\section{General Weyl and oscillatory-transfer details} \label{sec:supp-general} This section and the next prove the general results of Section~\ref{sec:normalformbridge} for an arbitrary even centered Wigner--Jacobi chain. Only (E1)--(E2) of Definition~\ref{def:even-centered-chain} and the results of Section~\ref{sec:general-second} are used; nothing specific to the Hahn or Racah coefficients enters. \subsection{Analytic oscillatory transfer} \begin{lemma}[Uniform analytic oscillatory transfer bound] \label{lem:transfer} Let $Z_{k+1}=T_k(t)Z_k$ and let $F_{k+1}=T_k(0)F_k$ be a naturally normalized oscillatory fundamental matrix with $F_k,F_k^{-1}$ uniformly bounded, $k\ge1$. Suppose $T_k(t)$ is analytic in $t$ on $k^2|t|\le c$ with \begin{equation} T_k(t)=T_k(0)+\sum_{r=1}^{R-1}t^rE_{k,r}+\mathcal R_{k,R}(t), \label{eq:transferTaylor} \end{equation} with \begin{equation} \|E_{k,r}\|\le C_rk^{2r},\qquad \|\mathcal R_{k,R}(t)\|\le C_R|t|^Rk^{2R}, \label{eq:transferTaylorBounds} \end{equation} and assume the fixed starting vector is analytic in $t$ on $|t|\le c$ with the analogous remainder (writing $Z_K(t)=\sum_{r\ge0}t^rZ_K^{(r)}$). Then for $K^3|t|\le\eta_0$, \begin{align} \left\|Z_K(t)-\sum_{r=0}^{R-1}t^rZ_K^{(r)}\right\|_{\rm nat}&\le C_R'(K^3|t|)^R, \label{eq:abstractremainder}\\ \|Z_K^{(r)}\|_{\rm nat}&\le C_r'K^{3r}. \label{eq:abstractcoeff} \end{align} \end{lemma} \begin{proof} Put $Y_k=F_k^{-1}Z_k$. Since $F_{k+1}=T_k(0)F_k$, one has $Y_{k+1}=[I+V_k(t)]Y_k$ with $V_k(t)=F_{k+1}^{-1}[T_k(t)-T_k(0)]F_k$, which inherits the bounds \eqref{eq:transferTaylor}--\eqref{eq:transferTaylorBounds}. Taking $R=2$ and shrinking the fixed analyticity disk if necessary gives $\|V_k(t)\|\le C|t|k^2$ uniformly whenever $k^2|t|\le c_1$. Hence on the disk $|t|K^3\le\eta$, with $\eta$ fixed sufficiently small, $\sum_{k0$, and the finitely many smaller ranks contribute at most $k_1M$. Moreover $|r_{k+1}-r_k|\le2|\delta_{k+1}-\delta_k|\le2\Lambda k^{-2}$. Thus $c_k=b_k/(1-r_k)$ satisfies $|c_k|\le M/c_{\theta_0}$ and $|c_{k+1}-c_k|\le|b_{k+1}-b_k|/c_{\theta_0}+2\Lambda Mk^{-2}/c_{\theta_0}^2$, so $\sum_k|c_{k+1}-c_k|\le C(V+M)$. Since $z_k=(z_k-z_{k+1})/(1-r_k)$, summation by parts gives $\sum_{k=a}^{b}b_kz_k=c_az_a-c_bz_{b+1}+\sum_{k=a+1}^{b}(c_k-c_{k-1})z_k$, which is bounded by $C'(M+V)$. \end{proof} \subsection*{Proof of Lemma~\ref{lem:critical-transfer}: Quantitative critical oscillatory transfer normal form} \begin{proof} In the natural amplitude gauge $\widehat y_k=\sqrt{2k+1}\,y_k$, the unperturbed Wigner recurrence is the Jacobi equation $x\widehat y_k=A_{k+1}\widehat y_{k+1}+A_k\widehat y_{k-1}+B_k\widehat y_k$ with $A_k=\frac12+O_K(k^{-2})$ and $B_k=O_K(k^{-2})$ for fixed $a,b$. Thus the unperturbed system is a summable perturbation of the constant elliptic transfer matrix with eigenvalues $e^{\pm\ii\theta}$. The two exact complex solutions are built from the Darboux branches. By Lemma~\ref{lem:wigner-darboux-differentiated-closure} at order $R\ge3$, $\sqrt{2k+1}\,d^k_{ab}(\theta)$ is a sum of two carrier-separated branches $e^{\pm\ii(k+\frac12)\theta}\mathfrak u^\pm_k$ with $\mathfrak u^\pm_k$ a polynomial in $\kappa^{-1}$ with coefficients smooth on $K$, up to $O(k^{-R})$. Since the Wigner recurrence is exact for every $k$ and $e^{2\ii\theta}\ne1$ on $K$, the residual of each branch separately is $O_K(k^{-R})$, so each branch is an approximate solution. Write $\mathfrak u^\pm_k:=e^{\pm\ii\kappa\theta}\sum_{rk_{\min}$, invertible. Each branch alone has residual $O_K(k^{-R})$, so $\mathfrak F_{k+1}=T^{(0)}_k\mathfrak F_k+\mathcal G_k$ with $\|\mathcal G_k\|\le C_Kk^{-R}$ (only the first row of $\mathcal G_k$ is nonzero). Moreover $\mathfrak F_k$ is bounded and, since $c^+_{ab}c^-_{ab}=\pi^{-1}$, $\det\mathfrak F_k=\mathfrak u^+_k\mathfrak u^-_{k-1}-\mathfrak u^-_k\mathfrak u^+_{k-1}=2\ii/\pi+O_K(k^{-1})$, so $\mathfrak F_k^{-1}$ is bounded for $k\ge k_2$. Look for exact solutions $\mathfrak F_kC_k$ of $Y_{k+1}=T^{(0)}_kY_k$. Then $C_{k+1}=(I+\eta_k)C_k$ with $\eta_k:=-\mathfrak F_{k+1}^{-1}\mathcal G_k$, $\|\eta_k\|\le Ck^{-R}$, which is summable since $R\ge3$. The ordered products $C^\pm_k:=\prod_{j\ge k}(I+\eta_j)^{-1}e_\pm$ converge and solve this recursion for $k\ge k_2$ (extend them to $k\ge k_0$ by the exact recurrence, which is possible because $T^{(0)}_k$ is invertible); they satisfy $\|C^\pm_k-e_\pm\|\le C\sum_{j\ge k}\|\eta_j\|\le Ck^{1-R}$ and $\|C^\pm_{k+1}-C^\pm_k\|=\|\eta_kC^\pm_k\|\le Ck^{-R}$, uniformly on $K$. The first components $u^\pm_k:=(\mathfrak F_kC^\pm_k)_1$ are therefore exact solutions, with $u^+_k=\mathfrak u^+_k(1+c^{(1)}_k)+\mathfrak u^-_kc^{(2)}_k$, where $c_k:=C^+_k-e_+$ obeys $|c_k|\le Ck^{1-R}$ and $|c_{k+1}-c_k|\le Ck^{-R}$, and $u^-_k=\overline{u^+_k}$ because the recurrence is real. Dividing by the leading term $A_+e^{\ii\varphi_k}$ of $\mathfrak u^+_k$, the bracket is $1+g_k$ with $g_k=\kappa^{-1}A^+_{1,ab}/A^+_{0,ab}+O_K(k^{-2})$ and $|g_{k+1}-g_k|\le C_Kk^{-2}$: the finite branch sum is smooth in $\kappa^{-1}$ (its terms of order $r\ge1$ have differences $O(k^{-r-1})$), and the correction $e^{-2\ii\varphi_k}c^{(2)}_k+c^{(1)}_k$ is $O(k^{1-R})$ with difference $O(k^{1-R})\le O(k^{-2})$, because $R\ge3$. (The difference-system theorem of Benzaid and Lutz~\cite{BenzaidLutz} by itself gives only $1+o(1)$; the extra regularity comes from the explicit branch.) This yields two conjugate exact solutions $u_k^\pm$ and a fundamental matrix $F_k=\bigl(\begin{smallmatrix}u_k^+&u_k^-\\u_{k-1}^+&u_{k-1}^-\end{smallmatrix}\bigr)$ such that \begin{equation} u_k^\pm =A_\pm(x) e^{\pm\ii\varphi_k(x)} \left[1+O_K(k^{-1})\right], \qquad \varphi_k=(k+\tfrac12)\theta+\varphi_0, \label{eq:complex-darboux} \end{equation} where $|A_\pm|$ is bounded above and below on $K$ and the discrete difference of the $O(k^{-1})$ coefficient is $O_K(k^{-2})$ (shown above). The ordinary Jacobi Darboux formula fixes the phase and normalization of the physical Wigner-$d$ combination uniformly in the interior~\cite{DLMF18}. Since $T_k^{(0)}F_k=F_{k+1}$, the perturbed coordinates $Y_k=F_k^{-1}Z_k$ satisfy \begin{equation} Y_{k+1}=\left[I+F_{k+1}^{-1} (T_k^{(N)}-T_k^{(0)})F_k\right]Y_k. \label{eq:Ycritical} \end{equation} Inverting the leading Darboux matrices as in \eqref{eq:E11exactlead}, with $x=\cos\theta$, one obtains \begin{align} F_{k+1}^{-1}E_{11}F_k ={}& \begin{pmatrix} -\dfrac{\ii}{2\sin\theta}& \beta(x)e^{-2\ii\varphi_k}\\[4pt] \overline{\beta(x)}e^{2\ii\varphi_k}& \dfrac{\ii}{2\sin\theta} \end{pmatrix} +O_K(k^{-1}), \label{eq:E11conjugated} \end{align} where $\beta$ is bounded and smooth; the $O(k^{-1})$ term again has discrete difference $O_K(k^{-2})$. Hence \eqref{eq:Ycritical} has the form \begin{equation} Y_{k+1}=[I+D_k+Q_k+S_k]Y_k, \qquad D_k=\frac{k^2}{N^2} \begin{pmatrix} -\ii\chi/\sin\theta&0\\ 0&\ii\chi/\sin\theta \end{pmatrix}, \label{eq:DQS}\eqalias{eq:Dk} \end{equation} where $Q_k$ is off diagonal with entries $O_K(k^2/N^2)e^{\pm2\ii\varphi_k}$, and \begin{equation} \sum_{k\le L}\|S_k\| =O_K\!\left(\frac{L^2}{N^2}+\frac{L^5}{N^4}\right) =o(1) \label{eq:Ssmall} \end{equation} under \eqref{eq:critical-transfer-window}. Here $Q_k=2\chi(x)\frac{k^2}{N^2}\bigl(\begin{smallmatrix}0&\beta e^{-2\ii\varphi_k}\\\overline\beta e^{2\ii\varphi_k}&0\end{smallmatrix}\bigr)$ is the off-diagonal part of the leading term, and $S_k$ collects the $O_K(k^{-1})$ remainder of \eqref{eq:E11conjugated} times $2\chi k^2/N^2$ and the term $F_{k+1}^{-1}\mathcal R_{k,N}F_k=O(k/N^2+k^4/N^4)$. We do not factor out the diagonal $D_k$ first: it is kept in the Harris--Lutz step below, and the diagonal product is extracted at the end. It remains to remove $Q_k$. The geometric partial sums of $e^{\pm2\ii(k+1/2)\theta}$ are uniformly bounded on $K$; in the notation of Lemma~\ref{lem:dirichletphase} the phase increment of the entries of $Q_k$ is exactly $\theta$ ($\delta_k=0$). The amplitudes of $Q_k$ are $b_k=2\chi\beta k^2/N^2$, so on $r\le k\le s\le L$ they have $M=O(L^2/N^2)$ and $V\le\sum_{k\le L}2|\chi\beta|(2k+1)/N^2=O(L^2/N^2)$. Abel summation (Lemma~\ref{lem:dirichletphase}) therefore gives \begin{equation} \sup_{r\frac12$ then $L^2/N^2$ is bounded below by a constant depending only on $K,\Lambda$ and \eqref{eq:critical-transfer-error} follows from these bounds, so we may assume $\eta_H\le\frac12$. Substitute $Y_k=(I+H_k)V_k$ in \eqref{eq:DQS}, keeping $D_k$. The matrix identity $(I+D_k+Q_k+S_k)(I+H_k)=(I+H_{k+1})(I+D_k)+\Pi_k$ holds with \begin{equation} \Pi_k:=\bigl(D_kH_k-H_{k+1}D_k\bigr)+Q_kH_k+S_k(I+H_k), \end{equation} because $I+H_k+Q_k=I+H_{k+1}$ cancels the terms linear in $Q_k$ and $H_k$. Therefore \begin{equation} V_{k+1}=\bigl[I+D_k+P_k\bigr]V_k,\qquad P_k:=(I+H_{k+1})^{-1}\Pi_k, \qquad \|P_k\|\le2\bigl[2\eta_H\|D_k\|+\eta_H\|Q_k\|+\tfrac32\|S_k\|\bigr]. \end{equation} The first term of $\Pi_k$ is the commutator-type contribution of the diagonal part; with $\sum\|D_k\|=O(\Lambda)$ it is the dominant one, and its total is $4\eta_H\sum_k\|D_k\|\le C_{K,\Lambda}\Lambda\,L^2/N^2$. With $\sum\|Q_k\|=O(\Lambda)$ and \eqref{eq:Ssmall}, \begin{equation} \sum_{k\le L}\|P_k\|\le C_{K,\Lambda}\Bigl(\eta_H\sum_k(\|D_k\|+\|Q_k\|)+\sum_k\|S_k\|\Bigr)\le C_{K,\Lambda}\Bigl(\frac{L^2}{N^2}+\frac{L^5}{N^4}\Bigr). \end{equation} It remains to extract the diagonal product. The matrices $D_k$ are diagonal with purely imaginary entries $\mp\ii\epsilon_k$, $\epsilon_k=\chi k^2/(N^2\sin\theta)$, so $\Delta_k:=\prod_{j=k_0}^{k-1}(I+D_j)$ is diagonal, $\|\Delta_k^{-1}\|\le1$ (each factor $1\mp\ii\epsilon_j$ has modulus $\ge1$) and $\|\Delta_k\|\le\exp(\frac12\sum\epsilon_j^2)\le C_{K,\Lambda}$. For $W_k:=\Delta_k^{-1}V_k$ one has $W_{k+1}=W_k+\Delta_{k+1}^{-1}P_k\Delta_kW_k$, hence $\|W_L-W_{k_0}\|\le C\sum_k\|P_k\|\exp(C\sum_k\|P_k\|)$. Since $H_L=0$, $Y_L=V_L=\Delta_LW_L$ and $Y_{k_0}=(I+H_{k_0})V_{k_0}$, so \begin{equation} \mathcal U_{L,N}=\Delta_L(I+H_{k_0})^{-1}+O\Bigl(\sum_k\|P_k\|\Bigr)=\Delta_L+O_{K,\Lambda}\Bigl(\frac{L^2}{N^2}+\frac{L^5}{N^4}\Bigr). \end{equation} Finally $1\mp\ii\epsilon_j=\sqrt{1+\epsilon_j^2}\,e^{\mp\ii\arctan\epsilon_j}$ and $\arctan\epsilon=\epsilon+O(\epsilon^3)$, so $\Delta_L=\operatorname{diag}(e^{-\ii\vartheta_{L,N}},e^{\ii\vartheta_{L,N}})+O\bigl(\sum_j\epsilon_j^2\bigr)$ with $\vartheta_{L,N}=\frac{\chi}{\sin\theta}\sum_{k=k_0}^{L-1}\frac{k^2}{N^2}$, an error $O(L^5/N^4)$. Thus the full ordered transfer is the diagonal phase product plus an error bounded by $C_{K,\Lambda}(L^2/N^2+L^5/N^4)$ in the normalized Wigner basis (within the window \eqref{eq:critical-transfer-window}, $L^5/N^4\le\Lambda L^2/N^2$, so the second term is redundant; it is kept to match the statement). Any mismatch of the fixed starting vector is added separately in applications, and \eqref{eq:k2sumcritical} is the elementary sum-of-squares formula. \end{proof} \subsection{Turning-free bulk: adiabatic basis and fixed-start stability} \subsection*{Proof of Lemma~\ref{lem:wigner-darboux-differentiated-closure}: Differentiated interior Wigner--Darboux expansion} \begin{proof} Fix $R,s$ and put $R'=R+s$. Enlarge $J$ to a compact interval $J_1\Subset(0,\pi)$ and write $w=(\sin\vartheta)^{1/2}\sqrt{2k+1}\,d^k_{ab}$. It is real and solves exactly \begin{equation} w''+(\kappa^2-V)w=0,\qquad V=\frac{a^2+b^2-\frac14-2ab\cos\vartheta}{\sin^2\vartheta}, \end{equation} with $V$ analytic and bounded on $J_1$ (this is the Jacobi equation in Liouville normal form, as used in Lemma~\ref{lem:bulk-zero-continuation}). \emph{Step 1: undifferentiated carrier-separated expansion.} The Frenzen--Wong Bessel-type expansion of the Jacobi polynomial with the Hankel expansion of the Bessel functions (DLMF 18.15.6 and 10.17.3; this is the expansion written out in the proof of Lemma~\ref{lem:darboux-parity}) gives smooth functions $\widetilde A^\sigma_r$ on $J_1$, independent of $k$, with \begin{equation} \sup_{J_1}\Bigl|w-\sum_{\sigma=\pm1}e^{i\sigma\kappa\vartheta}T^\sigma\Bigr|\le C\kappa^{-R'},\qquad T^\sigma:=\sum_{r0$. Taylor's formula with remainder therefore shows that the principal recurrence residual of each branch is bounded by $C_{\mathcal C}h^2|z_k^\pm|$. Restoring the exact coefficients through \eqref{eq:Akexactratio} gives the explicit defect \begin{equation} r_k^\pm :=x z_k^\pm-A_{k+1}z_{k+1}^\pm -A_kz_{k-1}^\pm-B_kz_k^\pm, \qquad |r_k^\pm| \le C_{\mathcal C}(N^{-2}+k^{-2})|z_k^\pm|. \label{eq:adiabaticresidual} \end{equation} Now write the exact one-step transfer as \begin{equation} T_k= \begin{pmatrix} (x-B_k)/A_{k+1}&-A_k/A_{k+1}\\ 1&0 \end{pmatrix}. \label{eq:adiabatic-transfer-matrix} \end{equation} The determinant of the branch matrix $P_k$ of \eqref{eq:adiabaticfundamental} is \begin{equation} \det P_k =2\ii\,D(q_k,x)^{-1/4}D(q_{k-1},x)^{-1/4} \sin\!\bigl(N[S(q_k,x)-S(q_{k-1},x)]\bigr), \end{equation} up to an irrelevant sign. Since \begin{equation} N[S(q_k,x)-S(q_{k-1},x)] =p(q_k-h/2,x)+O_{\mathcal C}(h^2), \end{equation} $P_k$ and $P_k^{-1}$ are uniformly bounded for all sufficiently large $N$. Equation \eqref{eq:adiabaticresidual} then gives \begin{equation} P_{k+1}^{-1}T_kP_k=I+E_k, \qquad \|E_k\|\le C_{\mathcal C}(N^{-2}+k^{-2}). \label{eq:adiabaticEk} \end{equation} Indeed $T_kP_k-P_{k+1}$ has zero second row and first row $(r_k^+/A_{k+1},r_k^-/A_{k+1})$. Initialize an exact fundamental matrix by $\mathcal F_K^{(N)}=P_K$ and write $\mathcal F_k^{(N)}=P_kC_k$. Then $C_{k+1}=(I+E_k)C_k$, $C_K=I$. Therefore \begin{equation} \sup_{K\le k\le L}\|C_k-I\| \le \exp\!\left(\sum_{j=K}^{L-1}\|E_j\|\right)-1 =O_{\mathcal C}(K^{-1}+L/N^2). \label{eq:adiabatic-product-bound} \end{equation} The term $L/N^2$ is retained so that the bound covers final ranks from mesoscopic to fixed ratio; on a fixed-ratio compact set it is simply $O(N^{-1})$. This proves \eqref{eq:adiabaticfundamental}. The argument is the finite-interval, turning-free discrete Liouville--Green construction; compare Geronimo--Smith~\cite{GeronimoSmithWKB}. \end{proof} \subsection*{Proof of Corollary~\ref{cor:general-fixed-start}: Fixed-start turning-free stability} \begin{proof} Choose an admissible intermediate rank $K=\min\{L,\lfloor N^{1/3}\rfloor+O(1)\}$ (if $L\le k_0$ there is nothing to prove). On the finite/mesoscopic interval $k_0\le k\le K$, the exact transfer is, by (E1), an analytic $t=h^2$ perturbation of the Wigner transfer with order-$t^r$ coefficients $O(k^{2r})$. Apply the interaction representation of Lemma~\ref{lem:transfer} to the two columns of a bounded Wigner-$d$ fundamental matrix initialized at $k_0$. Since $K^3h^2=O(N^{-1})$, both the interaction product and its inverse are uniformly bounded, so the exact fundamental matrix and inverse are bounded through $K$. If $L>K$, on $K\le k\le L$ Lemma~\ref{lem:general-adiabatic} applies because $K\to\infty$ and $K/N\to0$; its comparison matrix $P_k$ and inverse are uniformly bounded by turning-free interiority. The two exact fundamental matrices differ at $K$ by a constant invertible change of basis whose norm and inverse norm are bounded by the preceding estimates. This propagates the bound through $L$. \end{proof} \begin{lemma}[Shadow-phase pole] \label{lem:general-shadow-pole} On a compact set of $(q,x,\pi)$ with $00$, let $p_h$ be the exact frozen phase \eqref{eq:exact-shadow-dispersion} of an even centered chain, with the exact midpoint coefficients \begin{equation} a(s;h^2)=\frac{e_\pi(s^2,h^2)}2\left[\frac{(1-a^2h^2/s^2)(1-b^2h^2/s^2)}{1-h^2/(4s^2)}\right]^{1/2}, \qquad b(q;h^2)=\frac{\lambda_\pi(h^2)\,ab\,h^2}{q^2-h^2/4}+v_\pi(q^2,h^2). \end{equation} Then, uniformly while $h/q$ is small, \begin{equation} p_h=p+h^2p_2+O(h^4q^{-4}), \end{equation} where \begin{equation} p_2=\frac{1}{\sin p}\left[\frac{ab/q^2+\partial_tv_\pi(q^2,0)}{e_\pi(q^2)} +\cos p\left(\frac{(\log a_0)''(q)}{8}-\frac{a^2+b^2-\frac14}{2q^2}+\eta_\pi(q^2)\right)\right], \label{eq:general-shadow-p2} \end{equation} and \begin{equation} p_2(q,x)=-\frac{\mathfrak c_{ab}(\theta)}{q^2}+O(1)\qquad(q\downarrow0),\qquad\theta=\arccos x. \label{eq:general-shadow-darboux-pole-match} \end{equation} In particular a leading-order adiabatic basis that uses only $p=S_q$ accumulates an $O(K^{-1})$ fixed-index shadow-phase error between $q\asymp K/N$ and a fixed positive ratio, since $N\int_{K/N}^{q_0}h^2|p_2|\,\dd q=O(K^{-1})$. \end{lemma} \begin{proof} At fixed $(q,x)$, average $\log a(s;h^2)$ over $s=q\pm h/2$: the factor $e_\pi(s^2,h^2)/2=\frac12e_\pi(s^2)(1+h^2\eta_\pi(s^2)+O(h^4))$ contributes $\log a_0(q)+h^2\bigl(\frac18(\log a_0)''(q)+\eta_\pi(q^2)\bigr)+O(h^4)$, and the bracket contributes $-(a^2+b^2-\frac14)h^2/(2q^2)+O(h^4q^{-4})$. Hence $2\sqrt{a(q+h/2)a(q-h/2)}=e_\pi(q^2)[1+h^2(\frac18(\log a_0)''+\eta_\pi-(a^2+b^2-\frac14)/(2q^2))+O(h^4q^{-4})]$, while $x-b=x-b_0-h^2(ab/q^2+\partial_tv_\pi(q^2,0))+O(h^4q^{-4})$. Expanding \eqref{eq:exact-shadow-dispersion} and differentiating the analytic $\arccos$ branch gives \eqref{eq:general-shadow-p2}. As $q\downarrow0$, (E1) gives $e_\pi=1+O(q^2)$, $\cos p=x+O(q^2)$, $\sin p=\sin\theta+O(q^2)$, and $(\log a_0)''$, $\eta_\pi$ and $\partial_tv_\pi$ bounded, so the only pole is \begin{equation} \frac{ab-x(a^2+b^2-\frac14)/2}{q^2\sin\theta}=-\frac{\mathfrak c_{ab}(\theta)}{q^2} \end{equation} by \eqref{eq:wigner-darboux-phase-coeff}. The accumulated estimate follows from $Nh^2=h$ and $\int_{K/N}^{q_0}q^{-2}\dd q\le N/K$. \end{proof} \section{Second-order closure for even centered chains} \label{sec:supp-general-second} \subsection*{Proof of Proposition~\ref{prop:generic-second-weyl-eikonal}: Generic second midpoint eikonal} \begin{proof} Write \begin{equation} Y(q)=\exp\{g(q)+ih^{-1}(S(q)+h^2S_2(q))\}, \quad p=S_q,\quad \sigma_2=S_{2,q}. \end{equation} For either sign $\varepsilon\in\{+1,-1\}$, direct Taylor expansion gives \begin{equation}\begin{aligned} a(q+\varepsilon h/2;h^2) &=a_0+\varepsilon\frac h2a_0' +h^2\left(a_2+\frac18a_0''\right)+O(h^3),\\ \frac{Y(q+\varepsilon h)}{Y(q)} &=e^{i\varepsilon p} \left\{1+hA_\varepsilon +h^2\left(B_\varepsilon+\frac12A_\varepsilon^2\right) +O(h^3)\right\},\\ A_\varepsilon&=\varepsilon g'+\frac i2p',\qquad B_\varepsilon=\frac12g'' +i\varepsilon\left(\frac16p''+\sigma_2\right). \end{aligned}\end{equation} The $h^0$ coefficient yields $x-b_0=2a_0\cos p$. The coefficient of $ih$ is \begin{equation} 2a_0g'\sin p+a_0p'\cos p+a_0'\sin p=0, \end{equation} equivalently $(a_0e^{2g}\sin p)'=0$; hence $g=-\frac12\log(a_0\sin p)$ up to an additive constant. Using this value of $g$, the \emph{real} $h^2$ equation is \begin{equation} \begin{split} 2a_0\sin p\,\sigma_2={}& a_0(g''+g'^2)\cos p-a_0g'p'\sin p -\frac{a_0}{4}(p')^2\cos p-\frac{a_0}{3}p''\sin p\\ &+2a_2\cos p+a_0'g'\cos p -\frac12a_0'p'\sin p+\frac14a_0''\cos p+b_2. \end{split} \end{equation} This is \eqref{eq:generic-second-weyl-density}, including the curvature terms absent from a frozen dispersion expansion. All Taylor remainders and their finitely many $q$ derivatives are uniform on a fixed compact elliptic set where $a_0$ and $\sin p$ are bounded away from zero. Analyticity in $t=q^2$ at a matched endpoint does not follow from this formal expansion alone. For an even centered chain it is Proposition~\ref{prop:general-second-eikonal-pole}: the $t^{-1}$ part cancels against the derivative of the finite-rank Darboux phase. \end{proof} \subsection*{Proof of Proposition~\ref{prop:general-fixed-rank-contact}: Automatic second-order fixed-rank contact} \begin{proof} Write $s=1-x^2$, $\rho_{20}:=\rho_2(0,x)$ and $d_j=d^j_{ab}(\arccos x)$. Write $e_\pi(r,t)=1-\varepsilon_1r+\lambda_1t+\varepsilon_2r^2+\varepsilon_{11}rt+O(|r|^3+|t|^2)$ and $v_\pi(r,t)=\beta_1r+\beta_2t+\beta_{20}r^2+\beta_{11}rt+O(|r|^3+|t|^2)$. In \eqref{eq:generic-second-weyl-density} with \eqref{eq:class-a2b2}, the coefficients $\varepsilon_2,\beta_{20}$ enter $a_0,b_0,p,g$ only at order $q^4$, hence their first two $q$-derivatives at order $q^2$; $\beta_{11}$ enters $b_2=ab/q^2+\beta_2+\beta_{11}q^2+O(q^4)$ at order $q^2$, and $\varepsilon_{11}$ enters $\eta_\pi(q^2)=\lambda_1+(\varepsilon_{11}+\varepsilon_1\lambda_1)q^2+O(q^4)$ at order $q^2$; the