mathcal L(\mathcal C')\) be the set of projective labels \(b\) satisfying one of \[ \begin{gathered} \dim\{M\in\mathcal C':M\subset\ann b\}\geq3,\\ \{M:c\subset M\subset\ann b\}\subset\mathcal C' \quad\text{for some line }c\subset\ann b . \end{gathered} \] The family in the second condition is a \emph{restricted star}. Set \(\mathcal L_0=\mathcal L(\mathcal C_0)\), \(\mathcal F=\{N:\mathbb P(\ann N)\subset\mathcal L_0\}\) and \(\mathcal C=\mathcal C_0\cup\mathcal F\). \begin{lemma}\label{pl:derivedloci} Suppose \(\mathcal B\) satisfies Lemma~\ref{pl:ordinaryprep}(iii). Then the sets above are closed, also in coefficient families, \(\dim\mathcal E_{\rm bad}\leq2\), \(\dim\mathcal L_0\leq2\), and \(\mathcal F\) is finite. For every \(N\) outside \(\mathcal E_{\rm bad}\cup\mathcal F\), \[ \dim\{M\in\mathcal C:M\cap N\ne0\}\leq2,\qquad \mathbb P(\ann N)\not\subset\mathcal L_0 . \] \end{lemma} \begin{proof} We first bound \(\mathcal E_{\rm bad}\). Let \(\mathcal B_i\) be an irreducible component of \(\mathcal B\) of dimension \(d\leq4\). Its point-line incidence has dimension \(d+1\), and by the star exclusion all its point fibers have dimension at most two. Let \(S_2\subset\mathbb P^4\) be the locus of points whose fiber has dimension two. If \(d=4\) and the projection to \(\mathbb P^4\) is dominant, the generic fiber has dimension one and \(\dim S_2\leq2\) by upper semicontinuity. If \(d=4\) and the image \(Y\) has dimension three, the lines through a general point \(p\) of \(Y\) form a two-dimensional closed subset of the projective plane of tangent directions \(\mathbb P(T_pY)\), hence fill it. Then \(Y\) is a hyperplane and \(\mathcal B_i\) its whole line Grassmannian, which is excluded. An image of smaller dimension would violate the fiber bound. If \(d\leq3\), the generic fiber cannot have dimension two: the image would then have dimension at most two, and at a general smooth point it admits at most a one-dimensional family of tangent line directions. Hence \(\dim S_2\leq d-2\). In all cases \(\dim S_2\leq2\). Suppose a normal line \(N\) meets a family of lines of \(\mathcal B_i\) of dimension at least three. Discarding \(N\) itself, every remaining line meets \(N\) in one point, so by the fiber bound \(N\) is contained in the closure of \(S_2\). Lines in a projective surface form a family of dimension at most two, by the same incidence and tangent argument. Hence \(\dim\mathcal E_{\rm bad}\leq2\) on the whole Grassmannian of normal planes. In particular \(\mathcal C_0\) has dimension at most four and still contains no full star. Next let \(\mathcal C'\) be any closed family of normal planes of dimension at most four without a full star; we show \(\dim\mathcal L(\mathcal C')\leq2\). For the first condition, a component of dimension \(d\) has a label incidence \(\{(b,M):M\subset\ann b\}\) of dimension \(d+2\), since \(\mathbb P(\ann M)\) is a projective plane. If \(d\leq3\), subtracting three for the fiber gives the bound. If \(d=4\), the incidence is dominant over the label space \(\mathbb P^4\). Otherwise its image, of dimension at most three, would contain a four-dimensional family of projective planes; through a general smooth point the planes in the image give at most two dimensions of tangent planes, whereas the incidence requires at least three. Thus the generic fiber has dimension two, and the proper closed locus of larger fibers has image of dimension at most two. For the second condition, fix a center \(c\). The normals through \(c\) in \(\mathcal C'\) form a proper closed subset of \(\mathbb P^3\), with finitely many plane components. A restricted star is such a plane and determines the label; its center lies in the locus \(S_2\) of two-dimensional fibers estimated above. The only possible four-dimensional component with a larger center locus is the line Grassmannian of a hyperplane, which contributes only the label of that hyperplane. This proves the bound for both conditions. The incidences are closed in coefficient families by properness and the fiber-dimension theorem. Containment of a restricted star and containment of a plane are closed conditions, because the complement of the incidence has open projection in the corresponding projective-space bundles. A projective plane in a set of dimension two is one of its irreducible components, so \(\mathcal F\) is finite. Adding the finite closed set \(\mathcal F\) changes neither condition in the definition of \(\mathcal L\), so \(\mathcal L(\mathcal C)=\mathcal L_0\), and the two displayed properties follow from the definitions of \(\mathcal E_{\rm bad}\) and \(\mathcal F\). \end{proof} For a three-plane \(T(A)\subset V\), the map \(G_D(\Id\otimes T|_A)\) is injective on \(D\otimes A\) precisely when \(\ann(T(A)_\C)\notin\mathcal B\). Consider the open first-order relation on active derivatives \(T\colon A_4\to V\) given by rank four and injectivity of \[ \bigl(G_D(\Id\otimes T|_A),\sigma\bigr)\quad\text{on }D\otimes A , \] where the scalar row uses the original input \(a\in A\). We add the avoidance condition \(\ann(T(A)_\C)\notin\mathcal E_{\rm bad}\cup\mathcal F\), and call the resulting open relation \(\mathcal R_{\rm ord}\). \begin{proposition}\label{pl:ordinaryactivity} After an arbitrarily small change of \(G\) in the ordinary flat locus over \(P\), there are an active map on the product of \(P\) with the four original Hopf factors, with values arbitrarily close to \(z\), and a compact path of rank-four maps \(A_4\to V\) from the included derivative \(I\) to its Hopf derivative, along which the augmented test on \(D\otimes A\) is injective. Along the same path the test on \(Q\otimes A\), with the three rows \((\pi_Lh)\otimes a\), and its restriction to \(E_v\otimes A\) over the link are injective. \end{proposition} \begin{proof} \emph{Admission of the included derivative.} The included derivative can be brought into the avoidance condition by a small change of \(G\) in the ordinary flat locus and a small formal change of \(T\) over \(P\). Indeed, \(T\mapsto\ann(T(A)_\C)\) maps rank-four maps submersively onto the six-dimensional Grassmannian of normal planes, and the avoided family has dimension at most two for every coefficient by Lemma~\ref{pl:derivedloci}. Stratifying the algebraic incidence in these joint parameters and using the regular flat coordinate charts gives codimension at least four, which exceeds \(\dim P=3\). The augmented comparison at \(I\), the ordinary and exterior rank tests, and the absence of the full label and normal families of Lemma~\ref{pl:ordinaryprep} are open, so they persist under sufficiently small such changes. With the coefficient now fixed, the short path from \(I\) to the admitted \(T\) stays in the augmented comparison. \emph{Avoidance templates.} We verify the hypotheses of the avoidance template \cite[Definition~4.5]{MA} for \(\mathcal R_{\rm ord}\). For a finite list of principal covectors \(\tau_j\in A_4^*\) we require that any at most three of their restrictions to \(A\) be linearly independent. For \(J\) of size one or two put \[ G_J=T\Bigl(A\cap\bigcap_{j\in J}\ker\tau_j\Bigr). \] For a single index we require that only finitely many members of \(\mathcal C\) annihilate \(G_J\); for two indices we require that the label point \(G_J\) lie outside \(\mathcal L_0\). These open conditions are coherent under deletion of indices. For density, fix \(N=\ann(T(A)_\C)\) outside \(\mathcal E_{\rm bad}\cup\mathcal F\). By Lemma~\ref{pl:derivedloci} the members of \(\mathcal C\) meeting \(N\) form a family of dimension at most two. Each such member \(M\ne N\) annihilates exactly one two-plane of \(T(A)\), namely \(\ann(M+N)\), while \(N\) annihilates all of them and contributes one point to every fiber. Thus a generic two-plane in \(T(A)\) is annihilated by only finitely many members of \(\mathcal C\). Since \(\mathbb P(T(A)_\C)=\mathbb P(\ann N)\not\subset\mathcal L_0\), a generic line in \(T(A)\) avoids \(\mathcal L_0\). The maps \(T\mapsto G_J\) are submersive for independent covector restrictions, so any finite list of these conditions is dense. \emph{Ampleness.} Fix \(i\) and a vector \(e_i\) with \(\tau_i(e_i)=1\), and consider the principal slice in which \(T(e_i)\in V\) varies with \(T\) fixed on \(\ker\tau_i\). Rank failure has codimension two in \(V\). The frozen two-plane \(G_{\{i\}}\) is annihilated by finitely many members of \(\mathcal C\), and each possible failing three-plane \(T(A)\) is annihilated by one of them, which is a codimension-two condition on \(T(e_i)\). This also bounds the failure of the augmented test, since injectivity of \(G_D(\Id\otimes T|_A)\) alone implies it. The conditions with \(i\in J\) are fixed in the slice. For a one-index condition \(\{j\}\), \(j\ne i\), the moving two-plane \(G_{\{j\}}\) extends the fixed label line \(b=G_{\{i,j\}}\), which lies outside \(\mathcal L_0\); the extension is a point of \(\mathbb P(V/b)\simeq\mathbb P^3\). The members of \(\mathcal C\) annihilating \(b\) give a family of projective lines there of dimension at most two, containing no restricted star. The locus of points lying on a positive-dimensional family of these lines has dimension at most one. For a two-dimensional component whose lines sweep out \(\mathbb P^3\), this follows from its three-dimensional incidence by dropping one dimension and subtracting one. If the lines of a two-dimensional component sweep out only a surface, then at a general point their directions fill the tangent plane, so the surface is a plane and the component consists of all its lines; under annihilator duality this is a restricted star, which is excluded. A family of smaller dimension gives the bound directly. For a two-index condition \(\{j,l\}\) with \(i\notin\{j,l\}\), the moving label line ranges freely over \(\mathbb P^4\) and the excluded labels have dimension at most two. The maps from the moving vector to these projective spaces are submersive away from rank failure. Hence all failures lie in a finite union of sets of real codimension at least two. The allowed open slice is therefore connected, as any two points are joined by a path of two generic segments, and its convex hull is all of \(V\), by opposite displacements. This verifies ampleness for every finite configuration. \emph{Conclusion.