\subsection{A critical conic in every polynomial degree} \label{sec:hyperbolicsuspension} The conic and hyperbola rows of the preceding table are manifestations of the same calculation. It also works when one of the cubic terms is replaced by a polynomial of arbitrary degree. We prove that stronger statement to make clear what the mechanism actually uses. \begin{theorem}\label{thm:hyperbolicsuspension} Retain a full nondegenerate affine Gaussian quadratic \[ q=\tfrac12(Ax^2+By^2+mw^2+2Cxy+2Dxw+2Eyw) +\ell_x x+\ell_y y+\ell_w w+q_0. \] Let $AB\ne0$ and \[ P=wxy+F(w)+k,\qquad F(0)=0,\quad F'(0)=h\ne0. \] Allow any polynomial degree $d_F=\deg F\ge1$, requiring $m\ne0$ when $d_F=1$. Then $P$ is not pure. Every original marked Gaussian period is either zero or is not a finite exponential-polynomial. The generic top twisted rank is $d_F+3$ for $d_F\ge2$ and five for $d_F=1$. No Morse assumption on the Laurent phase obtained by restricting $q$ to the conic, or on the isolated critical points of $P$, is needed. \end{theorem} The nonzero quadratic poles at both ends of the conic are essential to the rank-four construction below. The condition $AB\ne0$ states this explicitly. Nondegeneracy of the full Gaussian alone would not imply it. This theorem concerns complex polynomials and valued-field formal frames; it does not assert an all-ring extension. \begin{proof} We first justify the full rank. Fix a generic $t\ne0$ and set $S=tP-q$. Its gradient has a linear lower bound at complex spatial infinity. To see this, suppose along an unbounded sequence $\|\nabla S\|=o(R)$, where $R=\|(x,y,w)\|$. If $w$ stays bounded, pass to a limit $w_0$ and normalize $(x,y)$ by $R$. The first two gradient equations give \[ -AX+(tw_0-C)Y=0,\qquad (tw_0-C)X-BY=0. \] At least one of $X,Y$ is nonzero, and $AB\ne0$ makes both nonzero. But the last gradient equation, divided by $R^2$, gives $tXY=0$, a contradiction. If $w$ is unbounded, the first two equations have a matrix with inverse $O(1/|w|)$. Hence $x,y=O(1)+o(R)/|w|$. First $R\asymp |w|$, then $x,y=O(1)$. The last equation is therefore \[ tF'(w)-mw+O(1)=o(|w|). \] This is impossible for $d_F\ge3$, for $d_F=2$ with $2t[w^2]F-m\ne0$, and for $d_F=1$ with $m\ne0$. The same proof is uniform on compact subsets of these coefficient conditions. Thus all phase-critical points stay in one ball during any compact coefficient deformation inside that domain. They are isolated, and their total Milnor number is constant by conservation of local gradient-intersection multiplicity. Pham's tame Milnor decomposition~\cite[Part~I, Sections~1--4]{Pham} identifies that number with the full twisted rank. For fixed $t$, the coefficient domain just used is the complement of finitely many complex algebraic hypersurfaces, and is path connected. We may compute the number at a generic diagonal Gaussian with zero means, retaining $ABm\ne0$, and a generic $F$. The critical equations there are \[ twy=Ax,\qquad twx=By,\qquad t(xy+F'(w))=mw. \] There are $d_F-1$ simple solutions with $x=y=0$ if $d_F\ge2$, and one if $d_F=1$. The other solutions have $w=\pm\sqrt{AB}/t$. At each of these two values, $x,y$ are proportional and their nonzero product is prescribed by the last equation; generically this gives two simple solutions. The counts are $d_F+3$ and five, respectively. Uniform coercivity and the connected deformation justify this specialization; a convenient critical-point count by itself would not justify the rank for a degenerate member. We now construct a full formal frame. Set $r=t^{-1/2}$ and write $F(w)=hw+bw^2+\sum_{j\ge3}f_jw^j$. Near the entire critical conic $w=0$, $xy=-h$, use the exact coordinates \[ y=(rV-h)/x,\qquad w=rU. \] They give \begin{align} tP-q&=k/r^2+UV+bU^2+\psi(x)+E_r(x,V,U), &E_r\in r\C[x,x^{-1},V,U][r],\label{eq:conicphase}\\ dx\,dy\,dw&=r^2\,\frac{dx}{x}\,dV\,dU,\label{eq:conicvolume}\\ \psi(x)&=-\frac A2x^2-\frac{Bh^2}{2x^2} -\ell_x x+\frac{h\ell_y}{x}+Ch-q_0.\notag \end{align} All Gaussian cross terms and means are in $\psi$ or $E_r$. The normal quadratic $UV+bU^2$ has Hessian determinant $-1$ for every $b$. Its normalized formal Gaussian functional is well defined and takes one to one. The logarithmic Laurent quotient for $\psi$ has rank four, with basis $1,x,x^2,x^3$. Here is a direct proof, including independence. Its relation on $x^j$ is \[ jx^j-Ax^{j+2}-\ell_xx^{j+1} -h\ell_yx^{j-1}+Bh^2x^{j-2}=0. \] The top coefficient eliminates all exponents at least four without creating a negative exponent; the bottom eliminates all negative exponents without creating an exponent above three. If the image of a nonzero Laurent polynomial under $x\partial_x+x\psi_x$ were supported in $[0,3]$, its highest exponent would be at most one and its lowest at least two, which is impossible. Let $\lambda_0,\ldots,\lambda_3$ be the dual basis on this quotient, tensored with the normal Gaussian functional. For a polynomial amplitude $a(x,y,w)$, define \begin{equation}\label{eq:coniccharacters} L_i(a)=e^{k/r^2}r^2\lambda_i\!