only $q^{-2}$ terms are $2a_2\cos p+b_2$. So the $q^0$ coefficient of $\Sigma_2$, and with it $\rho_{20}$, depends only on $(\varepsilon_1,\beta_1,\beta_2,\lambda_1)$, linearly, and expanding with \eqref{eq:class-small-q} and $\eta_\pi(0)\cot\theta=\lambda_1x/\sqrt s$ gives \eqref{eq:general-rho20}. At fixed rank $q^2=\kappa^2t$ and, by Proposition~\ref{prop:general-second-eikonal-pole}, $\Gamma_2/q=\rho_{20}+O(q^2)$. With \eqref{eq:class-small-q} and $\partial_\theta=-\sqrt s\,\partial_x$, \begin{equation} \widehat U_k^{[2]}=\widehat d_k+t\,\widehat u_k+O_k(t^2),\qquad \widehat u_k=\sqrt{2k+1}\,\mathcal Q_{2,k}d_k, \end{equation} \begin{equation} \mathcal Q_{2,k}=\Bigl(\frac{\kappa^2\chi}{3}-\sqrt s\,\rho_{20}\Bigr)\partial_x+\frac{\kappa^2\bigl((\varepsilon_1-x\beta_1)/2-x\chi/6\bigr)}{s}+\frac{x\rho_{20}}{2\sqrt s}. \end{equation} By (E1), $A_k(h)=A_k^{(0)}(1+(\lambda_1-\varepsilon_1k^2)t)+O_k(t^2)$ and $B_k(h)=B_k^{(0)}(1+\lambda_1t)+(\beta_1\kappa^2+\beta_2)t+O_k(t^2)$. Therefore the $t$ coefficient of $\mathcal R_k(\widehat U^{[2]})$ vanishes if and only if \begin{equation} \begin{aligned} &x\widehat u_k-A_{k+1}^{(0)}\widehat u_{k+1}-A_k^{(0)}\widehat u_{k-1}-B_k^{(0)}\widehat u_k\\ &\qquad=-\varepsilon_1\bigl[(k+1)^2A_{k+1}^{(0)}\widehat d_{k+1}+k^2A_k^{(0)}\widehat d_{k-1}\bigr]+(\beta_1\kappa^2+\beta_2)\widehat d_k+\lambda_1x\,\widehat d_k, \end{aligned} \label{eq:general-contact-identity} \end{equation} where the last term collects the $\lambda_1$ contributions, $\lambda_1[A_{k+1}^{(0)}\widehat d_{k+1}+A_k^{(0)}\widehat d_{k-1}+B_k^{(0)}\widehat d_k]=\lambda_1x\,\widehat d_k$ by the Wigner recurrence. Put $R_k=\sqrt{(k^2-a^2)(k^2-b^2)}$, so that $A_k^{(0)}=R_k/(k\sqrt{4k^2-1})$ and $B_k^{(0)}=ab/(k(k+1))$. The gauge factors cancel exactly, \begin{equation} A_{k+1}^{(0)}\sqrt{2k+3}=\frac{R_{k+1}}{(k+1)\sqrt{2k+1}},\qquad A_k^{(0)}\sqrt{2k-1}=\frac{R_k}{k\sqrt{2k+1}}, \end{equation} so dividing \eqref{eq:general-contact-identity} by $\sqrt{2k+1}$ and writing $M_k:=R_kd_{k-1}$, $P_k:=R_{k+1}d_{k+1}$ (the operators $\mathcal Q_{2,j}$ being linear), it becomes the gauge-free identity \begin{equation} \begin{aligned} &\Bigl(x-\frac{ab}{k(k+1)}\Bigr)\mathcal Q_{2,k}d_k-\frac{\mathcal Q_{2,k+1}P_k}{(k+1)(2k+1)}-\frac{\mathcal Q_{2,k-1}M_k}{k(2k+1)}\\ &\qquad=-\varepsilon_1\,\frac{(k+1)P_k+kM_k}{2k+1}+\bigl(\beta_1\kappa^2+\beta_2+\lambda_1x\bigr)d_k. \end{aligned} \label{eq:contact-gauge-free} \end{equation} The Wigner rank-shift (ladder) identities at ranks $k$ and $k+1$, \begin{equation} (1-x^2)d_k'=\Bigl(-kx+\frac{ab}{k}\Bigr)d_k+\frac{R_k}{k}d_{k-1},\qquad (1-x^2)d_{k-1}'=\Bigl(kx-\frac{ab}{k}\Bigr)d_{k-1}-\frac{R_k}{k}d_k, \end{equation} and the same pair with $k\to k+1$, express $M_k,M_k',P_k,P_k'$ through $(d_k,d_k')$ alone: \begin{equation} \begin{aligned} M_k&=k\,s\,d_k'+(k^2x-ab)d_k,\\ k\,s\,M_k'&=(k^2x-ab)M_k-(k^2-a^2)(k^2-b^2)\,d_k,\\ P_k&=\bigl((k+1)^2x-ab\bigr)d_k-(k+1)\,s\,d_k',\\ (k+1)\,s\,P_k'&=\bigl((k+1)^2-a^2\bigr)\bigl((k+1)^2-b^2\bigr)d_k-\bigl((k+1)^2x-ab\bigr)P_k, \end{aligned} \end{equation} in which $R_k$ enters only through $R_k^2$. After this substitution the coefficients of $d_k$ and $d_k'$ in the difference of the two sides of \eqref{eq:contact-gauge-free} are rational functions of $(a,b,k,x)$, linear in $(\varepsilon_1,\beta_1,\beta_2,\lambda_1)$, and they vanish identically (a finite rational-function identity, verified by exact symbolic computation). No separate treatment of the edge row $k=k_{\min}$ is needed: $R_k$ multiplies only the $d_{k-1}$ terms, and $R_{k_{\min}}=0$ gives $M_{k_{\min}}=0$, which is what the displayed formulas for $M_k$ and $M_k'$ return there (the first ladder identity at $k_{\min}$ reads $(1-x^2)d_k'=(-kx+ab/k)d_k$). (The conventions $A^{(0)}_{k_{\min}}:=0$ and $B^{(0)}_0:=0$ are needed only to define the coefficients; \eqref{eq:contact-gauge-free} has no $\sqrt{4k^2-1}$.) In the exceptional case $a=b=k=0$, where $d_0=1$, $d_1=x$, the identity is checked directly. The $\beta_2$ part is transparent: there $\chi=0$, $\rho_{20}=s^{-1/2}$ and $\mathcal Q_{2,k}=-\partial_x+x/(2s)$ is independent of $k$, so it commutes with the $x$-independent Wigner hoppings, and $[x,-\partial_x]=1$ produces the source $\beta_2\widehat d_k$; this is the shift $x\mapsto x-\beta_2t$. The $\lambda_1$ part is the same mechanism with $\beta_2$ replaced by $\lambda_1x$: it adds $\lambda_1x/\sqrt s$ to $\rho_{20}$, hence $-\lambda_1x\,\partial_x+\lambda_1x^2/(2s)$ to $\mathcal Q_{2,k}$. The multiplier $\lambda_1x^2/(2s)$ commutes with the $k$-hoppings and with $x-J^{(0)}$, where $J^{(0)}$ is the Wigner Jacobi operator of \eqref{eq:wigner-natural-jacobi}, so it drops out; and differentiating $x\widehat d_k=J^{(0)}\widehat d_k$ in $x$ gives $(x-J^{(0)})\partial_x\widehat d_k=-\widehat d_k$, so the term $-\lambda_1x\,\partial_x$ produces exactly the source $\lambda_1x\,\widehat d_k$ of \eqref{eq:general-contact-identity}. (This is the shift $x\mapsto x-\lambda_1xt$ of Remark~\ref{rem:E1-t}.) \end{proof} \begin{lemma}[Fixed-ratio cubic residual of the corrected pullback] \label{lem:general-cubic-residual} \label{lem:3j-second-pullback-residual} \label{lem:6j-second-pullback-residual} Fix $q_0>0$ and a closed turning-free corridor $\mathcal C_{q_0}$ of an even centered Wigner--Jacobi chain on which $q\ge q_0$. For the corrected pullback \eqref{eq:class-corrected-pullback} the naturally gauged scalar Jacobi residual obeys \begin{equation} |\mathcal R_k(\widehat U^{[2]})| \le C_{\mathcal C_{q_0}}N^{-3}, \label{eq:general-second-local-residual} \end{equation} uniformly whenever $(q_k,x,\pi)\in\mathcal C_{q_0}$. The proof uses only the Darboux expansion of Lemma~\ref{prop:wigner-darboux-first-phase} and Taylor expansion in $h$, so the bound holds for every sufficiently small $h>0$ with $q=\kappa h$, not only at $h=N^{-1}$ for admissible $N$. More precisely, the coefficients $C^+_0,C^+_1,C^+_2$ of the expansion of $e^{-i\Theta_q/z}\mathcal L_h\widehat U^{[2],+}$ in $z=h/q$ (Lemma~\ref{lem:cubic-residual-transport}) vanish. \end{lemma} \begin{proof} Fix $(x,\pi)$ and work at fixed $q\ge q_0$ with $h\to0$, $\kappa=q/h$, $z=h/q$. The margin $\delta_{\mathcal C}$ of the corridor keeps $q\pm h$ in the region where (E1) holds, and every function of $(q,h)$ below is smooth there, with derivatives bounded uniformly on $\mathcal C_{q_0}$; $O(h^j)$ means a quantity bounded by $Ch^j$ uniformly. Write $\mathcal L_h$ for the midpoint operator \eqref{eq:midpoint-operator}, so that $\mathcal R_k(\widehat U^{[2]})=(\mathcal L_h\widehat U^{[2]})(q_k)$ and $\widehat U^{[2]}(q)$ is regarded as a function of continuous $q$ with $\kappa=q/h$ (the ranks $k\pm1$ are the points $q\pm h$). \emph{Step 1: one Darboux branch.} By \eqref{eq:wigner-darboux-diff-closure} with $R=3$, applied at the angles $\Theta^{[2]}_{q'}$, $q'\in\{q-h,q,q+h\}$ (which lie in a fixed compact subinterval of $(0,\pi)$, since $\Theta^{[2]}_{q'}-\Theta_{q'}=O(h^2)$), and with \eqref{eq:corrected-branch}, $\widehat U^{[2]}=\widehat U^{[2],+}+\widehat U^{[2],-}+O(\kappa^{-3})$ at each of the three points, and $\widehat U^{[2],-}=\overline{\widehat U^{[2],+}}$. The coefficients of $\mathcal L_h$ are real and bounded, so $\mathcal L_h\widehat U^{[2],-}=\overline{\mathcal L_h\widehat U^{[2],+}}$ and \begin{equation} |\mathcal R_k(\widehat U^{[2]})|\le2\,|\mathcal L_h\widehat U^{[2],+}|+Ch^3 . \end{equation} By Lemma~\ref{lem:darboux-parity}, as formal series in $\kappa^{-1}$ at fixed angle $\vartheta$, \begin{equation} \sum_r\kappa^{-r}A^+_{r,ab}(\vartheta)=\omega_{ab}(\pi\sin\vartheta)^{-1/2}\,\mathfrak A_{ab}(\vartheta,\kappa)\,e^{i\mathfrak P_{ab}(\vartheta,\kappa)}, \end{equation} with $\mathfrak A_{ab}=1+\kappa^{-2}\mathfrak A_2+O(\kappa^{-4})$ and $\mathfrak P_{ab}=\kappa^{-1}\mathfrak c_{ab}+O(\kappa^{-3})$: there is no $\kappa^{-1}$ amplitude term and no $\kappa^{-2}$ phase term. Hence the truncated sum $\sum_{r<3}\kappa^{-r}A^+_{r,ab}$ differs by $O(\kappa^{-3})$ from $\omega_{ab}(\pi\sin\vartheta)^{-1/2}(1+\kappa^{-2}\mathfrak A_2)e^{i\mathfrak c_{ab}/\kappa}$, uniformly on compact angular sets. Replacing the former by the latter in \eqref{eq:corrected-branch} changes $\widehat U^{[2],+}$ by $O(\kappa^{-3})$ at each of the three points, hence changes $\mathcal L_h\widehat U^{[2],+}$ by $O(h^3)$, and does not change $C^+_r$ for $r\le2$. Since $W^{[2]}_q(\sin\Theta^{[2]}_q)^{-1/2}=D^{-1/4}$ exactly (definition of $W^{[2]}$ in \eqref{eq:class-second-data}), the branch to be studied is \begin{equation} u(q;h):=B_0(q)\Bigl(1+\frac{\mathfrak A_2(\Theta^{[2]}_q)}{\kappa^2}\Bigr)\exp\Bigl\{i\Bigl[\kappa\Theta^{[2]}_q+\frac{\mathfrak c_{ab}(\Theta^{[2]}_q)}{\kappa}\Bigr]\Bigr\},\qquad B_0:=\omega_{ab}\pi^{-1/2}D^{-1/4}. \label{eq:cubic-branch} \end{equation} The $h^2$ difference between $W^{[2]}_q$ and $W_q$ never appears separately: it is exactly what makes the principal amplitude $B_0$ independent of $h$. The only real $h^2$ amplitude correction is $\mathfrak A_2(\Theta^{[2]}_q)/\kappa^2=h^2\mathfrak A_2(\Theta^{[2]}_q)/q^2$. \emph{Step 2: phase and amplitude.} Put $S:=q\Theta_q$ (so $S_q=p$), $g:=-\frac14\log D$ (so $B_0(q')/B_0(q)=e^{g(q')-g(q)}$ and, since $D^{1/2}=2a_0\sin p$, $g=-\frac12\log(a_0\sin p)+\mathrm{const}$ is the amplitude exponent of Proposition~\ref{prop:generic-second-weyl-eikonal}), and \begin{equation} S_2:=\Gamma_2+\frac{\mathfrak c_{ab}(\Theta_q)}{q},\qquad S_{2,q}=\rho_2+\frac{d}{dq}\frac{\mathfrak c_{ab}(\Theta_q)}{q}=\Sigma_2 \label{eq:cubic-S2} \end{equation} by \eqref{eq:class-second-data}. This is where the two Darboux phases combine. Exactly, $\kappa\Theta^{[2]}_q=\kappa\Theta_q+\kappa\frac{h^2}{q}\Gamma_2=S/h+h\Gamma_2$, while the Darboux phase of $d^k_{ab}(\Theta^{[2]}_q)$ is $\mathfrak c_{ab}(\Theta^{[2]}_q)/\kappa=h\,\mathfrak c_{ab}(\Theta^{[2]}_q)/q$, and $\mathfrak c_{ab}(\Theta^{[2]}_q)-\mathfrak c_{ab}(\Theta_q)=O(h^2)$. So the Darboux phase $\mathfrak c_{ab}(\Theta_q)/\kappa$ restores, up to $O(h^3)$, precisely the term subtracted in $\Gamma_2$: \begin{equation} \kappa\Theta^{[2]}_q+\frac{\mathfrak c_{ab}(\Theta^{[2]}_q)}{\kappa}=\frac Sh+hS_2+\psi,\qquad \psi:=\frac hq\bigl[\mathfrak c_{ab}(\Theta^{[2]}_q)-\mathfrak c_{ab}(\Theta_q)\bigr]=O(h^3). \end{equation} Therefore \begin{equation} u=e^{iS/h}\,V\,E,\qquad V:=B_0\,e^{ihS_2},\qquad E:=\Bigl(1+\frac{h^2\mathfrak A_2(\Theta^{[2]}_q)}{q^2}\Bigr)e^{i\psi}=1+h^2m(q;h), \label{eq:cubic-factorization} \end{equation} with $m$ smooth and bounded together with its $q$-derivatives ($\psi$ is smooth and divisible by $h^3$). Thus $E-1=O(h^2)$: it contains the real amplitude correction and the $O(h^3)$ phase correction $\psi$, and no term of order $h$. \emph{Step 3: splitting the residual.} Put $a_\varepsilon:=a(q+\varepsilon h/2;h^2)$ and $\delta_\varepsilon:=[S(q+\varepsilon h)-S(q)]/h$ for $\varepsilon=\pm1$. Then \begin{equation} e^{-iS(q)/h}(\mathcal L_hu)(q)=E(q)\,T(q)-\sum_{\varepsilon=\pm1}a_\varepsilon e^{i\delta_\varepsilon}V(q+\varepsilon h)\bigl[E(q+\varepsilon h)-E(q)\bigr], \label{eq:cubic-split} \end{equation} where $T(q):=(x-b(q;h^2))V(q)-\sum_\varepsilon a_\varepsilon e^{i\delta_\varepsilon}V(q+\varepsilon h)$ is the residual of the pure ansatz $e^g\exp\{ih^{-1}(S+h^2S_2)\}$ of Proposition~\ref{prop:generic-second-weyl-eikonal} (times the constant $\omega_{ab}\pi^{-1/2}$ and $e^{\mathrm{const}}$). Since $E-1=h^2m$ with $m$ smooth, $E(q+\varepsilon h)-E(q)=h^2[m(q+\varepsilon h)-m(q)]=O(h^3)$; the coefficients $a_\varepsilon$ and $V$ are bounded, so the sum in \eqref{eq:cubic-split} is $O(h^3)$. The amplitude corrections are therefore invisible at orders $h^0,h^1,h^2$ for two reasons: the slowly varying factor $E=1+O(h^2)$ multiplies $T$, whose principal coefficients vanish (Step 4), and its variation across a step is $O(h^3)$. \emph{Step 4: the orders $h^0,h^1,h^2$ of $T$.} Taylor expansion with $S_q=p$, $S_{2,q}=\Sigma_2$ gives \begin{equation} \begin{aligned} \delta_\varepsilon&=\varepsilon p+\tfrac h2p_q+\varepsilon\tfrac{h^2}6p_{qq}+O(h^3),\\ g(q+\varepsilon h)-g(q)&=\varepsilon hg_q+\tfrac{h^2}2g_{qq}+O(h^3),\qquad ih\bigl[S_2(q+\varepsilon h)-S_2(q)\bigr]=i\varepsilon h^2\Sigma_2+O(h^3), \end{aligned} \end{equation} so that $V(q+\varepsilon h)/V(q)=\exp X_\varepsilon$ with $X_\varepsilon=i\varepsilon p+h\alpha_\varepsilon+h^2\beta_\varepsilon+O(h^3)$, $\alpha_\varepsilon=\varepsilon g_q+\frac i2p_q$, $\beta_\varepsilon=\frac12g_{qq}+i\varepsilon\bigl(\frac16p_{qq}+\Sigma_2\bigr)$. With $a_\varepsilon=a_0+\varepsilon\frac h2a_{0,q}+h^2(a_2+\frac18a_{0,qq})+O(h^3)$ and $b=b_0+h^2b_2+O(h^4)$ (from \eqref{eq:class-a2b2}), and the elementary sums $\sum_\varepsilon e^{i\varepsilon p}=2\cos p$, $\sum_\varepsilon\varepsilon e^{i\varepsilon p}=2i\sin p$, one finds \begin{equation} \frac{T}{V}=c_0+hc_1+h^2c_2+O(h^3), \end{equation} \begin{align} c_0&=x-b_0-2a_0\cos p,\\ c_1&=-i\bigl[2a_0g_q\sin p+a_0p_q\cos p+a_{0,q}\sin p\bigr],\\ c_2&=-b_2-\Bigl\{a_0(g_{qq}+g_q^2)\cos p-a_0g_qp_q\sin p-\tfrac{a_0}4p_q^2\cos p-\tfrac{a_0}3p_{qq}\sin p\notag\\ &\qquad\qquad+2a_2\cos p+a_{0,q}g_q\cos p-\tfrac12a_{0,q}p_q\sin p+\tfrac14a_{0,qq}\cos p-2a_0\Sigma_2\sin p\Bigr\}. \end{align} (The coefficient of $h^2$ in $\sum_\varepsilon a_\varepsilon e^{X_\varepsilon}$ is $\sum_\varepsilon e^{i\varepsilon p}\bigl[a_0(\beta_\varepsilon+\frac12\alpha_\varepsilon^2)+\frac12\varepsilon a_{0,q}\alpha_\varepsilon+a_2+\frac18a_{0,qq}\bigr]$, where $\beta_\varepsilon+\frac12\alpha_\varepsilon^2=\frac12(g_{qq}+g_q^2)-\frac18p_q^2+i\varepsilon\bigl(\frac16p_{qq}+\Sigma_2+\frac12g_qp_q\bigr)$; with $\sum_\varepsilon e^{i\varepsilon p}(R+i\varepsilon I)=2R\cos p-2I\sin p$ this is the brace in $c_2$.) Now $c_0=0$ is the principal eikonal $x-b_0=2a_0\cos p$; $c_1=0$ is the transport equation, because $g=-\frac12\log(a_0\sin p)+\mathrm{const}$ gives $2a_0g_q\sin p=-a_{0,q}\sin p-a_0p_q\cos p$; and $c_2=2a_0\sin p\,\bigl(\Sigma_2-\Sigma_2[a_0,a_2,b_2;p,g]\bigr)$ with $\Sigma_2[\cdot]$ the density \eqref{eq:generic-second-weyl-density}, which is $0$ because $S_{2,q}=\Sigma_2$ by \eqref{eq:cubic-S2}. The same three equations are the orders $h^0,h^1,h^2$ in the proof of Proposition~\ref{prop:generic-second-weyl-eikonal}; $c_1$ is purely imaginary and $c_2$ real, so a real phase correction $S_2$ alone closes the order $h^2$. Hence $T=V\cdot O(h^3)=O(h^3)$. \emph{Step 5: conclusion.} By \eqref{eq:cubic-split} and Step 4, $e^{-i\Theta_q/z}\mathcal L_hu=O(h^3)=O(z^3)$ at fixed $q$, with $S/h=\Theta_q/z$; that is, $C^+_0=C^+_1=C^+_2=0$, and $|\mathcal L_hu|\le Ch^3$. With Step 1, $|\mathcal R_k(\widehat U^{[2]})|\le Ch^3$ uniformly on $\mathcal C_{q_0}$, which is \eqref{eq:general-second-local-residual}. The first coefficient that can survive is $C^+_3$, to which both the $h^3$ terms of $T$ and the second term of \eqref{eq:cubic-split} contribute; Lemma~\ref{lem:cubic-residual-transport} shows that $C^+_3/B_0$ is imaginary. \end{proof} The next two lemmas give the corner argument behind the large-rank bound \eqref{eq:class-engine-large-rank}. The first is the model argument for a scalar remainder. The second, which is the one used here, repeats it carrier by carrier for the two-carrier Wigner--Darboux residual, where the carriers must first be separated. \begin{lemma}[Two-scale corner divisibility] \label{lem:two-scale-corner-divisibility} Let $p,s,R$ be positive integers with $R>p$. Suppose \begin{equation} A(q,z)=\sum_{r=0}^{R-1}z^rC_r(q)+E_R(q,z), \qquad 0\le q\le q_*,\quad 00$, and that for every fixed admissible discrete $z$, \begin{equation} A(q,z)=O_z(q^s)\quad(q\to0), \end{equation} where the admissible $z$-set contains a sequence tending to zero. Then \begin{equation} \boxed{|A(q,z)|\le C\{q^sz^p+z^R\}} \label{eq:two-scale-corner-divisibility-bound} \end{equation} uniformly on the closed corner. \end{lemma} \begin{proof} The fixed-positive-$q$ estimate first gives $C_0=\cdots=C_{p-1}=0$ on $(0,q_*]$, hence on $[0,q_*]$ by continuity. For $0\le mp$ be an integer, put $\kappa=k+\tfrac12$ and $z=\kappa^{-1}$, and treat $(q,z)$ as independent continuum variables with $h:=qz$ (so $q=\kappa h$ at the lattice points). Consider \begin{equation} \widehat U_k(h)=\sqrt{2k+1}\,W(q,h^2) d^k_{ab}\!\left(\Theta(q,h^2)\right). \end{equation} Assume on $0\le q\le q_*+\delta$ (for some $\delta>0$; the conclusion concerns $0\le q\le q_*$) that the exact Jacobi coefficients, $W$, and $\Theta$ are smooth in the blown-up variables $(q,z)$ after $h=qz$, with at least four $q$ derivatives and enough uniform inverse-rank derivatives for a finite Darboux expansion of order $R+3$ with three derivatives of its carrier-separated remainders (Lemma~\ref{lem:wigner-darboux-differentiated-closure} with order $R+3$ and $s=3$). Assume also the \emph{even-$h$ midpoint dependence} \begin{equation} \Theta(q,h^2)=\Theta_0(q)+h^2H(q,h^2) \end{equation} with smooth bounded $H$ on this closed blown-up corner, and that $\Theta_0(q):=\Theta(q,0)$ stays in a compact subinterval of $(0,\pi)$. Put \begin{equation} S_0(q):=q\Theta_0(q),\qquad p_0(q):=S_0'(q). \label{eq:corner-rank-phase} \end{equation} If (i) at fixed $q>0$ the scalar residual is $O(h^p)$ and (ii) at every fixed admissible rank $k$ the residual is analytic in $h^2$ with vanishing $h^0$ and $h^2$ coefficients, then \begin{equation} \boxed{|\mathcal R_k(\widehat U)|\le C(h^p+\kappa^{-R})} \label{eq:darboux-weyl-corner-bound} \end{equation} uniformly on the closed corner for all sufficiently large $k$. \end{lemma} \begin{proof} The two complex Darboux branches must be separated explicitly: cancellation for the \emph{real} Wigner solution does not by itself give cancellation in each branch. \emph{The independent variables.} The residual is regarded as a function of two independent continuum variables $(q,z)$ on the closed corner $0\le q\le q_*$, $00$, so that the neighbouring ranks below stay in the domain (this is the margin $\delta_{\mathcal C}$ in the definition of a turning-free corridor); the conclusion concerns $q\le q_*$. \emph{Carrier-separated residual.} Insert the differentiated expansion of Lemma~\ref{lem:wigner-darboux-differentiated-closure} of order $R+3$ with $s=3$ at the angle $\vartheta=\widetilde\Theta(q,z)$, which stays in a compact subinterval of $(0,\pi)$: \begin{equation} \widehat U_k=\sum_\sigma e^{i\sigma\Theta_0(q)/z}\,\mathfrak U^\sigma(q,z),\qquad \mathfrak U^\sigma=\widetilde We^{i\sigma zq^2H}\Bigl[\sum_{r\le R+2}z^rA^\sigma_{r,ab}(\widetilde\Theta)+\mathcal E^\sigma_{R+3,k}(\widetilde\Theta)\Bigr], \end{equation} using $\kappa\widetilde\Theta=\Theta_0/z+zq^2H$. The prefactor $\widetilde We^{i\sigma zq^2H}$ and the finite sum are smooth functions of $(q,z)$ on the corner (including $z=0$). The remainder part is $\widetilde We^{i\sigma zq^2H}\mathcal E^\sigma_{R+3,k}(\widetilde\Theta(q,z))$ at \emph{fixed} $k$, i.e.\ fixed $z$; by the chain rule (Fa\`a di Bruno, with $\partial_q^j\widetilde\Theta$ bounded) and \eqref{eq:wigner-darboux-diff-closure-rem} its $q$-derivatives of order $m\le3$ are $O(z^{R+3})$. The derivatives that fall on the carrier $e^{i\sigma\Theta_0/z}$ are not included here: they are taken explicitly and cost $z^{-1}$ each. \emph{Neighbouring ranks.