} We apply the foliated lifting and the relative \(C^0\)-close \(h\)-principle for avoidance \cite[Lemma~4.7 and Theorem~5.1]{MA}. It gives an active map on the four original Hopf factors and a compact formal path in \(\mathcal R_{\rm ord}\) from the admitted derivative to its Hopf derivative, with the augmented test on \(D\otimes A\) throughout. We prepend the short admission path. The test on \(Q\otimes A\) is injective along the path, with the three rows \((\pi_Lh)\otimes a\) on the original inputs \(a\in A\): their vanishing forces a kernel tensor to lie in \(D\otimes A\). The same holds on \(E_v\otimes A\). As in the exceptional case (Section~\ref{sec:angular}), precomposition with normalized powers on the active factors and the inverse scaling of the values place the value path arbitrarily close to \(z\) without changing the derivative tests. \end{proof} \section{Integration of the mixed coefficients and exact metric repair} \label{sec:mixedcompletion} \subsection{The old primitive on the exceptional link} On the exceptional link the old Hopf coefficients have rank \(r-1\) and source \(E_v\), while the slow bundle \(\mathcal H\) has rank \(r\). We first produce a map \(u\) with values in \(\mathcal H\) whose old derivative \(d_Eu\colon E_v\to\mathcal H\) is a monomorphism homotopic, through monomorphisms, to the inclusion \(E_v\to\mathcal H\). Consider the first-order open relation on sections requiring \(d_Eu\) to have rank \(r-1\). In an effective principal slice, transverse to an old hyperplane \(H_0\subset E_v\), the restriction \(d_Eu|_{H_0}\) is fixed and the remaining column must avoid the affine \((r-2)\)-plane \(d_Eu(H_0)\) in an \(r\)-dimensional space; this has codimension two. On the other slices the relation is constant. Hence the relation is ample, and the relative open ample theorem \cite{EM}, applied to the formal inclusion, gives a section \(u\) together with a homotopy of monomorphisms from the inclusion to \(d_Eu\). Using a finite chart partition, we approximate \(u\) in \(C^1\) by polynomials in the ambient sphere coordinates with smooth slow coefficients. Injectivity and the augmented boundary comparison persist along the short linear path. No global frame of \(E_v\) or of \(\mathcal H\) is needed. Let \(G\) be the exactly flat tensor of Section~\ref{sec:angularweight}, with the slow placement of the bank, let \(v_0\colon V\to W\) be a fixed slow frame of rank five, and put \(v=v_0+\lambda G(u,\cdot)\) with \(\lambda>0\) small. Since \(v_0\) and \(G\) are slow, \(d_Ev=\lambda G(d_Eu,\cdot)\) and \[ v^*d_Ev=\lambda\, v_0^*G(d_Eu,\cdot)+\lambda^2\,G(u,\cdot)^*G(d_Eu,\cdot). \] The last matrix is symmetric by the alternating Gram identity expressing the flatness of \(G\). Consequently the skew part of \(v^*d_Ev\) has the primitive \[ P=\skewp\bigl(v_0^*(v-v_0)\bigr),\qquad d_EP=\skewp(v^*d_Ev). \] All coefficient derivatives here keep the slow quantities fixed. The mixed tensor is \(d_Ev\), and the four bank rows in the formulas of Section~\ref{sec:physicalactive} are evaluated on \(d_Eu\). The homotopy from the inclusion to \(d_Eu\) preserves the augmented comparison: the comparison is precomposed with a monomorphism on the first factor, and the paired comparison is injective on all of \(\mathcal H\otimes A\). This gives the old variable on the exceptional link, in the original target and with zero passive term. \subsection{Derivatives of the old map on the exterior} \label{sec:actualold} Let \(G\) be the exterior tensor of Proposition~\ref{prop:flatmatching}, with the relative ordinary preparation of Section~\ref{sec:ordinaryselection} where it applies. On the collar band we use the generator and the placement of the bank constructed in Section~\ref{sec:pairedtest}. For the generating section \(u\) of \(\mathcal E\) on \(B\times\mathcal O_{\rm old}\) we impose two conditions: \(d_Eu\) is injective, and its projection to \(E\) has rank at least \(m-1\). In an effective principal slice the first condition excludes an affine subspace of codimension \((m+k)-(m-1)=k+1\). The second imposes no condition if the other projected columns have rank \(m-1\), and excludes an affine subspace of codimension two if they have rank \(m-2\). In either case the complement is connected and has full convex hull, and the formal inclusion satisfies the relation. On the inner band the prescribed generator and its derivative homotopy take values in \(E_v\) plus a single generating line, and there both rank conditions follow from injectivity. The relative open ample theorem therefore realizes the formal extension by a section \(u\). On the distant exterior we multiply the auxiliary coordinates of \(u\) by a small positive constant. Let the multiplier tend to zero. Any limit of the complex old image planes contains an \((m-1)\)-dimensional subspace of \(E_\C\): choose \(m-1\) domain vectors whose projected images are independent at the limiting old point. Hence the limiting plane lies in \(E_\C\) plus at most one auxiliary line, and all of its hyperplanes are good by Proposition~\ref{prop:flatmatching}. By compactness of the flag bundle and openness of goodness, one small positive multiplier works on the whole compact product. We interpolate to the multiplier one inside the product band, where the generator still takes values in \(E_v\) plus its generating line; for every positive multiplier all hyperplanes there are reduced planes of the required type. The multipliers are slow, so they enter the old derivative without a derivative term. Write \(A_u=d_Eu\). The condition that \(G(A\,\cdot\,,\cdot\,)\) is good is Zariski open in the matrix coordinates \((G,A)\). Choose finitely many polynomials defining its complement. At each slow value of \(G\) at least one of them, written \(g(G,A)\), is a nonzero polynomial in \(A\), since a good plane exists. On a slow chart carrying such a polynomial we arrange that \(f=g(G,A_u)\) has a nonzero old directional derivative of some finite order, possibly zero, at every point. This rests on the following jet estimate. In a commuting old frame with first field \(Y\), the truncated Taylor curve of \(A_u\) along \(Y\) through order \(J\) can be prescribed freely by the section jets of order \(J+1\): the first old column uses \(Y^{j+1}u\) and the other columns use \(Y^jY_iu\), \(0\leq j\leq J\), and these are distinct symmetric multi-indices. Let \(D_g\) bound the degrees of the defining polynomials. For fixed \(G\) and degree \(d>0\), select a scalar direction of \(A\) on which the highest homogeneous part of \(g(G,\cdot)\) is nonzero; finitely many directions suffice for all degrees at most \(D_g\). After division by a constant, the polynomial in this variable is monic of degree \(d\). Fix all other coefficients of the curve jet and write the resulting polynomial as \(q(s,v)\). Suppose that \(d+1\) scalar jets \(v_0(s),\dots,v_d(s)\), with pairwise distinct truncations modulo \(s^{\lfloor J/d\rfloor+1}\), satisfy \(\ord_s q(s,v_i(s))>J\). Lagrange interpolation over \(\C((s))\) recovers the leading coefficient: \[ 1=\sum_{i=0}^d\frac{q(s,v_i(s))}{\prod_{j\ne i}(v_i(s)-v_j(s))}. \] Each difference \(v_i-v_j\) has order at most \(\lfloor J/d\rfloor\), so every denominator has order at most \(d\lfloor J/d\rfloor\leq J\), and the right side has positive order, a contradiction. Thus the admissible truncations form at most \(d\) points, and vanishing through order \(J\) costs at least \(\lfloor J/D_g\rfloor+1\) jet conditions. When the polynomial is a nonzero constant there is no bad jet. We take \(J\) so large that this codimension exceeds the source dimension even after all independent matrix parameters \(G\) are allowed. The finitely stratified algebraic jet incidences can then be avoided by \(C^1\)-small jet perturbations on buffered charts. Since we project over the whole matrix parameter space before taking closures, no transversality of the particular slow section \(G\) is required. These perturbations are first made on the exceptional link, where all auxiliary coordinates are allowed and the compact boundary comparison is kept by \(C^1\)-smallness. The result is carried radially on the inner band and joined to the unperturbed generator before the auxiliary multiplier changes. Finite contact is an open condition on a compact buffer of the fixed band, witnessed by finitely many derivatives, and the remaining jet perturbations are made relative to this buffer. All hyperplane tests continue to hold by compactness and openness. With one sufficiently small \(\lambda>0\), the map \(v=v_0+\lambda G(u,\cdot)\) is an injective old frame and has the exact skew primitive \(P\) of the previous subsection on the double. The angular active map extends to the double as an activity of rank four. On the collar we extend the formal jets, using the given derivative path to the constant inclusion on the far side, and we extend the values and transverse formal derivatives freely. In an effective principal slice the rank-four relation excludes an affine three-plane in \(\R^5\), so the relative open ample theorem realizes the extension by a map \(a\). Later \(C^1\)-small perturbations on the collar band keep its boundary comparison by a short interpolation. For every such \(a\), the mixed tensor \((d_Ev)(\Id_E\otimes d_Aa)\) satisfies the tilted phase test everywhere. Choose any real old hyperplane. Its complexified image under \(A_u\) is good for \(G\), hence for \(v\) when \(\lambda\) is small, so its bad family does not contain the complex star centred at the annihilator line of the active four-plane \(d_Aa(A)\). The real normal planes of that star are Zariski dense in it; hence some real normal plane containing the annihilator line lies outside the bad family, and its annihilator is a real three-plane passing the injection test. This three-plane is the image of an active hyperplane under \(d_Aa\). Taking the maximum of the least singular value over the two real hyperplane choices gives a positive margin, uniform by compactness. \subsection{Differential identities across an old rank change} The next lemma concerns operators in the old and active variables. Write \(M\) for the active part of the mixed metric operator and \(\mathcal Q=M+d_E\) for the operator that also allows an active potential differentiated in the old variables. The old fields commute with the active fields; in particular multiplication by a function of the old variables alone commutes with \(M\). \begin{lemma}\label{lem:multipliers} Suppose that on a chart there is a finite-order differential operator \(R\) with \(\mathcal QR=f^p\Id\), where \(f\) is a smooth function of the old variables alone and \(p\) is a positive integer. Let \(Y\) be a fixed old coordinate field, and suppose that for the old covector \(\nu=d_Ef\) there is a conormal identity \[ MS_\nu+\nu\otimes V_\nu=\nu(Y)^b\Id ,\qquad b\geq0, \] where \(S_\nu,V_\nu\) are finite-order operators differentiating only in the active variables, with coefficients smooth through \(\nu=0\). Then \((Yf)^{(b+1)p}\Id=\mathcal QR'\) for a finite-order differential operator \(R'\). If such conormal identities are also available for the successive derivatives \(Yf,Y^2f,\dots\), and \(Y^jf\ne0\) at a point, then \(\mathcal Q\) has a finite-order differential right inverse near that point. \end{lemma} \begin{proof} Put \(t=Yf=\nu(Y)\); it depends only on the old variables. For every \(k\geq0\) and every datum \(h\), the conormal identity and the product rule give \[ f^kt^bh=M(f^kS_\nu h)+d_E\Bigl(\frac{f^{k+1}}{k+1}V_\nu h\Bigr) -\frac{f^{k+1}}{k+1}d_E(V_\nu h). \] The first two terms lie in the image of \(\mathcal Q\). We show by induction on \(k\leq p\) that \[ t^{(b+1)p}h=\mathcal Q(W_kh)+f^kt^{(b+1)(p-k)}L_kh \] for finite-order operators \(W_k,L_k\); for \(k=0\) take \(W_0=0\), \(L_0=\Id\). Given step \(k2k-2\) have been cancelled. We add a profile \(\tau^{j-k+2}I_xZ'\) in the complete normal direction. In the mixed rows, its quadratic cross term with the leading profile at order \(j\) is \[ d_\alpha^y\pair{d_i^yw}{Z'}+d_i^y\pair{d_\alpha^yw}{Z'}-2\pair{C_{i\alpha}}{Z'} . \] All derivatives falling on \(I_x\) or on the slow part of the profile carry at least one extra power of \(\tau\) at this step. We apply the inverse of \eqref{eq:mixedinput}, with \(x\) treated as a parameter, to the negative of the remaining mixed coefficient, and for its solution \((\eta,u,b)\) we choose \(Z'=-\eta/2\). The first-pairing lift at order \(j+1\) supplies \(u_i-\pair{d_i^yw}{Z'}\) and \(b_\alpha-\pair{d_\alpha^yw}{Z'}\), and the pure-pairing lift at order \(j\) cancels the remaining pure coefficients. Earlier orders are unchanged, because \(j-k+2\geq k+1\) and the new self-interaction and the cross terms with previous corrections have larger order. By induction we obtain finite profiles that solve the metric equation to any prescribed order \(J_{\mathrm{met}}\) before evaluation. As \(\tau\to0\), the first columns and the pure second columns converge uniformly to those of \(F_0\), and the mixed columns divided by \(\tau^{k-2}\) converge to \(I_xC(y)\). For a permitted unit pair of old and active covectors put \(\xi=(\xi_{\mathrm o}+\xi_{\mathrm a})/\sqrt2\). The kernel of the pure projection on \(\Sym^2\ker\xi\) is exactly \(\ker\xi_{\mathrm o}\otimes\ker\xi_{\mathrm a}\), because the contraction of a mixed tensor with \(\xi\) splits into its contractions with the two covectors. We use the independent reference first and pure columns on the orthogonal complement of this kernel and the scaled mixed columns on the kernel. Their limiting images are independent with a uniform lower bound, by the compact best margin. After restoring the scaling and projecting off the first span, this gives a phase margin of at least \(c_1\tau^{k-2}\). The normalized coordinatewise powers of a fixed complex sphere cost \(O(L_0^j)\) in every fixed \(C^j\) norm. Indeed their components are \[ P_{L_0}(z)_a=\frac{z_a^{L_0}}{\bigl(\sum_b\abs{z_b}^{2L_0}\bigr)^{1/2}}, \] and on the sphere some coordinate has modulus at least a fixed \(c_0>0\). Each derivative of a numerator divided by the largest modulus to the power \(L_0\) is bounded by \(C_jL_0^jc_0^{-j}\), using \(\abs{z_a}^{L_0-j}\) for \(L_0\geq j\). Likewise the derivatives of the sum in the denominator, relative to the sum itself, are \(O(L_0^j)\), and the reciprocal is handled by the chain rule. These estimates patch over a fixed finite atlas. Consequently the evaluated metric error is \(O(\tau^{J_{\mathrm{met}}+1-q})\) in every fixed \(C^q\) norm, and all finite derivatives of the reference map and of the metric are polynomially bounded in \(\tau^{-1}\). To remove this error we need the following quantitative form of Proposition~\ref{prop:repair}. On a fixed compact foliated source, suppose the initial data have ellipticity in a fixed interval, reference derivatives bounded by \(O(\tau^{-a_0})\) up to one fixed order, and best phase margin at least \(c_1\tau^{b_0}\). Then an error \(O(\tau^K)\) in full \(C^{440}\) can be corrected by a displacement of at most \(\tau^q\) in full \(C^2\), where \(K\) is finite and depends only on \(a_0,b_0,q\) and the fixed dimensions. We indicate the uniformity. For the transverse-minor polynomial used in the proof of Proposition~\ref{prop:repair}, the lower bound on the best phase gives a polynomial lower bound on its norm in the finite-dimensional space of polynomials. On any fixed small ball of directions, the values on a finite interpolation grid bound this norm from below by a uniform constant times the global norm. Hence each direction of the positive-square template can be replaced by a nearby good direction whose restricted singular value has a polynomial lower bound; the positive-square decomposition itself is uniformly stable by ellipticity. We freeze the phases on charts of radius \(\tau^A\), with \(A\) chosen from the finite reference jet bounds and the phase margin. A cover on this mesh has uniformly bounded enlarged overlap, and its cutoff derivatives are bounded by powers of \(\tau^{-1}\). The fixed number of phase labels is separated by successive polynomial weights in \(\tau^{-1}\), so that each paired phase keeps the margin of its dominant phase after normalization. Inverse Gram matrices, phase-coordinate changes and square roots in the finite formulas then obey polynomial bounds, by the polynomial bounds on their denominators and the product rule. The initial thresholds of stage ten of the iteration are finitely many inequalities with positive powers of its starting scale; we choose this scale to be \(\tau^\mu\) with \(\mu\) large, which meets the geometric margins and the prescribed displacement \(\tau^q\), and we choose \(K\) large enough for the normalized initial \(C^{440}\) defect. The subsequent finite waits are carried out for each fixed \(\tau\) as in the proof of the proposition and require no new uniform initial error order. The derivative order of the reference map required at the start is determined by the same fixed stage, independently of \(a_0,b_0,q\). We choose \(k\) larger than this finite reference order plus three. Then, for every truncation order, the reference maps have bounded required jets, since all new profile exponents are at least \(k\). The target \(h_0-\tau^{2k}d^2\abs{\dd_Dc_{\mathrm{ev}}}^2\) differs from \(h_0\) by \(O(\tau^{2k-2})\), so it has fixed ellipticity for small \(\tau\). With \(b_0=k-2\) we prescribe \(q=2k\) and choose \(J_{\mathrm{met}}\) so large that \(J_{\mathrm{met}}+1-440>K\). This gives an exact smooth retained map with the stated target, within \(\tau^{2k}\) in full \(C^2\) of the computed map; the phase margin and the target dimension are preserved. The completion uses a finite-order differential inverse of \eqref{eq:mixedinput} on the compact host; this input cannot be replaced by pointwise surjectivity of a symbol. \subsection{The active cofactor complement} \label{sec:boundarycolumns} On the attaching collar \(H=E_v\oplus A\). In the reference product we keep all complete old columns, the four active first columns, and the six pure second columns in \(\Sym^2A\). Their orthogonal complement is \(I_xW\oplus C_A(x)\), where \(C_A(x)\) is the rank-four complement of these columns inside the complete active span. The projections of the \(LA\) and \(LL\) second columns identify \(C_A(x)\) with \(L\otimes A\oplus\Sym^2L\). The normal isometry splits into an old block and an active block. In a slow coefficient frame the old derivative of \(I_x\) therefore has zero projection to \(C_A\), while the active derivative has the cofactor response of Section~\ref{sec:physicalactive}. These are the reference columns for the boundary comparison of Section~\ref{sec:wholeassembly}. \section{Characteristic directions of the active operator} \label{sec:actgeom} In this section we carry out the algebraic computation on graphs that is needed for the active parameter problem of Section~\ref{sec:actmodule}. All spaces and dimensions in this section are complex. Let \(\mathcal B\) be a closed algebraic family of projective lines in \(\mathbb P^4\), of dimension at most four and containing no full star. Let \(x\mapsto(x,f(x))\) be a local holomorphic graph over \(\C^4\), with affine normal \(\eta(x)=(-\dd f(x),1)\), represented in the affine chart \(\{p_0=1\}\subset\mathbb P^4\) by \(p(x)=\dd f(x)\). Its \emph{raw cotangent cone} is the set of \((x,\xi)\in T^*\C^4\) with \(\xi\neq0\) such that the projective line spanned by \((1,p(x))\) and \((0,\xi)\) belongs to \(\mathcal B\); thus it consists of the directions \(\xi\) of the lines of \(\mathcal B\) through the normal \(p=\dd f(x)\). We say that the Gauss map \(x\mapsto \dd f(x)\) is \emph{quasifinite} if its fibres are zero-dimensional. A complex submanifold \(\Lambda\) of \(T^*\C^4\setminus\{0\}\) is \emph{coisotropic} if \((T_q\Lambda)^{\omega}\subset T_q\Lambda\) for every \(q\in\Lambda\), where \(\omega\) is the canonical symplectic form; equivalently, the Poisson bracket of two local functions vanishing on \(\Lambda\) again vanishes on \(\Lambda\). A germ invariant under the scalings \((x,\xi)\mapsto(x,\lambda\xi)\), \(\lambda\in\C^*\), is called \emph{conic}. A conic coisotropic germ has dimension at least four, so its fibrewise projectivization has dimension at least three. \begin{lemma}\label{lem:chargeometry} Let \(J\) be a positive integer. If the raw cone of a graph with quasifinite Gauss map contains a nonzero conic coisotropic germ, then the graph satisfies one of finitely many algebraic \(J\)-point incidence conditions on jets of order at most five at \(J\) distinct source points. After the source points and the normal and line marks are allowed to vary, and the fibre dimension supplied by the germ is taken into account, each of these incidences has codimension at least \(J\) in any coefficient space that interpolates such jets independently at \(J\) distinct source points. These estimates hold fibrewise, at fixed parameters of the line family. \end{lemma} \begin{proof} We work with an irreducible component \(\mathcal B_0\) of the line family. Its marked incidence \[ I=\{(p,\ell)\in\mathbb P^4\times\mathcal B_0:\ p\in\ell\} \] has dimension \(\dim\mathcal B_0+1\). Since \(\mathcal B\) contains no full star, the lines of \(\mathcal B\) through a point \(p\) form a proper closed subset of the \(\mathbb P^3\) of all lines through \(p\); hence every projective direction fibre has dimension at most two. \emph{Dominant components: the common image.} Suppose first that \(I\to\mathbb P^4\) is dominant. Then \(\mathcal B_0\) has dimension \(d\in\{3,4\}\). At a general smooth line \(\ell\), the conormal space of \(\mathcal B_0\) in the Grassmannian is a \((6-d)\)-dimensional subspace of \(\Hom(\C^5/\ell,\ell)\). We claim that it cannot, for generic \(\ell\), consist of maps with one common image line \(c_\ell\subset\ell\). Indeed, suppose it does. Since \(6-d\in\{2,3\}\) and \(\dim\Hom(\C^5/\ell,c_\ell)=3\), this defines a rational focal map \(\ell\mapsto[c_\ell]\), and the conormal space consists exactly of the maps with image in \(c_\ell\) that vanish on a common kernel \(\Pi_\ell/\ell\subset\C^5/\ell\) of dimension \(3-(6-d)=d-3\). The preimage \(\Pi_\ell\subset\C^5\) is the vector space of a projective \((d-2)\)-plane containing \(\ell\). For a tangent variation \(\delta\colon\ell\to\C^5/\ell\) of \(\mathcal B_0\), the conormal equations require \(\delta(c_\ell)\in\Pi_\ell/\ell\). Differentiating \(c_\ell\subset\ell\) therefore shows that, at a general point, the projective tangent space of the focal image lies in \(\mathbb P(\Pi_\ell)\). Let \(s\le d-2\) be the dimension of the focal image. \begin{itemize} \item If \(s=0\), all lines of \(\mathcal B_0\) pass through one point and lie in its star. This is impossible when \(d=4\), and when \(d=3\) it makes \(\mathcal B_0\) a full star, which is excluded. \item If \(s=d-2>0\), then \(\mathbb P(\Pi_\ell)\) is the tangent plane of the focal image at \([c_\ell]\), and all lines with this focal point lie in the \((s-1)\)-dimensional family of lines through \([c_\ell]\) in that plane. Hence \(\dim\mathcal B_0\le s+(s-1)=2s-1\binom{a+b+4}{4}, \] for instance \(a=6b\). The linear system for the left coefficients making the two products agree then has a nonzero solution, and both multipliers are nonzero. This gives the division ring of left fractions. Gaussian elimination on \(P_J\) with nonzero principal pivots produces an inverse in this ring whose leading symbol is the commutative inverse, by the Schur complement formula at each step. The exact relations obtained by subtracting \(P_{\rm rem}P_J^{-1}\) times the selected rows from the remaining rows annihilate \(P\); we clear common left denominators. Together with the algebraic range constraints for \(e_\eta\), the resulting compatibility rows on \((g,h)\) have weighted principal symbols spanning the left cokernel of \((e_\eta,e_\xi)\) on \(K_U\), on a dense open set of directions. Indeed, the commutative row elimination gives a basis of the cokernel there, and each denominator multiplies its symbol by a nonzero scalar symbol. After substitution of \(h-Qg\), the first block of each row has order at most one more than the divergence block. The eliminated rows \(UX\) can be reinserted to obtain identities in the original ambient flux. For fixed matrix dimensions there is a bounded number of fraction operations and denominator clearings, so the argument for common multiples gives a bound on all orders independent of the polynomial degree of the activity. \paragraph{Elimination of the coherent part.