\subsection{A formal leading-coefficient criterion}\label{sec:stationaryballs} For a polynomial phase, the remaining shrinking argument can be replaced by a valuation calculation. The construction below includes critical points that collide, remain multiple, or escape to infinity. The key is to group them before making a Gaussian approximation. Write $v$ for the valuation of a formal parameter $r$, normalized by $v(r)=1$. The valuation of a polynomial is the least valuation of its coefficients. We use the formal Newton--Puiseux theorem: finitely many algebraic elements over $\C((r))$ belong to one field $\C((r^{1/N}))$~\cite[Corollary~1.10]{MannaaCoquand}. Characteristic zero makes each minimal polynomial separable; no convergence hypothesis is needed. \begin{lemma}[Polynomial separation of formal balls]\label{lem:ballseparation} Let finitely many disjoint balls have centers $c_B$ in a Puiseux field and radii specified by $v(x-c_B)\geq\rho_B$. After one common ramification, put $\delta_B=r^{\rho_B}$. For any one ball $B$ and integer $j\geq0$, there is a polynomial $q_{B,j}(x)$ over the Laurent field whose substitution $x=c_C+\delta_C y$ has integral coefficients for every ball $C$, and whose reduction is $y^j$ on $B$ and zero on every other ball. \end{lemma} \begin{proof} The finitely many valuations $v(c_B-c_C)$ give a finite tree of clusters of centers. At a node with scale $r^d$, the coordinate $u=(x-c)/r^d$ has distinct constant reductions $a_1,\ldots,a_s$ on its children. The Lagrange polynomial \[ e_i(u)=\prod_{k\ne i}\frac{u-a_k}{a_i-a_k} \] reduces to one on child $i$ and zero on every other child. Start at the chosen leaf with $((x-c_B)/\delta_B)^j$. Induct upward. Suppose a polynomial $h$ already has the required reductions on the chosen child. On each of the other child balls its coefficients have some finite lower valuation. The restriction of $e_i$ there has strictly positive valuation. Consequently $e_i^M h$, for a sufficiently large fixed integer $M$, has positive coefficient valuation on all those children. On the chosen child $e_i^M=1+$ positive-order terms, so the existing reductions are unchanged. There are finitely many nodes, and this constructs the required polynomial. A node with just one leaf requires no filter. This also explains why independently normalized Lagrange polynomials, with no control of their growth on other scales, would not suffice. \end{proof} \begin{theorem}[Formal stationary balls]\label{thm:stationaryballs} Let $\varphi(r,x)\in\C((r))[x]$ have degree $D\geq2$, and let $h(r)\in\C((r))\setminus\{0\}$. Put \[ M=\C((r))[x]/(\partial_x+\varphi_x)\C((r))[x],\qquad \nabla[q]=[q_r+(\varphi_r+h'/h)q]. \] After a finite ramification, this module has a full frame of rank-one formal exponential functionals. If $a_D$ is the leading coefficient of $\varphi$, every zero-exponential functional in this frame takes the seed $[1]$ to a series of valuation at least \begin{equation}\label{eq:valuationbound} v(h)-\frac{v(a_D)}D. \end{equation} In particular, if $Dv(h)>v(a_D)$, the differential submodule generated by $[1]$ has no horizontal quotient taking $[1]$ to the constant one. \end{theorem} \begin{proof} The two twisted derivatives commute, so the parameter connection is well defined on the quotient. A nonzero polynomial of degree $m$ has spatial twisted derivative of degree $m+D-1$. Reduction of successive leading terms therefore gives a basis $1,x,\ldots,x^{D-2}$ and dimension $D-1$. Split $\varphi_x$, retaining root multiplicities. For a critical root $\xi$, expand \[ \varphi(\xi+u)=c_\xi+\sum_{j=2}^D a_{\xi,j}u^j, \qquad \rho_\xi=\max_{2\leq j\leq D,\ a_{\xi,j}\ne0} \frac{-v(a_{\xi,j})}{j}. \] Enlarge the ramification so that all these scales are integral, and set $\delta_\xi=r^{\rho_\xi}$. The polynomial \[ \varphi(\xi+\delta_\xi y)-c_\xi \] has integral coefficients, no linear term, and at least one unit coefficient. Its reduction $F_\xi(y)$ has degree $d_\xi\geq2$. Here and below we may replace $r$ by a root of $r$ and renormalize the valuation; every displayed strict inequality is unchanged. As $a_{\xi,D}=a_D$, \begin{equation}\label{eq:balllower} \rho_\xi\geq-v(a_D)/D. \end{equation} \emph{The balls partition all critical roots.