} At the same $h$ the ranks $k\pm1$ correspond to the continuum points \begin{equation} (q,z)\longmapsto (q_\pm,z_\pm)=\left(q(1\pm z),\frac{z}{1\pm z}\right), \end{equation} since $\kappa\pm1=1/z_\pm$ and $(\kappa\pm1)h=q_\pm$; note $q_\pm z_\pm=qz=h$, so $\Theta(q_\pm,h^2)=\widetilde\Theta(q_\pm,z_\pm)$ and the same functions $\widetilde\Theta,\widetilde W,H$ apply at the shifted points. Hence \begin{equation} \widehat U_{k\pm1}=\sum_\sigma e^{i\sigma\Theta_0(q_\pm)/z_\pm}\,e^{i\sigma z_\pm q_\pm^2H(q_\pm,z_\pm)}\,\widetilde W(q_\pm,z_\pm)\,[\cdots], \end{equation} with the same bracket evaluated at $(q_\pm,z_\pm)$, and its remainder part is the above remainder composed with $(q,z)\mapsto(q_\pm,z_\pm)$. This composition preserves the bounds: the map is smooth, $\partial_q[q(1\pm z)]=1\pm z$, the fixed rank is now $k\pm1$ (so the same lemma applies with $\kappa\pm1\asymp\kappa$), and $z_\pm\asymp z$, so the $q$-derivatives of order $m\le3$ of the shifted remainder are $O(z^{R+3})$. The carrier quotients have smooth expansions: with $\Theta_0(q(1\pm z))=\Theta_0(q)\pm zq\Theta_0'(q)+O(z^2)$, \begin{equation} \frac{\Theta_0(q_\pm)}{z_\pm}-\frac{\Theta_0(q)}z =\pm\bigl(\Theta_0+q\Theta_0'\bigr)+zg_\pm^{(0)}(q,z) =\pm p_0(q)+zg_\pm^{(0)}(q,z), \end{equation} with $g^{(0)}_\pm$ smooth on the corner. Thus $e^{i\sigma\Theta_0(q_\pm)/z_\pm}=e^{i\sigma\Theta_0(q)/z}\,e^{\pm i\sigma p_0(q)+i\sigma zg^{(0)}_\pm}$, and the extra factor is smooth on the corner. The exact coefficients $A_{k\pm1},A_k,B_k$ and the neighbouring-rank shifts are smooth in $(q,z)$, so \begin{equation} \mathcal R_k(q,z)= e^{i\Theta_0(q)/z}\left[\sum_{r=0}^{R+2}z^r C_r^+(q)+E_+(q,z)\right] +e^{-i\Theta_0(q)/z} \left[\sum_{r=0}^{R+2}z^r C_r^-(q)+E_-(q,z)\right], \label{eq:corner-two-carrier-residual} \end{equation} where $\sum_rz^rC_r^\pm(q)$ is the Taylor polynomial in $z$ at fixed $q$ of the smooth finite parts (so $C_r^\pm\in C^4([0,q_*])$), and $E_\pm$ collects the Taylor remainder, of size $O(z^{R+3})$ with $q$-derivatives, and the three composed remainders of the carrier-separated expansions, so that $\|\partial_q^mE_\pm(\cdot,z)\|_\infty\le Cz^{R+3}$ for $0\le m\le3$. The carriers stand outside the brackets, so each $q$-derivative that falls on a carrier costs one factor $z^{-1}$; the expansion is therefore carried three orders beyond the order $z^R$ needed in \eqref{eq:darboux-weyl-corner-bound}. It is a finite-order statement and does not assert convergence of the Darboux series. All coefficient bounds are uniform because the angles stay in an interior compact set and the hopping and its reciprocal are smooth in the blown-up variables. Thus the one-step recurrence produces no uncontrolled factors $e^{i\,\mathrm{const}/z}$ beyond the two carriers in \eqref{eq:corner-two-carrier-residual}. We use the following elementary uniqueness fact. If $0<\vartheta<\pi$ and \begin{equation} e^{i(k+1/2)\vartheta}\sum_{r=0}^{M-1}a_r(k+1/2)^{-r} +e^{-i(k+1/2)\vartheta}\sum_{r=0}^{M-1}b_r(k+1/2)^{-r} =O(k^{-M}) \end{equation} for every sufficiently large rank $k$ of an admissible consecutive sequence $k\in k_{\min}+\mathbb Z_{\ge0}$ ($k_{\min}$ integer or half-integer), then $a_r=b_r=0$ for $0\le r0$ the assumed $O(h^p)=O(q^pz^p)$ scalar residual, together with this uniqueness argument applied to \eqref{eq:corner-two-carrier-residual}, forces $C_0^\pm=\dots=C_{p-1}^\pm=0$ (first for fixed $q$ and then at $q=0$ by continuity). Next fix $k$, hence $z$, and let $q\downarrow0$, with $h=qz$. Even analyticity in $h$ and the vanishing $h^0,h^2$ contact coefficients imply $\partial_q^m\mathcal R_k(0,z)=0$ for $0\le m<4$. We show by induction on $m\in\{0,1,2,3\}$ that \begin{equation} \partial_q^mC_r^\pm(0)=0 \qquad (p\le r\le R+2-m). \end{equation} Differentiate \eqref{eq:corner-two-carrier-residual} $m$ times at $q=0$. The remainder contributes $\partial_q^m(e^{\pm i\Theta_0/z}E_\pm)(0,z)=O(z^{R+3-m})$, since each derivative on the carrier costs at most $z^{-1}$. In the Leibniz expansion of $\partial_q^m(e^{\pm i\Theta_0/z}z^rC_r^\pm)$ at $q=0$, a term with $j\ge1$ derivatives on the carrier is $O(z^{r-j})\,\partial_q^{m-j}C_r^\pm(0)$; it vanishes by the induction hypothesis when $r\le R+2-(m-j)$, and otherwise $r-j\ge R+3-m$, so it is $O(z^{R+3-m})$. (For $m=0$ there are no such terms.) What remains is \begin{equation} e^{i\Theta_0(0)/z}\sum_{r=p}^{R+2-m}z^r\partial_q^mC_r^+(0) +e^{-i\Theta_0(0)/z}\sum_{r=p}^{R+2-m}z^r\partial_q^mC_r^-(0) =O(z^{R+3-m}), \end{equation} and the uniqueness fact with $\vartheta=\Theta_0(0)$ and $M=R+3-m$ gives $\partial_q^mC_r^\pm(0)=0$ for $p\le r\le R+2-m$. In particular $\partial_q^mC_r^\pm(0)=0$ for $p\le r0$ and $e_\pi$, $\sin p$, $\sin\Theta_q$ bounded below, and $\mathcal C$ contains the whole segment $\{(r,x,\pi):0\le r\le q\}$ with each of its points. By compactness and continuity there is a complex neighbourhood $\mathcal V$ of $\mathcal C$ in the variables $(q^2,x)$, with $(\sigma q^2,x)\in\mathcal V$ whenever $(q^2,x)\in\mathcal V$ and $0\le\sigma\le1$, and with the following properties for all $(q^2,x)\in\mathcal V$: the data $e_\pi(r)$, $v_\pi(r,0)$ of (E1) are analytic and $w:=(x-b_0(r))/e_\pi(r)$ stays in a fixed compact neighbourhood of a subset of $(-1,1)$ disjoint from the cuts $(-\infty,-1]\cup[1,\infty)$ of $\arccos$, equivalently $D(r,x)=e_\pi(r)^2(1-w^2)\ne0$ (here $r=q^2$ and all $r=\sigma q^2$); and $\Theta_q$ lies in a fixed complex neighbourhood of a compact subinterval of $(0,\pi)$, so that $\sin\Theta_q\ne0$. On $\mathcal V$, $p=\arccos w$ and $D$ are analytic in $(q^2,x)$, hence so are $\Theta_q=\int_0^1p(\sigma^2q^2,x)\,\dd\sigma$ (the substitution $s=\sigma q$ in $S=\int_0^qp(s^2,x)\,\dd s=q\Theta_q$), $S$, and $D^{-1/4}$ (on a small neighbourhood $D$ stays in a disc not containing $0$); $\Sigma_2$ is a fixed algebraic expression in $a_0,a_2,b_2,p,g$ and their $q$-derivatives, hence analytic in $(q^2,x)$ for $q\ne0$, and by Proposition~\ref{prop:general-second-eikonal-pole} $q^2\rho_2$ is analytic in $(q^2,x)$ and vanishes at $q^2=0$ identically in $x$, so $\rho_2$, $\Gamma_2/q$, $\Theta^{[2]}_q$, $W^{[2]}_q$ are analytic in $(q^2,x,h^2)$ on $\mathcal V$. The Darboux data $\mathfrak c_{ab}$ and the amplitudes $\mathfrak a_{r,ab}$ of \eqref{eq:darboux-parity} are analytic functions of $\vartheta\in(0,\pi)$, being obtained from the analytic potential $V$ of the proof of Lemma~\ref{lem:wigner-darboux-differentiated-closure} by the Liouville--Green recursion there (each step integrates an analytic function; the constants of integration are numbers), so their compositions with $\Theta_q$ and $\Theta^{[2]}_q$ are analytic on $\mathcal V$. Finally $a$, $b$ of \eqref{eq:midpoint-operator} do not involve $x$ at all, and $x$ enters $\mathcal L_h$ only through the diagonal term $x-b$. Each $C_r^+$ and $B_r$ is a fixed polynomial in these functions and their $q$-derivatives (the formal expansion described at the end of this proof), so it is analytic in $(q^2,x)$ on $\mathcal V$. The reason is thus not any special dependence of the recurrence on $x$: $p$, $D$, $\Theta_q$, $\Sigma_2$ and the Darboux amplitudes are analytic in $(q^2,x)$ wherever $D\ne0$ and $\sin\Theta_q\ne0$. \emph{Evenness and vanishing.} At fixed $k$ every factor of $\widehat U^{[2],+}_k$ is analytic and even in $q$ ($h^2=q^2z^2$; $\Theta^{[2]}$, $D$ are even analytic by Proposition~\ref{prop:general-second-eikonal-pole}), and so are the exact coefficients; the carriers of the neighbouring ranks are $e^{i\Theta_q/z}$ times functions analytic in $(q,z)$, as in the proof of Lemma~\ref{lem:darboux-weyl-corner}. This gives the even coefficients. The vanishing of $C^+_0,C^+_1,C^+_2$ is the direct expansion in the proof of Lemma~\ref{lem:general-cubic-residual} (Steps 3--5 there); the bound $C^+_3=O(q^4)$ is proved in the proof of Lemma~\ref{lem:darboux-weyl-corner} applied to $\widehat U^{[2]}$ (Corollary~\ref{cor:general-matched-remainder}), whose fixed-ratio hypothesis~(i) is Lemma~\ref{lem:general-cubic-residual}; and $B_0=W^{[2]}_qA^+_{0,ab}(\Theta^{[2]}_q)=\omega_{ab}\pi^{-1/2}D^{-1/4}$ by \eqref{eq:darboux-parity}, since $W^{[2]}_q=(\sin\Theta^{[2]}_q)^{1/2}D^{-1/4}$. For \eqref{eq:cubic-transport-type} regard $q>0$ as fixed and $h$ as a real parameter of either sign, with $\kappa=q/h$. This is a statement about \emph{formal finite branch expansions}, not about Wigner functions of negative rank: $\widehat U^{[2],\pm}$ denotes the explicit expression \eqref{eq:corrected-branch}, a carrier $e^{\pm i\kappa\Theta^{[2]}_q}$ times a finite sum $W^{[2]}_q\sum_{r0$, Lemma~\ref{lem:cubic-residual-transport} gives $\mathcal L_h\widehat U^{[2],+}=e^{i\Theta_q/z}[h^3q^{-3}C^+_3+O(h^4)]$. For a smooth $\alpha$, \begin{equation} \mathcal L_h(h^2\alpha u)=h^2\alpha\,\mathcal L_hu-h^2\sum_\pm a(q\pm\tfrac h2)\bigl(\alpha(q\pm h)-\alpha(q)\bigr)u(q\pm h), \end{equation} and with $u=\widehat U^{[2],+}$ one has $u(q\pm h)=e^{\pm ip}u(q)\,(1+O(h))$ and $a=\frac12e_\pi+O(h^2)$, so the sum is $ih\,\alpha_q\,e_\pi\sin p\,B_0e^{i\Theta_q/z}+O(h^2)$. Thus the $h^3$ coefficient of $\mathcal L_h\widehat U^{[3],+}$, with $\widehat U^{[3],+}=(1+h^2\alpha)\widehat U^{[2],+}$, is \begin{equation} q^{-3}C^+_3-i\,\alpha_q\,e_\pi\sin p\,B_0=0 \end{equation} by \eqref{eq:phase-exactness-alpha}. The same holds for the conjugate branch because $\alpha$ is real, and the Darboux remainder contributes $O(\kappa^{-R})$. Hence $|\mathcal R_k(\widehat U^{[3]})|\le Ch^4$ at fixed positive ratio. \emph{Fixed-rank residual, corner, and summation.} At fixed $k$, $h^2\alpha(q_k,x)$ is analytic in $t$ (since $\alpha(q,x)$ is even analytic in $q$ and $q_k^2=\kappa^2t$), so $\widehat U^{[3]}$ is analytic at fixed rank; the $t$ coefficient of its residual is that of $\widehat U^{[2]}$, which vanishes by Proposition~\ref{prop:general-fixed-rank-contact}, plus $\alpha(0,x)$ times the residual of $\widehat d$, which is zero. Cauchy's estimate on $|t|\le c\kappa^{-3}$ gives $|\mathcal R_k(\widehat U^{[3]})|\le Ch^4k^6$ for $k^3h^2\le\eta_0$, as in Proposition~\ref{prop:general-engine-inputs}, and Lemma~\ref{lem:darboux-weyl-corner} with $(p,R)=(4,22)$ (its blown-up hypotheses hold as in Corollary~\ref{cor:general-matched-remainder}, since $1+q^2z^2\alpha$ is analytic in $(q,z)$) gives $|\mathcal R_k(\widehat U^{[3]})|\le C(h^4+k^{-22})$ for $k\ge k_0$. At the fixed start rank $k_0$, $\widehat Y_{k_0}-\widehat U^{[3]}_{k_0}=t(\delta_{k_0}-\alpha(0,x)\widehat d_{k_0})+O(t^2)=O(h^4)$ by the choice $\alpha(0,x)=c(x)$, at both start entries. Applying Theorem~\ref{thm:second-weyl-darboux-engine} with $p=4$ (split rank $K_*=\min\{L,\lfloor h^{-1/7}\rfloor\}$) and Corollary~\ref{cor:general-fixed-start} gives $\sup_{k_0\le k\le L}|\widehat Y_k-\widehat U^{[3]}_k|=O(h^3)$; finitely many ranks below $k_0$ are covered by fixed-rank analyticity. \end{proof} \subsection*{Proof of Corollaries~\ref{cor:3j-phase-exactness} and~\ref{cor:6j-phase-exactness}: the fixed-rank datum} \begin{proof} It remains to show $c(x)=-\mathsf J_2(0,x)$. By the proof of Proposition~\ref{prop:general-fixed-rank-contact}, the coefficient of $t$ in $U^{[2]}_k$ is $\mathcal Q_{2,k}d_k$, a first-order operator in $x$ applied to $d_k=d^k_{ab}(\arccos x)$, so $c(x)d_k$ is the coefficient of $t$ in $Y_k$ minus $\mathcal Q_{2,k}d_k$, at any one rank $k$. For the Hahn chain \cite[Theorem~6.1]{LewisFixedRank} gives that coefficient as $(A_2d_k'+B_2d_k)/(12(x^2-1))$ with $A_2=2J(x^3-x)-4x(M^2+\sigma^2)+2M\sigma(x^2+3)+4x$, $B_2=J(x^2-3)+2M\sigma x$, $J=k(k+1)$. With \eqref{eq:hahn-class-data} and \eqref{eq:general-rho20} the $d_k'$ coefficients of the two agree identically in $(k,x)$, and the $d_k$ coefficients differ by $((8(M^2+\sigma^2)-5)x^2+4M\sigma x(x^2-5)-3)/(48(1-x^2)^2)=-\mathsf J_2(0,x)$. For the centered Racah chain \cite[Theorem~9.1]{LewisFixedRank} gives the coefficient as $P_kd_k'+Q_kd_k$ for its kernel $(-1)^{m-n}\Psi_k$, which has the same ratio to the Wigner function; with $A,B$ replaced by $\alpha,\beta$ and \eqref{eq:racah-class-data}, the $d_k'$ coefficients again agree and the $d_k$ coefficients differ by $-\mathsf J_2(0,x)$ from \eqref{eq:phase-exactness-J20}, identically in $(k,m,n,\alpha,\beta,x)$. Both identities are rational-function identities, checked by exact computer algebra. \end{proof} \section{Analytic continuation of bulk rank estimates and zero stability} \label{app:bulk-zero-continuation} \begin{lemma}[Polynomial continuation and complex-angle Darboux bounds] \label{lem:bulk-zero-continuation} For an even centered Wigner--Jacobi chain on a turning-free corridor $\mathcal C$ (satisfying the $k=0$ row of Theorem~\ref{thm:general-phase-exactness} when $a=b=0$), for all sufficiently large admissible $N$ the exact solution continues in the spectral coordinate $x$ as \begin{equation} Y_k(x;h)=G_N(x)P_{k,N}(x)\qquad(k_{\min}\le k\le L), \end{equation} where $G_N$ is analytic and zero-free on a fixed complex neighbourhood $U$ of the real compact set and $P_{k,N}$ is a polynomial in $x$ of degree $k-k_{\min}$ whose zeros are all real and simple; for the Hahn, centered Racah, and quantum $6j$ chains its zeros are those of the Hahn, Racah, or $q$-Racah polynomial. Moreover, for $\Re\vartheta$ in a compact subinterval of $(0,\pi)$ and $\kappa|\Im\vartheta|\le C$, one has $\sqrt{2k+1}\,|d^k_{ab}(\vartheta)|\le C'$. \end{lemma} \begin{proof} The minimal admissible rank recurrence has one vanishing backward hopping, $A_{k_{\min}}=0$, while on $\mathcal C$, $e_\pi(k^2h^2,h^2)>0$ and hence $A_k(h)>0$ for $k_{\min}0$; its eigenvalues, the zeros of $P_{k,N}$, are real and simple. In the Racah and quantum $6j$ cases $x$ is affine in $C^2=c(c+1)+\frac14$ or $\cosh\mathsf Z$, so these are precisely zeros in the standard spectral variable. For the complex-angle bound, the function $w=(\sin\vartheta)^{1/2}d^k_{ab}(\vartheta)$ solves $w''+(\kappa^2-V)w=0$ with $V=(a^2+b^2-\frac14-2ab\cos\vartheta)/\sin^2\vartheta$ analytic and bounded on the stated set, so the energy $|w'|^2+\kappa^2|w|^2$ grows at most by the factor $e^{(2\kappa+C/\kappa)|\Im\vartheta|}$ along vertical segments, and the real interior bounds $\sqrt{2k+1}\,|d^k_{ab}|\le C$, $\sqrt{2k+1}\,|\partial_\vartheta d^k_{ab}|\le C\kappa$ (Lemma~\ref{lem:wigner-darboux-differentiated-closure}) give $\sqrt{2k+1}\,|d^k_{ab}(\vartheta)|\le C'$. \end{proof} \begin{lemma}[Real-axis bounds continue to an $h$-strip] \label{lem:real-to-strip} Let $F$ be analytic on a neighbourhood of the closed disc $|x-x_0|\le r$, with $|F|\le\varepsilon$ on $[x_0-r,x_0+r]$ and $|F(x)|\le Me^{\Lambda|\Im x|}$ on the disc, $M\ge\varepsilon>0$. Then for real $|y|0$ such that, for $\Re x$ in a slightly enlarged real compact set, $|\Im x|\le r$ and every rank $L$ of the family, \begin{equation} |\widehat Y_L(x)|\le Ce^{CN|\Im x|},\qquad |\widehat U^{[2]}_L(x)|+|\widehat U^{[3]}_L(x)|\le CN^{1/2}e^{CN|\Im x|}, \end{equation} where $\widehat U^{[3]}=(1+h^2\alpha)\widehat U^{[2]}$ is the comparison of Theorem~\ref{thm:general-phase-exactness}. \end{lemma} \begin{proof} For real $\vartheta$, $d^k_{ab}(\vartheta)=\langle k,a|e^{-i\vartheta J_y}|k,b\rangle=\sum_{\mu}\langle k,a|v_\mu\rangle e^{-i\vartheta\mu}\langle v_\mu|k,b\rangle$ over an orthonormal eigenbasis of $J_y$ with $|\mu|\le k$; this formula continues to complex $\vartheta$, and Cauchy--Schwarz gives $|d^k_{ab}(\vartheta)|\le e^{k|\Im\vartheta|}$. The functions $\Theta^{[2]}_q(x)$, $W^{[2]}_q(x)$ and $\alpha(q,x)$ are analytic in $x$ on a fixed complex neighbourhood and real on the real axis, so $|\Im\Theta^{[2]}_q(x)|\le C|\Im x|$; this gives the pullback bounds. For the exact solution let $G_k(x_r)$, $x_r=\Re x$, be the bounded real fundamental matrix of Corollary~\ref{cor:general-fixed-start}. The exact transfer at $x$ differs from that at $x_r$ only by $i\Im x\,A_{k+1}^{-1}E_{11}$, so in the interaction representation each step is $I+V_k$ with $\|V_k\|\le C|\Im x|$, and the product over $L=O(N)$ ranks is at most $e^{CN|\Im x|}$. The finitely many ranks below the start rank are analytic and bounded on a fixed complex neighbourhood by (E2) (the real set $\{|x-X|\le r\}$ lies in $U$ for $r$ small, so the chain satisfies (E1)--(E2) on this slightly enlarged real compact set). \end{proof} \subsection*{Proof of Theorem~\ref{thm:general-bulk-zeros}: Bulk zero quantization} \begin{proof} The polynomial structure, and the reality and simplicity of the zeros of $P_{L,N}$, are proved in Lemma~\ref{lem:bulk-zero-continuation}, so it suffices to count real zeros. Put $F=\widehat Y_L-\widehat U^{[3]}_L$. By Theorem~\ref{thm:general-phase-exactness}, applied on a slightly larger turning-free set, $|F|\le Ch^3$ on the real axis; by Lemma~\ref{lem:strip-growth} the growth hypothesis of Lemma~\ref{lem:real-to-strip} holds with $\Lambda=CN$ and $M=CN^{1/2}$. Hence $|F|\le Ch^3$, $|F'|\le Ch^2$ and $|F''|\le Ch$ on the real compact set. The zeros of $\widehat U^{[3]}_L$ are those of $d^L_{ab}(\Theta^{[2]}_q(x))$, because $1+h^2\alpha>0$. By Lemma~\ref{lem:darboux-parity} (with differentiated remainder, Lemma~\ref{lem:wigner-darboux-differentiated-closure}), $\widehat U^{[3]}_L=\epsilon_{ab}(2/\sqrt\pi)(1+h^2\alpha)D^{-1/4}[\mathfrak A\cos\Phi+O(h^R)]$ with $\Phi=\kappa\Theta^{[2]}_q+p_{ab}\pi/2-\pi/4+\mathfrak P$ and $\partial_x\Phi=N\partial_xS+O(1)$ (each $x$-derivative on the carrier-separated remainder costs at most $O(N)$: since $\mathfrak A$ and $\mathfrak P$ are smooth in $(\vartheta,\kappa^{-1})$, the $O(\kappa^{-R})$ remainder of Lemma~\ref{lem:darboux-parity} inherits the $C^s$ bounds of Lemma~\ref{lem:wigner-darboux-differentiated-closure}, with a factor $\kappa$ per derivative); on a turning-free set $\partial_xS=-\int_0^q\dd s/(e_\pi\sin p)\le-cq$. So the comparison zeros $\widetilde x_j$ are simple, $|\partial_x\widehat U^{[3]}_L(\widetilde x_j)|\ge cN$, $|\partial_x^2\widehat U^{[3]}_L|\le CN^2$, consecutive zeros are $\asymp N^{-1}$ apart, and $|\widehat U^{[3]}_L(x)|\ge cN\operatorname{dist}(x,\{\widetilde x_j\})$ within distance $cN^{-1}$ of the zero set and $\ge c$ further away. Now count zeros. At $\widetilde x_j\pm C_1h^4$ Taylor's theorem gives $\widehat U^{[3]}_L$ opposite signs and size at least $cC_1h^3-CN^2h^8\ge\frac c2C_1h^3$, while $|F|\le Ch^3$; for $C_1$ large, $\widehat Y_L$ has a zero in between. Within distance $c'N^{-1}$ of $\widetilde x_j$, with $c'$ small, $|\partial_x\widehat U^{[3]}_L|\ge cN/2$ and $|F'|\le Ch^2$, so $\widehat Y_L$ is strictly monotone there and that zero is the only one. At distance between $C_1h^4$ and $c'N^{-1}$ from the zero set, $|\widehat U^{[3]}_L|\ge cC_1h^3>|F|$, and further away $|\widehat U^{[3]}_L|\ge c>|F|$; so $\widehat Y_L$ has no other zeros. Running this on a slightly larger compact turning-free set, every zero of $\widehat Y_L$ in the original set lies within $C_1h^4$ of exactly one comparison zero, which may lie in an $O(h^4)$ collar outside it. For the quantization condition, at $\widetilde x_j$ one has $\cos\Phi=O(h^R)$ and $\mathfrak A\ge\frac12$, so $\Phi=(\nu_j+\frac12)\pi+O(h^R)$. Here $\kappa\Theta^{[2]}_q=NS+h\Gamma_2$ because $\kappa h=q$, and $\mathfrak P=\mathfrak c_{ab}(\Theta^{[2]}_q)/\kappa+O(\kappa^{-3})=h\,\mathfrak c_{ab}(\Theta_q)/q+O(h^3)$. Moving from $\widetilde x_j$ to $x_j$ costs $O(N\cdot h^4)=O(h^3)$ in $NS+h\widetilde\Gamma$. This proves \eqref{eq:general-bulk-zeros}. Finally $\Theta^{[2]}_q-\Theta_q=h^2\Gamma_2/q$ moves the angle by $O(h^2)$, hence the zero of $d^L_{ab}(\Theta_q(x))$ with the same index by $O(h^2)$, and dropping $h\widetilde\Gamma$, or only its Darboux part $h\,\mathfrak c_{ab}(\Theta_q)/q$, changes the left side of the quantization condition by a quantity of exact order $h$ wherever that term is nonzero. \end{proof} \section{Rank-direction simple-turn Airy connection} \label{sec:supp-rank-airy} This section proves Theorem~\ref{thm:general-rank-airy}(i): it verifies the simple-turning hypotheses for an even centered chain, constructs the canonical Airy basis, and identifies the physical combination. Throughout, the chain has a simple rank turn \eqref{eq:general-simple-turn} of fixed type $\tau$ on a compact family, $u(q)=(x-b_0(q))/e_\pi(q^2)$, and $\zeta_\tau$ is the coordinate \eqref{eq:general-zeta}. \begin{lemma}[Simple-turning hypotheses for even centered chains] \label{lem:general-airy-applicability} On a fixed neighbourhood of $q_*$ the natural-gauge recurrence \eqref{eq:class-jacobi} has off-diagonal coefficient bounded away from zero and coefficient functions admitting uniform smooth asymptotic expansions in powers of $N^{-1}$. Its principal characteristic quotient is $u$, and \begin{equation} 1-u(q)^2=\frac{D(q,x)}{e_\pi(q^2)^2}, \qquad \partial_q(1-u^2)(q_*)=\frac{\partial_qD(q_*,x)}{e_\pi(q_*^2)^2}<0. \label{eq:general-airy-simple-turn-identity} \end{equation} Consequently the two principal characteristic roots coalesce simply at $\lambda=\tau$. When $\tau=-1$, the exact gauge $Y_k=(-1)^{k-k_{\min}}\widetilde