} The support of the relative module, which is coherent over the functions, is saturated along the connected fibres. Indeed, its restriction to each fibre is a coherent module with integrable connection, hence locally free there by horizontal transport, and Nakayama's lemma identifies the points of the support with the points where this restriction is nonzero. Thus the parameter projection of the complement of the support is the open complement of the projected support. We stratify that algebraic support into finitely many smooth pieces over which the module, restricted to algebraic local sections of the orbit projection, is locally free. Such sections exist by the local triviality of the torus bundle, after shrinking around a point of a constant torus. Generic freeness and induction on closed complements give the stratification. On each nonzero stratum, flat relative transport over a small source ball gives a nonzero holomorphic family of solutions \(X_\lambda\) of the adjoint equations. At a fixed parameter of a nonzero stratum, choose \(J'\) distinct source points at which the solution is nonzero and the normal is general, away from the meromorphic poles used in the elimination. This is possible because the quasifinite Gauss map is locally open and the nonzero solution is nonzero on a dense open subset of the source. At each of these points, forcing detection and the compatibility rows give an exact row whose first principal block detects \(e_\xi X\) for some good direction \(\xi\). Differentiate the equations along a tangent vector to the parameter stratum and apply this exact row. It annihilates the term involving the differentiated solution, and so gives a linear scalar constraint on the tangent vector. To see that these constraints are independent, vary only the jets of the graph height at the chosen points: keep all derivatives through the order \(b\) of the first block equal to zero, and use the homogeneous jet \(\xi^{\otimes(b+1)}\) at the next order. In the differentiated equations the divergence row is unchanged in graph coordinates, and \(U\) is held fixed. The product rule in the first block leaves exactly the nonzero symbol applied to \(e_\xi X\), while the derivatives of the solution and the lower derivatives of the graph drop out of this test. Independent interpolation of the ambient coefficients gives \(J'\) independent tangent constraints. The required order of interpolation is bounded before \(D\) is chosen, by the fraction elimination above. Hence the projected exceptional set of coefficients for these nonzero strata has codimension at least \(J'-p_s-t_{\rm orb}\), and this remains true for its algebraic closure. \paragraph{Choice of the activity.} Finite complex Gauss germs at all real points can be arranged by a preparation of finite jets. For a centred scalar height jet, the loci where the Jacobian algebra modulo the \((j+1)\)-st power of the maximal ideal has dimension at least \(j+1\) are closed determinantal loci in the spaces of \((j+1)\)-jets. They are nested under inverse images, because each nonzero successive quotient of powers of the maximal ideal has positive dimension, while a zero successive quotient terminates the filtration by Nakayama's lemma. Their codimensions tend to infinity. Otherwise the components of maximal dimension would eventually be full jet cylinders over components at the preceding level; since there are finitely many components at the starting level, one such cylinder would persist at all higher orders. However, every fixed Taylor jet extends to a polynomial with an isolated critical point by adding \(\sum_ix_i^q\) in strictly higher degree, because the homogeneous gradient terms have no common projective zero. This contradiction gives arbitrarily large codimension. For a real nonzero height covector, varying the activity in that scalar height direction surjects onto the scalar jets. Finite-jet transversality, with the projective height marks allowed to vary, therefore gives isolated critical Gauss germs at every real point. By compactness we obtain a fixed jet order and an open margin. We approximate in the corresponding norm by ambient polynomials with smooth slow coefficients, preserving the prescribed \(C^1\) comparison. This gives quasifiniteness where it is needed on the complexified orbits. Choose \(J\) and \(J'\) so that both codimension bounds exceed the dimension of the slow base, at both old ranks, and choose a single degree \(D\) large enough for both interpolations and for the approximation. By real stratified transversality, the coefficient map of the compact slow base can be moved by an arbitrarily small constant real translation of the coefficients so as to avoid the finite union of the closed exceptional sets for the two ranks. The complex conjugates may be added to this finite union, and the real traces have real codimension at least the stated complex codimensions. All rank-four and finite-Gauss conditions persist, as do the compact paired comparisons. On the joint good open sets of parameters and orbits there is now no support of the coherent module at either rank. Hence the order-zero flux rows are finite left differential combinations of the intrinsic adjoint equations on algebraic charts. Since the good order filtration and all coefficient incidences are algebraic, these identities hold on those charts before the smooth slow coefficients are substituted. They are the required active parameter identities at both ranks. \end{proof} \section{Relative filling of the hyperplane metric} \label{sec:relativefill} For a distribution \(D\) of rank \(d\) we write \[ j_d=d+\frac{d(d+1)}{2}=\frac{d(d+3)}{2} \] for the rank of the leafwise second-jet bundle \(J_2(D)=D\oplus\Sym^2D\). A map, or a map defined on an open subset, is called \emph{\(D\)-free} if its first and symmetric second leaf derivatives (its first and second leaf columns) define a monomorphism \(J_2(D)\to\R^N\) at every point. \begin{proposition}\label{prop:relativefill} Let \(\pi\colon X\to B_0\) be a compact smooth submersion, where \(B_0\) may have boundary and \(\partial X=\pi^{-1}(\partial B_0)\). Let \(D\subset\ker\dd\pi\) be an integrable vertical distribution of rank \(d\) with a smooth positive metric \(h\), and let \(N\geq j_d+5\). Let \(K\subset B_0\) be a closed subset containing \(\partial B_0\). Suppose that \(F_{\mathrm{in}}\) is a \(D\)-free \(h\)-isometric map into \(\R^N\), defined on \(\pi^{-1}(U)\) for an open neighbourhood \(U\) of \(K\), and that its columns \(J_2(D)F_{\mathrm{in}}\) extend to a formal monomorphism \(\Phi_0\colon J_2(D)\to X\times\R^N\) over all of \(X\). Then there is a smooth \(D\)-free map \(F_{\mathrm{out}}\colon X\to\R^N\) with \(\metric_D(F_{\mathrm{out}})=h\), equal to \(F_{\mathrm{in}}\) on \(\pi^{-1}(U')\) for a smaller neighbourhood \(U'\) of \(K\), such that \(J_2(D)F_{\mathrm{out}}\) is homotopic to \(\Phi_0\) through monomorphisms relative to \(\pi^{-1}(U')\). \end{proposition} \begin{proof} \emph{Step 1: a free map in the given class.} Using a connection and a complement of \(D\), extend \(\Phi_0\) to a formal second jet over \(X\) which equals the jet of \(F_{\mathrm{in}}\) near \(\pi^{-1}(K)\). The open \(D\)-free relation and its first-jet projection are ample. Indeed, let \(\tau\) be a principal covector. If \(\tau|_D\neq0\), choose a leaf basis \(e_1,\dots,e_d\) with \(e_1,\dots,e_{d-1}\in\ker\tau|_D\). The principal slice varies only the last pure second column, the other \(j_d-1\) columns being fixed and linearly independent, and freeness excludes their linear span, which has codimension \(N-j_d+1\geq6\). If \(\tau|_D=0\), the relation is constant on the slice. The first-jet projection is the partial immersion relation, and the same slice argument with the \(d\) first columns applies; the excluded span has codimension \(N-d+1\geq2\). Thus every slice is open and connected with full convex hull. The relative higher-order ample theorem \cite[Theorem~3.5]{MA} gives a smooth \(D\)-free map \(F\colon X\to\R^N\), equal to \(F_{\mathrm{in}}\) near \(\pi^{-1}(K)\), together with a homotopy of monomorphisms from \(\Phi_0\) to \(J_2(D)F\) relative to a neighbourhood of \(\pi^{-1}(K)\). \smallskip\noindent \emph{Step 2: supported fourth-power primitives.} We describe the device used to adjust the metric, recording the local analytic results it relies on. Work in a contractible foliation box \(\Omega\times Q\), with \(\Omega\) a leaf ball, \(Q\) a compact transverse parameter ball, \(D\) spanned by the leaf coordinate fields and \(\pi\) constant on the plaques. A smooth compactly supported amplitude \(a\geq0\) and a linear leaf coordinate \(\phi\) determine the increment \(a^4(\dd_D\phi)^2\). We use the free-map primitive construction \cite[Sections~4.1--4.3, Theorem~4.1]{ZerLocal} and the supported perturbation theorem \cite[Theorem~2.12]{ZerStab}. In leaf coordinates and in target dimension at least \(j_d+5\), the first construction gives supported free maps whose leaf metric approximates the desired increment to arbitrarily high order in the wavelength \(\eps\), measured in leaf \(C^{2,\alpha}\). More precisely, the finite correction through order \(r\) has metric error \(O(\eps^{r-2-\alpha})\) in leaf \(C^{2,\alpha}\). Its determinant estimate uses normalized free columns and gives a lower bound independent of \(\eps\) for fixed \(r\), while the leaf derivatives of the columns of each fixed order are bounded by powers of \(\eps^{-1}\). Consequently, by the Gram inverse formula and the product rule, the free pairing inverse has \(C^{2,\alpha}\) norm \(O(\eps^{-b})\) for some \(b\) independent of \(r\). These local results hold with the transverse parameters of the box. To obtain the relative and homotopy assertions, choose five orthonormal sections of the normal bundle \((J_2(D)G)^\perp\) of the current free map \(G\) over a buffered product box, by orthogonal transport along radial segments. The finite primitive formula uses only these sections, linear pairings, and zero-mean integration in an auxiliary periodic variable \(\theta\). For a fixed finite order, the normalized column determinants and all cutoff and inverse estimates are uniform on the compact parameter box. At a zero of the amplitude, the only apparent quotient by \(a^2\) occurs in the first step and has numerator \(\abs{\partial_\theta(aR)}^2=a^2\abs{\partial_\theta R}^2\), where \(R\) is the fixed periodic curve; it is therefore extended as \(\abs{\partial_\theta R}^2\). The later pairing systems divide by \(a\) only on the large-amplitude side of the smooth partition in \(a^2\) used in that construction. Inductively from the initial oscillator \(a^2R\), all displacement terms vanish at a transverse parameter where \(a\) vanishes identically on the plaque. Choose one outer leaf cutoff equal to one on the support of the remainder, and apply the supported perturbation theorem with this cutoff throughout the compact transverse family. If \(e\) is the prescribed metric remainder, the smallness required for the contraction is the product of the \(C^{2,\alpha}\) norm of the pairing inverse with the \(C^{2,\alpha}\) norm of its value on \((0,e)\), which is \(O(\eps^{r-2-\alpha-2b})\); the displacement is \(O(\eps^{r-2-\alpha-b})\). We first fix the finite order \(r\) so that, for small \(\eps\), this displacement also lies within the freeness tolerance given by the normalized column bounds, and then fix one small \(\eps\), uniformly on the compact box. The contraction uses the fixed leaf Dirichlet inverse and has a common contraction constant less than one. Its solution depends smoothly on the parameters, by the implicit function theorem in leaf \(C^{2,\alpha}\) and the higher leaf regularity of the same supported equation \cite[Sections~3.1--3.3]{ZerStab}. For mixed derivatives, a derivative in a further parameter satisfies an equation with the same contractive linearization acting on the highest parameter derivative; the higher leaf estimates absorb the highest leaf derivative in that equation and bound the lower derivatives inductively, and the fixed leaf Dirichlet inverse commutes with parameter derivatives. This gives joint smoothness for one fixed choice of \(\eps\). On the leaf and transverse boundary buffers, where the amplitude vanishes on whole plaques, the zero correction is the unique small solution; hence the map is unchanged off the buffered chart. Adding a compact parameter \(t\in[0,1]\) with amplitude \(ta\) gives a smooth path of \(D\)-free maps from the given free map to the one realizing the increment. \smallskip\noindent \emph{Step 3: shortening and decomposition of the defect.