} If another critical root $\zeta$ satisfies $v(\zeta-\xi)\geq\rho_\xi$, translate the integral scaled polynomial by $(\zeta-\xi)/\delta_\xi$. All its coefficients remain integral, and the coefficient of degree $d_\xi$ remains a unit: higher-degree terms have positive valuation. The linear term vanishes because $\zeta$ is critical. It follows from the definition of the maximum that $\rho_\zeta=\rho_\xi$. Thus their balls are identical. If two of the balls intersect, one contains the other; the same argument makes them identical. Otherwise they are disjoint. The scaled derivative has reduction $F_\xi'$ of degree $d_\xi-1$. It therefore has exactly $d_\xi-1$ integral roots, with multiplicity. Here is a direct check of that count: factor it in a Puiseux field. An integral root contributes a linear factor to the reduction; for a root of negative valuation, factor out that root and the remaining factor reduces to a constant. Multiplicativity of the coefficient valuation leaves a nonzero constant scalar. Hence the degree of the reduction counts exactly the integral roots. The ball $B$ thus contains $d_B-1$ critical roots. Since $\deg\varphi_x=D-1$, \begin{equation}\label{eq:ballrankcount} \sum_B(d_B-1)=D-1. \end{equation} \emph{A functional for every dimension in a ball.} The fixed polynomial $F_B$ has twisted quotient \[ H_B=\C[y]/(\partial_y+F_B')\C[y] \] with basis $1,y,\ldots,y^{d_B-2}$. Let $\ell^0_{B,j}$ be its dual basis. Write \[ \varphi(\xi_B+\delta_B y)=c_B+F_B(y)+E_B(r,y), \qquad E_B\in r\C[[r]][y]. \] For a polynomial amplitude $q$ define \begin{equation}\label{eq:ballfunctional} \ell_{B,j}(q)=h\delta_B e^{c_B} \ell^0_{B,j}\!\left(e^{E_B(r,y)}q(\xi_B+\delta_B y)\right). \end{equation} Each coefficient uses finitely many polynomial terms. The exponential in this formula is purely formal. The product rule conjugates the spatial twisted derivative to $\partial_y+F_B'$, so the functional annihilates the defining relations of $M$. We spell out the parameter check because it retains the moving volume. At fixed $y$, put $w=\xi_B'+\delta_B'y$. Differentiating \eqref{eq:ballfunctional}, then subtracting its value on $q_r+(\varphi_r+h'/h)q$, leaves, inside the fixed functional, \[ e^{E_B}\left(wq_x+\frac{\delta_B'}{\delta_B}q+\varphi_xwq\right) =\frac1{\delta_B}(\partial_y+F_B') \left(e^{E_B}wq\right). \] It is zero. Thus the functionals are horizontal, including the Jacobian $\delta_B$ and normalization $h$. \emph{Completeness and the diagonal connection.} Choose the polynomials $q_{B,j}$ of Lemma~\ref{lem:ballseparation}, for $0\leq j0, \end{equation} its seed submodule cannot have a rational horizontal quotient with seed value one. In particular any horizontal formal functional at another point cannot take that seed to one. \end{corollary} \begin{proof} Apply the theorem at infinity. A rational horizontal quotient remains a quotient on extension to the formal Laurent field. A formal functional with seed value one restricts to such a rational quotient: its values on successive seed derivatives are $1,0,0,\ldots$. \end{proof} The inequality must be strict. For example $\varphi=-x^2/(2t^2)$ and $h=1/t$ have $v_\infty(h^2/a_2)=0$, and their normalized Gaussian period is a nonzero constant. Rescaling that functional gives seed value one, so the conclusion would fail if equality were allowed.