\subsection{Resolving a Laurent phase}\label{sec:laurentresolution} The logarithmic estimate needs two ingredients: a complete description of the regular part, and a bound for the distinguished differential $d\log x$. The first is a nearby-cycle calculation. The second is a calculation on annuli. Keeping these ingredients separate prevents a perfect period pairing from being mistaken for a growth estimate. Throughout this subsection the coefficients are rational functions of the parameter. In particular their Puiseux expansions converge. The nonconvergent formal statements proved earlier have a different scope. The two-ended twisted quotient is finite dimensional without a genericity assumption. If the actual Laurent orders are $p,q\geq1$, the classes $x^{-q},\ldots,x^{p-1}$ are a basis. The top term of $x\varphi_x$ eliminates every exponent at least $p$, without crossing below $-q$; the bottom term eliminates every exponent below $-q$, without crossing above $p-1$. An exact expression supported in this interval would require its preimage to have highest exponent at most $-1$ and lowest exponent at least zero. This proves independence and rank $p+q$. Coefficient specializations use their actual orders. \begin{lemma}[A marked model]\label{lem:markedmodel} Let $\varphi(r,x)\in\C(r)[x,x^{-1}]$ be nonconstant in $x$. After a finite ramification of $r$ there is a smooth surface proper over a parameter disc, unchanged over its punctured disc, with these properties: \begin{enumerate} \item Its generic fibre is $\mathbb P^1$. Its central fibre is a reduced normal-crossing curve, with local equations $r=u$ or $r=uv$. \item The phase extends to a morphism to $\mathbb P^1$. Its zero and pole divisor and the divisor of $x$ have normal crossings with the central fibre. \item The horizontal sections $x=0,\infty$ are disjoint and meet smooth points of the central fibre. Locally $x=u^a v^b U$, where $a,b$ are integers and $U$ is holomorphic and nonvanishing. \end{enumerate} Additional markings used in the construction are not additional punctures of the exponential connection. \end{lemma} \begin{proof} Mark $0,\infty$ and every point of the generic fibres $\varphi^{-1}(0),\varphi^{-1}(1),\varphi^{-1}(\infty)$. Retain multiplicities in the divisors, but identify identical sections. Convergent Newton--Puiseux makes the finitely many marked branches meromorphic sections after one finite ramification. Start with $\mathbb P^1$ over that disc and blow up collisions of distinct sections. This terminates: the contact order of each pair is a finite integer and decreases on blowing up their common point. Every centre is a smooth point of a single central component. Indeed a section cannot pass through a node $uv=r$: two positive integral orders cannot have sum one. Thus each new exceptional component has multiplicity one and all nodes still have the equation $uv=r$. The final marked sections are disjoint and miss the nodes. The divisor of the meromorphic function $\varphi$ is supported on the central components and its marked zero and pole sections. If a point were an indeterminacy, its local divisor would have a zero branch and a pole branch meeting there. After cancelling common factors its expression is a unit times $u^a/v^b$, $a,b>0$. The equation $\varphi=1$ then has a horizontal branch through that point. At a node this contradicts the section property; at a smooth point it contradicts the separation from the zero or pole section. These are all possible intersections. Hence $\varphi$ is a morphism. The asserted normal crossings and the expression for $x$ follow from its divisor: horizontally $x$ has just the two marked sections. Only those original sections are removed from the generic fibre; marking a zero or a point with phase value one does not remove it. \end{proof} The marking of phase value one is necessary. For $\varphi=r/x^2$, the zero and pole sections are already disjoint, but the map is undefined at $r=x=0$. Its value-one branches $x=\pm\sqrt r$ force the cover $r=s^2$ and the chart $x=sy$, where the phase is $1/y^2$. The marked construction therefore detects a transition that merely separating zeros and poles would miss. We use the following precise comparison facts. For a meromorphic connection on a punctured disc, its