} Before applying this device, shorten \(F\) by a positive basic function \(\lambda\) (that is, a function pulled back from \(B_0\)) equal to one near a smaller closed neighbourhood of \(\pi^{-1}(K)\). Take \(\lambda^2=1-\beta^4\), where the basic function \(\beta\geq0\) vanishes near that set and equals a positive constant outside a region on which \(F\) is already \(h\)-isometric. Choose the constant so that \(\lambda^2\metric_D(F)0\), and the passive profile \[ f=-\lambda\,u_L\otimes z^*,\qquad z^*\in V^*,\quad z^*(z)=1, \] which takes values in the original active line \(L\). At the comparison point \(\delta_0=0\), \(a=z\), and for any derivative \(T\) on the path of active derivatives, the paired mixed map on \(Q\otimes A\), up to the common positive factor \(\lambda\), is \[ h\otimes b\longmapsto \Bigl(G(\pi_Dh,Tb),\ \sigma(\pi_Dh,b)\,Z_x+M_x\bigl((\pi_Lh)\otimes b\bigr)\Bigr). \] It is injective by the path of active derivatives and the three \(LA\)-rows. This comparison at \(\delta_0=0\) concerns columns only. We choose \(\delta_0>0\) sufficiently small after the compact comparison and the bank placement have been fixed. The generator then has a monomorphic old derivative \(E_v\to Q\), and by compactness the paired map keeps its rank and class, also along the derivative path. Restriction along the old derivative \(E_v\to Q\) preserves rank and is homotopic to restriction along the original inclusion. In this way we avoid having to realize an old generator whose line component vanishes identically. On exceptional collars we use instead an old generator into \(\mathcal H_v\), as in Section~\ref{sec:actualold}, and set the passive profile identically zero. Both generators take values in \(E_v\) plus their single generating line in \(\mathcal E\), and all reduced tests hold there. The classes of these derivatives are preserved by these homotopies of monomorphic old derivatives. After the compact tests have been fixed, we shrink the active values toward \(z\) by normalized powers and the inverse scaling of values. The paired comparisons for all independent samples of the points and of the coefficients persist at the endpoints. We then use Section~\ref{sec:actualold} to extend and prepare a single old generator on the double. The required rank of its projection to \(E\) holds on each link, since both generators take values in \(E_v\) plus one line, and it holds along the required monomorphic formal homotopies to the inclusion. The slow contraction of the auxiliary coordinates away from these bands and the small finite-contact preparations preserve every required reduced test. On the band the changes of coefficients can be taken arbitrarily small in \(C^1\), so that the four cofactor comparisons are retained. The passive profiles may be kept fixed during these small changes, since they already satisfy all the compact open comparisons at the initial generators. We extend them off the split-normal collars by slow cutoffs in their original \(L\)-slots and by zero elsewhere. They have the same fixed rank-one label reader \(z^*\) everywhere, so \(f^*d_Ef=\lambda^2\pair{u_L}{d_Eu_L}\,z\otimes z^*\) is symmetric and \(\skewp(f^*d_Ef)=0\). Choose an independent slow reference frame \(v_0\) on the double by rank avoidance, and one sufficiently small \(\lambda>0\). Then \(v=v_0+\lambda G(u,\cdot)\) is an independent old frame, and \(\skewp\bigl(v_0^*(v-v_0)\bigr)\) is its exact old skew primitive, valid for every basic passive weight. Extend the active maps relative to the collars over the double, using their formal rank-four paths and the open ample argument of Section~\ref{sec:actualold}. Choose the simultaneous activity of Theorem~\ref{thm:activeparameter} so close on this compact double that all the boundary comparisons on the boundary band persist. The finite old contact and Lemma~\ref{lem:multipliers} give the differential right inverse \eqref{eq:mixedinput} on the same product, and the reduced tests give its compact tilted phase margin. On the boundary band, with \[ U(h,b')=(d_h^yv)b',\qquad S(h,b')=(d_h^yf)b', \] the paired map on \(E_v\otimes A\) \[ h\otimes b\longmapsto \Bigl(U(h,d_b^ya),\ J_x(b)U(h,a)-M_x\bigl(S(h,a)\otimes b\bigr)\Bigr) \tag{P}\label{eq:pairedactual} \] is uniformly injective. At \(a\approx z\) one has \(-M_x(S(h,a)\otimes b)\approx\lambda M_x((d_h^yu_L)\otimes b)\), which carries the same factor \(\lambda\) as \(U(h,d_b^ya)=\lambda G(d_h^yu,d_b^ya)\). After adjoining the old columns and the ten active columns, the class of \eqref{eq:pairedactual} agrees with \(J_\Sigma\). The argument gives this class while the independent old and active samples are still separate, so it is preserved by the subsequent power substitution. \section{Completing the surviving line} \label{sec:line} We first construct the periodic family used to complete the metric along the surviving line. The construction is intrinsic for a metric real line bundle $L$, including a nonorientable one: all formulas are written in a local unit $\ell$ of $L$ and are checked to be independent of the choice of $\ell$. Throughout this section, for a function $f$ on a sphere bundle with Hopf flow $\rho_t$ we write \[ \overline f(z)=\int_0^1 f(\rho_t z)\,\dd t \] for its orbit mean, and $J$ denotes a fixed quarter turn of $\R^2$. \begin{lemma}\label{lem:linecontact} Let $L$ be a metric line bundle over a compact slow base, and let $d_*\geq1$ and $b_*\geq2$ be integers. In a local unit of $L$, let $X$ denote the period-one Hopf generator on the unit sphere bundle $S(\triv^{b_*}\oplus L^{b_*})$. There is a smooth family $a(q,p,\theta)$ of sections of $2L=L\oplus L$ over this sphere bundle, defined for $q$ in a small ball of $\R^{d_*}$, $p$ in a small ball of $L^{d_*}$, and $\theta$ in a small interval of $L$, with the following properties. \begin{enumerate} \item Each member of the family has unit speed $\abs{Xa}=1$, zero orbit mean $\overline a=0$, and zero signed orbit area $\overline{\langle a,JXa\rangle}=0$. \item The orbit-mean parameter one-form of the family is \[ \overline{\langle \dd a,Xa\rangle}=\dd\theta-\sum_{j=1}^{d_*}p_j\,\dd q_j , \] where the differentials are taken in the family parameters only. \end{enumerate} The form in (ii) has the parity of $L$. The family and its parameter derivatives are smooth also in the sphere and slow variables. \end{lemma} \begin{proof} In a local unit of $L$ write the sphere coordinates as $z=x+\ii y\in\C^{b_*}$, so that the Hopf flow is $\rho_t(z)=e^{2\pi\ii t}z$. Changing the unit conjugates $z$ and reverses the sign of $X$. The antipodal map $\sigma(z)=-z=\rho_{1/2}(z)$ is a globally defined free involution. On the quotient by $\sigma$ the Hopf tangent line is the pullback of $L$; let $\gamma$ be the sign line of the double cover. We first find a section of $L\oplus(L\otimes\gamma)$ over the quotient whose derivative along the Hopf line vanishes nowhere. A formal solution is given by the derivative that maps the Hopf line isometrically onto the first summand, and the relation is open and ample: each principal slice is the complement of one point in a plane. By the relative parametric $h$-principle for open ample relations \cite[Theorem~18.4.1]{EM} such a section exists. Lifted to the sphere bundle, it becomes a $2L$-valued section $a$ satisfying $a(\sigma z)=\mathcal R a(z)$ with $\mathcal R=\diag(1,-1)$. For a function $f$ with $\overline f=0$ define \[ \mathcal I_X f(z)=\int_0^1 t\, f(\rho_t z)\,\dd t . \] Integration by parts gives $X\mathcal I_X f=f$, and Fubini's theorem gives $\overline{\mathcal I_X f}=0$. Reversing the unit of $L$ reverses this primitive. Let $v_0=\abs{Xa}>0$ be the speed of the section and put $\lambda_0=\overline{v_0}$. The orbit map with variable time $\mathcal I_X(v_0/\lambda_0-1)$ has positive derivative $v_0/\lambda_0$ along the orbits and degree one on every circle, so it is a diffeomorphism of each orbit. Composing the section with its inverse and dividing by $\lambda_0$ makes the speed equal to one. Since $v_0$ is $\sigma$-invariant, this operation commutes with the reflection symmetry. Finally we subtract the orbit mean. The resulting section is centered, and its signed area has mean zero, because $\mathcal R^{T}J\mathcal R=-J$ implies that the area integrand $\langle a,JXa\rangle$ changes sign under $\sigma=\rho_{1/2}$. Write $v=Xa$ and $j=Jv$ in a unit of $L$. Since $a$ and $X$ both change sign under unit reversal, $v$ and $j$ are unchanged, and $Xv=\kappa j$ for an $L$-valued curvature $\kappa$. The mean curvature $\overline\kappa$ vanishes: the sphere fibre has dimension $2b_*-1\geq3$ and is simply connected, so the unit tangent map $v$ sends each Hopf circle to a null-homotopic loop in $S^1$. Moreover $b_2=\overline{\kappa^2}>0$, since zero curvature on an orbit would give a constant unit tangent $v$, contradicting $\overline v=0$. Both statements hold on every orbit. We next produce enough unit-speed tangent pairs for the parameter family. On every orbit $\max v_1>0$, since otherwise $\overline v=0$ and $\abs v=1$ would be incompatible. Choose $\eta_0>0$ such that this maximum exceeds $2\eta_0$ on every orbit, and a smooth function $\eta(v_1)\geq0$ supported in $\{v_1>\eta_0\}$ and positive on $\{v_1>2\eta_0\}$. Then \[ b_0=\overline{a_2^2\eta^2v_1}>0 . \] Indeed, if $a_2=0$ throughout $\{v_1>2\eta_0\}$, then $v_2=Xa_2=0$ and hence $v_1=1$ there. This set would then be a nonempty open and closed subset of the orbit circle, contradicting $\overline{v_1}=0$. Choose distinct positive integers $c_1,\dots,c_{d_*}$ and, for a large positive integer $K_*$, set \[ h_{q,i}=a_2\eta\cos(K_*c_i a_1),\qquad h_{p,i}=a_2\eta\sin(K_*c_i a_1). \] For either type of input $h$ put \[ \beta_h=h-\overline h-\kappa\,\overline{\kappa h}/b_2,\qquad \alpha_h=\mathcal I_X(\kappa\beta_h),\qquad u_h=\alpha_hv+\beta_h j-\overline{\alpha_hv+\beta_h j}. \] The primitive is defined since $\overline{\kappa\beta_h}=0$; also $\overline{\beta_h}=0$. From $Xu_h=(\kappa\alpha_h+X\beta_h)j$ we get $\langle v,Xu_h\rangle=0$, and $\overline{\langle u_h,v\rangle}=\overline{\alpha_h}=0$. The $q$-tangents have the parity of $2L$, and the $p$-tangents are ordinary two-vectors. Both have the same reflection symmetry as $a$: under $\sigma$, $v$ transforms by $\mathcal R$ and $j$ by $-\mathcal R$, while $\kappa,\beta_h$ are odd and $\alpha_h$ is even. The skew pairing of these tangent pairs is nondegenerate for $K_*$ large. On the support of the oscillations $Xa_1=v_1>\eta_0$, so repeated integration by parts against $(\ii K_*c\,v_1)^{-1}X$ shows that every mean of nonzero frequency tends to zero faster than any inverse power of $K_*$. One such antiderivative, together with a bounded zero-mean primitive of its remainder, gives $\alpha_h=O(K_*^{-1})$. Consequently \[ \overline{\langle u_{q,i},Xu_{p,i}\rangle}=\tfrac12K_*c_i b_0+o(K_*). \] Indeed, the term of order $K_*$ is $\beta_{q,i}X\beta_{p,i}$, whose nonoscillatory part is $\tfrac12K_*c_i a_2^2\eta^2v_1$. All pairs of distinct frequencies, and all pairs of the same type, have zero mean at order $K_*$. These estimates hold uniformly on the compact orbit quotient. For the parameter form $\lambda_a=\overline{\langle \dd a,Xa\rangle}$, the exterior derivative on two variations $u,u'$ is $-2\overline{\langle u,Xu'\rangle}$. After rescaling each pair by $(K_*c_ib_0)^{-1/2}$ and reversing