Y_k$ converts the collision at $\lambda=-1$ into the same simple $+1$ turning-point normal form. \end{lemma} \begin{proof} Since $q_*^20$, (E1) gives a fixed neighbourhood of $q_*$ on which $e_\pi$ is bounded below and $e_\pi$, $v_\pi$ are analytic; there $k\asymp N$, so $A_k^{(0)}$ and $B_k^{(0)}$ are analytic functions of $k^{-1}=O(N^{-1})$ with $2A_k^{(0)}=1+O(N^{-2})$. This gives the positive lower bound and the uniform expansions. Freezing the principal coefficients gives $\lambda+\lambda^{-1}=2u(q)$. The first identity in \eqref{eq:general-airy-simple-turn-identity} is the definition $D=e_\pi^2-(x-b_0)^2$, and the second follows from $D(q_*,x)=0$ and \eqref{eq:general-simple-turn}. Hence $u^2-1$ has a simple zero at $q_*$ and $u(q_*)=\tau$. The alternating gauge changes the signs of both characteristic roots and therefore reduces the $-1$ collision to the $+1$ case without changing its order. \end{proof} \begin{theorem}[Canonical Airy basis at a simple rank turn] \label{thm:general-canonical-airy} \label{thm:racahCanonicalAiryBasis} In the natural gauge $\widehat U_L=\sqrt{2L+1}\,U_L$, there is a real canonical Airy pair $\widehat U_{\mathrm{Ai}}^{(N)},\widehat U_{\mathrm{Bi}}^{(N)}$ of exact solutions of \eqref{eq:class-jacobi} on a fixed neighbourhood of $q_*$ such that, with $t=N^{2/3}\zeta_\tau(q)$ and $q=(L+\tfrac12)/N$, \begin{align} \widehat U_{\mathrm{Ai}}^{(N)}(L) &=2\chi_\tau(L)N^{1/6} \left|\frac{\zeta_\tau(q)}{D(q,x)}\right|^{1/4} \left[\operatorname{Ai}(t)+O_{A_0}(N^{-1})\right], \label{eq:general-canonical-Ai}\\ \widehat U_{\mathrm{Bi}}^{(N)}(L) &=2\chi_\tau(L)N^{1/6} \left|\frac{\zeta_\tau(q)}{D(q,x)}\right|^{1/4} \left[\operatorname{Bi}(t)+O_{A_0}(N^{-1})\right], \label{eq:general-canonical-Bi} \end{align} for $|t|\le A_0$. Here $\chi_+(L)=1$, while $\chi_-(L)=(-1)^{L-k_{\min}}$ is the exact alternating gauge, real for integer and half-integer ranks. The displayed remainders are additive inside the square brackets; no relative-error claim is intended at zeros of $\operatorname{Ai}$ or $\operatorname{Bi}$. On the entire fixed turning neighbourhood, before restricting to $|t|\le A_0$, the same exact pair satisfies the uniform estimate of Geronimo--Bruno--Van Assche Theorem~3.8 in the form of Geronimo's Theorem~2.6: each bracket error is at most $CN^{-1}(\operatorname{Ai}^2+\operatorname{Bi}^2)^{1/2}(t)$. On the forbidden side $t>0$ this weight has the size of $\operatorname{Bi}$, so the $\operatorname{Ai}$ member is not asserted to be recessive there; a recessive exact solution is constructed in Lemma~\ref{lem:airy2-recessive}. Undoing the natural gauge gives the physical scale \begin{equation} U_{\mathrm{Ai}}^{(N)}(L) =\chi_\tau(L)\sqrt{\frac{2}{q}}\,N^{-1/3} \left|\frac{\zeta_\tau(q)}{D(q,x)}\right|^{1/4} \left[\operatorname{Ai}(t)+O_{A_0}(N^{-1})\right], \label{eq:general-canonical-physical-scale} \end{equation} with the analogous formula for $\operatorname{Bi}$. In an oscillatory overlap $N^{-2/3}\ll|q-q_*|\ll1$, the $\operatorname{Ai}$ member is the canonical WKB cosine branch and the $\operatorname{Bi}$ member is its quadrature branch. \end{theorem} \begin{proof} Set $\varepsilon=N^{-1}$ and $s=k\varepsilon$; the midpoint variable is $q=s+\varepsilon/2$. In the natural gauge the exact equation \eqref{eq:class-jacobi} is a smoothly varying Jacobi recurrence \begin{equation} xY_k=A_{k+1}Y_{k+1}+B_kY_k+A_kY_{k-1}. \end{equation} On a fixed turning neighbourhood, (E1) and $k\asymp N$ give \begin{equation} A_k=a_0(s)+O(\varepsilon^2),\qquad B_k=b_0(s+\varepsilon/2)+O(\varepsilon^2), \quad a_0(s)=\tfrac12e_\pi(s^2),\quad b_0(q)=v_\pi(q^2,0), \label{eq:general-airy-coeff-expansion} \end{equation} with the corresponding discrete differences one power of $\varepsilon$ smaller. The lower bound on $e_\pi$ keeps $a_0$ uniformly positive. To match Geronimo--Bruno--Van Assche Eq.~(2.6) exactly, distinguish their smooth comparison coefficients from the coefficients of the actual recurrence. Set \begin{equation} \varepsilon^{-1}=N,\qquad y=x,\qquad a(s,\varepsilon)=a_0(s),\qquad b(s,\varepsilon)=b_0(s+\varepsilon/2), \end{equation} and use (E1) to regard \begin{equation} a_1(k\varepsilon,\varepsilon)=A_k,\qquad b_1(k\varepsilon,\varepsilon)=B_k \end{equation} as samples of smooth actual-coefficient functions on the fixed neighbourhood. Then \eqref{eq:general-airy-coeff-expansion} is literally their perturbation condition (3.52), \begin{equation} \sup|a-a_1|+\sup|b-b_1|=O(\varepsilon^2). \end{equation} Moreover their characteristic quotient, Eq.~(2.7), is exactly \begin{equation} \frac{y-b(s,\varepsilon)}{2a(s+\varepsilon/2,\varepsilon)} =\frac{x-b_0(q)}{e_\pi(q^2)} =u(q),\qquad q=s+\frac{\varepsilon}{2}. \end{equation} By \eqref{eq:general-airy-simple-turn-identity}, $u^2-1$ has a simple zero with positive derivative in the outward $q$ direction. After shrinking the fixed neighbourhood if necessary, the squared characteristic exponent $k^2$ of Geronimo--Bruno--Van Assche is therefore increasing in $s$ with one zero, as their Theorems~3.3 and~3.8 require: their hyperbolic side $t\ge t_0$ (Eqs.~(2.7)--(2.8)) is the forbidden side $q\ge q_*$ here. Because $e_\pi$ and $v_\pi$ are analytic on a fixed complex neighbourhood and $A_k^{(0)},B_k^{(0)}$ are analytic in $k^{-1}$, the $N^{-1}$ expansions of the coefficients and the finitely many scaled derivatives required by Geronimo--Bruno--Van Assche are uniform in $N$ and in the parameter family. Thus the compact family satisfies their uniform coefficient hypotheses, not merely the pointwise turning geometry. In the $\tau=-1$ case the exact transformation $Y_k=(-1)^{k-k_{\min}}\widetilde Y_k$ replaces the characteristic quotient by $-u$, changing the colliding root from $-1$ to $+1$ while leaving the coefficient bounds and $O(\varepsilon^2)$ comparison estimates unchanged. With the uniform hypotheses verified in Lemma~\ref{lem:general-airy-applicability}, Geronimo--Bruno--Van Assche Theorem~3.3, using their explicit Airy functions (2.10)--(2.11), gives an $O(\varepsilon^2)$ residual with Airy envelopes, and their Theorem~3.8 converts those parametrices to exact solutions with $O(\varepsilon)$ uniform envelope error under Eq.~(3.52)~\cite{GeronimoBrunoVanAssche}. In particular this control holds on the whole fixed turning neighbourhood, not only when the scaled Airy argument is bounded. In the form restated by Geronimo~\cite[Theorem~2.6, Eqs.~(2.22), (2.25)]{Geronimo2009}, the exact solutions $f_i$ and the parametrices $\psi_1=g\operatorname{Ai}(\varepsilon^{-2/3}\varrho)$, $\psi_2=g\operatorname{Bi}(\varepsilon^{-2/3}\varrho)$ satisfy $|f_i-\psi_i|\le c\varepsilon\,g\,|\tilde u_i(\varepsilon^{-2/3}\varrho)|$ at every lattice point of the neighbourhood (Eq.~(2.26) there after the alternating gauge), with the Hankel envelope $\tilde u_1(x)=(-x/3)^{1/2}e^{i\pi/6}H^{(1)}_{1/3}(\tfrac23(-x)^{3/2})$ and $\tilde u_2(x)=(-x/3)^{1/2}e^{-i\pi/6}H^{(2)}_{1/3}(\tfrac23(-x)^{3/2})$. For $x<0$, $\tilde u_1=\operatorname{Ai}-i\operatorname{Bi}$ and $\tilde u_2=\overline{\tilde u_1}$, so $|\tilde u_i|=(\operatorname{Ai}^2+\operatorname{Bi}^2)^{1/2}$; for $x>0$, evaluating $(-x)^{1/2}=i\sqrt x$ with the principal branches of $H^{(1,2)}_{1/3}$ at $-i\tfrac23x^{3/2}$ gives $\tilde u_1=\operatorname{Ai}+i\operatorname{Bi}$ and $\tilde u_2=-2\operatorname{Ai}$. Only the bound $|\tilde u_i(x)|\le2(\operatorname{Ai}(x)^2+\operatorname{Bi}(x)^2)^{1/2}$ for real $x$ is used here, that is, the weight $(\operatorname{Ai}^2+\operatorname{Bi}^2)^{1/2}$ up to a factor $2$. We use only this non-recessive envelope; the recessive solution is constructed separately (Lemma~\ref{lem:airy2-recessive}). The Wang--Wong theory provides an equivalent general simple-turning framework~\cite{WangWong}. For the comparison equation the Langer coordinate is exactly the coordinate \eqref{eq:general-zeta} at the lattice points. The Langer variable $\varrho$ of Geronimo--Bruno--Van Assche integrates, between the turning point $t_0$ and the lattice point $t$, the function $\operatorname{arcosh}$ on the hyperbolic side $t\ge t_0$ and $\arccos$ on the oscillatory side $t\le t_0$ of $(y-b(r,\varepsilon))/(2a(r+\varepsilon/2,\varepsilon))$ over $r$~\cite[Eqs.~(2.7)--(2.8), (2.13)]{GeronimoBrunoVanAssche}. (Geronimo~\cite[Eqs.~(2.1)--(2.2)]{Geronimo2009} places the hyperbolic side at $t\le t_p$; the reflection $n\mapsto-n$ of the lattice maps one convention to the other and leaves the quotient, with its midpoint $a(t+\varepsilon/2)$, unchanged.) With the comparison coefficients above this quotient is $u(r+\varepsilon/2)$, so the oscillatory integrand is $p$ (after the alternating gauge, $\pi-p$ on the $-1$ branch). The substitution $q'=r+\varepsilon/2$ maps the lattice point $t=k\varepsilon$ to $q_k$ and $t_0$ to $q_*$, and turns these integrals into the defining actions in \eqref{eq:general-zeta}, including their real hyperbolic continuations. Hence $\varepsilon^{-2/3}\varrho(k\varepsilon)=N^{2/3}\zeta_\tau(q_k)$ exactly, with no $O(\varepsilon)$ shift of the turning point. For half-integer ranks one indexes the lattice by the integer $n=k-k_{\min}$ and shifts both comparison coefficients by $k_{\min}\varepsilon$, which changes none of these identities. Since $(\operatorname{Ai}^2+\operatorname{Bi}^2)^{1/2}$ is bounded above and below on $|t|\le A_0$, the envelope estimate gives the bounded-window displays \eqref{eq:general-canonical-Ai}--\eqref{eq:general-canonical-Bi} with additive $O_{A_0}(N^{-1})$ bracket error. The transport factor is equally explicit. Geronimo--Bruno--Van Assche Eq.~(2.9) (Eq.~(2.3) of~\cite{Geronimo2009}) gives, on the oscillatory side, \begin{equation} g=\left|\frac{\zeta_\tau}{a(s+\varepsilon/2,\varepsilon)^2\sin^2p}\right|^{1/4}. \end{equation} Since $a(s+\varepsilon/2,\varepsilon)=e_\pi(q^2)/2$ and $D=e_\pi^2\sin^2p$ by \eqref{eq:class-principal-data}, this simplifies exactly to \begin{equation} g=\sqrt2\left|\frac{\zeta_\tau}{D}\right|^{1/4}. \end{equation} Multiplying this canonical pair by $\sqrt2\,N^{1/6}$ makes its oscillatory Airy expansion have natural-gauge WKB amplitude $2\pi^{-1/2}D^{-1/4}$, exactly the normalization used in the fixed-ratio basis. Finally, \begin{equation} \frac{2N^{1/6}}{\sqrt{2L+1}} =\sqrt{\frac2q}\,N^{-1/3},\qquad 2L+1=2Nq, \end{equation} which gives \eqref{eq:general-canonical-physical-scale}. The same Airy asymptotics identify the $\operatorname{Ai}$ member with the cosine WKB branch and the $\operatorname{Bi}$ member with its quadrature branch in the overlap $N^{-2/3}\ll|q-q_*|\ll1$. \end{proof} \subsection*{Proof of Theorem~\ref{thm:general-rank-airy}(i)} \begin{proof} Theorem~\ref{thm:general-canonical-airy} constructs a canonically normalized local $\operatorname{Ai}/\operatorname{Bi}$ pair in the coordinate $\zeta_\tau$, including the alternating gauge on the $-1$ branch and the physical transport normalization. Thus the only remaining task is to identify the physical linear combination and its lattice sign. Choose a fixed oscillatory matching point $q_00$ on $[0,q_*)$, the swept path of $q_0$ is a turning-free corridor. This point lies outside the bounded-$A_0$ display as $N\to\infty$, so the matching uses the fixed-neighbourhood Airy-envelope estimate of Geronimo--Bruno--Van Assche Theorem~3.8, verified in Theorem~\ref{thm:general-canonical-airy}, rather than only the bounded-scaled-argument display. At that point, the fixed-ratio case of Corollary~\ref{thm:general-first-order}, $W_q=(\sin\Theta_q)^{1/2}D^{-1/4}$, $\kappa\Theta_q=NS$ and \eqref{eq:wigner-darboux-sign} give \begin{equation} Y_L=\epsilon_{ab} \sqrt{\frac{2}{\pi Nq}}D(q,x)^{-1/4} \cos\!\left(NS(q,x)+\frac{p_{ab}\pi}{2}-\frac\pi4\right) +O(N^{-3/2}), \label{eq:general-airy-bulk-match} \end{equation} For fixed negative $\zeta$, \begin{equation} \operatorname{Ai}(N^{2/3}\zeta)= \frac{N^{-1/6}|\zeta|^{-1/4}}{\sqrt\pi} \left[ \cos\!\left(N\mathcal A-\frac\pi4\right)+O(N^{-1}) \right], \qquad \mathcal A=\frac23(-\zeta)^{3/2}. \label{eq:general-AiBulk} \end{equation} Using the exact transport factor of the comparison parametrix, the canonical $\operatorname{Ai}$ member therefore has physical oscillatory amplitude $\sqrt{2/(\pi Nq)}D^{-1/4}$ at the matching point, and this matches \eqref{eq:general-airy-bulk-match} exactly. For $\tau=+1$, $\mathcal A=S(q_*,x)-S(q,x)$, and the quantization \eqref{eq:general-turn-quantization}, $NS(q_*,x)=\pi K_+-\pi(p_{ab}-1)/2$, gives \begin{equation} NS(q,x)+\frac{p_{ab}\pi}{2}-\frac\pi4 =\pi K_+ +\frac\pi4-N\mathcal A, \label{eq:general-plus-phase-identity} \end{equation} so the bulk cosine is $(-1)^{K_+}\cos(N\mathcal A-\pi/4)$. For $\tau=-1$, \begin{equation} \mathcal A=\int_q^{q_*}[\pi-p(s,x)]\,\dd s =\pi(q_*-q)-S(q_*,x)+S(q,x), \end{equation} and $Nq=\kappa$ together with \eqref{eq:general-turn-quantization} yields the exact physical-lattice identity \begin{equation} NS(q,x)=\pi\Bigl[K_--\frac{p_{ab}}2\Bigr]+N\mathcal A. \label{eq:general-minus-action-identity} \end{equation} Hence \begin{equation} NS(q,x)+\frac{p_{ab}\pi}{2}-\frac\pi4 =\pi K_-+N\mathcal A-\frac\pi4. \label{eq:general-minus-phase-identity} \end{equation} To identify the two coefficients of the \emph{exact} Airy basis one must match a two-component Jacobi vector, not a single value. Choose consecutive ranks $L_0,L_0-1$ at the fixed oscillatory matching point $q_00$ and satisfies the eikonal $\cosh(\zeta^{1/2}\zeta')=z_0$ (with $\zeta^{1/2}\in i\mathbb R_+$ on the oscillatory side). Let $C(s;\zeta,h)$, $S(s;\zeta,h)$ solve $Y''=(\zeta+hs)Y$ with $C=1$, $C'=0$, $S=0$, $S'=1$ at $s=0$, and put $s_\pm(q,h)=(\zeta(q\pm h)-\zeta(q))/h$. \begin{lemma}[Airy shift, parity and first-order operators] \label{lem:airy2-shift-parity} (i) For every solution $\chi$ of $\chi''=X\chi$, $X=h^{-2/3}\zeta(q)$, one has $\chi(X+h^{1/3}s)=\chi(X)C(s)+h^{1/3}\chi'(X)S(s)$ and $\chi'(X+h^{1/3}s)=h^{-1/3}\chi(X)C'(s)+\chi'(X)S'(s)$. Consequently, for $\psi=P\chi(h^{-2/3}\zeta)+h^{1/3}Q\chi'(h^{-2/3}\zeta)$, \begin{equation} L\psi=\chi\,\mathfrak A[P,Q]+h^{1/3}\chi'\,\mathfrak B[P,Q], \label{eq:airy2-AB} \end{equation} \begin{equation}\begin{aligned} \mathfrak A&=\sum_\pm a(q\pm\tfrac h2)\bigl[P_\pm C(s_\pm)+Q_\pm C'(s_\pm)\bigr]-2az_hP,\\ \mathfrak B&=\sum_\pm a(q\pm\tfrac h2)\bigl[P_\pm S(s_\pm)+Q_\pm S'(s_\pm)\bigr]-2az_hQ, \end{aligned}\end{equation} $P_\pm=P(q\pm h)$, independently of $\chi$. (ii) If $P$ is even and $Q$ odd in $h$, then $\mathfrak A$ is even and $\mathfrak B$ is odd in $h$; if $P$ is odd and $Q$ even, then $\mathfrak A$ is odd and $\mathfrak B$ is even. (iii) For $h$-independent $p,q$: $\mathfrak A[p,0]=O(h^2)$, $\mathfrak B[0,q]=O(h^2)$, and \begin{equation} \mathfrak B[p,0]=h\,\frac{(Vp^2)'}{p}+O(h^3),\qquad \mathfrak A[0,q]=h\,\frac{(Wq^2)'}{q}+O(h^3), \end{equation} with $V=a_0S_0(\zeta')$, $W=\zeta V$, $S_0(s)=\sinh(\zeta^{1/2}s)/\zeta^{1/2}$. $V$ has no zero on $I$, $W$ has exactly one zero, simple, at $q_*$, and $g:=|V|^{-1/2}=\sqrt2\,|\zeta/D|^{1/4}$ satisfies $(Vg^2)'=0$. \end{lemma} \begin{proof} (i) $Y(s)=\chi(X+h^{1/3}s)$ solves $Y''=(\zeta+hs)Y$ with $Y(0)=\chi(X)$, $Y'(0)=h^{1/3}\chi'(X)$; and $h^{-2/3}\zeta(q\pm h)=X+h^{1/3}s_\pm$. (ii) Under $h\mapsto-h$ the two links exchange, $s_\pm\mapsto-s_\mp$, and uniqueness for the ODE gives $C(s;\zeta,-h)=C(-s;\zeta,h)$, $S(s;\zeta,-h)=-S(-s;\zeta,h)$; substitution gives $\mathfrak A[P,Q](-h)=\mathfrak A[\check P,-\check Q](h)$ and $\mathfrak B[P,Q](-h)=-\mathfrak B[\check P,-\check Q](h)$, $\check P(h)=P(-h)$. (iii) At $h=0$, $C_0=\cosh(\zeta^{1/2}s)$ is even and $S_0$ odd, $s_\pm=\pm\zeta'$, and the eikonal gives $[h^0]\mathfrak A[p,0]=2a_0p(\cosh(\zeta^{1/2}\zeta')-z_0)=0$ and likewise $[h^0]\mathfrak B[0,q]=0$; the $h^1$ coefficients of these two vanish by (ii). For the $h^1$ coefficients of the other two use $s_\pm=\pm\zeta'+\frac h2\zeta''+O(h^2)$ and $S_1=\frac s2\partial_\zeta S_0$, $C_1'=\frac s2\partial_\zeta C_0'$ (both from the variational equation of $Y''=(\zeta+hs)Y$); then $[h^1]\mathfrak B[p,0]=(a_0'p+2a_0p')S_0+a_0p\frac{\dd}{\dd q}S_0(\zeta'(q);\zeta(q))=(Vp^2)'/p$, and the same computation with $C_0'=\zeta S_0$ gives $(Wq^2)'/q$. Finally $|V|=a_0\sqrt{|z_0^2-1|}/|\zeta|^{1/2}=\sqrt D/(2|\zeta|^{1/2})$ by \eqref{eq:general-airy-simple-turn-identity}, which is bounded and nonzero through the turn, and $W=\zeta V$. \end{proof} \begin{theorem}[Formal Airy parametrix through relative order $h^2$] \label{thm:airy2-formal} There are functions $q_1$, $p_2$, real-analytic on $I$, with analyticity domains and bounds uniform on the compact family, such that \begin{equation} \psi^{[2]}:=(g+h^2p_2)\operatorname{Ai}(h^{-2/3}\zeta)+h^{4/3}q_1\operatorname{Ai}'(h^{-2/3}\zeta) \end{equation} has $\mathfrak A[g+h^2p_2,hq_1]=O(h^4)$ and $\mathfrak B[g+h^2p_2,hq_1]=O(h^5)$ uniformly on $I$; hence $|(L\psi^{[2]})(q_k)|\le Ch^4\,\mathsf M(X_k)/\mathsf E(X_k)$ at the lattice points, with Olver's modulus $\mathsf M$ and weight $\mathsf E$ for the Airy functions. Here $q_1$ is the unique solution analytic at $q_*$ of $(Wq_1^2)'/q_1=-[h^2]\mathfrak A[g,0]$. \end{theorem} \begin{proof} By Lemma~\ref{lem:airy2-shift-parity}, $\mathfrak A[g,0]$ is even with vanishing $h^0$ term and $\mathfrak B[g,0]$ is odd with vanishing $h^1$ term (since $(Vg^2)'=0$). Adding $Q=hq_1$ changes $[h^2]\mathfrak A$ by $(Wq_1^2)'/q_1$ and changes $\mathfrak B$ only at order $h^3$. The equation $2Wq_1'+W'q_1=f$ with $W=(q-q_*)\omega$, $\omega$ analytic and nonzero, has the analytic solution $q_1(q)=\omega(q)^{-1/2}\int_0^1f(q_*+(q-q_*)v^2)\,\omega(q_*+(q-q_*)v^2)^{-1/2}\dd v$, while the homogeneous solution $|W|^{-1/2}$ is singular; so $q_1$ is unique. Now $\mathfrak A$ is $O(h^4)$ by parity, and $\mathfrak B=h^3B_3+O(h^5)$. Adding $h^2p_2$ changes $[h^3]\mathfrak B$ by $(Vp_2^2)'/p_2$ and $\mathfrak A$ only at order $h^4$; the regular equation $(\sqrt{|V|}p_2)'=-\operatorname{sgn}(V)B_3/(2\sqrt{|V|})$ fixes $p_2$ up to a multiple of $g$. Taylor remainders of functions analytic on a compact set are uniform. No differentiability or analyticity in the parameter $\pi$ is asserted, since (E1) assumes only continuity in $\pi$. For the envelope bound use $|\operatorname{Ai}|\le\mathsf M/\mathsf E$, $|\operatorname{Ai}'|\le\mathsf N/\mathsf E$ and $h^{1/3}\mathsf N(X)/\mathsf M(X)\le C(h^{2/3}+|\zeta|)^{1/2}$. \end{proof} \begin{lemma}[Recessive solution] \label{lem:airy2-recessive} Put $\psi_1=g\operatorname{Ai}(h^{-2/3}\zeta)$, $\psi_2=g\operatorname{Bi}(h^{-2/3}\zeta)$ and $X_k=h^{-2/3}\zeta(q_k)$ at the lattice points $q_k$ of $I$, let $f_1$ be the exact solution of $Lf=0$ on the lattice of $I$ with $f_1=\psi_1$ at the last two lattice points, and let $J$ be the set of lattice points with $X_k\ge-1$. Then \begin{enumerate}[label=(\roman*)] \item $|f_{1,k}-\psi_{1,k}|\le Ch\,\psi_{1,k}$ for $k\in J$; \item $|f_1|\le C\mathsf M/\mathsf E$ on the lattice of $I$; the exact canonical $\operatorname{Bi}$ solution $f_2$ of Theorem~\ref{thm:general-canonical-airy}, normalized to $\psi_2$, satisfies $|f_2|\le C\mathsf M\mathsf E$ there; and the Casoratian $\mathcal W=a(q_k+\tfrac h2)(f_{1,k+1}f_{2,k}-f_{1,k}f_{2,k+1})$ satisfies $|\mathcal W|\ge ch^{1/3}$. \end{enumerate} \end{lemma} \begin{proof} (i) Since $\operatorname{Ai}>0$ on $[-1,\infty)$, $\psi_1>0$ on $J$, and $J$ is a final segment of the lattice because $\zeta'>0$. Write $f_1=\psi_1(1+\sigma)$ on $J$ and $D_k=a(q_k+\tfrac h2)\psi_{1,k}\psi_{1,k+1}(\sigma_{k+1}-\sigma_k)$. Then $Lf_1=0$ reads $D_k-D_{k-1}=-\psi_{1,k}(L\psi_1)_k(1+\sigma_k)$, and $\sigma$ vanishes at the last two lattice points, so summing twice gives the final-value Volterra equation \begin{equation} \sigma_k=-\sum_{i>k}K_{ki}(1+\sigma_i),\qquad K_{ki}=(L\psi_1)_i\sum_{j=k}^{i-1}\frac{\psi_{1,i}}{a(q_j+\tfrac h2)\psi_{1,j}\psi_{1,j+1}}, \end{equation} both sums running over $J$ up to the second-last lattice point. By Lemma~\ref{lem:airy2-shift-parity} with $P=g$, $Q=0$, $\mathfrak A[g,0]=O(h^2)$ and $\mathfrak B[g,0]=O(h^3)$; with \eqref{eq:airy2-AB} and $|\operatorname{Ai}'(X)|\le C(2+X)^{1/2}\operatorname{Ai}(X)$ on $[-1,\infty)$, $h^{1/3}(2+X)^{1/2}\le C$ on $I$, this gives $|(L\psi_1)_i|\le Ch^2\psi_{1,i}$ on $J$. Next $\operatorname{Ai}(X)\asymp(2+X)^{-1/4}e^{-\xi(X)}$ on $[-1,\infty)$ with $\xi=\frac23X_+^{3/2}$, and $g\asymp1$, so $\psi_{1,i}\le Ce^{-(\xi_i-\xi_j)}\psi_{1,j}$ for $j\le i$ in $J$; and $X_{j+1}-X_j\asymp h^{1/3}$ because $\zeta'\asymp1$. Hence $|K_{ki}|\le Ch^2\sum_{j4$, the terms with $X_{j+1}\ge X_i/4$ satisfy $\xi_i-\xi_{j+1}\ge\frac12X_i^{1/2}(X_i-X_{j+1})$ and form a geometric sum $\le1+Ch^{-1/3}X_i^{-1/2}$, and the at most $Ch^{-1/3}(1+X_i)$ others are $\le e^{-X_i^{3/2}}$ each; recall $X_i\le Ch^{-2/3}$. So $|K_{ki}|\le Ch^{5/3}(1+X_i^+)^{-1/2}$, and $\sum_{i\in J}(1+X_i^+)^{-1/2}\le Ch^{-1/3}\int_{-1}^{Ch^{-2/3}}(1+X_+)^{-1/2}\dd X\le Ch^{-2/3}$. Thus $\|K\|_\infty:=\max_k\sum_i|K_{ki}|\le Ch$, and $\max_J|\sigma|\le\|K\|_\infty(1+\max_J|\sigma|)$ gives $\max_J|\sigma|\le2\|K\|_\infty$ for $h$ small. (ii) For $X\ge c_0$ ($c_0\approx-0.37$, where $\operatorname{Ai}=\operatorname{Bi}$) Olver's functions give $\mathsf M/\mathsf E=\sqrt2\operatorname{Ai}$ and $\mathsf M\mathsf E=\sqrt2\operatorname{Bi}\ge(\operatorname{Ai}^2+\operatorname{Bi}^2)^{1/2}$, while $\mathsf E=1$, $\mathsf M=(\operatorname{Ai}^2+\operatorname{Bi}^2)^{1/2}$ for $X\le c_0$; on $[-1,c_0]$ also $\operatorname{Ai}\asymp\mathsf M$. So (i) gives $|f_1|\le C\mathsf M/\mathsf E$ on $J$, and the envelope estimate of Theorem~\ref{thm:general-canonical-airy}, with weight $g(\operatorname{Ai}^2+\operatorname{Bi}^2)^{1/2}\le Cg\mathsf M\mathsf E$, gives $|f_2|\le C\mathsf M\mathsf E$ and, for the canonical $\operatorname{Ai}$ solution $F_1$, $|F_1-\psi_1|\le Ch\,\mathsf M\mathsf E$. Take consecutive lattice points $k_0,k_0+1$ with $X\in[-1,1]$. There $f_1-\psi_1$, $F_1-\psi_1$ and $f_2-\psi_2$ are $O(h)$ while $\psi_1,\psi_2=O(1)$, so the constant Casoratians satisfy $\mathcal W(f_1,f_2),\mathcal W(F_1,f_2)=\mathcal W_{k_0}(\psi_1,\psi_2)+O(h)$ and $\mathcal W(F_1,f_1)=O(h)$. By Lemma~\ref{lem:airy2-shift-parity}(i), $\mathcal W_{k_0}(\psi_1,\psi_2)=a(q_{k_0}+\tfrac h2)g_{k_0}g_{k_0+1}\bigl(\operatorname{Ai}(X+h^{1/3}s_+)\operatorname{Bi}(X)-\operatorname{Ai}(X)\operatorname{Bi}(X+h^{1/3}s_+)\bigr)=-a\,g_{k_0}g_{k_0+1}h^{1/3}S(s_+)/\pi$ with $X=X_{k_0}$, and $|S(s_+)|\ge c$ on $I$; so $|\mathcal W(f_1,f_2)|\ge ch^{1/3}$, and $f_1=\alpha F_1+\beta f_2$ with $\alpha=\mathcal W(f_1,f_2)/\mathcal W(F_1,f_2)=1+O(h^{2/3})$, $\beta=\mathcal W(F_1,f_1)/\mathcal W(F_1,f_2)=O(h^{2/3})$. At lattice points with $X_k\le0$, $\mathsf E\le\mathsf E(0)$ and $|F_1|+|f_2|\le C\mathsf M$, so $|f_1|\le C\mathsf M\le C\mathsf M/\mathsf E$ there. \end{proof} \begin{proposition}[Exact solution near the parametrix] \label{prop:airy2-exact} Let $f_1$ be the recessive solution and $f_2$ the canonical $\operatorname{Bi}$ solution of Lemma~\ref{lem:airy2-recessive}, so that $|f_1|\le C\mathsf M/\mathsf E$ and $|f_2|\le C\mathsf M\mathsf E$ on the lattice of $I$. If a lattice function $\psi$ on $I$ has $|(L\psi)_k|\le\eta\,\mathsf M(X_k)/\mathsf E(X_k)$, the recurrence has an exact solution $f$ with $f=\psi$ at the last two lattice points of $I$ and $|f_k-\psi_k|\le C\eta h^{-1}\mathsf M(X_k)/\mathsf E(X_k)$. In particular there is an exact solution $f_h$ with $|f_h-\psi^{[2]}|\le Ch^3\mathsf M/\mathsf E$ on $I$. \end{proposition} \begin{proof} The Casoratian $\mathcal W=a(q_k+\tfrac h2)(f_{1,k+1}f_{2,k}-f_{1,k}f_{2,k+1})$ is constant, and $|\mathcal W|\ge ch^{1/3}$ by Lemma~\ref{lem:airy2-recessive}(ii). Let $n$ be the last lattice point of $I$. The final-value Green function $e_k=\sum_{kk$, $\mathsf E(X_k)\le\mathsf E(X_i)$, so both terms $f_{2,k}f_{1,i}r_i$ and $f_{1,k}f_{2,i}r_i$ of the summand are bounded by $C\eta\,(\mathsf M_k/\mathsf E_k)\mathsf M_i^2$, and $\sum_i\mathsf M(X_i)^2\le Ch^{-2/3}$ because $\mathsf M(X)^2\le C(1+|X|)^{-1/2}$ and $X_{i+1}-X_i\asymp h^{1/3}$. Hence $|e_k|\le C\eta h^{-1}\mathsf M_k/\mathsf E_k$. With $\eta=Ch^4$ from Theorem~\ref{thm:airy2-formal} this gives $f_h$. \end{proof} \begin{lemma}[Olver form of the Airy coefficient] \label{lem:airy2-olver} Put $B_\tau:=q_1/g$ on $I$ and $\psi_\tau:=-g\,[h^2]\mathfrak A[g,0]/\zeta'$. (i) $\psi_\tau$ is real-analytic on $I$, and $B_\tau$, as a function of $\zeta=\zeta_\tau(q)$, is the solution analytic at $\zeta=0$ of Olver's transport equation \begin{equation} 2\zeta\frac{\dd B}{\dd\zeta}+B=\psi_\tau,\qquad B_\tau=\frac{1}{2\zeta^{1/2}}\int_0^{\zeta}\psi_\tau(v)\,v^{-1/2}\,\dd v=\int_0^1\psi_\tau(\zeta y^2)\,\dd y; \label{eq:airy2-olver-transport} \end{equation} in particular $B_\tau(q_*)=\psi_\tau(q_*)$. (ii) $\psi_\tau(q_*)$ equals \eqref{eq:general-airy-varpi-local} and \eqref{eq:general-airy-varpi-olver}. \end{lemma} \begin{proof} (i) $[h^2]\mathfrak A[g,0]$ is a Taylor coefficient of a function analytic in $(q,h)$, and $\zeta'>0$. By Lemma~\ref{lem:airy2-shift-parity}(iii) $V>0$ on $I$ (for $\zeta<0$, $\sqrt{-\zeta}\,\zeta'\in(0,\pi)$ is $p$ or $\pi-p$), so $Vg^2=1$, and $W=\zeta V$. For $q_1=gB$, using $(Vg^2)'=g(2Vg'+V'g)=0$, \begin{equation} \frac{(Wq_1^2)'}{q_1}=2\zeta VgB'+\zeta'VgB+\zeta B(2Vg'+V'g)=\frac{\zeta'}{g}\Bigl(2\zeta\frac{\dd B}{\dd\zeta}+B\Bigr), \end{equation} so the transport equation of Theorem~\ref{thm:airy2-formal} is \eqref{eq:airy2-olver-transport}; the homogeneous solution $\zeta^{-1/2}$ is singular, which leaves the displayed integral. (ii) The first identity is the evaluation at the end of the proof of Theorem~\ref{thm:general-rank-airy}(ii) below. The eikonal $\cosh(\zeta^{1/2}\zeta')=\tau u$ gives $F=\operatorname{arcosh}^2(\tau u)=\zeta\zeta'^2$, with $F_1=2u_1$, $F_2=2u_2-\frac13u_1^2$, $F_3=2u_3-\frac23u_1u_2+\frac4{45}u_1^3$; substituting these, $x-b_0(q_*)=\tau e_0$ and $(2e_0e_2-e_1^2)/e_0^2=(\log e_\pi(q^2))''(q_*)$ turns \eqref{eq:general-airy-varpi-local} into \eqref{eq:general-airy-varpi-olver}, an identity checked by exact computer algebra. \end{proof} \subsection*{Proof of Theorem~\ref{thm:general-rank-airy}(ii)} \begin{proof} Work in the gauge of the neighbourhood and write the physical solution as $\widehat Y=\kappa_hf_h+\mu_hf_2$, with $f_1,f_2$ as in Lemma~\ref{lem:airy2-recessive}; this is possible because $\mathcal W(f_h,f_2)=\mathcal W(f_1,f_2)+O(h)$ (evaluate at the points $k_0,k_0+1$ of that proof, where $f_h-f_1=O(h)$) and $|\mathcal W(f_1,f_2)|\ge ch^{1/3}$. \emph{Recessive selection.} $\mu_h=\mathcal W(f_h,\widehat Y)/\mathcal W(f_h,f_2)$. At the last lattice points of $I$, with $\xi_{\rm end}=\frac23N\zeta(q_*+\delta_0)^{3/2}$, $|f_h|\le C\mathsf M/\mathsf E\le Ce^{-\xi_{\rm end}}$ (Proposition~\ref{prop:airy2-exact}; at these two points $\xi_{\rm end}-\xi=O(1)$, since $\xi'=N\Upsilon$ and $\xi$ increases by at most $\max_I\Upsilon$ per rank) and by (R) $|\widehat Y|\le CN^m$, so $|\mu_h|\le CN^{m+1/3}e^{-\xi_{\rm end}}$. On $I'=\{\zeta\le4^{-2/3}\zeta(q_*+\delta_0)\}$, $\mathsf E(X)^2\le Ce^{\frac43X_+^{3/2}}\le Ce^{\xi_{\rm end}/2}$, hence $|\mu_hf_2|\le CN^{m+1/3}e^{-\xi_{\rm end}}\mathsf M\mathsf E\le CN^{m+1/3}e^{-\xi_{\rm end}/2}\,\mathsf M/\mathsf E=O(N^{-\infty})\,\mathsf M/\mathsf E$ (for the Hahn, centered Racah, and quantum $6j$ chains the larger range of Proposition~\ref{prop:forbidden-decay}(c) holds). \emph{The scalar $\kappa_h$.} Fix $q_0\in(q_*-\delta_0,q_*)$; its swept path is a turning-free corridor, so Theorem~\ref{thm:general-second-order}, Lemma~\ref{lem:darboux-parity} and $W^{[2]}(\sin\Theta^{[2]})^{-1/2}=D^{-1/4}$ give, at the lattice points $k$ near $q_0N$ (in the physical frame: for $\tau=-1$ the alternating gauge is restored, which adds $\pi(k-k_{\min})$ to the phases $\phi^f_k$ below), \begin{equation} \widehat Y_k=\frac{2}{\sqrt\pi}\epsilon_{ab}D(q_k)^{-1/4}\bigl[(1+O(h^2))\cos\phi^Y_k+O(h^2)\bigr], \qquad \phi^Y_k-\phi^Y_{k-1}=p(q_k-\tfrac h2)+O(h^2), \end{equation} with $\phi^Y_k=NS(q_k)+h\widetilde\Gamma(q_k)+p_{ab}\pi/2-\pi/4+O(h^3)$, $\widetilde\Gamma:=\Gamma_2+\mathfrak c_{ab}(\Theta_q)/q$, where the amplitude is even in $h$ by Lemma~\ref{lem:darboux-parity}. On the other side, $X=h^{-2/3}\zeta(q_0)\to-\infty$, and the large-argument expansions $\operatorname{Ai}(-y)=\pi^{-1/2}y^{-1/4}[\cos(\xi-\frac\pi4)(1+O(\xi^{-2}))+\frac{5}{72\xi}\sin(\xi-\frac\pi4)+O(\xi^{-3})]$, $\operatorname{Ai}'(-y)=\pi^{-1/2}y^{1/4}[\sin(\xi-\frac\pi4)+O(\xi^{-1})]$, $\xi=\frac23y^{3/2}$ (\cite[9.7.9--9.7.10]{DLMFAiryBessel}), together with Proposition~\ref{prop:airy2-exact}, give \begin{equation}\begin{aligned} f_{h,k}&=\sqrt{\tfrac2\pi}\,h^{1/6}D(q_k)^{-1/4}\bigl[(1+O(h^2))\cos\phi^f_k+O(h^2)\bigr],\\ \phi^f_k&=N\mathcal A(q_k)-\frac\pi4-h\Bigl[\frac{5}{72\mathcal A}+\sqrt{-\zeta}\,\frac{q_1}{g}\Bigr](q_k)+O(h^2), \end{aligned}\end{equation} with $\mathcal A=\frac23(-\zeta)^{3/2}$; the $h^{4/3}q_1\operatorname{Ai}'$ term is of size $h$ and in quadrature, so it shifts the phase and changes the amplitude only at order $h^2$. With the alternating gauge restored, the increments of $\phi^f$ are $-\tau p(q_k-\frac h2)+O(h^2)$, since $N\mathcal A$ decreases by $p$ per rank for $\tau=+1$ and by $\pi-p$ for $\tau=-1$. For a phase sequence with increments $p_k+O(h^2)$ the vector $v_k=(\cos\phi_k,\cos\phi_{k-1})$ satisfies $\mathcal Q_k(v_k)=1+O(h^2)$, where $\mathcal Q_k(v)=(v_1^2+v_2^2-2v_1v_2\cos p_k)/\sin^2p_k$. Applying this to the exact identity $\widehat Y=\kappa_hf_h+\mu_hf_2$ at two consecutive ranks, after dividing each row $j$ by the amplitude factor $D(q_j,x)^{-1/4}$ common to $\widehat Y_j$ and $f_{h,j}$ (these differ by $1+O(h)$ between the rows, so they must be removed row by row), gives $|\kappa_h|\,h^{1/6}/\sqrt2=1+O(h^2)$, and then, since the map $\phi\mapsto(\cos\phi,\cos(\phi-p))$ is an embedding of the circle with derivative bounded below, $\phi^Y_k\equiv-\tau\phi^f_k\pmod\pi$ up to $O(h^2)$ (the pair $(\cos\phi^f_k,\cos\phi^f_{k-1})$ is $(\cos\psi,\cos(\psi-p_k))$ with $\psi=-\tau\phi^f_k$). The sign of $\kappa_h$ is that of Theorem~\ref{thm:general-rank-airy}. \emph{The bracket.} By the scalar step, $\kappa_h=\pm\sqrt2N^{1/6}(1+c_h)$ with $c_h$ real and $|c_h|\le Ch^2$, and by Proposition~\ref{prop:airy2-exact}, $f_h=\psi^{[2]}+O(h^3)\mathsf M/\mathsf E$ on $I$. Since $g\ge c$ on $I$, dividing $\widehat Y=\kappa_hf_h+\mu_hf_2$ by $\pm\sqrt2N^{1/6}g$ gives on $I'$ \begin{equation} (1+c_h)\Bigl[\Bigl(1+h^2\frac{p_2}{g}\Bigr)\operatorname{Ai}(X)+h^{4/3}\frac{q_1}{g}\operatorname{Ai}'(X)\Bigr]+O(h^3)\frac{\mathsf M}{\mathsf E}(X). \end{equation} Here $c_hh^{4/3}(q_1/g)\operatorname{Ai}'=O(h^{10/3})\mathsf N/\mathsf E=O(h^3)\mathsf M/\mathsf E$, because $|\operatorname{Ai}'|\le\mathsf N/\mathsf E$, $\mathsf N\le C(1+|X|)^{1/2}\mathsf M$ and $|X|\le Ch^{-2/3}$ on $I$; likewise $c_hh^2(p_2/g)\operatorname{Ai}=O(h^4)\mathsf M/\mathsf E$. With $\mathfrak a_N=h^2p_2/g+c_h$, $B_\tau=q_1/g$, $\sqrt2g=2|\zeta/D|^{1/4}$ and $\sqrt{2L+1}=\sqrt{2Nq}$ this is \eqref{eq:general-rank-airy-uniform}. For the shifted form put $s=h^{4/3}B_\tau$: $\operatorname{Ai}(X+s)=\operatorname{Ai}+s\operatorname{Ai}'+\frac12s^2X\operatorname{Ai}+\frac16s^3\operatorname{Ai}'''(\tilde X)$, where $\frac12s^2X=\frac12h^2\zeta B_\tau^2=O(h^2)$ joins the amplitude, and $\operatorname{Ai}'''=\operatorname{Ai}+X\operatorname{Ai}'$ bounds the last term by $Ch^4(1+|X|)^{3/2}\mathsf M/\mathsf E\le Ch^3\mathsf M/\mathsf E$ (on the forbidden side $(\mathsf M/\mathsf E)(\tilde X)\le e^{C|s|X^{1/2}}(\mathsf M/\mathsf E)(X)$ and $|s|X^{1/2}\le Ch$). On $|t|\le A_0$, $\mathsf M/\mathsf E\le C$ and $q-q_*=O(h^{2/3})$, so $B_\tau(q)=\varpi_\tau+O(h^{2/3})$ with $\varpi_\tau:=B_\tau(q_*)$, which gives \eqref{eq:general-rank-airy-second}. \emph{The coefficient.} On the oscillatory side \eqref{eq:general-plus-phase-identity} and \eqref{eq:general-minus-phase-identity} give $NS+p_{ab}\pi/2-\pi/4\equiv-\tau(N\mathcal A-\pi/4)\pmod\pi$ for both turn types, so $\phi^Y\equiv-\tau(N\mathcal A-\pi/4-\tau h\widetilde\Gamma)\pmod\pi$. Comparing with $-\tau\phi^f$ at every $q_0\in(q_*-\delta_0,q_*)$, the coefficients of $h$ must agree: $\tau\widetilde\Gamma=5/(72\mathcal A)+\sqrt{-\zeta}\,q_1/g$. This is \eqref{eq:general-airy-B}, since $5/(72\mathcal A\sqrt{-\zeta})=5/(48\zeta^2)$; as $q_1/g$ is analytic on $I$, $B_\tau$ is the analytic continuation of its right side, and \eqref{eq:general-airy-varpi} follows. For \eqref{eq:general-airy-varpi-local}, evaluate the regular equation $2Wq_1'+W'q_1=-[h^2]\mathfrak A[g,0]$ of Theorem~\ref{thm:airy2-formal} at $q_*$, where $W(q_*)=0$ and $W'(q_*)=\zeta'(q_*)V(q_*)=\frac12e_0\zeta'(q_*)^2\neq0$: so $q_1(q_*)=-[h^2]\mathfrak A[g,0](q_*)/W'(q_*)$ and $\varpi_\tau=q_1(q_*)/g(q_*)$, with $g(q_*)=(2/(e_0\zeta'(q_*)))^{1/2}$. At $q_*$ one has $\zeta=0$, so $C(s;0,h)=1+\frac16hs^3+\frac1{180}h^2s^6+O(h^3)$ and $s_\pm=\zeta(q_*\pm h)/h$; the eikonal $\cosh(\zeta^{1/2}\zeta')=\tau u$ determines the jets of $\zeta$ at $q_*$ from those of $\tau u$ (in particular $\zeta'(q_*)=(2u_1)^{1/3}$), and $V=\frac12e_\pi S_0(\zeta')$ gives those of $g$. Expanding $\mathfrak A[g,0](q_*)=\sum_\pm a(q_*\pm\frac h2;h^2)g(q_*\pm h)C(s_\pm)-\tau\bigl(x-b(q_*;h^2)\bigr)g(q_*)$ with $a(s;h^2)=\frac12e_\pi(s^2)\bigl(1+h^2\bigl((\frac18-\frac{a^2+b^2}2)/s^2+\eta_\pi(s^2)\bigr)\bigr)+O(h^4)$ and $\tau(x-b(q_*;h^2))=e_0-\tau h^2(ab/q_*^2+v_1)+O(h^4)$, the coefficients of $h^0$ and $h^1$ vanish and the coefficient of $h^2$ yields \eqref{eq:general-airy-varpi-local}, an identity checked by exact computer algebra. The Olver form is Lemma~\ref{lem:airy2-olver}(ii). \end{proof} \subsection{Flux of the Airy-form branch and the amplitude} \label{sec:supp-airy2-amplitude} Work in the setting of Lemma~\ref{lem:airy2-shift-parity} and Theorem~\ref{thm:airy2-formal}. Put $\chi=\operatorname{Ai}+i\operatorname{Bi}$, $X=h^{-2/3}\zeta(q)$, $P=g+h^2p_2$, $Q=hq_1=hgB_\tau$ (real, analytic), $\psi^{\rm c}:=P\chi(X)+h^{1/3}Q\chi'(X)$, so $\Re\psi^{\rm c}=\psi^{[2]}$. With $\hat s_\epsilon=(\zeta(q+\epsilon\frac h2)-\zeta(q))/h$, $\hat P_\epsilon=P(q+\epsilon\frac h2)$, $\hat Q_\epsilon=Q(q+\epsilon\frac h2)$ ($\epsilon=\pm$) and $C,S$ as in Lemma~\ref{lem:airy2-shift-parity} (parameter $\zeta(q)$), put $\mathcal A_\epsilon=\hat P_\epsilon C(\hat s_\epsilon)+\hat Q_\epsilon C'(\hat s_\epsilon)$, $\mathcal B_\epsilon=\hat P_\epsilon S(\hat s_\epsilon)+\hat Q_\epsilon S'(\hat s_\epsilon)$ and $\mathfrak F(q;h):=a(q;h^2)(\mathcal A_-\mathcal B_+-\mathcal B_-\mathcal A_+)$. \begin{lemma}[Flux of the Airy-form branch] \label{lem:airy2-flux} Let $J^{\rm c}(s):=\pi a(s;h^2)\Im\bigl(\overline{\psi^{\rm c}(s-\frac h2)}\psi^{\rm c}(s+\frac h2)\bigr)$. \begin{enumerate}[label=(\roman*),leftmargin=2.2em,itemsep=1pt] \item $J^{\rm c}=h^{1/3}\mathfrak F$ and $\pi\Im(\overline{\psi^{\rm c}}L\psi^{\rm c})(q)=J^{\rm c}(q+\frac h2)-J^{\rm c}(q-\frac h2)$. \item $\mathfrak F$ is even in $h$, $\mathfrak F=1+h^2\bigl(2p_2/g+\mathfrak j(q)\bigr)+O(h^4)$ uniformly on $I$, with $\mathfrak j:=[h^2]\mathfrak F|_{p_2=0}$ real-analytic ($p_2$-independent). \item For $p_2$ of Theorem~\ref{thm:airy2-formal}, $2p_2/g+\mathfrak j$ is constant on $I$. \item With $\zeta^{(k)}$ the derivatives of $\zeta_\tau$ at $q_*$, $\varpi_\tau=B_\tau(q_*)$, $B_{\tau,\zeta}(0)=\psi_\tau'(\zeta{=}0)/3$ (by \eqref{eq:airy2-olver-transport}) and $a_0=\frac12e_\pi(q^2)$, \begin{equation} \mathfrak j(q_*)=B_{\tau,\zeta}(0)+\frac{\zeta'^2}{6}\varpi_\tau+\frac{13\,\zeta'^6}{10080}-\frac{\zeta'^2\zeta''}{12}-\frac{\zeta'''}{12\zeta'}+\frac{\zeta''^2}{8\zeta'^2} +\frac{\frac18-\frac{a^2+b^2}{2}}{q_*^2}+\eta_\pi(q_*^2)-\frac18\bigl(\log e_\pi(q^2)\bigr)''\Big|_{q_*}. \label{eq:airy2-j-turn} \end{equation} \item For the Hahn chain with $x\neq0$, $\mathfrak j(q_*)$ is the rational function \eqref{eq:airy2-j-hahn}. \end{enumerate} \end{lemma} \begin{proof} (i) By Lemma~\ref{lem:airy2-shift-parity}(i) at the links $q\pm\frac h2$, $\psi^{\rm c}(q+\epsilon\frac h2)=\chi\mathcal A_\epsilon+h^{1/3}\chi'\mathcal B_\epsilon$ with real $\mathcal A_\epsilon,\mathcal B_\epsilon$, so $\Im(\overline{\psi^{\rm c}(q-\frac h2)}\psi^{\rm c}(q+\frac h2))=h^{1/3}(\mathcal A_-\mathcal B_+-\mathcal B_-\mathcal A_+)\Im(\bar\chi\chi')$, $\Im(\bar\chi\chi')=\operatorname{Ai}\operatorname{Bi}'-\operatorname{Bi}\operatorname{Ai}'=\frac1\pi$. As $L$ has real coefficients, $\Im(\bar\psi L\psi)(q)=a(q+\frac h2)\Im(\bar\psi(q)\psi(q+h))-a(q-\frac h2)\Im(\bar\psi(q-h)\psi(q))$. (ii) By Lemma~\ref{lem:airy2-shift-parity}(ii), $\mathcal A_\epsilon(-h)=\mathcal A_{-\epsilon}(h)$, $\mathcal B_\epsilon(-h)=-\mathcal B_{-\epsilon}(h)$ ($\hat Q_\epsilon$ is odd in $h$), so $\mathfrak F$ is even. At $h=0$, $\mathcal A_\pm=g\cosh(\zeta^{1/2}\zeta'/2)$, $\mathcal B_\pm=\pm gS_0(\zeta'/2)$, so $\mathfrak F|_{h=0}=a_0g^2S_0(\zeta')=Vg^2=1$. $p_2$ enters through $\hat P_\pm$ at order $h^2$, linearly in each of $\hat P_\pm$, giving $2(p_2/g)$ times the leading flux. (iii) $L\psi^{\rm c}=\chi\mathfrak A+h^{1/3}\chi'\mathfrak B$ with the real $\mathfrak A=O(h^4)$, $\mathfrak B=O(h^5)$ of Theorem~\ref{thm:airy2-formal}, so $\pi\Im(\overline{\psi^{\rm c}}L\psi^{\rm c})=h^{1/3}(P\mathfrak B-Q\mathfrak A)=O(h^{16/3})$. By (i)--(ii) it is $h^{10/3}\partial_q(2p_2/g+\mathfrak j)+O(h^{16/3})$, so $\partial_q(2p_2/g+\mathfrak j)=0$. (iv) At $q_*$, $\zeta=0$, $C(s;0,h)=1+\frac h6s^3+\frac{h^2}{180}s^6+O(h^3)$, $S(s;0,h)=s+\frac h{12}s^4+\frac{h^2}{504}s^7+O(h^3)$, and $g$ has the jets of $(a_0S_0(\zeta'))^{-1/2}$. Expanding $\mathfrak F(q_*;h)$ in the generic jets $(\zeta',\zeta'',\zeta''',a_0,a_0',a_0'',a_2,B_\tau,B_\tau')$ is a finite polynomial computation, done by exact computer algebra: $1+0\cdot h+h^2(\cdots)$ with $(\cdots)$ the right side of \eqref{eq:airy2-j-turn} ($B_\tau'(q_*)=\zeta'B_{\tau,\zeta}(0)$, $a_2/a_0=(\frac18-\frac{a^2+b^2}2)/q_*^2+\eta_\pi(q_*^2)$). (v) By \eqref{eq:airy2-olver-transport}, $B_\tau(q_*)=\psi_\tau(q_*)$, $B_\tau'(q_*)=\frac13\psi_\tau'(q_*)$, $\psi_\tau=-g[h^2]\mathfrak A[g,0]/\zeta'$. The truncated-series computation of (iv), one order further in $q-q_*$, gives $[h^2]\mathfrak A[g,0](q_*+\delta)$ to first order in $\delta$ in the jets of $\zeta$ (to order 4), $a_0$ (order 3), $a_2$, $\tau b_2$ (order 1); at $\delta=0$ it reproduces \eqref{eq:general-airy-varpi-local}. For Hahn, $e_\pi=\sqrt{1-q^2}$, $\tau u=|x|(1-q^2)^{-1/2}$, $a_2=a_0(\frac18-\frac{M^2+\sigma^2}2)/q^2$, $b_2=M\sigma/q^2$, and with $x=2w/(1+w^2)$, $q_*=(1-w^2)/(1+w^2)$ all jets (those of $\zeta$ from $\cosh(\zeta^{1/2}\zeta')=\tau u$, $\zeta'^3=2u_1$) are rational in $w$; \eqref{eq:airy2-j-turn} reduces modulo $\zeta'^3=2u_1$ to a cube-root-free rational function of $w$ equal to \eqref{eq:airy2-j-hahn}, ($\tau=-1$: via $\tau M\sigma$, $|x|$). \end{proof} \begin{proof}[Proof of Proposition~\ref{prop:airy2-amplitude}] Fix an oscillatory $q_0\in(q_*-\delta_0,q_*)$. As in the proof of Theorem~\ref{thm:general-rank-airy}(ii), $\widehat Y=\kappa_hf_h+O(N^{-\infty})\mathsf M/\mathsf E$, $\kappa_h=\pm\sqrt2N^{1/6}(1+c_h)$, $f_h=\psi^{[2]}+O(h^3)\mathsf M/\mathsf E$; so near $q_0$, $\widehat Y=2\Re u_A+O(h^3)$ with $u_A=\frac12\kappa_h\psi^{\rm c}$ ($\tau=+1$), $\frac12\kappa_h(-1)^{k-k_{\min}}\overline{\psi^{\rm c}}$ ($\tau=-1$, alternating gauge restored), and also $\widehat Y=2\Re u_D+O(h^3)$, $u_D=(1+h^2\alpha)\widehat U^{[2],+}$ (Theorem~\ref{thm:general-phase-exactness}). Both are $+$ carriers: $\chi(-y)\sim\pi^{-1/2}y^{-1/4}e^{-i(\xi-\pi/4)}$ has phase increasing with $q$ ($\xi=N\mathcal A(q)$ decreases), and for $\tau=-1$ the factor $(-1)^k\overline{(\cdot)}$ makes it increasing in the physical frame. Put $\eta_k:=u_A(q_k)/u_D(q_k)-1$ at the lattice points of a compact interval $J_0\subset(q_*-\delta_0,q_*)$ around $q_0$. Then $\Re(u_D\eta)=O(h^3)$ at every lattice point of $J_0$, but $\eta_k$ is a priori only a sequence: two consecutive real equations $\Re(u_{D,k}\eta_k)=O(h^3)$, $\Re(u_{D,k+1}\eta_{k+1})=O(h^3)$ involve two different complex numbers, and smoothness of $\eta$ does not bound $\eta_{k+1}-\eta_k$ (the sequence could alternate in sign). What is needed, and proved next, is the slow variation \begin{equation} \eta_{k+1}-\eta_k=O(h^3)\qquad(q_k,q_{k+1}\in J_0). \label{eq:airy2-eta-slow} \end{equation} \emph{Proof of \eqref{eq:airy2-eta-slow}: carrier-amplitude expansions of both representations.