moderate nearby cycles describe its formal regular summand, with monodromy. Nearby cycles commute with proper direct image. For a rank-one exponential on a surface with normal crossings, a component on which the phase has a vertical pole contributes no moderate nearby cycles. At a node between a finite-phase component and a pole component, the surviving object is the exponential connection on the finite component, localized at the node; it is not an extra point class. At a node between two finite components, the local complex is the usual nearby annulus complex, including its monodromy. These are standard comparison and normal-crossing facts, not new theorems here; see Roucairol~\cite[Section~1.2 and Lemma~1.4]{RoucairolFormal} and Sabbah~\cite[Section~1.b, especially Proposition~1.5]{SabbahStationary}. For clarity, their local exponential models are a holomorphic unit (finite phase), $\exp(u^{-p})$ at a finite--pole boundary, and $\exp(u^{-p}v^{-q})$, $p,q>0$, when both components are poles. An analytic unit in a pole monomial is absorbed in a coordinate. This may multiply the equation of the base by a nonvanishing unit; locally that change does not change the vanishing assertion or the nearby-cycle object up to the usual unit identification. In the $f^s$ description the identification is multiplication by $\exp(s\log U)$, with a local analytic choice of $\log U$. One keeps the actual gluing maps and monodromy, rather than just adding the ranks of the components. \begin{lemma}[Lifting the regular part]\label{lem:regularlifting} For the marked model, the complete nearby-cycle space of the exponential direct image is represented, on a sufficiently small parameter sector, by flat rapid-decay chains made from the following pieces: \begin{enumerate} \item bounded deformations of finitely many paths and cycles in the finite-phase central components, away from their nodes and poles; \item a bounded number of bridges and circles in the necks $uv=r$ between finite-phase components; \item descending phase tails at irregular ends and circles about regular punctures. \end{enumerate} The coefficients of each chosen flat chain are fixed constants. Their nearby-cycle classes exhaust the formal regular part, with its actual gluing maps and monodromy. \end{lemma} \begin{proof} Apply proper nearby-cycle comparison to the compactified surface, with the exponential connection localized at $r=0$ and at the two original punctures. The normal-crossing rules just stated kill the interiors of the pole components. Along a finite component they leave its ordinary exponential connection, with localization at its pole ends and at any original regular puncture. A finite--pole node gives that localized end. A finite--finite node retains the nearby annulus. Thus no pole component is silently counted as an additional regular class, and no annulus or puncture residue is discarded. Here is a chain description of this calculation. Remove small discs about the finitely many nodes and punctures in each finite component. Triangulate the remaining compact surfaces, keeping the boundary circles. At an irregular boundary attach the finitely many descending sectors. At a regular puncture keep the circle. The one-dimensional exponential period comparison identifies the resulting relative chain complexes with the component de Rham complexes; for a constant phase this is ordinary homology. At each finite--finite node glue by the actual annulus chain complex, with both its degree-zero and degree-one terms. Here is the total complex and the degree in which it is used. Let $U$ be the disjoint union of these component pieces with their descending ends, $V$ the disjoint union of the finite--finite necks, and $W$ their boundary-collar overlaps. Write $C_\bullet$ for chains with the stated rapid boundary condition at the ends, and with the exponential local coefficients. Set \[ C^{\mathrm{gl}}_\bullet= \operatorname{Cone}\!