the sign of the $p$-member, the resulting form is close to \[ \omega_0=\sum_{i=1}^{d_*}\dd q_i\wedge \dd p_i . \] A fibrewise linear change makes it equal to $\omega_0$ at the origin: interpolate linearly between the two nondegenerate forms, $\omega_t$, and use the linear transport generated by $-\tfrac12\omega_t^{-1}\dot\omega_t$. Form the affine family with these tangents, and normalize its speed and center it exactly as above. At the origin the tangents are unchanged by this normalization, since their linearized speed and mean already vanish. The reflection symmetry gives zero area for all nearby parameters. Including the time shift $\theta$ along each Hopf orbit, the parameter form becomes $\dd\theta+\lambda(q,p)$ with $\lambda(0)=0$ and $\dd\lambda(0)=\omega_0$. On a common smaller ball, the radial homotopy on the parameter fibres writes $\dd\lambda-\omega_0=\dd\mu$ with $\mu=O(\abs{(q,p)}^2)$. The vector field $Z_t$ defined by $\iota_{Z_t}(\omega_0+t\,\dd\mu)=-\mu$ vanishes to second order at the origin, so it has a smooth time-one flow on a smaller ball, and pullback by this flow converts $\dd\lambda$ into $\omega_0$. Now $\lambda+\sum_ip_i\,\dd q_i$ is closed, hence exact by the radial homotopy, and changing $\theta$ by its negative primitive gives the required formula. All orbit means and all these parameter changes are smooth over the orbit quotient. The forms and times have the parity of $L$, so their defining vector fields and flows agree under unit reversal. This proves the lemma. \end{proof} \subsection{Line completion on a compact model} \label{sec:linecompletion} Let $B$ be a closed manifold and let $V_0$ be a rank-$n$ metric bundle on $B$ with an injective coefficient map $A_0\colon V_0\to\triv^a$. On an open tube $U\subset B$ fix an orthogonal splitting $V_0|_U=H\oplus L$. The old host is $B\times(S^{2h-1})^a$, the product of $B$ with the original $a$ Hopf spheres. Denote by $Y_0$ the old coefficient field of a local unit $\ell$ of $L$, and by $Y_1,\dots,Y_{n-1}$ those of a local slow orthonormal frame of $H$. Fix smooth basic functions $c,\chi\geq0$ with $c^2+\chi^2=1$ and $\supp\chi$ compact in $U$, and fix a slow open neighbourhood of $\supp\chi$ with closure in $U$. \begin{proposition}\label{prop:linecompletion} Let $N\geq s_n+1$. Suppose there are an old map $F$ into $\R^N$ and an $L$-valued function $s$ on the old host whose slow support is compact in $U$, with $\abs{\dd_{V_0}s}<1/10$. Suppose that on a neighbourhood of $\supp\chi$ the map $F$ is $H$-free, and there is a line monomorphism $e\colon L\to\R^N$, independent of the columns $J_2(H)F$, satisfying $cY_0F=c\,e(\ell)$. Suppose that on the same neighbourhood there is an isometric inclusion $\nu\colon2L\to\R^N$, slow or fast varying, orthogonal to $J_2(H)F\oplus e$. In a local unit of $L$ write $\alpha=\dd_Hs$ and $p=Y_0s$. Assume that the old metric has the blocks \begin{equation*} \gamma_F|_{HH}=1_H+c^2\alpha^2,\qquad \gamma_F|_{HL}=(1+c^2p)\alpha,\qquad \gamma_F|_{LL}=1+2p+c^2p^2 \tag{L}\label{eq:oldlinemetric} \end{equation*} where $c>0$ in $U$, that $\gamma_F|_{HH}=1_H$ where $c=0$, and that outside $U$ the old metric is the specified unit metric on $V_0$. Finally, suppose that $F$ is phase regular on $V_0$ outside a relatively compact neighbourhood of $\supp\chi$ on which $e$ and $\nu$ are defined. Fix a slow isometric inclusion $\iota_L\colon L\to\triv^6$ on $U$ and a basic cutoff $\beta$, equal to one near $\supp s\cup\supp\chi$ and supported in $U$. Define the coefficient map $A_1\colon V_0\to\triv^{a+6}$ by \[ A_1(v_H+v_L)=\bigl(A_0(v_H+cv_L),\ \beta\,\iota_L(v_L)\bigr) \] on $U$, and by $A_1=(A_0,0)$ outside $U$. Then on the old product with six further ordinary Hopf factors of the same sphere size there is a smooth map into the same $\R^N$ which induces the unit metric of $V_0$ along the distribution of $A_1$ and is phase regular along it. \end{proposition} \begin{proof} All estimates in this proof are for one fixed original sphere size $h$. Choose two nested slow buffers around $\supp\chi$ on which $e$ and $\nu$ are defined; they have compact closures in the region of definition. Choose finitely many smooth functions $q=(q_1,\dots,q_{d_*})$ of the old variables whose differential $\dd_Hq$ is injective on the larger buffer. For example, use cutoffs times the coordinates of a finite atlas of the compact old host, together with the cutoffs themselves. Scale $q$ into the small parameter ball of Lemma~\ref{lem:linecontact}. For the moment we work on the auxiliary twisted sphere bundle of that lemma; its size $b_*$ will be fixed below, independently of all frequencies. In a local unit of $L$ its generator $X$ commutes with the old fields. Put \[ a=a(q,0,0),\qquad v=Xa,\qquad Z=\nu a . \] Then \[ \overline Z=0,\qquad \abs{XZ}=1,\qquad \overline{\langle YZ,XZ\rangle}=0 \] for every old field $Y$, including the field of the old line. To see the last identity, write $YZ=(Y\nu)a+\nu\sum_j\partial_{q_j}a\,(Yq_j)$. Since $\nu^*\nu=\Id$, the connection matrix $\nu^*Y\nu$ is skew, hence a multiple $\omega_YJ$ of the quarter turn, and its contribution is $-\omega_Y\overline{\langle a,JXa\rangle}=0$ by the zero signed area. The remaining terms vanish because the parameter form of Lemma~\ref{lem:linecontact} restricted to $p=\theta=0$ is zero. \paragraph{Periodic coefficients.} We use the formal independent new derivative $cY_0+\eps^{-1}X$ along the line and consider the formal series \[ F+\sum_{r\geq1}\eps^rF_r,\qquad F_1=\chi Z . \] The leading line column is \[ w=c\,e(\ell)+\chi\,\nu v, \] where the local unit is understood in the last term. Its leading pairings with the old columns, together with its squared length, give the metric with blocks \begin{equation*} k_{HH}=1_H+c^2\alpha^2,\qquad k_{HL}=c(1+c^2p)\alpha,\qquad k_{LL}=(1+c^2p)^2 . \tag{K}\label{eq:newlinemetric} \end{equation*} Where $c=0$ this uses only the $H$-metric and the unit speed. Elsewhere it follows from \eqref{eq:oldlinemetric} and the orthogonality of the normal plane; for instance $\abs w^2=c^2(1+2p+c^2p^2)+\chi^2=(1+c^2p)^2$. The first variation of the $H$-metric in $F_1$ vanishes, by orthogonality of $F_1$ to the first and second $H$-columns. Suppose that $F_1,\dots,F_r$ have been constructed so that the $(H,L)$ and $(L,L)$ residuals vanish below order $\eps^r$ and the $(H,H)$ residuals vanish below order $\eps^{r+1}$. We keep the factorizations \[ F_j\in\chi\, C^\infty,\qquad \overline{F_j}\in\chi^2C^\infty , \] that is, $F_j=\chi\widetilde F_j$ and $\overline{F_j}=\chi^2\widehat F_j$ with smooth $\widetilde F_j$ and smooth $X$-invariant $\widehat F_j$. At the next orders the mixed, line and $H$ equations are, respectively, \begin{equation*} \begin{split} X\langle Y_iF,F_{r+1}\rangle+C_{r,i}&=0,\\ 2\langle w,XF_{r+1}\rangle+D_r&=0,\\ \mathcal L_HF_{r+1}+Q_{r+1}&=0, \end{split} \tag{R}\label{eq:linerecurrence} \end{equation*} where the first two equations are of degree $r$ and the third of degree $r+1$. Here $\mathcal L_H$ is the linearized metric pairing of the old map, \[ (\mathcal L_HW)_{ij}=Y_i\langle Y_jF,W\rangle+Y_j\langle Y_iF,W\rangle -2\langle Y_iY_jF,W\rangle , \] and $Q_{r+1,ij}=\sum_{m=1}^r\langle Y_iF_m,Y_jF_{r+1-m}\rangle$. These equations follow from the product rule for the quadratic metric expression, in which only the new derivative costs a factor $\eps^{-1}$. The known terms $C_r,D_r$, which depend only on $F_1,\dots,F_r$, have a factor $\chi$ and their means have a factor $\chi^2$, while $Q_{r+1}$ has a factor $\chi^2$. Indeed, each known quadratic term contains two factors of $\chi$, and the means of the terms that are linear in old derivatives of $F_r$ acquire the second factor from the induction hypothesis, since old derivatives commute with orbit averaging. At $r=1$ the mixed mean is zero. Explicitly, \[ C_{1,i}=-2c\chi\langle Y_ie(\ell),Z\rangle+\chi^2\langle Y_iZ,XZ\rangle ; \] this follows by differentiating $\langle Y_iF,Z\rangle=\langle e,Z\rangle=0$ and using $cY_0F=c\,e(\ell)$. Both terms have zero mean, by $\overline Z=0$ and by the third identity for $Z$ above. For $r\geq2$ we replace $F_r$ by \[ F_r+V_r,\qquad V_r=\chi\,\nu\sum_{j=1}^{d_*}\partial_{p_j}a(q,0,0)\,\zeta_j , \] where the $L$-valued functions $\zeta_j$ are $X$-invariant. This preserves the earlier equations, since $V_r$ is centered and normal to $J_2(H)F\oplus e$, and $\langle w,XV_r\rangle=0$ by the differentiated unit-speed identity $\langle v,X\partial_{p_j}a\rangle=0$. At degree $r$ the mixed mean changes by $-\chi^2(Y_iq)\cdot\zeta$. Indeed, the linear old-derivative terms in $V_r$ have zero mean because $V_r$ is centered. The two remaining cross terms with $F_1$ form the variation of $\chi^2\,\overline{\langle Y_iZ,XZ\rangle}$ as $p$ moves to $t\zeta$; by the contact formula of Lemma~\ref{lem:linecontact} this variation is $-\chi^2(Y_iq)\cdot\zeta$, and the area remains zero. A self-product of $V_r$ would have degree $2r-1>r$. Since $\dd_Hq$ is injective, the adjoint $(\dd_Hq)^*$ has a smooth right inverse; we use it to choose $\zeta$ making the mean of $C_r$ zero. For this we use the smooth coefficient of $\overline{C_r}$ after the factor $\chi^2$, which is available also where $\chi$ vanishes, and we then recompute the other known terms. Put $a_i=-\mathcal I_XC_{r,i}$ and $b=-D_r/2$, and write $\overline b=\chi^2b_m$ with the smooth mean factor $b_m$. Let $e_N$ be the component of $e(\ell)$ orthogonal to the first $H$-columns, and set \[ w_N=c\,e_N+\chi\,\nu v,\qquad Q_N=\abs{w_N}^2=c^2\abs{e_N}^2+\chi^2 . \] The function $Q_N$ is $X$-invariant and has a positive minimum on the compact buffer, since $e$ is independent of the $H$-columns. With $Xv=\kappa j$ as in Lemma~\ref{lem:linecontact}, the line correction is \[ b_1=-\kappa\, b_m/\overline{\kappa^2},\qquad u=Q_N^{-1}\mathcal I_X\bigl(b+\chi^2\kappa b_1\bigr),\qquad W_{\rm line}=u\,w_N+\chi b_1\,\nu j . \] Its pairings with the first $H$-columns are zero. The primitive has an input of zero mean, and differentiation gives \[ \langle w,XW_{\rm line}\rangle=Q_N\,Xu-\chi^2\kappa b_1=b . \] The correction $W_{\rm line}$ has the factor $\chi$ and its mean has the factor $\chi^2$: the mean of $u\,c\,e_N$ vanishes, the term containing $u\chi$ has both factors, and $\overline{b_1j}=-(b_m/\overline{\kappa^2})\,\overline{Xv}=0$. Using the pairing inverse of the independent columns $J_2(H)F\oplus e$, choose $W_H$ in their span with pairings \[ \begin{split} \langle Y_iF,W_H\rangle&=a_i,\qquad \langle e(\ell),W_H\rangle=0,\\ \langle Y_iY_jF,W_H\rangle&=\tfrac12\bigl(Y_ia_j+Y_ja_i+Q_{r+1,ij}\bigr) -\langle Y_iY_jF,W_{\rm line}\rangle . \end{split} \] Then $F_{r+1}=W_{\rm line}+W_H$ solves all three equations \eqref{eq:linerecurrence}. The line pairing is not affected by $W_H$, since the pairings of $W_H$ with $e$ and $\nu$ vanish. The pairing inverse does not depend on the new orbit variable, so the prescribed pairings give $W_H$ the same factor and mean factor. This completes the induction, to any finite depth. All coefficients extend by zero off the buffer of the wave. Under a change of unit of $L$, the coordinates of $a$ in that unit are odd while $Z=\nu a$ is unchanged, the coordinate vectors $v,j$ are even, $\kappa,u,b_1$ have the parity of $L$, and $w,w_N$ have this parity as columns evaluated in that unit. Hence the recurrence and all resulting ambient displacements $F_r$ are well defined independently of the unit. For truncation at depth $R_*$, the error in the formal first Gram matrix is $O(\eps^{R_*})$. After substitution of a twisted normalized power of integer frequency $K_0=\eps^{-1}$, the bounds in derivative order $J$ are $O(\eps^{R_*-J})$ for the error and $O(1+\eps^{1-J})$ for the map; this uses the power derivative estimate of the metric production argument. The limiting augmented columns for the good hyperplane $H$ are $J_2(H)F$ together with $w$. They are independent, because the normal projection of $c\,e$ and the component $\chi\,\nu v$ are orthogonal and cannot vanish simultaneously. They give a positive lower bound, uniform on the compact buffer, for the normal second form on $\Sym^2H$. In this test the first and pure $H$-columns and the first line column have errors $O(\eps)$. Away from the support the approximation is the old map. Where the old augmented columns are available and $c=1$, they themselves give a good hyperplane; on the rest of the complement the phase hypothesis on $F$ supplies the bound. \paragraph{Six ordinary Hopf factors and exact correction.