} All functions of $q$ below are smooth on $J_0$ and independent of $h$. (D) By Lemma~\ref{lem:darboux-parity} and the proof of Lemma~\ref{lem:pullback-flux}(ii), with $\kappa\Theta^{[2]}_q=NS+h\Gamma_2$, $\mathfrak P_{ab}=h\,\mathfrak c_{ab}(\Theta_q)/q+O(h^3)$ and $\mathfrak A_{ab}=1+h^2\mathfrak A_2(\Theta_q)/q^2+O(h^4)$, \begin{equation} u_D(q_k)=e^{iNS(q_k)}\beta(q_k)\bigl[1+ih\widetilde\Gamma(q_k)+h^2d_2(q_k)\bigr]+O(h^3),\qquad\beta:=\omega_{ab}\pi^{-1/2}D^{-1/4}, \end{equation} with $\widetilde\Gamma=\Gamma_2+\mathfrak c_{ab}(\Theta_q)/q$ and $d_2=\alpha+\mathfrak A_2/q^2-\frac12\widetilde\Gamma^2$ (expand the exponential; $1+h^2\alpha$ is real). There is no order-$h$ amplitude term because $\mathfrak A_{ab}$ is even. (A) At $q\in J_0$ the Airy argument is $X=-y$, $y=h^{-2/3}(-\zeta(q))\gg1$, and $\xi=\frac23y^{3/2}=N\mathcal A(q)$, $\xi^{-1}=h/\mathcal A$. The large-argument expansions of $\operatorname{Ai}$ and $\operatorname{Bi}$ (DLMF 9.7.9 and 9.7.11) combine to $\chi(-y)=\operatorname{Ai}(-y)+i\operatorname{Bi}(-y)=\pi^{-1/2}y^{-1/4}e^{-i(\xi-\pi/4)}\bigl[1+iu_1\xi^{-1}-u_2\xi^{-2}+O(\xi^{-3})\bigr]$, $u_1=\frac5{72}$, and $\chi'(-y)=i\pi^{-1/2}y^{1/4}e^{-i(\xi-\pi/4)}[1+O(\xi^{-1})]$, so in $\psi^{\rm c}=(g+h^2p_2)\chi(X)+h^{1/3}hq_1\chi'(X)$ the $\chi'$ term has relative size $h^{4/3}y^{1/2}=h\sqrt{-\zeta}$ and is imaginary at leading order. Using $\tfrac12\kappa_h\,g\,y^{-1/4}=\pm(1+c_h)D^{-1/4}$ (from $\kappa_h=\pm\sqrt2N^{1/6}(1+c_h)$ and $g=\sqrt2|\zeta/D|^{1/4}$), \begin{equation} \tfrac12\kappa_h\psi^{\rm c}(q_k)=\pm(1+c_h)\,e^{i\pi/4}e^{-iN\mathcal A(q_k)}\pi^{-1/2}D^{-1/4}\bigl[1+ihA_1+h^2A_2\bigr](q_k)+O(h^3), \end{equation} with $A_1=u_1/\mathcal A+\sqrt{-\zeta}\,B_\tau$ real and $A_2$ smooth. There is no order-$h$ amplitude term here either. For $\tau=+1$ this is $u_A$. For $\tau=-1$, $u_A=(-1)^{k-k_{\min}}\overline{\tfrac12\kappa_h\psi^{\rm c}}$ has the bracket $1-ihA_1+h^2\overline{A_2}$. By the exact phase identities \eqref{eq:general-plus-phase-identity}, \eqref{eq:general-minus-phase-identity}, and the rank-independence of $\sigma_\tau=(-1)^{K_\tau(L)}/\chi_\tau(L)$, the product of the carrier of $u_A$ and the lattice factor is a constant times $e^{iNS(q_k)}$: $e^{-iN\mathcal A(q_k)}=e^{-iNS(q_*)}e^{iNS(q_k)}$ for $\tau=+1$ ($\mathcal A=S(q_*)-S$), and $(-1)^{k-k_{\min}}e^{iN\mathcal A(q_k)}=\text{const}\cdot e^{iNS(q_k)}$ for $\tau=-1$. Hence \begin{equation} u_A(q_k)=\Lambda_N\,e^{iNS(q_k)}\pi^{-1/2}D^{-1/4}(q_k)\bigl[1+ih\gamma_1+h^2\gamma_2\bigr](q_k)+O(h^3), \end{equation} with a constant $\Lambda_N$ independent of $k$ ($|\Lambda_N|\le C$), $\gamma_1=\tau A_1$ and $\gamma_2$ smooth. Finally $\tau\widetilde\Gamma=A_1$ is exactly the order-$h$ phase matching \eqref{eq:general-airy-B} (note $5/(48\zeta^{2})\sqrt{-\zeta}=u_1/\mathcal A$), so $\gamma_1=\widetilde\Gamma$. Dividing (A) by (D), the carriers $e^{iNS}$ and the amplitudes $D^{-1/4}$ cancel, and the order-$h$ terms cancel because $\gamma_1=\widetilde\Gamma$: \begin{equation} 1+\eta_k=\Lambda'_N\bigl[1+h^2\rho(q_k)\bigr]+O(h^3),\qquad\rho:=\gamma_2-d_2, \end{equation} with $\Lambda'_N=\Lambda_N/\omega_{ab}$ independent of $k$ and bounded. Since $\rho$ is Lipschitz on $J_0$ and $q_{k+1}-q_k=h$, $\eta_{k+1}-\eta_k=\Lambda'_Nh^2\bigl(\rho(q_{k+1})-\rho(q_k)\bigr)+O(h^3)=O(h^3)$, which is \eqref{eq:airy2-eta-slow}. $\square$ \emph{Conclusion.} By \eqref{eq:airy2-eta-slow}, $\Re(u_{D,k}\eta_k)=O(h^3)$ and $\Re(u_{D,k+1}\eta_k)=\Re(u_{D,k+1}\eta_{k+1})+O(h^3)=O(h^3)$: the real vector $(\Re\eta_k,\Im\eta_k)$ is annihilated, up to $O(h^3)$, by the matrix with rows $(\Re u_{D,k},-\Im u_{D,k})$ and $(\Re u_{D,k+1},-\Im u_{D,k+1})$, whose determinant is $\pm\Im(\overline{u_{D,k}}u_{D,k+1})=\allowbreak\pm(1+h^2\alpha_k)(1+h^2\alpha_{k+1})\allowbreak\mathsf J(q_{k+1/2})/(\pi a(q_{k+1/2}))=\allowbreak\pm(1+O(h^2))/(2\pi a)$ by Lemma~\ref{lem:pullback-flux}(ii), bounded away from zero; here and below $q_{k+1/2}:=q_k+\frac h2$ is the link midpoint and $\alpha_j:=\alpha(q_j)$. As $|u_D|\le C$ on $J_0$, this gives $\eta_k=O(h^3)$, and the fluxes of $u_A$ and $u_D$ agree to relative $O(h^3)$. The flux of $u_D$ across the link $(q_k,q_{k+1})$ is $(1+h^2\alpha_k)(1+h^2\alpha_{k+1})\mathsf J(q_{k+1/2})$. Since $\alpha$ is smooth, $(1+h^2\alpha_k)(1+h^2\alpha_{k+1})=(1+h^2\alpha(q_{k+1/2}))^2+O(h^4)$, so with Lemma~\ref{lem:pullback-flux}(ii) the flux is $(1+h^2\alpha)^2(\frac12+h^2\mathsf J_2)\big|_{q_{k+1/2}}+O(h^4)=\frac12+h^2\delta_{\rm fl}+O(h^4)$, all evaluated at the midpoint, where $\alpha+\mathsf J_2(q)=c+\mathsf J_2(0,x)$ is constant; that of $u_A$ is $\frac14\kappa_h^2h^{1/3}\mathfrak F=\frac12(1+c_h)^2\mathfrak F$ (Lemma~\ref{lem:airy2-flux}(i); for $\tau=-1$, $\Im(\overline{(-1)^k\bar\psi_k}(-1)^{k+1}\bar\psi_{k+1})=\Im(\bar\psi_k\psi_{k+1})$). Equating, with Lemma~\ref{lem:airy2-flux}(ii)--(iii) and $2p_2/g+\mathfrak j\equiv0$, \begin{equation} c_h=h^2\delta_{\rm fl}+O(h^3),\qquad \mathfrak a_N=h^2p_2/g+c_h=h^2\bigl(\delta_{\rm fl}-\tfrac12\mathfrak j\bigr)+O(h^3), \end{equation} with the $O(h^3)$ absorbed in $\mathfrak r_N$; $\delta_{\rm fl}=0$ for Hahn and Racah since $c=-\mathsf J_2(0,x)$, and $\delta_{\rm fl}=-\frac1{12}$ for the quantum $6j$ chain (Corollary~\ref{cor:q6j-consequences}). \end{proof} \subsection{Hyperbolicity and the physical solution on the forbidden side} \label{sec:supp-forbidden-side} Proofs of Propositions~\ref{prop:exact-hyperbolicity} and~\ref{prop:forbidden-decay}. Write $\sigma(q)=q^2$ for Hahn and Racah and $\sigma(q)=\tanh^2q$ for one-small quantum $6j$, with $\sigma_k=\sigma(kh)$, $\bar\sigma_k=\frac12(\sigma_k+\sigma_{k+1})$, $q_k=(k+\frac12)h$, and offsets $(a,b)$. By Propositions~\ref{prop:hahn-class}, \ref{prop:racah-class} and~\ref{prop:q6j-class}, with $\mathrm s_z=\sinh(zh)$ and $\mathrm c_z=\cosh(zh)$, \begin{equation} A_k=\tfrac12r_k\widetilde e(\sigma_k),\qquad \widetilde e(\sigma)=\sqrt{(1-\sigma/\sigma_1)(1-\sigma/\sigma_2)}, \end{equation} where $(\beta_0,\sigma_1,\sigma_2^{-1})=(0,1,0)$ for Hahn, $\bigl(\frac1{4\alpha\beta},\allowbreak 4\alpha^2,\allowbreak (4\beta^2)^{-1}\bigr)$ for Racah, and $(\coth\mathsf X\coth\mathsf Y,\allowbreak\tanh^2\mathsf X,\allowbreak\coth^2\mathsf Y)$ for quantum $6j$, so $\beta_0=(\sigma_1\sigma_2)^{-1/2}$ throughout (the $q$-Hahn limit $\coth\mathsf Y=1$, $1/\sinh\mathsf Y=0$ of Remark~\ref{rem:qhahn-limit} is included: $\sigma_2=1$, and all constants below stay finite); here \begin{equation} r_k^2=\frac{(1-a^2/k^2)(1-b^2/k^2)}{1-1/(4k^2)}\quad\text{(Hahn, Racah)},\qquad r_k^2=\frac{(1-\mathrm s_a^2/\mathrm s_k^2)(1-\mathrm s_b^2/\mathrm s_k^2)}{1-\mathrm s_1^2/\mathrm s_{2k}^2}\quad\text{(quantum $6j$)}, \end{equation} using $\mathrm s_{2k}^2\mathrm s_{2k+1}\mathrm s_{2k-1}=16\mathrm s_k^4\mathrm c_k^4(1-\mathrm s_1^2/\mathrm s_{2k}^2)$ and $(1-\mathrm s_k^2/\sinh^2\mathsf X)/\mathrm c_k^2=1-\sigma_k/\sigma_1$, while \begin{equation} B_k=\beta_0\bar\sigma_k-\frac{\beta_0h^2}{2}(a^2+b^2)+\frac{ab}{k(k+1)}\ \text{(Hahn, Racah)},\quad B_k=\beta_0\Bigl(1-\frac{\mathrm c_a\mathrm c_b}{\mathrm c_k\mathrm c_{k+1}}\Bigr)+\frac{\mathrm s_a\mathrm s_b}{\mathrm s_k\mathrm s_{k+1}}\ \text{(quantum $6j$)}. \end{equation} Put $\sigma_{\max}=\min(\sigma_1,\sigma_2)$ ($1$ for Hahn) and $q_m=\sigma^{-1}(\sigma_{\max})$ ($\sqrt{\sigma_{\max}}$ for Hahn and Racah, $\min(\mathsf X,\mathsf Y)$ for quantum $6j$); the hard edge is $\sigma_{k_{\max}+1}=\sigma_{\max}$, $\widetilde e(\sigma_{\max})=0$. Put $\sigma_*=\sigma(q_*)$, $m_*=(1-\tau x)/\sigma_*>0$, and $\Phi(\sigma)=\tau(x-\beta_0\sigma)-\widetilde e(\sigma)$, which is $\tau\bigl(x-\Psi_\tau(\sigma)\bigr)$ for $\Psi_\tau=\beta_0\sigma+\tau\widetilde e$; a simple turn means $\Phi(\sigma_*)=0$. (F1) $\widetilde e$ is concave and decreasing, $0\le\widetilde e\le1$ on $[0,\sigma_{\max}]$: $\widetilde e''=-(\sigma_1^{-1}-\sigma_2^{-1})^2/(4\widetilde e^3)\le0$. (F2) $\Phi$ is convex on $[0,\sigma_{\max})$ for either sign of $\tau$, being an affine function minus the concave $\widetilde e$. Since $\widetilde e(0)=1$, $\Phi(0)=\tau x-1$, so the secant slope of $\Phi$ on $[0,\sigma_*]$ is $-\Phi(0)/\sigma_*=m_*$. The secant slopes of a convex function are nondecreasing, so beyond $\sigma_*$ both the difference quotient and the derivative are at least $m_*$: \begin{equation} \Phi'(\sigma)\ge m_*,\qquad\Phi(\sigma)\ge m_*(\sigma-\sigma_*)\qquad(\sigma_*\le\sigma<\sigma_{\max}). \label{eq:forbidden-slope} \end{equation} (F3) Set $N_0:=\max\bigl(2\max(|a|,|b|,1)/q_*,\max(|a|,|b|,1)\bigr)$. Since $\sinh u/u$ increases on $(0,\infty)$, for $k\ge q_*N/2$ and $N\ge N_0$ all three chains satisfy $1-\frac{a^2+b^2}{k^2}\le r_k^2\le1+\frac1{3k^2}$, so $|r_k-1|\le\rho_1h^2$ with $\rho_1:=4(a^2+b^2+\frac13)/q_*^2$. For Hahn and Racah $\bar\sigma_k=q_k^2+\frac{h^2}4$ gives $|B_k-\beta_0q_k^2|\le\rho_2h^2$ with $\rho_2:=\beta_0(1+2(a^2+b^2))/4+4|ab|/q_*^2$; for quantum $6j$, since $1-(\mathrm c_k\mathrm c_{k+1})^{-1}-\bar\sigma_k=\frac12(\mathrm c_k^{-1}-\mathrm c_{k+1}^{-1})^2\in[0,\frac18h^2]$, $0\le\mathrm c_a\mathrm c_b-1\le(a^2+b^2)h^2$ ($|a|h,|b|h\le1$), and $|\mathrm s_a\mathrm s_b/(\mathrm s_k\mathrm s_{k+1})|\le|ab|/(k(k+1))$, one has $|B_k-\beta_0\bar\sigma_k|\le\rho_2h^2$ with $\rho_2:=\beta_0(\frac18+a^2+b^2)+4|ab|/q_*^2$. (F4) For $\sigma>\sigma_*$, $\tilde u-1=\Phi/\widetilde e$ with $\tilde u(\sigma):=\tau(x-\beta_0\sigma)/\widetilde e(\sigma)$ is positive and increasing, so $\Upsilon(q)=\operatorname{arcosh}\tilde u(\sigma(q))$ is positive and increasing on $(q_*,q_m)$, while $\Upsilon(q):=0$ for $0\le q\le q_*$. \begin{proof}[Proof of Proposition~\ref{prop:exact-hyperbolicity}] Let $q_k\ge q_*$, $c_k=\tau(x-B_k)$. By (F3) and concavity of $\widetilde e$, $A_k+A_{k+1}\le\frac12(1+\rho_1h^2)(\widetilde e(\sigma_k)+\widetilde e(\sigma_{k+1}))\le(1+\rho_1h^2)\,\widetilde e(\bar\sigma_k)\le\widetilde e(\bar\sigma_k)+\rho_1h^2$ ($k_{\max}$ included): the middle step is concavity of $\widetilde e$ and the last is $\widetilde e\le1$ from (F1). For Hahn and Racah $\widetilde e(\bar\sigma_k)\le\widetilde e(q_k^2)$ and $q_k^2-q_*^2\ge2q_*(q_k-q_*)$, so \eqref{eq:forbidden-slope} gives \begin{equation} c_k-A_k-A_{k+1}\ge\Phi(q_k^2)-(\rho_1+\rho_2)h^2\ge2m_* q_*(q_k-q_*)-(\rho_1+\rho_2)h^2\ge0 \end{equation} for $q_k-q_*\ge C_0h^2$, $C_0:=(\rho_1+\rho_2)/(2m_* q_*)$. For quantum $6j$, $\sigma'(q)=2\tanh q\operatorname{sech}^2q$ is unimodal on $[0,\infty)$ with $\sigma''(q)=2\operatorname{sech}^2q(1-3\tanh^2q)\ge-\frac23$, so $c_\sigma:=\min_{[q_*,q_m]}\sigma'=\min(\sigma'(q_*),\sigma'(q_m))>0$ and $\bar\sigma_k\ge\sigma(q_k)-\frac1{12}h^2\ge\sigma_*+c_\sigma(q_k-q_*)-\frac1{12}h^2$; hence $c_k-A_k-A_{k+1}\ge\Phi(\bar\sigma_k)-(\rho_1+\rho_2)h^2\ge0$ for $q_k-q_*\ge C_0h^2$ with $C_0:=(\frac1{12}m_*+\rho_1+\rho_2)/(m_* c_\sigma)$. \end{proof} \begin{proof}[Proof of Proposition~\ref{prop:forbidden-decay}] Let $c_k=\tau(x-B_k)$, $k_c$ the first rank with $q_k\ge q_*+C_0N^{-2}$, $k_1$ the first rank with $q_k\ge q_*+N^{-2/3}$ (so that $N^{2/3}\zeta_\tau(q_{k_1})=O(1)$, as in Remark~\ref{rem:airy2-R}(b)), and $k_e$ the last rank in $I$. At a physical node the physical solution satisfies row $k_{\max}$ ($A_{k_{\max}+1}=0$): if $\widetilde Y_{k_{\max}}$ were $0$, then since $A_k>0$ for $k_{\min}0$), contradicting (E2). Thus $\widetilde Y_{k_{\max}}\ne0$. \emph{Comparison with constant ratios.} Let $y>0$ and $j\ge k_c$, and suppose \begin{equation} F_i(y):=A_{i+1}y+A_i/y\le c_i\qquad(j\le i\le k_{\max}). \label{eq:forbidden-const-ratio} \end{equation} Then $\widetilde Y_i\ne0$, of the sign of $\widetilde Y_{k_{\max}}$, for $j-1\le i\le k_{\max}$, and $0<\rho_i\le y$ for $j-1\le i0$, so $\widetilde Y_{i-1}\ne0$ has the sign of $\widetilde Y_i$ and \begin{equation} \rho_{i-1}=\frac{A_i}{c_i-A_{i+1}\rho_i}\in(0,y], \label{eq:forbidden-backward} \end{equation} because $\rho\mapsto A_i/(c_i-A_{i+1}\rho)$ is increasing and $A_i/(c_i-A_{i+1}y)\le y$. Backward induction from $i=k_{\max}$ gives the claim. \emph{(a)} is the case $y=1$: $F_i(1)=A_i+A_{i+1}\le c_i$ for $i\ge k_c$ is Proposition~\ref{prop:exact-hyperbolicity}. \emph{(b)} The inequality \eqref{eq:forbidden-const-ratio} is verified for explicit $y$ as follows. Let $y=e^{-\vartheta}$, $\cosh\vartheta\le\Lambda:=\tilde u(\sigma(q_*+\delta_0))$. By (F3) and Jensen's inequality for $\widetilde e$ (weights $\lambda=1/(1+y^2)$, $1-\lambda$), $F_i(y)\le(1+\rho_1h^2)\cosh\vartheta\,\widetilde e(\tilde\sigma_i)\le\cosh\vartheta\,\widetilde e(\tilde\sigma_i)+\rho_1\Lambda h^2$ (using $\widetilde e\le1$ and $\cosh\vartheta\le\Lambda$) with $\tilde\sigma_i=\lambda\sigma_i+(1-\lambda)\sigma_{i+1}=\bar\sigma_i-\frac12(\sigma_{i+1}-\sigma_i)\tanh\vartheta$. Put $\Phi_\vartheta(\sigma):=\tau(x-\beta_0\sigma)-\cosh\vartheta\,\widetilde e(\sigma)$; with $\sigma_\vartheta>\sigma_*$ defined by $\tilde u(\sigma_\vartheta)=\cosh\vartheta$ we have $\Phi_\vartheta(\sigma_\vartheta)=0$, and $\Phi_\vartheta'=\Phi'-(\cosh\vartheta-1)\widetilde e'\ge\Phi'\ge m_*$ on $[\sigma_\vartheta,\sigma_{\max})$ by \eqref{eq:forbidden-slope}, since $\cosh\vartheta\ge1$ and $\widetilde e'\le0$. For Hahn and Racah ($\sigma_{i+1}-\sigma_i=2q_ih$), $\tilde\sigma_i=q_i^2-hq_i\tanh\vartheta+\frac{h^2}4\ge q_i^2-q_mh$ and $c_i\ge\tau(x-\beta_0\tilde\sigma_i)-(\beta_0q_m+\frac{\beta_0}4+\rho_2)h$ ($h\le1$ since $N\ge N_0$), so $c_i-F_i(y)\ge\Phi_\vartheta(\tilde\sigma_i)-D_3h$ with $D_3:=\beta_0q_m+\frac{\beta_0}4+\rho_2+\rho_1\Lambda$; setting $D_0:=q_m+D_3/m_*$ and $\sigma_\vartheta=q_j^2-D_0h$ whenever $q_j^2-D_0h>q_*^2$ gives $\tilde\sigma_i\ge\sigma_\vartheta+(D_0-q_m)h$ for $i\ge j$, hence $c_i-F_i(y)\ge0$ and \eqref{eq:forbidden-const-ratio} for $y=\exp(-\Upsilon(\sqrt{q_j^2-D_0h}))$. For quantum $6j$, $0\le\sigma'\le1$ gives $\sigma_{i+1}-\sigma_i\le h$, $\tilde\sigma_i\ge\sigma_i\ge\sigma(q_i)-\frac12h$ and $|\bar\sigma_i-\tilde\sigma_i|\le\frac12h$, so $c_i-F_i(y)\ge\Phi_\vartheta(\tilde\sigma_i)-D_3h$ with $D_3:=\frac12\beta_0+\rho_2+\rho_1\Lambda$. Put $D_\sigma:=\frac12+D_3/m_*$, $c_\sigma':=\min(\sigma'(q_*),\sigma'(q_*+\delta_0))>0$, and $D_0:=\frac{2(q_*+\delta_0)}{c_\sigma'}D_\sigma$: whenever $q_j\in[q_*,q_*+\delta_0]$ has $\tilde q_j:=\sqrt{q_j^2-D_0h}>q_*$, the mean value theorem on $[\tilde q_j,q_j]$ gives $\sigma(q_j)-\sigma(\tilde q_j)\ge c_\sigma'(q_j^2-\tilde q_j^2)/(2(q_*+\delta_0))=D_\sigma h$, so $\sigma_\vartheta:=\sigma(q_j)-D_\sigma h\ge\sigma(\tilde q_j)>\sigma_*$ satisfies $\tilde\sigma_i\ge\sigma_\vartheta+(D_\sigma-\frac12)h$ for $i\ge j$, giving $c_i-F_i(y)\ge0$ and $\rho_{j-1}\le\exp(-\operatorname{arcosh}\tilde u(\sigma_\vartheta))\le\exp(-\Upsilon(\tilde q_j))$. When $q_j^2-D_0h\le q_*^2$, $\Upsilon(\sqrt{(q_j^2-D_0h)_+})=0$ and $\rho_{j-1}\le1$ by (a); thus in all three chains, for every $j\ge k_1$ ($j\ge k_c$ for large $N$) with $q_j\le q_*+\delta_0$, \begin{equation} \rho_{j-1}\le\exp\bigl(-\Upsilon(\sqrt{(q_j^2-D_0h)_+})\bigr). \end{equation} \emph{The product.} Put $m:=\lceil D_0/(2q_*)\rceil$ and $\Upsilon_{\max}:=\Upsilon(q_*+\delta_0)$. Since adjacent $q_j^2$ are spaced by at least $2q_*h$, for $j\ge k_1+m+1$ one has $q_j^2-D_0h\ge q_{j-m}^2\ge q_{k_1+1}^2>q_*^2$, so $\Upsilon(\sqrt{(q_j^2-D_0h)_+})\ge\Upsilon(q_{j-m})$; for $k_10$ and $C_1:=7/(4c)$. Since $q_*+\delta_0-q_{k_e}(C_1\log N-1)h$. For $q\ge q_*+\frac{\delta_0}2$ one has $\Upsilon\ge c$ on $[q,q_{k_e}]$, so $\xi(q_{k_e})-\xi(q)=N\int_q^{q_{k_e}}\Upsilon\ge cN(q_{k_e}-q)\ge\frac74\log N-c$, and the factor $e^{2c}$ lost against $N^{-7/2}$ is absorbed in $C$; while for $q\le q_*+\frac{\delta_0}2$ monotonicity of $\xi$ ($\xi=0$ for $q\le q_*$) gives $\xi(q_{k_e})-\xi(q)\ge\xi(q_{k_e})-\xi(q_*+\frac{\delta_0}2)\ge cN(q_{k_e}-q_*-\frac{\delta_0}2)\ge\frac{c\delta_0}4N$ for large $N$. The rest of the proof of Theorem~\ref{thm:general-rank-airy}(ii) is unchanged ($\kappa_h$ is fixed at an oscillatory $q_0$; the other remainders were estimated on all of $I$); the amplitude is Proposition~\ref{prop:airy2-amplitude}. \end{proof} \section{Hahn/\texorpdfstring{$3j$}{3j} chain data} \label{sec:supp-hahn} This section records the Hahn specialization of the fixed-rank contact identity, the explicit shadow phase, the principal secular terms, and the explicit higher phase horizons of Section~\ref{sec:growing3j}. \begin{corollary}[Fixed-rank second-order contact for Hahn] \label{lem:3j-fixed-rank-second-contact} For the Hahn chain $(\varepsilon_1,\beta_1,\beta_2)=(\tfrac12,0,0)$, so Proposition~\ref{prop:general-fixed-rank-contact} gives, with $s=1-x^2$, \begin{equation} \rho_{20}^{(3)}(x)=\frac{9x-x^3-8x(M^2+\sigma^2)+4M\sigma x^2+12M\sigma}{24s^{3/2}}, \end{equation} the contact source $-\frac12\bigl[(k+1)^2A_{k+1}^{(0)}\widehat d_{k+1}+k^2A_k^{(0)}\widehat d_{k-1}\bigr]$ in \eqref{eq:general-contact-identity}, and \begin{equation} \mathcal R_k(\widehat U^{[2]})=O_k(h^4)\qquad\text{at every fixed admissible rank }k. \label{eq:3j-fixed-rank-contact-residual} \end{equation} \end{corollary} \begin{proposition}[Hahn shadow phase] \label{prop:3j-shadow-p2} For the Hahn chain, on a compact subset of $00$, \begin{equation} p_h(q,x)=p(q,x)+h^2p_2^{(3)}(q,x)+O(h^4), \label{eq:3j-shadow-expansion} \end{equation} where \begin{equation} p_2^{(3)}(q,x)= -\frac{4(1-q^2)^2\bigl[(M^2+\sigma^2)x-2M\sigma\bigr] +x(3q^2-1)} {8q^2(1-q^2)^2\sqrt{1-x^2-q^2}}. \label{eq:3j-shadow-p2} \end{equation} In particular \begin{equation} p_2^{(3)}(q,x) =-\frac{\mathfrak c_{M\sigma}(\theta)}{q^2}+O(1), \qquad q\downarrow0. \label{eq:3j-shadow-darboux-pole-match} \end{equation} \end{proposition} \begin{proof} This is \eqref{eq:general-shadow-p2} with $(a,b)=(M,\sigma)$, $e_\pi(s)=\sqrt{1-s}$, $v_\pi=0$ and $\sin p=\sqrt{1-x^2-q^2}/\sqrt{1-q^2}$: here $(\log a_0)''=-(1+q^2)/(1-q^2)^2$, and clearing the denominator $8q^2(1-q^2)^2$ gives the displayed numerator, since $(1-q^2)^2-q^2(1+q^2)=1-3q^2$. The pole is \eqref{eq:general-shadow-darboux-pole-match}. \end{proof} \begin{proposition}[Principal secular terms at orders $N^{-2}$ and $N^{-4}$] \label{prop:D2D4match} For fixed $M,\sigma$ and $x$ in a compact set $K\Subset(-1,1)$, let $f=d^L_{M\sigma}(\arccos x)$ and $\kappa=L+\tfrac12$, and let $f_2,f_4$ be the exact fixed-rank coefficients of $N^{-2},N^{-4}$ in $\Phi_L^{M,\sigma}$. Then \begin{equation} f_2=\kappa^2\mathcal P_2f+O_K(L^{1/2}), \qquad \mathcal P_2=\frac{x}{6}\partial_x+\frac{3-x^2}{12(1-x^2)}, \label{eq:P2} \end{equation} and the exact fourth-order coefficient $f_4=\mathcal D_4f$ satisfies \begin{equation} \boxed{f_4=\kappa^4\mathcal P_4f+O_K(L^{7/2}),} \label{eq:f4match} \end{equation} with $\mathcal P_4$ from \eqref{eq:P4}. \end{proposition} \begin{proof} The first statement is the large-$L$ part of the exact fixed-rank $f_2$ formula of Ref.~\cite[Theorem~6.1]{LewisFixedRank}. For the second, the same reference gives $f_4=\mathcal D_4f$ with $\mathcal D_4=A_4(x)\partial_x+B_4(x)$ and $A_4,B_4$ rational in $(x,J,\Sigma,\tau_\star)$, $J=L(L+1)$, $\Sigma=M^2+\sigma^2$, $\tau_\star=M\sigma$~\cite[Theorem~6.3 and Eqs.~(323)--(324)]{LewisFixedRank}. Reducing $f''$ in $\kappa^4\mathcal P_4f$ by the Wigner equation $(1-x^2)^2f''=2x(1-x^2)f'+[-J(1-x^2)+\Sigma-2\tau_\star x]f$ gives a first-order form $\widetilde A_4f'+\widetilde B_4f$, and exact rational algebra gives $\deg_J(A_4-\widetilde A_4)=1$ and $\deg_J(B_4-\widetilde B_4)=2$: the $J^2f'$ and $J^3f$ terms cancel. Since $f=O_K(L^{-1/2})$ and $f'=O_K(L^{1/2})$ on compact bulk sets, $(\mathcal D_4-\kappa^4\mathcal P_4)f=O_K(L^{5/2}+L^{7/2})$. \end{proof} \subsection{Exact Hahn representation} \label{app:hahn-representation} In the fundamental Hahn chamber $\sigma\ge|M|$, the standard $3j$--Clebsch--Gordan relation and Koornwinder's Hahn identity, Eq.~(4.5) of~\cite{KoornwinderHahn}, give the exact parameter dictionary \begin{equation} n_H=L-\sigma,\qquad N_H=N-1-\sigma,\qquad \alpha_H=\sigma-M,\qquad \beta_H=\sigma+M, \label{eq:hahnExactDictionaryA} \end{equation} \begin{equation} k_H=\frac{N-1-\sigma-2\mu+M}{2}. \label{eq:hahnExactDictionaryB} \end{equation} Writing \begin{align} w_H(k)&=\frac{(\alpha_H+1)_k}{k!