\left( e^{E_r}a(x,(rV-h)/x,rU)\right). \end{equation} Every coefficient is a finite Laurent-polynomial calculation, so this is a well-defined formal series. Multiplication by $e^{E_r}$ conjugates the fixed twisted differential to the perturbed one. Consequently these four functionals annihilate all spatial exact forms. The moving volume is retained in the factor $r^2$. For an explicit parameter check the coordinate velocity in the original variables is \[ Z=\left(0,\frac{xy+h}{rx},\frac wr\right), \qquad \operatorname{div}Z=2/r. \] Cartan's identity $\mathcal L_Z=d\iota_Z+\iota_Zd$ shows that the pullback derivative differs from the original parameter connection only by a spatial exact form and this volume derivative. Thus \eqref{eq:coniccharacters} are horizontal for the original full module. In particular, for every fixed polynomial $a$, \begin{equation}\label{eq:conicseedorder} L_i(a)\in e^{kt}r^2\C[[r]]. \end{equation} This bound holds for all original markers, not just for the unit. Suppose first $d_F\ge2$. The other critical points of $P$ are $x=y=0$, $F'(w_p)=0$. Each $w_p$ is nonzero because $h\ne0$. If it has multiplicity $\mu_p$, analytic coordinates give \[ P=k+F(w_p)+s z+y_0^{\mu_p+1}. \] Indeed absorb the nonvanishing factor $w$ into $x$ or $y$, then use the one-variable power coordinate for $F$ at $w_p$. The explicit characters \eqref{eq:ordinarycharacters}, including their integration-by-parts and time recurrences, apply to this exact local model. Their prefactors are \[ e^{(k+F(w_p))t}t^{-1-j/(\mu_p+1)}, \qquad 1\le j\le\mu_p. \] The Gaussian and coordinate Jacobian are analytic units in that construction. These characters retain multiple roots and equal critical values. Their total number is $\sum_p\mu_p=d_F-1$. The $d_F+3$ characters just written are independent on polynomial amplitudes. Put $B_0(w)=\prod_p(w-w_p)^{\mu_p}$. For the four conic rows use $x^jB_0(w)$, $0\le j\le3$. Their leading conic matrix is $B_0(0)I_4$. Their leading finite jets vanish. For the remaining rows choose polynomials in $w$ with value zero at $w=0$ and the separate Hermite jets of orders $0,\ldots,\mu_p-1$ at the nonzero $w_p$. The Chinese remainder theorem supplies them. Their conic values have an additional positive power of $r$. Their normalized finite evaluation matrix is the invertible local jet map, multiplied by analytic units. After dividing each character column by its scalar prefactor, the leading matrix is therefore block diagonal and invertible. Since its size is the proved full rank, these are a full formal frame. No uncomputed image kernel or infinity block remains. For $d_F=1$ there is instead one further character. In the exact coordinates \[ x=X/t,\qquad y=Y/t,\qquad w=(h/m)t+W, \] the phase is \begin{align*} tP-q={}&\frac{h^2}{2m}t^2+(k-h\ell_w/m)t+H(X,Y,W)+O(t^{-1}),\\ H={}&\frac hm XY-\frac hm(DX+EY)-\frac m2 W^2-\ell_w W-q_0. \end{align*} The correction is polynomial in $t^{-1},X,Y,W$, with positive $t^{-1}$ order. The Hessian determinant of $H$ is $h^2/m\ne0$. Formal Gaussian evaluation gives a horizontal character with nonzero quadratic action and volume prefactor $t^{-2}$. To check independence set $w_*=(h/m)t$. Use $x^j(1-w/w_*)$, $0\le j\le3$, and $w/w_*$ as five polynomial amplitudes with rational parameter coefficients. The former have the original conic leading matrix and zero leading escape values; the latter has zero leading conic values and nonzero leading escape value. This gives the full rank-five frame. In either degree case, after finite ramification the frame splits the module into rank-one factors. Every zero-exponential functional has positive order on every original polynomial marker, by \eqref{eq:conicseedorder} and the displayed finite prefactors. Lemma~\ref{lem:positiveformal} rules out a nonzero constant original marked period. A constant-coefficient operator reduces any nonzero finite exponential-polynomial to $c e^{\beta t}$; it replaces the marker by another original polynomial in $P$. Shifting $k$ by $-\beta$ and normalizing the marker gives the same contradiction. This proves the theorem. \end{proof} \begin{corollary}\label{cor:ordinaryconics} All the critical-conic and critical-hyperbola rows in the ordinary-node table are not pure. Proposition~\ref{prop:axisexclusion} treats its remaining center-axis row. \end{corollary} \begin{proof} For $h\ne0$ and $a=b=c=d=e=f=0$ the residual is $T(uv+h)+k$. Apply the $d_F=1$ case with $w=T$, $x=u$, $y=v$. The no-frame conditions give $ABm\ne0$. For $h=0$, $a=b=e=0$, $f\ne0$, permute the coordinates and write the residual as \[ v(Tu+d v^2+c v+f)+k. \] Take $w=v$, $x=T$, $y=u$, and $F(w)=d w^3+c w^2+f w$. The quadratic coefficients along $T,u$ are nonzero: the original $T$ square coefficient is nonzero, and the $u$ node has $F_2=0$, so its Gaussian square coefficient is nonzero by the no-frame condition. If this $F$ is linear, the remaining $v$ node also has $F_2=0$, giving the extra $w$ square coefficient required in degree one. The symmetric hyperbola row is identical after swapping $u,v$. These are all the indicated rows. \end{proof}