} This step uses only six new input factors. The coefficient map $A_1$ is injective: where $c<1$ we are in $\supp\chi$, where $\beta=1$ and the new line block $\iota_L$ is isometric; where $\beta\neq1$ we have $c=1$ and the old block is $A_0$. Fix a finite atlas of slow charts trivializing $L$ over the compact tube containing the supports. For an integer $T$, subdivide these charts in every slow coordinate at scale $1/T$, and take a squared partition of unity $\sum_\lambda\rho_\lambda^2=1$ on a neighbourhood of the wave support, with all cutoffs in the trivializing charts. The enlarged boxes have bounded overlap and their intersection graph has bounded degree, so a fixed number $d_c$ of colours suffices for every $T$. At a box centre $b_\lambda$, round the six coordinates of $T\iota_L(\ell_\lambda)(b_\lambda)$ to an integer row $Q_\lambda\in\Z^6$, where $\ell_\lambda$ is the local unit of $L$ on that box. On the six new spheres of size $h$, let $\Psi_{Q_\lambda}$ be the tensor product of the normalized coordinate powers of degrees $\abs{Q_{\lambda j}}$, conjugating negative powers and using one fixed unit vector for zero powers. It is a unit vector in $\C^{h^6}$, and the generator of factor $j$ acts on it by $2\pi\ii\,Q_{\lambda j}\Psi_{Q_\lambda}$. On each colour take the sum of \[ \rho_\lambda\bigl(\Re\Psi_{Q_\lambda},\ \ell_\lambda\Im\Psi_{Q_\lambda}\bigr) \] over its boxes. Since the supports within one colour are disjoint, the sum over the orthogonal colour blocks is a unit section $Z_T$ of the twisted sphere bundle of size $b_*=d_ch^6$. The cutoff supports allow an intrinsic extension of each summand by zero. This fixes the auxiliary sphere size used in the periodic coefficients before $T$ varies. In a local unit of $L$, let $Y_1$ be the new coefficient field of that unit given by $\iota_L$. Then $Z_T$ satisfies the Hopf equation with speed $T$ up to an error \[ T^{-1}Y_1Z_T-X(Z_T)=O(T^{-1}), \] where $X(Z_T)$ denotes the twisted generator at the image point. The estimate holds in all fixed scaled $C^J$ norms. Indeed, on one summand the difference is \[ 2\pi\ii\,\rho_\lambda\,\bigl\langle \iota_L(\ell_\lambda),\,Q_\lambda/T-\iota_L(\ell_\lambda) \bigr\rangle\,\Psi_{Q_\lambda}, \] up to conjugation, and $Q_\lambda/T-\iota_L(\ell_\lambda)=O(T^{-1})$ on the box by the rounding and the variation of $\iota_L(\ell_\lambda)$ over a slow box of size $O(T^{-1})$. Scaled slow derivatives preserve this bound, as do scaled derivatives of the cutoffs and of the normalized powers, by their $O(T^j)$ bounds in each fixed derivative order $j$. The sums have bounded overlap. The sign of the unit gives the same intrinsic estimate on each overlap. Use the ordinary $T$-th powers on \emph{all} old input factors, and denote the powered sample by $x_T$. Use the twisted $K_0$-th power $P_{K_0}$ on $Z_T$, with $K_0=\eps^{-1}$. Evaluating every coefficient at these samples gives the finite map \[ F(x_T)+\sum_{r=1}^{R_*}\eps^rF_r\bigl(x_T,P_{K_0}Z_T\bigr). \] It is smooth across the boundary of the wave support, since all added coefficients vanish outside its compact slow buffer. Along the normalized coefficient fields $Y_{A_1(\cdot)}/T$, the old derivative is exactly the derivative of the old profile, and where the coefficients are supported we have $\beta=1$, so the new line part obeys the approximate transport above. Thus the formal first Gram matrix \eqref{eq:newlinemetric}, with $\alpha,p$ evaluated at $x_T$, approximates the first Gram matrix of this map. Outside the wave support the map depends only on the old inputs, $c=1$, and the old metric is given exactly by the same formula. On the transition region of $\beta$ the potential $s$ also vanishes. To quantify the estimates we use local torus angle charts, projective charts and slow charts, all scaled by $1/T$ on compact buffered nets. At each slow point, a local frame of $A_1(V_0)$ and a complementary frame in the angle coordinates straighten the coefficient fields; these frames are chosen on finitely many slow charts before rescaling. The coefficient map has a positive lower bound for its least singular value on the compact base, so these changes of coordinates and their inverses have bounded jets in the scaled charts. The chart transitions and their inverses are uniformly bounded at every fixed order, with uniformly bounded overlap of the enlarged charts. In these charts the derivatives of order $J$ of the reference map are $O(1+\eps^{1-J})$, and the defect of the first Gram matrix is \[ O(\eps^{R_*-J})+O(\eps^{-B_J}/T) \] in scaled order $J$, for a finite exponent $B_J$. This follows by the chain rule from the finite recurrence and from the scaled transport estimate together with its differentiated versions. The same comparison of first and second columns retains a positive phase margin for small $\eps$ and large $T$, since the compact limiting test on $\Sym^2H$ is independent of $\eps$ and $T$. On the compact complement we use the positive phase margin of the old map. The target metric \eqref{eq:newlinemetric} has an exact description by a change of source. In a unit of $L$, translate the angle torus along the line generator $Y_c=cY_0+\beta Y_1$ by the variable time $c\,s(x_T)/T$, and extend this map by the identity off the support of the potential. The map is intrinsic under reversal of the unit, since $s$ and $Y_c$ change sign together. Since the old power multiplies old derivatives by $T$ and the new fields annihilate the potential, its differential along $H$ and $L$ adds $c\alpha(x_T)$ and $c^2p(x_T)$ times the unit line, respectively. In particular, the rank-one determinant is $1+c^2p(x_T)>9/10$, because $\abs p\leq\abs{\dd_{V_0}s}<1/10$. On each compact angle torus the map is a local diffeomorphism of degree one, since scaling the variable time gives a homotopy to the identity. Hence it is a diffeomorphism $\Theta_T$, and it preserves the coefficient distribution. The pullback by $\Theta_T$ of the unit metric of $V_0$ is exactly \eqref{eq:newlinemetric}. This target metric is uniformly elliptic and has uniformly bounded scaled derivatives of every fixed order, by the bound on $\dd_{V_0}s$ and the estimates for the old powers. We now apply the polynomial-size entrance estimate of Proposition~\ref{prop:repair}, in the form detailed in Section~\ref{sec:metricproduction}, in these uniformly controlled charts. The fixed reference derivative order is chosen first. The exponent of the reference bound in $\eps^{-1}$ is independent of the truncation depth, and the phase lower bound has exponent zero. Thus an error $O(\eps^K)$ in order $440$, for some finite $K$, suffices for a phase-preserving exact correction. Take $R_*>K+440$, then $\eps$ a sufficiently small reciprocal integer, and then $T$ sufficiently large for all of the finitely many entrance bounds. In the unscaled coefficient fields the corrected map has metric $T^2\Theta_T^*h_0$, where $h_0$ is the specified unit metric. Divide it by $T$ and precompose with $\Theta_T^{-1}$. The resulting map is exactly isometric along $A_1$. Since $\Theta_T^{-1}$ preserves the leaves, the normal second form is transported by its first leaf derivative: the chain-rule term involving its second derivative lies in the span of the first columns and is removed by the normal projection. Phase regularity is therefore retained, and the target remains $\R^N$ throughout. \end{proof} \section{Global assembly} \label{sec:wholeassembly} We choose the slow radii on each tube before choosing the common Hopf size. The attaching sphere lies in an inner plateau. Farther out, choose a smooth slow angle $\omega$ which decreases from $\pi/2$ to zero within the band of Section~\ref{sec:pairedtest}. Set $c=\cos\omega$ and $\chi=\sin\omega$ on that band, and continue them by $c=0$, $\chi=1$ on the filled side and by $c=1$, $\chi=0$ on the far side. We keep an outer buffer of the band beyond $\supp\chi$. Choose a slow cutoff $\rho$ equal to one on $\supp\chi$ and on its buffer of fixed $H$, and turning off on a farther collar on which $\chi=0$ and $L$ is still defined. The components of the core are disjoint, so all such cutoffs can be chosen simultaneously. For each fixed Hopf size we apply the exact metric production of Section~\ref{sec:metricproduction} on the doubled exterior, with passive weight $d=\chi$. On the annuli the passive value $fa$ takes values in the old line $L$. We use the same exponent $k\geq4$ and parameter $\tau=L_0^{-1}$ as in that production, and write $b_\tau$ for the evaluation of $fa$ after the old and active powers of frequency $L_0$. The resulting exact map $F_\tau$ has partial metric \[ h_0-\tau^{2k}\chi^2\abs{\dd_{V_0}b_\tau}^2 \] on the exterior, where $V_0=TM$, and it has a positive phase margin by the exterior construction. In a unit $\ell$ of $L$, let $Y_0$ be its old line generator, and define the forward orbit translation along $Y_0$ by the variable time $t_\tau=\rho\,\tau^kb_\tau$, extended by the identity off the collar of the potential. Its first angle derivatives are $O(\tau^{k-1})$. The derivative of the translation on the angle torus is the identity plus a rank-one map, with determinant $1+Y_0t_\tau>0$ for small $\tau$. By degree one, it is a diffeomorphism of each torus. Denote its inverse by $\Psi_\tau$ and put $s=-t_\tau\circ\Psi_\tau$. This is a well-defined $L$-valued time, since reversing the unit reverses $t_\tau$ and $Y_0$ together. The leaf differential of $\Psi_\tau$ is $v\mapsto v+(\dd_{V_0}s(v))\ell$, so \[ \Psi_\tau^*h_0=h_0+2\ell^*\odot\dd_{V_0}s+(\dd_{V_0}s)^2 . \] Throughout the support of the passive weight we have $\rho=1$, so $\tau^{2k}\Psi_\tau^*\abs{\dd_{V_0}b_\tau}^2=\abs{\dd_{V_0}s}^2$, and $1-\chi^2=c^2$. Thus the old exterior map $\widetilde F_\tau=F_\tau\circ\Psi_\tau$ has exactly the metric \[ h_0+2\ell^*\odot\dd_{V_0}s+c^2(\dd_{V_0}s)^2 . \] This is \eqref{eq:oldlinemetric}. In the inner plateau its $H$-metric is exactly the unit metric. Its phase regularity on the exterior is preserved by the line translation, since $\Psi_\tau$ preserves the leaves and the chain-rule term of its second derivative is tangential. On the collar where only the potential is present, $c=1$ and the same formula holds, including where the potential is turned off. \subsection{Boundary comparison with the passive term} We verify the augmented rank throughout the wave buffer, including the transition region $0<\chi<1$. \begin{lemma}\label{pl:boundaryclass} Let $k\geq4$. For all sufficiently small $\tau$, the columns $J_2(H)\widetilde F_\tau$ and the surviving first column $\dd_L\widetilde F_\tau$ are linearly independent throughout the wave buffer. Their augmented columns lie in the collar class of \eqref{eq:pairedactual}, and on the attaching sphere in the class $J_\Sigma$. \end{lemma} \begin{proof} We work at the inverse-translated point and its independently powered sample. Modulo the first columns, the Hessian of the composition is the Hessian of $F_\tau$ evaluated on the two moved inputs; the term from the second derivative of the translation lies in the first