} \frac{(\beta_H+1)_{N_H-k}}{(N_H-k)!},\\ h_H(n)&=\frac{(-1)^n(n+\alpha_H+\beta_H+1)_{N_H+1} (\beta_H+1)_n n!} {(2n+\alpha_H+\beta_H+1)(\alpha_H+1)_n(-N_H)_nN_H!}, \label{eq:hahnWeightNormBridge} \end{align} Koornwinder's exact identity and the standard $3j$--Clebsch--Gordan phase can be written in the orthonormal form \begin{equation} \widehat\Phi_L^{M,\sigma}=\sqrt{2L+1}\,\Phi_L^{M,\sigma} =\varsigma_H\sqrt{\frac{Nw_H(k_H)}{h_H(n_H)}}\, Q_{n_H}(k_H;\alpha_H,\beta_H,N_H), \label{eq:hahnExactOrthonormalBridge} \end{equation} where $\varsigma_H=+1$ in the fundamental chamber $\sigma\ge|M|$ ($k_H=0$ gives $m_-=j_-$, $Q_{n_H}(0)=1$, and a single Racah term of sign $(-1)^{4L}=+1$); it is not an undetermined connection scalar. All other fixed $(M,\sigma)$ chambers are finite exact Wigner/Regge symmetry images of the fundamental chamber, up to a fixed sign $\varsigma_H\in\{\pm1\}$. \begin{remark}[Higher Hahn phase horizons] \label{rem:hahn-higher-horizons} \label{sec:higherhorizons} Expanding the arcsine in \eqref{eq:pq}, which converges on every compact turning-free set, \begin{equation} NS(q,x)=\kappa\arccos x-\sum_{r\ge1}\frac{\binom{2r}{r}}{4^r(2r+1)}\frac{\kappa^{2r+1}}{N^{2r}}\int_0^x\frac{\dd u}{(1-u^2)^{r+1/2}}, \label{eq:phasehierarchy} \end{equation} so every coefficient of the hierarchy \eqref{eq:even-phase-horizon-hierarchy} is explicit and nonzero for $x\ne0$, and the closed action resums it. The first correction displaces a bulk zero by $O(L^2/N^2)$ in $\theta$, one zero spacing (of order $1/\kappa$) when $L^3/N^2\asymp1$ and $x\ne0$. \end{remark} \section{Hahn Airy connection as a rank-direction turn} \label{app:hahn-airy-dictionary} This section proves Theorem~\ref{thm:hahnAiryConnection} from Theorem~\ref{thm:general-rank-airy}. At a fixed node $x\ne0$ the Hahn chain turns in the rank direction exactly where the node meets the band edge, and the two Langer coordinates coincide. \begin{lemma}[Hahn rank turn and Langer coordinate] \label{lem:hahn-rank-turn} Let $X$ be a compact subset of $(0,1)$ or of $(-1,0)$, and take the Hahn chain of Proposition~\ref{prop:hahn-class} with any $s_1\in(\max_X(1-x^2),1)$. For $x\in X$ put $q_*(x)=\sqrt{1-x^2}$. Then $q_*^20$ for $0\le q0),\qquad S(q_*,x)=\pi q_*-\frac\pi2(1+x)\quad(x<0). \label{eq:SupperTurn} \end{equation} \end{lemma} \begin{proof} The turning statements are immediate from $D=1-x^2-q^2$, and $u(q)=x/\sqrt{1-q^2}$ gives $u(q_*)=\operatorname{sgn}x$. Take $x>0$. On the oscillatory side $q\le q_*$, equivalently $x\le x_+(q)$, write $p(s,x)=\arccos(x/\sqrt{1-s^2})=\int_x^{\sqrt{1-s^2}}(1-s^2-u^2)^{-1/2}\,\dd u$. Fubini's theorem on $\{(s,u):s\ge q,\ u\ge x,\ s^2+u^2\le1\}$ gives \begin{equation} \int_q^{q_*}p(s,x)\,\dd s =\int_x^{x_+(q)}\Bigl[\arcsin\frac{s}{\sqrt{1-u^2}}\Bigr]_{s=q}^{s=\sqrt{1-u^2}}\dd u =\int_x^{x_+(q)}\Bigl(\frac\pi2-p_q(u)\Bigr)\dd u, \end{equation} which is \eqref{eq:hahn-zeta-xi} there. On the forbidden side $q\ge q_*$, equivalently $x\ge x_+(q)$, the rank action $I_{\rm r}(q)=\int_{q_*}^q\operatorname{arcosh}(x/\sqrt{1-s^2})\,\dd s$ and the spatial action $I_{\rm s}(q)=\int_{x_+(q)}^x\operatorname{arcosh}(q/\sqrt{1-u^2})\,\dd u$ both vanish at $q=q_*$, and, since the integrand of $I_{\rm s}$ vanishes at $u=x_+(q)$, \begin{equation} \partial_qI_{\rm s}=\int_{x_+(q)}^x\frac{\dd u}{\sqrt{u^2-(1-q^2)}}=\operatorname{arcosh}\frac{x}{\sqrt{1-q^2}}=\partial_qI_{\rm r}. \end{equation} Hence $I_{\rm r}=I_{\rm s}$. For $x<0$, $p(s,-x)=\pi-p(s,x)$ and $u\mapsto-u$, so $\zeta_-(q;x)=\zeta_+(q;-x)$, while $\xi_-(q,x)=\xi_+(q,-x)$ by definition. Finally, \eqref{eq:Fclosed}--\eqref{eq:Theta} at $x=x_\pm(q)$ give $F=\frac\pi2(q+x-1)$ for $x>0$ and $F=\frac\pi2(x+1-q)$ for $x<0$ (the arctangent tends to $\pm\pi/2$), and $S=\pi q/2-F$ is \eqref{eq:SupperTurn} with $q=q_*(x)$. \end{proof} \subsection*{Proof of Theorem~\ref{thm:hahnAiryConnection}: Weyl-normalized fixed-ratio Hahn Airy connection} \begin{proof} Upper turn. By \eqref{eq:xiPlusLinear}, $\xi_+(q,\cdot)$ is continuous and increasing with derivative bounded below near $x_+(q)$, uniformly for $q\in[q_0,q_1]$. Hence for $N\ge N_0(q_0,q_1,A)$ every pair with $q\in[q_0,q_1]$ and $|N^{2/3}\xi_+(q,x)|\le A$ has $x$ in the compact set $X=[\frac12\sqrt{1-q_1^2},\frac12(1+\sqrt{1-q_0^2})]\subset(0,1)$. Lemma~\ref{lem:hahn-rank-turn} gives a simple rank turn of type $+1$, uniformly on $X$. At a physical node $x=1-(p_++1+2\nu_+)/N$ of \eqref{eq:upperInwardAiry}, \eqref{eq:SupperTurn} gives $NS(q_*,x)/\pi=(p_++1+2\nu_+)/2$, so \begin{equation} K_+=\frac{NS(q_*,x)}{\pi}+\frac{p_+-1}{2}=\nu_++p_+\in\mathbb Z, \end{equation} with $p_{ab}=p_+=|M-\sigma|$. Theorem~\ref{thm:general-rank-airy} with $(a,b)=(M,\sigma)$, $Y_L=\Phi_L^{M,\sigma}$ and $\epsilon_{ab}=\epsilon_{M\sigma}$ applies on the compact family $X$; by \eqref{eq:hahn-zeta-xi} its window $|N^{2/3}\zeta_+|\le A$ is the stated one, and it gives \eqref{eq:upperHahnAiry} with $(-1)^{K_+}=(-1)^{\nu_++p_+}$ and additive $O_A(N^{-1})$ bracket error. Condition~(R) of Theorem~\ref{thm:general-rank-airy}(ii) holds for every rank: by $3j$ orthogonality the $3j$ symbol $w_L$ of \eqref{eq:PhiDef} satisfies $\sum_L(2L+1)w_L^2=1$, so $|\widehat\Phi_L|=\sqrt{(2L+1)N}\,|w_L|\le N^{1/2}$. Theorem~\ref{thm:general-rank-airy}(ii) therefore gives the bracket $\operatorname{Ai}(t)+N^{-4/3}\varpi_+(x)\operatorname{Ai}'(t)+O_A(N^{-2})$. For $e_\pi=\sqrt{1-s}$, $v_\pi=0$ and $q_*=\sqrt{1-x^2}$ one has $\tau u(q)=|x|(1-q^2)^{-1/2}$, $e_0=|x|$, $u_1=q_*/x^2$ and $v_1=0$, and \eqref{eq:general-airy-varpi-local} reduces to \eqref{eq:hahn-varpi} (exact computer algebra). In the window $x-x_\pm(q)=O(N^{-2/3})$, so $\varpi^{\mathrm H}$ may equally be evaluated at the band edge, at cost $O(N^{-2})$. Lower turn. The same argument applies on $-X$, with type $\tau=-1$. At a node $x=-1+(p_-+1+2\nu_-)/N$ of \eqref{eq:lowerInwardAiry}, \eqref{eq:SupperTurn} gives \begin{equation} K_-=\kappa+\frac{NS(q_*,x)}{\pi}-Nq_*+\frac{p_+}2=\kappa-\frac{N(1+x)}2+\frac{p_+}{2}=L-\nu_-+\frac{p_+-p_-}{2}. \end{equation} This is an integer: $(p_+-p_-)/2=-\operatorname{sgn}(M\sigma)\min(|M|,|\sigma|)$ and $L\pm M$, $L\pm\sigma$ are integers. Since $K_-$ and $\nu_-+L+(p_+-p_-)/2$ differ by the even integer $2\nu_-$, Theorem~\ref{thm:general-rank-airy} gives \eqref{eq:lowerHahnAiry}. Equivalently, the lower turn follows from the upper one by the exact reflection $\Phi_L^{M,\sigma}(-x;N^{-1})=(-1)^{L-\sigma}\Phi_L^{-M,\sigma}(x;N^{-1})$ (a $3j$ symmetry), which exchanges $p_+\leftrightarrow p_-$ and $\nu_+\leftrightarrow\nu_-$, takes $\xi_+(q,-x)$ to $\xi_-(q,x)$, and leaves $D$ and $\varpi^{\mathrm H}$ unchanged ($\varpi^{\mathrm H}$ depends on $x$ only through $x^2$ and $M\sigma x$); the signs agree because $\epsilon_{M\sigma}\epsilon_{-M,\sigma}=(-1)^{\sigma+\max(|M|,|\sigma|)}$. \end{proof} The proof uses no parameter translation to a printed Hahn-polynomial normalization: the exact representation \eqref{eq:hahnExactOrthonormalBridge} is needed only for the bulk zeros (Corollary~\ref{cor:hahn-bulk-zero-quantization}). \section{Centered Racah chain data} \label{sec:supp-racah} This section records the Racah specialization of the fixed-rank contact identity, the first pullback coefficient, the Hahn degeneration and second secular phase of the principal symbol, and the proofs of the dual, tetrahedral and flat-turn statements of Section~\ref{sec:growing6j}. \begin{corollary}[Fixed-rank second-order contact for centered Racah] \label{lem:6j-fixed-rank-second-contact} Put $s=1-x^2$, $\mathfrak e_6=\alpha^{-2}+\beta^{-2}$, $\mathfrak p_6=(\alpha\beta)^{-1}$ and $\mathfrak q_{mn}=\frac18-\frac{m^2+n^2}2$. For the centered Racah chain $(\varepsilon_1,\beta_1,\beta_2)=(\mathfrak e_6/8,\ \mathfrak p_6/4,\ (\frac12-m^2-n^2)\mathfrak p_6/8)$, so Proposition~\ref{prop:general-fixed-rank-contact} gives \begin{equation}\begin{aligned} \rho_{20}^{(6)}={}&\frac{1}{96s^{3/2}}\Bigl[ 16\mathfrak e_6\mathfrak q_{mn}x +4\mathfrak e_6mnx^2+12\mathfrak e_6mn -\mathfrak e_6x^3+7\mathfrak e_6x\\ &\hspace{14mm} -24\mathfrak p_6\mathfrak q_{mn}x^2 -8\mathfrak p_6\mathfrak q_{mn} -32\mathfrak p_6mnx -7\mathfrak p_6x^2-5\mathfrak p_6\Bigr], \end{aligned}\end{equation} the contact source $-\frac{\mathfrak e_6}{8}\bigl[(k+1)^2A_{k+1}^{(0)}\widehat d_{k+1}+k^2A_k^{(0)}\widehat d_{k-1}\bigr]+\delta_k^{(6)}\widehat d_k$ in \eqref{eq:general-contact-identity}, with $\delta_k^{(6)}=\beta_1\kappa^2+\beta_2=(2k(k+1)+1-m^2-n^2)/(8\alpha\beta)$, and \begin{equation} \mathcal R_k(\widehat U^{[2]})=O_k(h^4)\qquad\text{at every fixed admissible rank }k, \label{eq:6j-fixed-rank-contact-residual} \end{equation} including $k=0$ in the channel $m=n=0$, whose exact two-start seed is \eqref{eq:6jexceptionalstart}. \end{corollary} \begin{proposition}[First centered-Racah pullback coefficient] \label{prop:6j-first-pullback-coeff} Put \begin{equation} \mathfrak a_6:=\alpha^{-2}+\beta^{-2},\qquad \chi_6(x):=\frac{x\mathfrak a_6}{8}-\frac1{4\alpha\beta}. \label{eq:6j-a6-chi6} \end{equation} Then, as $\rho\to0$, \begin{align} \Theta_6(\rho) &=\theta-\frac{\chi_6(x)}{3\sin\theta}\rho^2+O(\rho^4),\\ W_6(\rho) &=1+\rho^2 \frac{(3-x^2)\mathfrak a_6/48-x/(12\alpha\beta)}{1-x^2} +O(\rho^4). \label{eq:6j-W6-smallrho} \end{align} Hence the first pullback operator is \begin{equation} \boxed{ \mathcal P_2^{(6)}= \left(\frac{x\mathfrak a_6}{24}-\frac1{12\alpha\beta}\right)\partial_x +\frac{(3-x^2)\mathfrak a_6/48-x/(12\alpha\beta)}{1-x^2}.} \label{eq:6j-P2-pullback} \end{equation} If $g_2^{(f)}$ denotes the coefficient of $t=h^2=N^{-2}$ in the centered Racah fixed-rank series of $\Psi_f$, then on compact strict-triangle families \begin{equation} \sqrt{f+1}\, \left|g_2^{(f)}-\kappa_f^2\mathcal P_2^{(6)}d^f_{mn}(\theta)\right| \le C(f+1). \label{eq:6j-g2-pullback-bound} \end{equation} \end{proposition} \begin{proof} The expansions are \eqref{eq:class-small-q} with \eqref{eq:racah-class-data}, and the bound \eqref{eq:6j-g2-pullback-bound} is Lemma~\ref{lem:general-first-coefficient}. \end{proof} Hahn also appears as a principal-symbol degeneration of Racah, a useful unification at the Hamilton--Jacobi level. Set $\alpha=\tfrac12$ and let $\beta\to\infty$ while $x$ is held fixed through the law of cosines, so $\gamma^2=\alpha^2+\beta^2-2\alpha\beta x$. Then \eqref{eq:6jsymbol} tends locally uniformly before the turning point to \begin{equation} \cos\omega_{\rm H}(\rho)=\frac{x}{\sqrt{1-\rho^2}}, \label{eq:hahn-from-racah-symbol} \end{equation} which is exactly the rank-direction Hahn symbol associated with the action \eqref{eq:Theta}. In particular \begin{equation} \frac{x}{\sqrt{1-\rho^2}} =x\sum_{m\ge0}\frac{\binom{2m}{m}}{4^m}\rho^{2m}, \label{eq:hahn-all-secular-symbol-coeffs} \end{equation} so every even Hahn principal-symbol coefficient is the boundary value \begin{equation} \chi_{2m}^{(\mathrm H)}=x\frac{\binom{2m}{m}}{4^m}. \end{equation} Moreover the Racah turning radius tends to $\sqrt{1-x^2}$, while $\partial_\rho S_6=\omega$ tends to \begin{equation} \partial_qS=\arccos\!\left(\frac{x}{\sqrt{1-q^2}}\right). \end{equation} Thus the Hahn and centered-Racah principal actions are two members of the same two-parameter symbol family. This is a principal-symbol degeneration only; no uniform $\beta\to\infty$ theorem for the full offset-dependent recurrence is asserted. \begin{remark}[Second centered-Racah secular phase and the $N^{4/5}$ horizon] \label{rem:6j-second-secular-horizon} With $E_6=\alpha^{-2}+\beta^{-2}$, $P_6=(\alpha\beta)^{-1}$, $\chi_2^{(6)}=xE_6/8-P_6/4$ and $\chi_4^{(6)}=x(3E_6^2/128-P_6^2/32)-E_6P_6/32$, the principal symbol is $\cos\omega=x+\chi_2^{(6)}\rho^2+\chi_4^{(6)}\rho^4+O(\rho^6)$, so with $s_\theta=\sin\theta$ \begin{equation} NS_6(\rho)=\kappa_f\theta-\frac{\chi_2^{(6)}}{3s_\theta}\frac{\kappa_f^3}{N^2}-\frac15\Bigl[\frac{\chi_4^{(6)}}{s_\theta}+\frac{x(\chi_2^{(6)})^2}{2s_\theta^3}\Bigr]\frac{\kappa_f^5}{N^4}+O\Bigl(\frac{\kappa_f^7}{N^6}\Bigr). \label{eq:6j-second-secular-phase} \end{equation} Once the first term is resummed, the heuristic rule of Remark~\ref{rem:even-phase-hierarchy} gives $f\asymp N^{4/5}$ as the next generic horizon (not proved here); on the locus $\chi_2^{(6)}=0$ it is the first, unless the bracket also vanishes. \end{remark} \subsection*{Proof of the statements of Remark~\ref{prop:6j-dual-principal-symbol}: dual principal symbol and bispectral action} \begin{proof} For the exact self-adjoint $c$-lattice operator of Ref.~\cite[Theorem~8.2]{LewisFixedRank}, let $a_c^{\pm}$ be the two hopping coefficients and $b_c$ the diagonal term. At $C=c+\tfrac12=N\gamma$ with fixed $m,n$, expanding its four-factor edge coefficient gives $N^{-2}a_c^{\pm}=a_C^{(0)}+O(N^{-1})$ with $a_C^{(0)}=\Delta_C(\gamma)/(4\gamma^2)$ and $N^{-2}b_c=b_C^{(0)}+O(N^{-1})$ with $b_C^{(0)}=\Delta_C(\gamma)/(2\gamma^2)$. Since the exact eigenvalue is $f(f+1)=N^2\rho^2+O(1)$, the principal WKB ansatz in the $c$ direction gives $b_C^{(0)}-2a_C^{(0)}\cos p_C=\rho^2$, proving \eqref{eq:racah-dual-HJ}. The law of cosines $2\alpha\beta x=\alpha^2+\beta^2-\gamma^2$ gives $\Delta_C=4\alpha^2\beta^2(1-x^2)$, and hence the common caustic by \eqref{eq:D6linear}. Since $\eta(\rho)$ is independent of $\gamma$, differentiating \eqref{eq:6jsymbol} at fixed $\rho$ gives $\partial_\gamma\omega=\gamma/(\alpha\beta\eta\sin\omega)=2\gamma/\sqrt{\Delta_C-\rho^2\gamma^2}$, because $4\alpha^2\beta^2\eta^2\sin^2\omega=\Delta_C-\rho^2\gamma^2$ by \eqref{eq:D6adiabatic}; differentiating \eqref{eq:racah-dual-HJ} in $\rho$ gives the same expression. Since $S_6(\rho,\gamma)=\int_0^\rho\omega(s,\gamma)\,\dd s$ and $p_C(0,\gamma)=0$, $\partial_\gamma S_6=\int_0^\rho\partial_sp_C(s,\gamma)\,\dd s=p_C(\rho,\gamma)$, proving \eqref{eq:racah-bispectral-action}. \end{proof} \subsection*{Proof of Lemma~\ref{lem:racah-cayley-menger}: The Racah turning curve is the flat-tetrahedron locus} \begin{proof} Order the vertices $1,2,3,4$ with $d_{12}=d_{14}=A$, $d_{23}=d_{34}=B$, $d_{13}=C$, $d_{24}=F$, and write $a=A^2$, $b=B^2$, $c=C^2$, $f=F^2$. The Cayley--Menger matrix, bordered by the vector of ones and indexed $0,1,2,3,4$, is \begin{equation} \mathrm{CM}=\begin{pmatrix} 0&1&1&1&1\\ 1&0&a&c&a\\ 1&a&0&b&f\\ 1&c&b&0&b\\ 1&a&f&b&0 \end{pmatrix}. \label{eq:cm-matrix} \end{equation} The following computation is a polynomial identity valid for all real $N,\alpha,\beta,\gamma,\rho$: only the squares $a,b,c,f$ enter, and no tetrahedron is assumed to exist. \emph{Direct expansion.} Regard $\mathrm{CM}$ as a linear map on $\mathbb R^5$ with standard basis $e_0,\dots,e_4$. Exchanging the vertices $2$ and $4$ leaves it invariant, so it preserves $v=e_2-e_4$ and the span of $e_0,e_1,e_3,u$ with $u=e_2+e_4$. Indeed $\mathrm{CM}\,v=-f\,v$, and in the basis $(e_0,e_1,e_3,u)$ of the complement the images $\mathrm{CM}\,e_0=e_1+e_3+u$, $\mathrm{CM}\,e_1=e_0+c\,e_3+a\,u$, $\mathrm{CM}\,e_3=e_0+c\,e_1+b\,u$, $\mathrm{CM}\,u=2e_0+2a\,e_1+2b\,e_3+f\,u$ give \begin{equation} \det\mathrm{CM}=-f\det\begin{pmatrix}0&1&1&2\\1&0&c&2a\\1&c&0&2b\\1&a&b&f\end{pmatrix}. \label{eq:cm-reduced} \end{equation} Subtracting the second row from the third and fourth and expanding along the first column, the $4\times4$ determinant in \eqref{eq:cm-reduced} equals \begin{equation} -\det\begin{pmatrix}1&1&2\\c&-c&2(b-a)\\a&b-c&f-2a\end{pmatrix} =-2\bigl[4ab-(a+b-c)^2-cf\bigr]. \label{eq:cm-minor} \end{equation} With $a=N^2\alpha^2$, $b=N^2\beta^2$, $c=N^2\gamma^2$, $f=N^2\rho^2$, \eqref{eq:cm-reduced} and \eqref{eq:cm-minor} give \begin{equation} \det\mathrm{CM}=2N^6\rho^2\bigl[4\alpha^2\beta^2-(\alpha^2+\beta^2-\gamma^2)^2-\gamma^2\rho^2\bigr]. \label{eq:cm-poly} \end{equation} \emph{The two forms of \eqref{eq:racah-cayley-menger}.} Heron's formula gives $4\alpha^2\beta^2-(\alpha^2+\beta^2-\gamma^2)^2=16T^2$, which with \eqref{eq:cm-poly} is the second equality. For $\alpha,\beta\ne0$, $2\alpha\beta x=\alpha^2+\beta^2-\gamma^2$ and \eqref{eq:D6linear} give $4\alpha^2\beta^2D_6=4\alpha^2\beta^2-(\alpha^2+\beta^2-\gamma^2)^2-\gamma^2\rho^2$, which is the first. \emph{Sign trichotomy.} The prefactor $8N^6\rho^2\alpha^2\beta^2$ is strictly positive for $\rho>0$, so $\det\mathrm{CM}$ has the sign of $D_6$. A tetrahedron with prescribed edge lengths exists in $\mathbb R^3$ if and only if $\det\mathrm{CM}\ge0$ and every face satisfies the triangle inequalities. The four faces are $(\alpha,\beta,\gamma)$ twice, $(A,A,F)$ and $(B,B,F)$, and $D_6>0$ already makes all of them admissible. Indeed $D_6>0$ gives $\gamma^2\rho^2<16T^2$, so $T^2>0$ and the $(\alpha,\beta,\gamma)$ faces are nondegenerate triangles; and \begin{equation} 4\alpha^2\gamma^2-16T^2=(\alpha^2-\beta^2+\gamma^2)^2\ge0 \label{eq:cm-face} \end{equation} gives $\gamma^2\rho^2<16T^2\le4\alpha^2\gamma^2$, that is $\rho<2\alpha$, the triangle inequality for $(A,A,F)$; exchanging $\alpha$ and $\beta$ gives $\rho<2\beta$ for $(B,B,F)$. So $D_6>0$ is the nondegenerate Euclidean tetrahedron, $D_6=0$ the flat one, and for $D_6<0$ the determinant is negative and the six lengths are not those of any Euclidean tetrahedron. \end{proof} \begin{remark}[Geometric check for genuine tetrahedra] For an actual tetrahedron $\det\mathrm{CM}=288\mathcal V^2$, and $\mathcal V=Fr_Ar_B\sin\omega/6$ with $r_Ar_B=N^2\alpha\beta\,\eta$ and $\sin\omega=\sqrt{D_6}/\eta$ by \eqref{eq:6jsymbol} and \eqref{eq:D6adiabatic}. Hence $288\mathcal V^2=8N^6\rho^2\alpha^2\beta^2D_6$, which is the first equality of \eqref{eq:racah-cayley-menger} again. This route covers only $D_6\ge0$, where the tetrahedron exists; the expansion above is the statement for all real parameters. \end{remark} \begin{remark}[Relation to the full six-edge wall] The lemma identifies the one-parameter turning curve of this paper with a slice of the full six-edge Cayley--Menger wall $\{\det\mathrm{CM}=0\}$ in the space of squared edge lengths $x_{ij}$. At a point of the wall where the bordered distance matrix has corank one, with kernel vector $(v_0,\lambda)$, its adjugate is $c\,(v_0,\lambda)(v_0,\lambda)^{\mathsf T}$ for a scalar $c\ne0$, and differentiating the determinant gives $\partial\det\mathrm{CM}/\partial x_{ij}=2c\lambda_i\lambda_j$. The Airy turn treated here is therefore the pullback of the full wall along the symmetric family $(A,B,C;B,A,F)$; this gradient supplies the normal and the transversality of the wall in directions the one-parameter recurrence does not see. Conversely, the normalized Airy connection of Theorem~\ref{thm:racahAiryConnection} is the analytic local model on this slice. \end{remark} \begin{remark}[Volume and the equilateral family] First, $D_6$ decreases monotonically in $\rho^2$, but the tetrahedral volume does \emph{not}: \eqref{eq:racah-cayley-menger} gives \begin{equation} \mathcal V^2=\frac{N^6}{144}\rho^2(16T^2-\gamma^2\rho^2) =\frac{N^6\gamma^2}{144}\rho^2(\rho_*^2-\rho^2). \end{equation} Thus, on $0\le\rho\le\rho_*$ where the tetrahedron exists, $\mathcal V$ grows from zero at $\rho=0$, reaches its unique maximum at $\rho=\rho_*/\sqrt2$, and then decreases to zero at the positive flattening radius $\rho_*=4T/\gamma$. This distinguishes the monotonicity of the \emph{characteristic discriminant} from that of the geometric volume. Second, for the equilateral family $\alpha=\beta=\gamma=L$ one gets $\det\mathrm{CM}=2N^6L^2\rho^2(3L^2-\rho^2)$ and $\rho_*=\sqrt3\,L$, recovering the classical five-equal-edge caustic ratio. Since $\rho_*/L=\sqrt3$ is irrational whereas the ratio of two nonzero admissible Weyl edge lengths is rational, no exactly equilateral admissible half-integer label set lies on that caustic. The turning curve is met only through the $\rho-\rho_*=O(N^{-2/3})$ layer of Theorem~\ref{thm:racahAiryConnection}, equivalently a layer of width $O(N^{1/3})$ in the edge length $F$, consistent with the simple-turn hypothesis used there. \end{remark} \begin{lemma}[Closed Racah Airy coefficient] \label{lem:racah-varpi} For the data \eqref{eq:racah-class-data} at a simple turn, \eqref{eq:general-airy-varpi-local} equals \eqref{eq:racah-varpi}. \end{lemma} \begin{proof} Write $z=\rho^2$, $E(z)=e(z)^2$ and $K=\gamma^2/(4\alpha^2\beta^2)$, so $D_6=\sin^2\theta-Kz$ by \eqref{eq:D6linear} and $K\rho_*^2=\sin^2\theta$. At the turn $E(\rho_*^2)=(x-s_*)^2=\varepsilon^2$, so $e_0=|\varepsilon|$, $\tau=\operatorname{sgn}\varepsilon$ and $\tau/e_0=1/\varepsilon$. Since $\tau u_6=(1+w)^{1/2}$ with $w=K(z-\rho_*^2)/E(z)$ and $z-\rho_*^2=2\rho_*\delta+\delta^2$, the jets $u_j$ are rational in $(\alpha,\beta,x,\rho_*)$, and $u_1=K\rho_*/\varepsilon^2=\sin^2\theta/(\rho_*\varepsilon^2)$; hence $\rho_*^2u_1^{2/3}=(\rho_*\sin\theta)^{4/3}|\varepsilon|^{-4/3}$. With $v_1=(\frac12-m^2-n^2)/(8\alpha\beta)$ and $(2e_0e_2-e_1^2)/e_0^2=\frac12(\log E(\rho^2))''(\rho_*)$, the bracket of \eqref{eq:general-airy-varpi-local} reduces modulo $\rho_*^2=4\alpha^2\beta^2\sin^2\theta/\gamma^2$ to $[140\varepsilon^3((m^2+n^2)x-2mn)+\mathcal P(x,s_*)]/\varepsilon^4$; the offset terms combine through $1/\rho_*^2+1/(4\alpha\beta\varepsilon)=x/(\rho_*^2\varepsilon)$. This is an identity of rational functions of $(\alpha,\beta,x)$, checked by exact computer algebra. \end{proof} \subsection*{Proof of Lemma~\ref{lem:racahFlatTurnAction}: Flat centered-Racah turning action} \begin{proof} At the turning radius $R=\rho_*$ one has $\phi=\pi/2$ in \eqref{eq:S6closed}, so each $\operatorname{atan2}$ term equals $\frac\pi2\operatorname{sgn}c$, and \begin{equation} S_6(R)=R\,\omega(R)+\pi\gamma-\pi\alpha\operatorname{sgn}c_\beta-\pi\beta\operatorname{sgn}c_\alpha. \end{equation} By the law of cosines, $x-R^2/(4\alpha\beta)=-c_\alpha c_\beta$, and $\eta(R)>0$, so $\omega(R)=0$ when $c_\alpha c_\beta<0$ and $\omega(R)=\pi$ when $c_\alpha c_\beta>0$. If $\tau=+1$ the characteristic roots meet at $+1$, so $\omega(R)=0$: one base angle is obtuse, necessarily the one opposite the longer of $\alpha,\beta$, and $S_6(R)=\pi(\gamma-|\alpha-\beta|)$. If $\tau=-1$ they meet at $-1$, so $\omega(R)=\pi$, both base angles are acute, and $S_6(R)=\pi[R-(\alpha+\beta-\gamma)]$. This is \eqref{eq:racahFlatTurnAction}. \end{proof} \section{One-small quantum \texorpdfstring{$6j$}{6j} chain: proofs} \label{sec:supp-q6j} Two facts about the quantum $6j$ symbol are cited, both known. (K1) The Kirillov--Reshetikhin formula~\cite{KirillovReshetikhin1989}, which we take as the definition of the symbol, \begin{equation} \sixj{j_1}{j_2}{j_{12}}{j_3}{j}{j_{23}}_{\!q}=\prod_{\text{4 triads}}\Delta(\cdot)\ \sum_z\frac{(-1)^z[z+1]!