\subsection{It is enough to complete the regular part} \label{sec:boundedregular} The cusp and tangent cases share a substantial simplification. We need not construct every escaping Gaussian character. It is enough to exclude escaping contributions to the regular part, compute that part's full rank, and construct precisely that many local functionals. This keeps the original marker while avoiding an unnecessary classification of the other exponential blocks. \begin{lemma}[A positive seed on the regular part] \label{lem:regularpositiveseed} Let $M$ and a seed $a$ satisfy the formal Gaussian hypotheses of Lemma~\ref{lem:seedfunctional}. After a finite cover at infinity, write $M_0$ for the zero-exponential summand. Evaluate its dual functionals on the canonical projection of $a$ to $M_0$. Suppose $M_0$ has a full horizontal dual frame over $\C((\rho))$, and every member takes $a$ into $\rho\C[[\rho]]$. Then the formal Gaussian functional cannot take $a$ to one. No semisimplicity assumption on the nonzero-exponential summands is required. The same conclusion holds if the regular frame exists over $\C((\rho))(\log\rho)$ and its seed values are finite polynomials in $\log\rho$ with coefficients of strictly positive $\rho$ order. \end{lemma} \begin{proof} The rational seed restriction supplies a quotient $U\to\C(t)$. The formal exponential decomposition is functorial and exact on submodules and quotients; different exponential types have no horizontal morphisms. These are the classical formal decomposition facts, recalled in Sabbah~\cite[Section~3, Corollary~3.3 and its proof]{SabbahStationary}. Consequently the quotient can only use $U_0\subset M_0$. Since $M_0$ is trivial over the displayed field, it is semisimple there, and this quotient extends to a horizontal functional on $M_0$. It is a constant linear combination of its full dual frame. Positive seed orders cannot sum to one. Adjoining $\log\rho$, with derivative $1/\rho$, does not add constants or negative-power exponential solutions. The same argument applies over this differential field. A finite sum of positive powers times logarithmic polynomials still cannot be one. This also explains why a full regular frame, rather than semisimplicity of the whole ambient module, is the relevant hypothesis. For an elementary check, expand this field in $\C(\log\rho)((\rho))$. The operator $\rho\partial_\rho$ preserves its valuation. A nonzero exponential twist has a negative-valuation scalar term in that operator, which cannot be cancelled by a regular matrix. Its horizontal maps to the trivial module therefore vanish even over this field. Its constants are $\C$, since a rational function $a(z)$ satisfying $a'+na=0$ is zero for integer $n\ne0$ and constant for $n=0$. \end{proof} \begin{lemma}[No finite-target contribution from spatial infinity] \label{lem:nospatialinfinity} Fix a small nonzero parameter $\epsilon_0$. Suppose in a neighbourhood of $\epsilon_0$ that $\|\nabla_x S\|\ge c\|x\|$ outside a compact set, where $S=P/\epsilon-q$ and $q$ is quadratic. The outer fibre pairs of $(\epsilon,S)$ are locally trivial over bounded target neighbourhoods. Consequently no additional finite-target discriminant or conormal multiplicity arises from spatial infinity there. \end{lemma} \begin{proof} On $S=g$ with $g$ bounded, $\partial_\epsilon S=-(g+q)/\epsilon=O(1+\|x\|^2)$, locally uniformly in $\epsilon$. A target-lifting vector field obtained by dividing the conjugate gradient by its squared norm has size $O(1/\|x\|)$. Correcting a parameter-lifting vector field by this field has size $O(1+\|x\|)$. After a cutoff inside a larger ball, these fields are complete for finite transport times: the radial growth inequality is at most linear. They lift the coordinate paths in the target and parameter neighbourhood and identify the exterior fibre pairs. All changes at a finite target are therefore supported inside a compact set. This is a fixed nonzero-parameter statement; it does not assert a uniform compact set as $\epsilon\to0$. \end{proof} \begin{proposition}[Bounded critical branches determine the regular rank] \label{prop:boundedregularrank} Let $P$ have degree $D\ge2$ and leading form $F_D$, and let $q$ be quadratic. Suppose, locally uniformly for small nonzero $\epsilon$, the phase $S=P/\epsilon-q$ has a complex linear gradient lower bound outside a compact set. The constants and compact set may depend on $\epsilon$. Assume also \begin{equation}\label{eq:nonisoleading} q_2(w)\ne0\quad\hbox{whenever }[w]\in\mathbb P^{n-1} \hbox{ and }\nabla F_D(w)=0. \end{equation} Then the rank of the full zero-exponential summand equals the sum of the generic local Milnor numbers of bounded phase-critical branches whose limiting $P$ value is zero. An escaping branch never has action polar part $\alpha/\epsilon$ for any constant $\alpha$. \end{proposition} \begin{proof} At a stationary branch the chain rule gives $S'=-P/\epsilon^2$, hence $q=-S-\epsilon S'$. A bounded branch has bounded $q$, so its action