\left( C_\bullet(W)\xrightarrow{(i_*,-j_*)} C_\bullet(U)\oplus C_\bullet(V)\right). \] Thus a class includes its matching boundary chains; taking the separate component homology groups would lose these data. The neck complex retains both $C_0$ and $C_1$. Let $\mathcal K_\psi$ denote the moderate nearby-cycle de Rham complex on the central curve of the marked model, before proper pushforward. We use the unshifted curve de Rham normalization (functions in degree zero, one-forms in degree one), and take its hypercohomology over that central curve. The integration comparison and proper specialization give \[ H_1(C^{\mathrm{gl}}_\bullet) \xrightarrow[\text{integration}]{\ \simeq\ } \mathbb H^1(\mathcal K_\psi)^\vee \xrightarrow[\text{proper comparison}]{\ \simeq\ } \operatorname{Sol}(M_0). \] Here $M_0$ is the formal zero-exponential summand, and $\operatorname{Sol}$ denotes its horizontal dual on the chosen parameter sector, with nearby-cycle monodromy retained. The first arrow is obtained on $U,V,W$ from the curve period pairing and then on their mapping cones. The second is the moderate-nearby-cycle/proper-image comparison just cited, in fibre cohomological degree one. This also specifies the specialization map used here: lift the glued chains, integrate, and take their zero-exponential asymptotic classes. On a non-Stokes sector the latter forgets possible exponentially decaying terms in a lift. Naturality of integration under the two collar inclusions makes these maps agree. In particular this comparison is of complexes with their gluing maps and monodromy; equality of Euler characteristics or a count of local saddles is not being used to infer surjectivity. Choose a finite basis of this finite-dimensional homology. Smooth pieces and their boundary points lift by the implicit function theorem; an annulus is lifted by $uv=r$. Descending ends are continued by the inverse phase coordinate. On a fixed small sector these choices are isotopies, so the homology coefficients are constants. No Morse condition on the phase is used. Degenerate critical points stay inside the triangulated pieces, and a constant phase on a component is allowed. For the tail continuation, choose the ray of $r$ so that no unbounded critical value lies on the negative real phase axis for small $r$. There are finitely many excluded directions: critical values are convergent Puiseux functions. Bounded critical values fit in a fixed disc. From a sufficiently negative phase value every chosen inverse branch therefore continues to $-\infty$ without meeting a critical value. In the two-ended Laurent case it ends at $x=0$ or $\infty$. The same choice can avoid the finitely many limiting Stokes directions of the parameter connection. This constructs flat lifts whose nearby-cycle classes exhaust the space computed by proper comparison. No quantitative estimate has been used in that completeness calculation. \end{proof} \begin{lemma}[The logarithmic bound on the lifted pieces] \label{lem:logpieces} For the chains of Lemma~\ref{lem:regularlifting}, the period of $e^{\varphi}d\log x$ is $O(1+|\log r|)$ along the chosen ray. The bound is uniform for a fixed finite collection of generators. With a fixed Laurent amplitude the same construction has at most an additional algebraic power of $|r|^{-1}$. \end{lemma} \begin{proof} On a compact finite-phase piece the phase and the coefficients of $d\log x$ are bounded, so its integral is bounded. In a finite--finite neck write \[ r=uv,\qquad x=u^a v^b U(u,v). \] At fixed $r$, \[ d\log x=\left(a-b+\frac{uU_u-vU_v}{U}\right)\frac{du}{u}. \] The coefficient is uniformly bounded in a small bidisc. A radial bridge across $|r|/\eta\leq|u|\leq\eta$, with bounded angular detours, has logarithmic length $O(1+|\log r|)$. Circles have bounded logarithmic length. The phase is holomorphic on this bidisc and is uniformly bounded there. This treats every finite--finite neck. Near an irregular end of a finite component, use a local root of the phase unit to write $\varphi=w^{-p}$, $p\geq1$. At a finite--pole node one can simultaneously write $r=wv$ after changing the other coordinate by a unit. On $|v|,|w|<\eta$, the same calculation bounds $d\log x$ by a constant times $|dw|/|w|$. Along a descending phase ray, $s=-\varphi>0$, its integral is bounded by a constant times \[ \int_{s_0}^{\infty}e^{-s}\frac{ds}{s}. \] If the node chart ends, truncate it first at $|w|=|r|^\alpha$, where $0<\alpha<1$. Its phase is already of size $-|r|^{-\alpha p}$. Continue that inverse branch along the negative phase axis as in Lemma~\ref{lem:regularlifting}. We give a uniform bound for that