span. For $h\in E_v$, \[ \dd\Psi_\tau(h)=h-\frac{\dd_ht_\tau}{1+Y_0t_\tau}\,\ell =h-\tau^{k-1}S(h,a)+O(\tau^{2k-2}), \] where we have used that $\rho=1$ on this buffer whenever the passive term is nonzero. For $b\in A$ the change is $O(\tau^{k-1})\ell$. Therefore the first-order change in the mixed column of the reference map is $-\tau^{k-1}D^2B(S(h,a),b)$. The mixed old--active columns of the reference map vanish, so the change coming from the moved $b$ alone vanishes. The cross term from both substitutions is $O(\tau^{2k-2})$, and changing the inputs of the perturbed second columns costs $O(\tau^{2k-3})$. In the unmoved mixed Hessian the leading retained term is $\tau^{k-2}I_xU(h,d_b^ya)$. The first term visible in the four cofactor rows is $\tau^{k-1}J_x(b)U(h,a)$: the other term with one fast derivative has zero cofactor projection by the separate reference blocks of Section~\ref{sec:boundarycolumns}. Every higher normal profile starts at order $k+1$; its term with two fast derivatives lies in $I_xW$, so after projection to the four rows it has order at least $k$. The lifts of the pure pairings start at order $2k-2$, so after two derivatives their error in the four rows is $O(\tau^{2k-4})$. The lifts of the first pairings and the exact metric correction are of higher order still. The retained projection of the translation term of the reference map vanishes, since that term is an active second column; its cofactor projection is $-\tau^{k-1}M_x(S(h,a)\otimes b)$. Replacing the reference projector by the projector onto the complement of the first and pure columns of $\widetilde F_\tau$ costs $O(\tau^{2k-4})$: by the formula for the Gram inverse the projector moves by $O(\tau^{k-2})$, and the mixed columns being projected are $O(\tau^{k-2})$. These estimates are uniform over the buffer, since the slow weight is annihilated by all leaf derivatives. Scale the projected retained rows by $\tau^{-(k-2)}$ and the four cofactor rows by $\tau^{-(k-1)}$. Since $k\geq4$ gives $2k-4\geq k$, the resulting mixed columns are within $O(\tau)$ of the uniform paired injection \eqref{eq:pairedactual}. Moreover, $\dd_L\widetilde F_\tau$ is close to the fourth active first column of the reference map, which is independent of the remaining reference columns. This proves the simultaneous independence of $J_2(H)\widetilde F_\tau$ and $\dd_L\widetilde F_\tau$ through the whole buffer. The same estimate gives the collar class of \eqref{eq:pairedactual}, by linear interpolation over the $O(\tau)$ error after scaling, by undoing the positive row scalings, and by undoing the short transport of the base projector. On the attaching sphere the resulting comparison is with $J_\Sigma$. The inverse translation can be removed from the formal endpoint by its time isotopy. The powers and shifts of the independent samples are allowed throughout the comparison, so no new identification of normal bundles is needed after substitution. \end{proof} \subsection{Filling the core and completing the line} On the attaching plateau, where the $H$-metric is the unit metric, orthogonalize the surviving first column against the first $H$-columns and normalize it positively. This does not change the class of the injection. Extend the augmented columns of $\widetilde F_\tau$ formally over an inward collar: make the radial coordinate stationary at a comparison sphere, run the class homotopy of Lemma~\ref{pl:boundaryclass} to $J_\Sigma$, and then use the fixed disk extension of Section~\ref{sec:referenceproduct}. With smooth stationary end parameters this gives a smooth extension which is pulled back from the core $P$ near the zero section and agrees with the columns of $\widetilde F_\tau$ near the outer collar. Corollary~\ref{cor:diskfill} now gives an $H$-free filling with unit $H$-metric which equals the exterior map on a smaller open collar, together with an isometric augmented line and two normal copies of $L$ orthogonal to all augmented columns. In particular, the old map obtained by gluing is a single smooth map on the original Hopf product over $M$. On the inner collar write the surviving derivative as \[ \dd_L\widetilde F_\tau=\dd_H\widetilde F_\tau\, T+\mu\, e_\perp , \] where $e_\perp$ is the positively orthogonalized unit line and $\mu>0$. After a radial retraction, cut off $T$ and $\log\mu$ to zero inwards by a slow cutoff, and use the same formula with the filled $H$-columns and their augmented unit line. The augmented span is unchanged. This gives a column $e$ which agrees with the old line column on the exterior side, as required where $c>0$; inside, $c=0$. The two normal copies extend outwards through the wave buffer by orthogonal transport in the rank-five complements of the augmented spans. To match them smoothly on an open overlap, we first trivialize those complements by radial transport, and reparametrize the given family radially so that it is stationary before the end of the overlap. Thus both copies are smooth isometric inclusions on a neighbourhood of the wave support. Extend $s$ inwards by a slow radial retraction and a cutoff in the region $c=0$, where only the $H$-metric is prescribed. The first original angle derivatives of its incoming germ are $O(\tau^{k-1})$ on the compact old host, by the bounds for the powers and the translation. The radial extension has derivatives of the same size and all fixed coefficients are bounded, so the hypothesis $\abs{\dd_{V_0}s}<1/10$ holds for the extended potential when $\tau$ is small. Its support lies in a fixed slow tube. Away from the wave neighbourhood the old map is phase regular, since it is the corrected, possibly translated, exterior map. Its metric there is exactly the prescribed formula, and it is the unit metric outside the tube of the potential. All hypotheses of Proposition~\ref{prop:linecompletion} therefore hold on this single old host, with $N=s_n+5\geq s_n+1$. Take $\iota_L\colon L\to\R^6$ to be the original inclusion of $L$ into the four active coefficient slots followed by two zeros, and choose the slow cutoff $\beta$ once on the tubes. Since there are finitely many components, all constants, frequencies and slow cutoffs can be chosen simultaneously. Proposition~\ref{prop:linecompletion} gives a single exact phase-regular unit partial isometry over $M$ into $\R^{s_n+5}$, defined on a product of ordinary Hopf spheres. The final injective coefficient map is fixed independently of their common size, and its coefficient bundle is isomorphic to $TM$. If the core is empty, the exterior production already supplies this compact model, with zero passive term and no additional inputs. All these arguments apply for every sufficiently large common sphere size. Finally we apply Theorem~\ref{thm:model} with slow base $M$, with the metric on the coefficient bundle transported from the chosen auxiliary Riemannian metric, and with the identity as classifying map; the dimension condition $N\geq s_n+2$ holds for $N=s_n+5$. (In its proof, the passage from the auxiliary metric to $g$ uses Lemma~\ref{p03:lem:factorization} and Lemma~\ref{p03:lem:hopf-pullback}, followed by Propositions~\ref{prop:repair} and~\ref{prop:embed}.) The theorem gives the asserted smooth isometric embedding for the independently prescribed metric $g$, in the same $s_n+5$ coordinates. The prescribed image ball and the phase regularity of the embedding also follow from that theorem. This proves Theorem~\ref{thm:main}. \section*{Acknowledgements} Generative AI was used in developing the proofs. \begin{thebibliography}{BCGGG91} \bibitem[C27]{Cartan27} \'E.~Cartan, \emph{Sur la possibilit\'e de plonger un espace riemannien donn\'e dans un espace euclidien}, Ann. Soc. Polon. Math. \textbf{6} (1927), 1--7. \href{https://mccuan.math.gatech.edu/courses/6455/Cartan1927.pdf}{Original volume, pp.~1--7}. \bibitem[EM02]{EM} Y.~Eliashberg and N.~Mishachev, \emph{Introduction to the $h$-Principle}, Graduate Studies in Mathematics, vol.~48, American Mathematical Society, Providence, RI, 2002. \href{https://doi.org/10.1090/gsm/048}{doi:10.1090/gsm/048}. \bibitem[G70]{Greene70} R.~E.~Greene, \emph{Isometric embeddings of Riemannian and pseudo-Riemannian manifolds}, Mem. Amer. Math. Soc., no.~97 (1970). \href{https://doi.org/10.1090/memo/0097}{doi:10.1090/memo/0097}. \bibitem[G86]{GromovPDR} M.~Gromov, \emph{Partial Differential Relations}, Ergebnisse der Mathematik und ihrer Grenzgebiete (3), vol.~9, Springer-Verlag, Berlin, 1986. \href{https://doi.org/10.1007/978-3-662-02267-2}{doi:10.1007/978-3-662-02267-2}. \bibitem[G17]{Gromov17} M.~Gromov, \emph{Geometric, algebraic, and analytic descendants of Nash isometric embedding theorems}, Bull. Amer. Math. Soc. (N.S.) \textbf{54} (2017), no.~2, 173--245. \href{https://doi.org/10.1090/bull/1570}{doi:10.1090/bull/1570}. \bibitem[GR70]{GR70} M.~L.~Gromov and V.~A.~Rokhlin, \emph{Embeddings and immersions in Riemannian geometry}, Russian Math. Surveys \textbf{25} (1970), no.~5, 1--57. \href{https://doi.org/10.1070/RM1970v025n05ABEH003801}{doi:10.1070/RM1970v025n05ABEH003801}. \bibitem[G\"u89]{Gunther89} M.~G\"unther, \emph{Zum Einbettungssatz von J.~Nash}, Math. Nachr. \textbf{144} (1989), 165--187. \href{https://doi.org/10.1002/mana.19891440113}{doi:10.1002/mana.19891440113}. \bibitem[G\"u91]{Gunther91} M.~G\"unther, \emph{Isometric embeddings of Riemannian manifolds}, Proceedings of the International Congress of Mathematicians, Vol.~I, II (Kyoto, 1990), Math. Soc. Japan, Tokyo, 1991, pp.~1137--1143. \bibitem[J26]{Janet26} M.~Janet, \emph{Sur la possibilit\'e de plonger un espace riemannien donn\'e dans un espace euclidien}, Ann. Soc. Polon. Math. \textbf{5} (1926), 38--43. \href{https://rcin.org.pl/Content/15034/PDF/WA35_31746_cz159-r1926-T5_Annales-art-4.pdf}{Original article}. \bibitem[MAdP21]{MA} J.~Mart\'inez-Aguinaga and \'A.~del~Pino, \emph{Convex integration with avoidance and hyperbolic $(4,6)$ distributions}, preprint (2021), \href{https://arxiv.org/abs/2112.14632}{arXiv:2112.14632}. \bibitem[MS74]{MilnorStasheff} J.~W.~Milnor and J.~D.~Stasheff, \emph{Characteristic Classes}, Annals of Mathematics Studies, no.~76, Princeton University Press, Princeton, NJ, 1974. \href{https://doi.org/10.1515/9781400881826}{doi:10.1515/9781400881826}. \bibitem[N56]{Nash56} J.~Nash, \emph{The imbedding problem for Riemannian manifolds}, Ann. of Math. (2) \textbf{63} (1956), no.~1, 20--63. \href{https://doi.org/10.2307/1969989}{doi:10.2307/1969989}. \bibitem[WY26]{YW-28} Y.~Wang and W.~Yang, \emph{Every closed smooth $n$-manifold embeds isometrically in $\R^{n(n+1)/2+28}$}, Hexagon Math (2026), \href{https://hexagonmath.org/2611.00153v1}{hexagon:2611.00153v1}. \bibitem[Wu22]{Wu} L.~Wu, \emph{Characteristic cycles associated to holonomic $\mathcal D$-modules}, Math. Z. \textbf{301} (2022), no.~2, 2059--2098. \href{https://doi.org/10.1007/s00209-022-02974-0}{doi:10.1007/s00209-022-02974-0}. \href{https://arxiv.org/abs/2011.14527}{arXiv:2011.14527}. \bibitem[YW26a]{YW-local} W.~Yang and Y.~Wang, \emph{Smooth local isometric and conformal embeddings in optimal dimension}, Hexagon Math (2026), \href{https://hexagonmath.org/2610.00378v1}{hexagon:2610.00378v1}. \bibitem[YW26b]{YW-parallel} W.~Yang and Y.~Wang, \emph{Isometric embeddings of parallelizable and stably parallelizable Riemannian manifolds}, Hexagon Math (2026), \href{https://hexagonmath.org/2610.00690v1}{hexagon:2610.00690v1}. \bibitem[Z16]{ZerLocal} N.~Zerg\"ange, \emph{Isometrische Einbettung von geschlossenen Riemannschen Mannigfaltigkeiten}, preprint (2016), \href{https://arxiv.org/abs/1607.04259}{arXiv:1607.04259}. \bibitem[Z17]{ZerStab} N.~Zerg\"ange, \emph{Stability analysis of isometric embeddings}, preprint (2017), \href{https://arxiv.org/abs/1712.02766}{arXiv:1712.02766}. \end{thebibliography} \end{document}