has only the possible polar term $P(x(0))/\epsilon$. If a branch escapes, divide $\nabla P=\epsilon\nabla q$ by $\|x\|^{D-1}$. Its limiting direction is singular for $F_D$. Condition~\eqref{eq:nonisoleading} makes the quadratic term of $q$ along that direction nonzero. Thus $q(x(\epsilon))$ has a pole, whereas $-S-\epsilon S'$ would be bounded for either zero polar part or a single term $\alpha/\epsilon$. This excludes both possibilities, without an escape-scale or Morse assumption. We spell out the global rank comparison. Fixed-parameter tameness gives full twisted rank equal to the total critical Milnor number by Pham's decomposition. For the regular algebraic input $\mathcal O$ on $\C^*_{\epsilon}\times\C^n$, form the ordinary direct image by $(\epsilon,S)$. Its relevant cohomology is the middle reduced Gauss--Manin module; the constant fiber cohomology has zero exponential integral in the affine $S$ coordinate. Lemma~\ref{lem:nospatialinfinity} excludes any finite-target contribution from spatial infinity, using the locally uniform gradient hypothesis. Every nonvertical discriminant curve at finite target comes from an affine critical value, with conormal multiplicity equal to the sum of its local Milnor numbers. First ramify all critical-value curves into graphs. Roucairol's nonramified rank formula, Theorem~0.1(2), Section~1.2, together with Theorem~0.3~\cite{RoucairolFormal}, computes every nonzero exponential-block rank from those graph multiplicities. Count all conjugate graphs even when their polar parts agree. No distinct-action assumption or monodromy formula is used. More explicitly, on the ramified parameter disc let $\ell$ index the distinct critical-value graphs $c_\ell(\epsilon)$ and put \[ \mu_{\rm tot}=\sum_x\mu_x(S),\qquad m_\ell=\sum_{x:\,S(x)=c_\ell}\mu_x(S),\qquad Q_\ell=\operatorname{pp}_{\epsilon=0}c_\ell. \] The first sum is over the generic affine critical scheme, and the second over the branches on that value graph. Tameness gives $\operatorname{rank}M=\mu_{\rm tot}=\sum_\ell m_\ell$. The conormal multiplicities entering Roucairol are exactly $m_\ell$: the affine Milnor decomposition gives them locally and the preceding lemma excludes an extra term at spatial infinity. In the graph case the rank formula reads \[ \operatorname{rank}M_Q=\sum_{\ell:Q_\ell=Q}m_\ell\quad(Q\ne0), \qquad \operatorname{rank}M_0 =\mu_{\rm tot}-\sum_{Q\ne0}\operatorname{rank}M_Q =\sum_{\ell:Q_\ell=0}m_\ell. \] All conjugate graphs are included after the cover; equal polar parts are grouped only in the sum for $M_Q$. Exactly the bounded-action branches remain. By the first part they are precisely the bounded branches with limiting $P$ value zero. This proves the assertion about the entire regular rank. \end{proof} We use the following explicit way to construct that regular part. Suppose a bounded group at $P$ value $\alpha$ has coordinates, after a cover, in which \begin{equation}\label{eq:boundedfixedmodel} S=\alpha t+H_0(z)+E_\rho(z),\qquad \omega=\rho^\kappa J_\rho(z)\,dz, \quad E_\rho\in\rho\C[z][[\rho]],\quad\kappa>0, \end{equation} where $J_\rho$ is a unit with polynomial coefficients at each order. If $\lambda$ is a dual functional on the fixed twisted quotient of $H_0$, then \[ L(A)=e^{\alpha t}\rho^\kappa\lambda\!\left( e^{E_\rho}J_\rho A(x(\rho,z))\right) \] is a formal functional on the original module. The product rule proves spatial exactness. Cartan's formula for the coordinate velocity, together with the displayed moving Jacobian, proves parameter horizontality. Every coefficient uses finitely many polynomial terms. Fixed original markers with nonnegative powers in these coordinates have positive seed order. The rank check has two directions. First, every bounded critical branch in the group must have bounded scaled coordinates $z$. Its limiting $z$ then belongs to the critical scheme of $H_0$, since $d_zE_\rho$ has positive order. Conversely, every isolated critical point of $H_0$ deforms with its full local intersection multiplicity in $H_0+E_\rho$. In the applications below the coordinate map $x(\rho,z)$ is regular at $\rho=0$ near these points, so all the deformed points give bounded original critical branches. No root of the fixed model is merely counted without being realized in the original family. Conservation of local intersection multiplicity therefore counts exactly the whole group, including collisions. One must also check that the fixed critical scheme is finite and that its total Milnor number equals the fixed twisted rank. A polynomial basis of the fixed quotient then gives independent functionals at leading order. Distinct groups are separated by polynomial jets, with arbitrarily high vanishing orders to overcome the finite negative powers introduced by the scaled representatives. When their number equals the regular rank of Proposition~\ref{prop:boundedregularrank}, they exhaust $M_0$. These checks will be made explicitly below; the formal perturbation formula alone does not establish completeness.