\subsection{A curve-length estimate} \begin{lemma}\label{lem:curvelength} A real algebraic curve of degree at most $D$ has length at most $\pi D\rho$ in a disk of radius $\rho$. Consequently its logarithmic length in $a\leq|x|\leq b$ is at most a constant times $D(1+\log(b/a))$. \end{lemma} \begin{proof} Average the number of intersections with lines over directions and apply the one-dimensional coarea formula. Outside exceptional line components the number is at most $D$; the components themselves satisfy the same estimate. In each dyadic annulus $\rho\leq|x|\leq2\rho$, divide the disk estimate by $\rho$ to bound $\int|dx|/|x|$ by $2\pi D$. Sum the dyadic annuli. \end{proof} \subsection{The logarithmic branch, including its removed component} For $\Delta\ne0$, $f\ne0$, choose $\eta^2=\Delta$ and set \[ \rho=\frac{e+\eta}{2f},\quad \sigma=\frac{e-\eta}{2f},\quad u=\frac{\rho c-b+1/t}{\eta},\quad v=\frac{b-1/t-\sigma c}{\eta},\quad k=uv-a/f. \] On the circular constraint $z=th(z,T)$, put $x=T+\rho z+u$, so that $T+\sigma z+v=k/x$. Then \[ z=\frac f\eta(x-k/x-u+v),\quad T=x-u-\rho z, \qquad \frac{dT}{1-th_Z}=\frac{dx}{t\eta x}. \] The coefficient of $t^{-2}$ in $k$ is $-1/\eta^2$, so the conic is generically smooth. The exponent $tG(z,T)-T^2/2$ is Laurent polynomial in $x$. On the formal Gaussian branch, $x=1/(\eta t)+O(1)$, which makes every Laurent power legitimate. If $f=0$, then $e\ne0$. Put $z_0=-c/e$ and $x=z-z_0$. With $A(z)=a+bz+dz^2$, the constraint gives \[ T=\frac{z/t-A(z)}{ex},\qquad \frac{dT}{1-th_Z}=\frac{dx}{tex}. \] This localization is valid if $c\ne0$, or if $c=0$ and $a\ne0$. If $a=c=0$, the chosen fixed point is instead $g=0$ identically; inverting $g$ would discard the very branch being tested. In that case the exact unit series is \[ \E_T\left[\frac{e^{tG(0,T)}}{1-t(b+eT)}\right]. \] Use $x=T-(1/t-b)/e$, now a valid Laurent unit. Its volume is $-dx/(tex)$ and its exponent is polynomial in $x$. These formulas cover both the smooth conic and the removed-component branch. We next connect these coordinates to the formal Gaussian functional. This will close every one-ended phase without an analytic estimate. \begin{lemma}[Formal logarithmic transport]\label{lem:logtransport} In each of the three charts above, write $X=X(t,T)$ for the formal coordinate, $D$ for the displayed circular denominator, and $B$ for $tG(g,T)$ or $tG(0,T)$ on the removed component. Put $\kappa=\eta,e,-e$, respectively. Then \[ J(q)=\E_T\!\left[\frac{q(X)e^B}{D}\right] \] is horizontal on the logarithmic twisted quotient with phase $\varphi$ and normalization $h=1/t$, and $J(1)=\Phi_1$. \end{lemma} \begin{proof} All Laurent powers of $X$ are legitimate by the unit checks above. The coordinate identities give \[ X_T=t\kappa X/D,\qquad \varphi(t,X)=B-T^2/2. \] Consequently Gaussian integration by parts gives $J((x\partial_x+x\varphi_x)q)=0$: the corresponding integrand, multiplied by $t\kappa$, is $(\partial_T-T)\{q(X)e^B\}$. For parameter horizontality set $w=1/D$ and $R=X_t/X_T$. Differentiating $w=X_T/(t\kappa X)$ gives the exact identity \[ w_t+w/t=\partial_T(wR). \] Subtracting $J(q_t+(\varphi_t-1/t)q)$ from $\partial_tJ(q)$ now gives the Gaussian expectation of $(\partial_T-T)\{wR q(X)e^B\}$, hence zero. These are formal coefficientwise identities, without an integral-convergence hypothesis. The seed equality is the circular contraction formula. \end{proof} \begin{proposition}\label{prop:oneendgaussian} Every chart above whose actual Laurent phase has at most one irregular end is not pure. The removed-component branch $a=c=f=0$, $e\ne0$, is therefore excluded for every polynomial $G$, without a degree bound. \end{proposition} \begin{proof} At most one irregular end means that $\varphi$ is polynomial in $x$ or in $x^{-1}$, including the constant case. Purity would give the horizontal functional in Lemma~\ref{lem:logtransport} seed value one, contrary to Corollary~\ref{cor:oneend}. The removed-component phase is always polynomial in $x$, irrespective of $\deg G$. \end{proof} \input{laurent_logarithmic_bound.tex} \begin{proposition}\label{prop:nondeglog} If $e^2-4df\ne0$, no polynomial $G$, of any degree, makes $P=Wh(Z,T)+G(Z,T)$ pure. \end{proposition} \begin{proof} The smooth-conic, linear-conic and removed-component charts are exhaustive. Their Laurent-unit checks and Lemma~\ref{lem:logtransport} give the formal Gaussian functional with seed value $\Phi_1$ and normalization $1/t$. Theorem~\ref{thm:laurentlog} excludes $\Phi_1=1$ in every chart, using its actual Laurent orders. No image-kernel calculation is used. \end{proof} \subsection{Why normalization and full rank matter} The following exact examples prevent several tempting shortcuts. For positive real $t$, \[ \frac1t\int_{\mathbb R}e^{-x^2/(2t^k)}\,dx =\sqrt{2\pi}\,t^{k/2-1}. \] The unique critical point is bounded and has zero action for every $k$, yet the normalized period decays for $k=1$, is constant for $k=2$, and grows for $k=3$. A critical-point count or a zero-action label alone gives no decay estimate. For the fixed phase $tx^k$, $k>2$, the critical point is degenerate and the decay scale is $t^{-1/k}$, not a Gaussian Hessian scale. The fixed quotient in Theorem~\ref{thm:stationaryballs} retains its full multiplicity $k-1$. For $tV_nx^n-x^4/2$, $1\leq n\leq3$, the $n-1$ bounded and $4-n$ escaping critical dimensions total three. Counting only bounded points loses actual rank; the stationary-ball construction retains all roots and tests its full matrix using polynomial separation. Finally, $t^{-3/2}=r^3$ under $r=t^{-1/2}$: a claim about a noninteger exponent before ramification must not be reused after ramification. Lemma~\ref{lem:positiveformal} uses positive seed order and an actual full frame instead.