continuation, rather than declaring it negligible from a local saddle formula. Choose $M$ large enough to include all coefficient transition scales in $|r|^M\leq|x|\leq|r|^{-M}$. Such $M$ exists because the coefficients are rational and there are finitely many Laurent powers. In this annulus each descending path lies on $\operatorname{Im}\varphi=0$, a real algebraic curve of bounded degree after clearing its Laurent denominator. The curve-length bound used in Lemma~\ref{lem:curvelength} gives $\int|dx|/|x|=O(1+|\log r|)$, by dyadic annuli. Along the entire tail the exponential is at most $\exp(-|r|^{-\alpha p})$. Outside that annulus the leading Laurent monomial dominates uniformly, so $|d\log x/d\varphi|\leq C/|\varphi|$ and its descending integral is bounded by the same exponential tail integral. This proves the required uniform tail bound. A fixed Laurent amplitude contributes only a power of $|r|^{-1}$ on the annulus and algebraic growth in $s$ outside; the exponential still dominates it. At a regular puncture use a small fixed circle in its smooth marked chart. Both its phase and logarithmic length are bounded. There are finitely many pieces and their chain coefficients are fixed. Adding their bounds proves the lemma. \end{proof} \begin{theorem}[Logarithmic seed bound]\label{thm:laurentlog} Let $\varphi(t,x)\in\C(t)[x,x^{-1}]$. On a generic ray at $t=\infty$, the zero-exponential part of every scalar flat period of $e^{\varphi(t,x)}dx/x$ is $O(1+\log|t|)$. This includes all critical degeneracies, coincident actions and regular punctures. Consequently, for a rational normalization $h(t)$ with $v_\infty(h)>0$, its differential seed submodule admits no rational horizontal quotient taking the seed to one. The same exclusion holds for a horizontal formal Gaussian functional at another point. \end{theorem} \begin{proof} If the phase is independent of $x$, the sole residue period is a constant times $e^{\varphi(t)}$. Its zero-exponential part is bounded when $\varphi$ is regular at infinity, and absent otherwise. For a nonconstant phase put $r=1/t$ and use Lemmas~\ref{lem:markedmodel}--\ref{lem:logpieces}. The construction includes regular puncture circles when there is just one irregular end. Finite ramification changes $\log|r|$ only by a constant factor. The lifted chains have at most algebraic growth on every Laurent amplitude by Lemma~\ref{lem:logpieces}. On the chosen generic ray, a full period vector with such growth has no exponentially growing factor: after formal exponential decomposition a largest growing factor could not cancel on every basis amplitude. The other nonzero factors are exponentially decaying there. The nearby-cycle basis therefore gives a full basis of regular functionals modulo those decaying terms, with constant coefficients in the flat basis. Every one of their seed periods has the asserted logarithmic bound. Thus their zero-exponential seed expansions have no negative algebraic powers. Multiplication by $h$ makes all of them tend to zero. For a constant phase the rank-one period already gives the same contradiction directly. For a nonconstant phase, a quotient with normalized seed one would extend locally to a horizontal functional on the full module, by the seed restriction and analytic extension argument of Lemma~\ref{lem:seedfunctional}. Its comparison hypotheses hold on $(\mathbb A^1\setminus S)\times\mathbb G_m$: remove the finitely many parameter poles, zeros of $h$, and comparison exceptions. The base projection and nonconstant phase are regular and algebraically independent there. Its scalar seed is constant. Distinct formal exponential factors separate, so its zero-exponential seed is also one. This contradicts the bound after normalization. A formal Gaussian functional with unit seed first restricts to the same rational seed quotient, as in the first part of Lemma~\ref{lem:seedfunctional}. \end{proof} The comparison theorems supply completeness here. The bounds come from the logarithmic differential on the marked model and the explicit descending-tail estimate. Neither part follows from a finite symbolic calculation. This proof uses convergence of the rational parameter family; it makes no assertion for arbitrary nonconvergent Laurent families or arbitrary coefficient rings.