}{\prod_i[z-\alpha_i]!\ \prod_i[\beta_i-z]!}, \end{equation} \begin{equation} \Delta(a,b,c)^2=\frac{[a+b-c]!\,[a-b+c]!\,[-a+b+c]!}{[a+b+c+1]!}, \end{equation} with $[z]!=\prod_{j=1}^z[j]$, the four triads $(j_1,j_2,j_{12}),(j_3,j,j_{12}),(j_1,j,j_{23}),(j_3,j_2,j_{23})$ with sums $\alpha_i$, the three quadrilateral sums $\beta_1=j_1{+}j_2{+}j_3{+}j$, $\beta_2=j_2{+}j_{12}{+}j{+}j_{23}$, $\beta_3=j_{12}{+}j_1{+}j_{23}{+}j_3$, and $z$ running over $\max\alpha_i\le z\le\min\beta_i$ in steps of $1$ (with the convention $1/[k]!=0$ for $k<0$ the sum may be taken over all of $\mathbb Z$). (K2) $q$-orthogonality: $\sum_f[2f+1][2c+1]W_f^2=1$ for the family $W_f$ of Section~\ref{sec:q6j}; it supplies an independent global bound for (R) (alongside Proposition~\ref{prop:exact-hyperbolicity}) and the Gauss-weight identity of Remark~\ref{rem:flux-gauss}. The three-term recurrence (a) is \emph{not} cited: it is proved from (K1) in Lemma~\ref{lem:q6j-recurrence} below. \begin{lemma}[Three-term recurrence of the Kirillov--Reshetikhin sum] \label{lem:q6j-recurrence} For the family $W_f$ in the fundamental chamber $m\ge|n|$, put $\alpha_1=a{+}b{+}c$, $\alpha_2=\alpha_1{+}m{+}n$, $\alpha_3=2a{+}n{+}f$, $\alpha_4=2b{+}m{+}f$, $\beta_1=2a{+}2b{+}m{+}n$, $\beta_2=a{+}b{+}c{+}n{+}f$, $\beta_3=a{+}b{+}c{+}m{+}f$, \begin{equation} t(z,f)=\frac{(-1)^z[z+1]!}{\prod_{i=1}^4[z-\alpha_i]!\prod_{i=1}^3[\beta_i-z]!},\qquad S_f=\sum_{z\in\mathbb Z}t(z,f), \end{equation} so that $W_f=P_0\,P(f)\,S_f$ with $P_0=\Delta(a,b,c)\Delta(b{+}m,a{+}n,c)$ independent of $f$ and $P(f)=\Delta(a,a{+}n,f)\Delta(b{+}m,b,f)$. Then $S_f=0$ unless $m\le f\le f_{\max}:=\min(2a{+}n,2b{+}m)$, and for every $f\ge\tfrac12$ with $f-m\in\mathbb Z$ \begin{equation} (x-B_f)\,S_f=\rho_+(f)\,S_{f+1}+\rho_-(f)\,S_{f-1}, \label{eq:q6j-R3} \end{equation} with, for $\mathrm s_z=\sinh(z/N)$ and $2A=2a{+}n{+}1$, $2B=2b{+}m{+}1$, \begin{equation} \rho_+(f)=-\frac{2\,\mathrm s_{f+1-n}\mathrm s_{f+1+n}\mathrm s_{f+1-m}\mathrm s_{f+1+m}}{\mathrm s_{2A}\mathrm s_{2B}\,\mathrm s_{2f+2}\mathrm s_{2f+1}},\qquad \rho_-(f)=-\frac{2\,\mathrm s_{2A-f}\mathrm s_{2A+f}\mathrm s_{2B-f}\mathrm s_{2B+f}}{\mathrm s_{2A}\mathrm s_{2B}\,\mathrm s_{2f}\mathrm s_{2f+1}}, \end{equation} and $B_f$ of \eqref{eq:q6j-AkBk}. For $m=n=0$ the row $f=0$ reads $(x-B_0)S_0=\rho_+(0)S_1$ with $B_0=\coth\mathsf X\coth\mathsf Y(1-1/\cosh\frac1N)$. \end{lemma} \begin{proof} \emph{Support.} For $ff_{\max}$ one of $\alpha_3,\alpha_4$ exceeds $\beta_1$; the sum is empty. \emph{Termwise relations.} In this appendix $q$ inside $q$-numbers, powers $q^z$ and the Kirillov--Reshetikhin sum denotes the deformation parameter $q_{\rm def}=e^{1/N}$ of Section~\ref{sec:q6j}, while $q_k$, $q_*$ and the argument of $p(q)$ remain the rank coordinate. Put $q=e^{1/N}$ and write each $q$-number as $[k]=(u-u^{-1})/(q-q^{-1})$ with $u=q^k$. Let \begin{equation} r(z)=\frac{(-1)^z[z+1]!}{[z{-}\alpha_1]![z{-}\alpha_2]![z{-}\alpha_3{+}1]![z{-}\alpha_4{+}1]![\beta_1{-}z]![\beta_2{+}1{-}z]![\beta_3{+}1{-}z]!} \end{equation} (with the same $f$). Since $[k]/[k]!=1/[k-1]!$ for every integer $k$ (both sides vanish for $k\le0$), \begin{equation}\begin{aligned} t(z,f{+}1)&=r(z)\,[z{-}\alpha_3{+}1][z{-}\alpha_3][z{-}\alpha_4{+}1][z{-}\alpha_4],\\ t(z,f)&=r(z)\,[\beta_2{+}1{-}z][\beta_3{+}1{-}z][z{-}\alpha_3{+}1][z{-}\alpha_4{+}1],\\ t(z,f{-}1)&=r(z)\,[\beta_2{+}1{-}z][\beta_2{-}z][\beta_3{+}1{-}z][\beta_3{-}z]. \end{aligned}\end{equation} Let $G(z)=r(z)\,[z{-}\alpha_1][z{-}\alpha_2][z{-}\alpha_3{+}1][z{-}\alpha_4{+}1]\,\mathrm{Pol}(q^z)$, where $\mathrm{Pol}(w)=p_{-2}w^{-2}+p_0+p_2w^2$ has coefficients to be fixed below. $G$ has finite support, and by the same rule $G(z{+}1)=-r(z)\,[z{+}2][\beta_1{-}z][\beta_2{+}1{-}z][\beta_3{+}1{-}z]\,\mathrm{Pol}(q^{z+1})$. Hence the telescoping identity \begin{equation} \rho_-t(z,f{-}1)+\rho_0\,t(z,f)+\rho_+t(z,f{+}1)=G(z{+}1)-G(z),\qquad \rho_0:=-(x-B_f), \label{eq:q6j-tele} \end{equation} is, after division by $r(z)$ (it is polynomial in the displayed $q$-numbers, so holds also where $r(z)=0$), the identity \begin{multline} \rho_-[\beta_2{+}1{-}z][\beta_2{-}z][\beta_3{+}1{-}z][\beta_3{-}z]+\rho_0[\beta_2{+}1{-}z][\beta_3{+}1{-}z][z{-}\alpha_3{+}1][z{-}\alpha_4{+}1]\\ +\rho_+[z{-}\alpha_3{+}1][z{-}\alpha_3][z{-}\alpha_4{+}1][z{-}\alpha_4]\\ =-[z{+}2][\beta_1{-}z][\beta_2{+}1{-}z][\beta_3{+}1{-}z]\mathrm{Pol}(q^{z+1})-[z{-}\alpha_1][z{-}\alpha_2][z{-}\alpha_3{+}1][z{-}\alpha_4{+}1]\mathrm{Pol}(q^z). \end{multline} Both sides are Laurent polynomials in $w=q^z$ whose coefficients are rational functions of $q$ and of $u_a=q^a,\ u_b=q^b,\ u_c=q^c,\ u_f=q^f,\ u_m=q^m,\ u_n=q^n$ ($\alpha_i,\beta_i$ are monomials in these, $x$ and $B_f$ rational in them). Linear algebra over $\mathbb Q(q,u_a,u_b,u_c,u_f,u_m,u_n)$ on the coefficients of $w^{-4},\dots,w^{4}$ determines $\rho_0$ (equal to $-(x-B_f)$, a check) and $p_{-2},p_0,p_2$ (explicit rational functions; their denominators are products of $\mathrm s_f,\mathrm s_{f+1},\mathrm s_{2f+1},\cosh\frac fN,\cosh\frac{f+1}N,\sinh\mathsf X,\sinh\mathsf Y$ and monomials, all strictly nonzero for $f>0$, so they are finite for every $f\ge\tfrac12$, with $S_{-1/2}=0$ by support when $m=f=\tfrac12$), and then \eqref{eq:q6j-tele} holds as an identity of rational functions in all eight symbols, verified by exact symbolic computation; it holds for every integer $z$ whenever the denominators do not vanish. Summing \eqref{eq:q6j-tele} over $z\in\mathbb Z$, the right side telescopes to $0$ because $G$ has finite support, which is \eqref{eq:q6j-R3} (with the sign convention $\rho_0=-(x-B_f)$). \emph{The row $f=0$ for $m=n=0$.} Here $\mathrm s_f=0$ is a pole of the certificate, and the row is checked directly: $S_0=t(a{+}b{+}c,0)$ has one term, and $S_1=t(z_0,1)+t(z_0{+}1,1)$, $z_0=a{+}b{+}c$, with $S_1/S_0=[b{+}c{-}a][a{+}c{-}b]-[a{+}b{+}c{+}2][a{+}b{-}c]$ (the factors $[b{+}c{-}a]$, $[a{+}b{-}c]$ vanish exactly when the corresponding term is outside the range). The identity $(x-B_0)=\rho_+(0)\,S_1/S_0$ is an identity of rational functions in $(q,q^a,q^b,q^c)$, verified by exact symbolic computation. \end{proof} \begin{lemma}[Closed seed] \label{lem:q6j-seed} For $m\ge|n|$ and $f=m$ the sum (K1) has exactly one term, and with $\Sigma_L:=\{-\frac{L-1}2,\dots,\frac{L-1}2\}$ (a symmetric multiset of $L$ shifts), \begin{equation} \begin{aligned} W_m^2={}&\frac{[2m]!}{[m+n]!\,[m-n]!}\;\prod_{j=-m}^{m}\frac1{[2A+j]\,[2B+j]}\\ &\times\prod_{\sigma\in\Sigma_{m+n}}[A{+}B{-}C{+}\sigma]\,[A{+}B{+}C{+}\sigma]\prod_{\sigma\in\Sigma_{m-n}}[B{+}C{-}A{+}\sigma]\,[A{-}B{+}C{+}\sigma], \end{aligned} \label{eq:q6j-seed} \end{equation} where $j$ runs over $-m,-m+1,\dots,m$ (the $2m+1$ values). \end{lemma} \begin{proof} For $f=m$ the triad sums are $\alpha=(a{+}b{+}c,\ a{+}b{+}c{+}m{+}n,\ 2a{+}n{+}m,\ 2b{+}2m)$ and the quadrilateral sums $\beta=(2a{+}2b{+}m{+}n,\ a{+}b{+}c{+}m{+}n,\ a{+}c{+}b{+}2m)$. So $\alpha_2=\beta_2=:z_0$ and $\max\alpha\le z\le\min\beta$ forces $z=z_0$. Write the squared one-term expression as a product of $q$-factorials. Collecting the exponents of the four $\Delta^2$ and of the square of the summand, the factorials $[a{+}b{-}c]!,[b{+}c{-}a]!,[a{-}b{+}c{+}n{-}m]!$ occur to the power $-1$, and $[a{-}b{+}c]!$, $[a{+}b{+}m{+}n{-}c]!$, $[b{-}a{+}m{-}n{+}c]!$, $[z_0{+}1]!$ to the power $+1$, $[a{+}b{+}c{+}1]!$ to the power $-1$; the triads containing $m$ contribute $[2m]!\,[m{-}n]!\,[m{+}n]!\,[2a{+}n{-}m]!\,[2b]!/([2a{+}n{+}m{+}1]!\,[2b{+}2m{+}1]!)$ and the square contributes $[m{+}n]!^{-2}[m{-}n]!^{-2}$. With $2A=2a+n+1$, $2B=2b+m+1$, \begin{equation} \frac{[2a{+}n{-}m]!}{[2a{+}n{+}m{+}1]!}=\prod_{j=-m}^{m}[2A+j]^{-1},\qquad \frac{[2b]!}{[2b{+}2m{+}1]!}=\prod_{j=-m}^{m}[2B+j]^{-1}, \end{equation} \begin{equation} \frac{[z_0{+}1]!}{[a{+}b{+}c{+}1]!}=\prod_{\sigma\in\Sigma_{m+n}}[A{+}B{+}C{+}\sigma],\quad \frac{[a{+}b{+}m{+}n{-}c]!}{[a{+}b{-}c]!}=\prod_{\sigma\in\Sigma_{m+n}}[A{+}B{-}C{+}\sigma], \end{equation} \begin{equation} \frac{[b{-}a{+}m{-}n{+}c]!}{[b{+}c{-}a]!}=\prod_{\sigma\in\Sigma_{m-n}}[B{+}C{-}A{+}\sigma],\quad \frac{[a{-}b{+}c]!}{[a{-}b{+}c{+}n{-}m]!}=\prod_{\sigma\in\Sigma_{m-n}}[A{-}B{+}C{+}\sigma], \end{equation} where the shifts are matched by \begin{equation} \begin{gathered} A{+}B{-}C=a{+}b{-}c+\tfrac{m+n+1}2,\qquad A{+}B{+}C=a{+}b{+}c+\tfrac{m+n+3}2,\\ B{+}C{-}A=b{+}c{-}a+\tfrac{m-n+1}2,\qquad A{-}B{+}C=a{-}b{+}c-\tfrac{m-n-1}2 . \end{gathered} \end{equation} The remaining prefactor is $[2m]!/([m{+}n]!\,[m{-}n]!)$. (The identity is also verified in exact rational arithmetic for integer and half-integer families.) \end{proof} \paragraph{Proof of Proposition~\ref{prop:q6j-class}.} \emph{Symmetry reduction to $m\ge|n|$.} In (K1), swapping $(j_1,j_3)\leftrightarrow(j_2,j)$ leaves the triads and quadrilateral sums invariant and sends $(a,b,m,n)\mapsto(b,a,n,m)$ with $(A,B,\mathsf X,\mathsf Y)\leftrightarrow(B,A,\mathsf Y,\mathsf X)$; composing this with the row flip $(j_1,j_2)\leftrightarrow(j_3,j)$ sends $(a,b,m,n)\mapsto(a+n,b+m,-m,-n)$ at fixed $(A,B,C)$. These two involutions, together with their composition (the row flip $(j_1,j_2)\leftrightarrow(j_3,j)$ sending $(a,b,m,n)\mapsto(b+m,a+n,-n,-m)$ with $(A,B,\mathsf X,\mathsf Y)\leftrightarrow(B,A,\mathsf Y,\mathsf X)$), map every admissible pair $(m,n)$ into the fundamental chamber $m\ge|n|$, leave $x,\mu,\Delta_0$, $f_{\min}=\max(|m|,|n|)$, $f_{\max}=\min(2A-1,2B-1)$, and all coefficients and class data of (a)--(d) invariant, and multiply both $\widehat\Psi_k$ of \eqref{eq:q6j-Psi} and $\widehat d_k(\theta)=\sqrt{2k+1}\,d^k_{mn}(\theta)$ by the same $k$-independent sign ($+1$ for the row flip and $(-1)^{m-n}$ for the two generators, using $d^k_{mn}=(-1)^{m-n}d^k_{nm}=(-1)^{m-n}d^k_{-m,-n}=d^k_{-n,-m}$). It therefore suffices to prove (a)--(d) in the fundamental chamber $m\ge|n|$, where $f_{\min}=m$. \emph{(a)} By \eqref{eq:q6j-Psi}, \begin{equation} \widehat\Psi_f=K_0(-1)^{f-m}\sqrt{[2f+1]}\,P(f)\,S_f,\qquad K_0=(-1)^{a+b+c+2m}\sqrt{[2A][2B]}\,P_0\ \text{independent of }f. \end{equation} Multiply \eqref{eq:q6j-R3} by $K_0(-1)^{f-m}\sqrt{[2f+1]}P(f)$. For $m\le f0$ for $f>m\ge0$; the $q$-factorials in $P$ have non-negative arguments on $[m,f_{\max}]$; $\mathrm s_{2A\pm k},\mathrm s_{2B\pm k}>0$ for $k\le f_{\max}$ except at $k=f_{\max}+1$, which is not needed; and $\mathrm s_{k\pm m},\mathrm s_{k\pm n}>0$ for $k>m$), so it suffices that their squares agree. From $P(f)^2=[2a{+}n{-}f]!\allowbreak[2b{+}m{-}f]!\allowbreak[f{-}n]!\allowbreak[f{+}n]!\allowbreak[f{-}m]!\allowbreak[f{+}m]!\allowbreak/\bigl([2a{+}n{+}f{+}1]!\allowbreak[2b{+}m{+}f{+}1]!\bigr)$, \begin{equation}\begin{gathered} \frac{P(f{+}1)^2}{P(f)^2}=\frac{[f{+}1{-}n][f{+}1{+}n][f{+}1{-}m][f{+}1{+}m]}{[2A{-}f{-}1][2B{-}f{-}1][2A{+}f{+}1][2B{+}f{+}1]},\\ \frac{P(f{-}1)^2}{P(f)^2}=\frac{[2A{-}f][2B{-}f][2A{+}f][2B{+}f]}{[f{-}n][f{+}n][f{-}m][f{+}m]}, \end{gathered}\end{equation} and $1-\mathrm s_k^2/\sinh^2\mathsf X=\mathrm s_{2A-k}\mathrm s_{2A+k}/\mathrm s_{2A}^2$ (similarly for $\mathsf Y$), so the factors $\mathrm s_{2A\mp k}$ cancel against $P(f\pm1)^2/P(f)^2$ and $\rho_+(f)^2=A_{f+1}^2\frac{[2f+3]}{[2f+1]}\frac{P(f+1)^2}{P(f)^2}$, $\rho_-(f)^2=A_f^2\frac{[2f-1]}{[2f+1]}\frac{P(f-1)^2}{P(f)^2}$, an identity of rational functions (verified symbolically). At $f=m\ge\tfrac12$ the term $A_m\widehat\Psi_{m-1}$ is absent ($A_m=0$ and $S_{m-1}=0$ by Lemma~\ref{lem:q6j-recurrence}), and at $f=f_{\max}$ the term $A_{f_{\max}+1}\widehat\Psi_{f_{\max}+1}$ is absent ($A_{f_{\max}+1}=0$ and $S_{f_{\max}+1}=0$); $W_f=0$ outside $[m,f_{\max}]$ by the triangle conditions. This gives \eqref{eq:q6j-AkBk}. For $m=n=0$ the row $k=0$ is row $f=0$ of Lemma~\ref{lem:q6j-recurrence}, with $\widehat\Psi_0^2=[1][2A][2B]W_0^2=1$ by \eqref{eq:q6j-seed}. \emph{(b)} \emph{Evenness and analyticity of $e$.} With $K=hk$, $H=h$: $\sinh h(k-n)\sinh h(k+n)=\Xi(K,Hn)(K^2-H^2n^2)$, $\sinh^22hk=4K^2\,\Xi(2K,0)$ and $\mathrm s_{2k+1}\mathrm s_{2k-1}=\Xi(2K,H)(4K^2-H^2)$, so that $A_k^2=\frac{(k^2-m^2)(k^2-n^2)}{k^2(4k^2-1)}\,e(K^2,H^2)^2=A_k^{(0)2}e^2$ with the stated $\Xi$-quotient $e^2$. $\Xi(u,w)$ is analytic in $(u^2,w^2)$ with $\Xi(0,0)=1$, so $e^2$ is jointly analytic, equals $1$ at $(0,0)$, and is positive on $[0,s_1]\times\{0\}$ iff $\sinh K<\sinh\mathsf X,\sinh\mathsf Y$, that is $s_1<4\min(\alpha,\beta)^2$. At $t=0$, $\Xi(K,0)=\sinh^2K/K^2$ and $\Xi(2K,0)=\sinh^2K\cosh^2K/K^2$, so $e(s)^2=(1-\sinh^2K/\sinh^2\mathsf X)(1-\sinh^2K/\sinh^2\mathsf Y)/\cosh^4K=(1-\tanh^2K\coth^2\mathsf X)(1-\tanh^2K\coth^2\mathsf Y)$. At $K=0$, $\Xi(0,w)=\sinh^2w/w^2$ gives $e(0,t)=\lambda(t)$. Differentiating $\log\Xi(K,Hn)=\log\bigl((\sinh^2K-n^2t)/(K^2-n^2t)\bigr)+O(t^2)$ and $\log\Xi(2K,H)=\log\bigl((\sinh^22K-t)/(4K^2-t)\bigr)+O(t^2)$ in $t$ at $t=0$ gives \begin{equation} \eta(K^2)=\tfrac12\Bigl[(m^2+n^2)\Bigl(\frac1{K^2}-\frac1{\sinh^2K}\Bigr)+\frac1{\sinh^22K}-\frac1{4K^2}\Bigr], \end{equation} which is \eqref{eq:q6j-data}, and $\eta(0)=\frac12[\frac{m^2+n^2}3-\frac13]=\lambda_1$. \emph{The split of $B$.} With $U=h(k+\frac12)$, $B_k=\coth\mathsf X\coth\mathsf Y\bigl(1-\frac{2\cosh Hm\cosh Hn}{\cosh2U+\cosh H}\bigr)+\frac{2\sinh Hm\sinh Hn}{\cosh2U-\cosh H}$. The first term is analytic in $(U^2,H^2)$. In the second, $\cosh2U-\cosh H=w\varphi$ with $w=U^2-H^2/4$ and $\varphi$ analytic, $\varphi(0,0)=2$, $\varphi|_{w=0}=2\sinh H/H$; therefore $2/(w\varphi)-(H/\sinh H)/w=(2-(H/\sinh H)\varphi)/(w\varphi)$ has a numerator divisible by $w$ and is analytic. Since $B_k^{(0)}=mnH^2/w$ and $\lambda(t)=\sinh Hm\sinh Hn/(mnH\sinh H)$, $B_k-\lambda(h^2)B_k^{(0)}=v(q_k^2,h^2)$ is analytic, jointly with $e$. Expanding $v(0,t)$ to first order gives $\beta_2$, and $\partial_sv(0,0)=\beta_1$, $v(s,0)=\coth\mathsf X\coth\mathsf Y\tanh^2\sqrt s$ gives $b_0$. Without $\lambda$, the coefficient of $h^4/q^2$ in $B_k-B_k^{(0)}$ is $mn(m^2+n^2-1)/6$, nonzero when $mn\ne0$. The identities \eqref{eq:q6j-D}: with $\Delta_0=\sin^2\theta+\mu^2$ (hyperbolic identity $\coth^2\mathsf X+\coth^2\mathsf Y-2x\coth\mathsf X\coth\mathsf Y=1-x^2+\mu^2$), $e^2-(x-b_0)^2=\sin^2\theta-\tanh^2q\,\Delta_0=(\sin^2\theta-\mu^2\sinh^2q)/\cosh^2q$. These are identities between rational functions of $e^{\mathsf X},e^{\mathsf Y},e^{\mathsf Z},e^{q}$, verified by exact symbolic computation. \emph{(c)} \emph{Seed (E2).} Put $\ell_\pm:=\frac12(\mathsf X+\mathsf Y)\pm\frac12\mathsf Z$ and $\ell'_\pm:=\frac12(\mathsf Z\pm(\mathsf Y-\mathsf X))$, and note $[A{+}B{\mp}C{+}\sigma]=\sinh(\ell_\mp+h\sigma)/\sinh h$, $[B{+}C{-}A{+}\sigma]=\sinh(\ell'_++h\sigma)/\sinh h$, $[A{-}B{+}C{+}\sigma]=\sinh(\ell'_-+h\sigma)/\sinh h$ and $[2A+j]=\sinh(\mathsf X+hj)/\sinh h$. In \eqref{eq:q6j-seed} the powers of $\sinh h$ cancel against $[2A][2B]$, so \begin{equation} \begin{aligned} \widehat\Psi_m^2={}&[2m+1]\,\frac{[2m]!}{[m{+}n]![m{-}n]!}\,\frac{\sinh\mathsf X\sinh\mathsf Y}{\prod_{j=-m}^{m}\sinh(\mathsf X{+}hj)\sinh(\mathsf Y{+}hj)}\\ &\times\prod_{\Sigma_{m+n}}\sinh(\ell_-{+}h\sigma)\sinh(\ell_+{+}h\sigma)\prod_{\Sigma_{m-n}}\sinh(\ell'_-{+}h\sigma)\sinh(\ell'_+{+}h\sigma), \end{aligned} \label{eq:q6j-seed-sinh} \end{equation} Each product over a symmetric multiset is even in $h$, the $q$-binomial and $[2m+1]$ are even analytic functions of $h$ with values $\binom{2m}{m+n}$ and $2m+1$ at $h=0$, and by the symmetry of $\Sigma_L$ the factors pair into $\sinh(\ell_-{+}h\sigma)\sinh(\ell_+{+}h\sigma)=\frac12[\cosh(\mathsf X+\mathsf Y+2h\sigma)-\cosh\mathsf Z]$ and (re-pairing $\sigma$ with $-\sigma$) $\sinh(\frac{\mathsf Z+\mathsf Y-\mathsf X}2+h\sigma)\sinh(\frac{\mathsf Z+\mathsf X-\mathsf Y}2-h\sigma)=\frac12[\cosh\mathsf Z-\cosh(\mathsf X-\mathsf Y+2h\sigma)]$. Thus $\widehat\Psi_m^2=G(x;h^2)$, where $G$ is a polynomial of degree $2m$ in $\cosh\mathsf Z$, hence in $x=(\cosh\mathsf X\cosh\mathsf Y-\cosh\mathsf Z)/(\sinh\mathsf X\sinh\mathsf Y)$, with coefficients analytic in $t=h^2$ and real zeros at $x=-1+O(h)$ (from $\Sigma_{m+n}$) and $x=1+O(h)$ (from $\Sigma_{m-n}$); in particular $G$ is jointly analytic in $(x,t)$. At $h=0$ the hyperbolic half-angle formulas $\cos^2\frac\theta2=\sinh s\sinh(s-\mathsf Z)/(\sinh\mathsf X\sinh\mathsf Y)$, $\sin^2\frac\theta2=\sinh(s-\mathsf X)\sinh(s-\mathsf Y)/(\sinh\mathsf X\sinh\mathsf Y)$, $s=\frac12(\mathsf X+\mathsf Y+\mathsf Z)$, give $G(x;0)=(2m+1)\binom{2m}{m+n}\cos^{2(m+n)}\frac\theta2\sin^{2(m-n)}\frac\theta2=\widehat d_m(\theta)^2$. Hence $\widehat\Psi_m=(-1)^{m-n}\sqrt{G(x;h^2)}$ (the sign being $(-1)^{a+b+c+2m}(-1)^{z_0}=(-1)^{m-n}$, from \eqref{eq:q6j-Psi} and $\operatorname{sgn}W_m=(-1)^{z_0}$; it is the sign of $d^m_{mn}$) is analytic, bounded, and nonvanishing in $(x,t)$ on $U\times\{|t|\le t_0\}$, for a complex neighbourhood $U$ of $X$ (because $G>0$ on $X\times\{0\}$, where it equals $\widehat d_m^{\,2}$), with value $\widehat d_m$ at $t=0$. By (a) at $k=m$, where $A_m=0$ ($\mathrm s_{k-m}=0$), $\widehat\Psi_{m+1}=(x-B_m)\widehat\Psi_m/A_{m+1}$, jointly analytic because $A_{m+1}=A_{m+1}^{(0)}e((m+1)^2t,t)$ is analytic and nonzero, with value $\widehat d_{m+1}$ at $t=0$ by the Wigner recurrence; the exceptional case $m=n=0$ is the row $k=0$, with $\widehat\Psi_0=1$ and $\widehat\Psi_1=(x-B_0)/A_1$ affine in $x$. This gives (E2) (and polynomiality in $x$ for Theorem~\ref{thm:general-bulk-zeros}). \emph{(R).} At a simple turn (R) holds intrinsically by Proposition~\ref{prop:exact-hyperbolicity}; independently, by (K2), $\widehat\Psi_k^2=[2k+1][2A][2B]W_k^2\le[2A][2B]/[2c+1]=\sinh\mathsf X\sinh\mathsf Y/(\sinh\frac1N\sinh\mathsf Z)=O(N)$ for all $k$. \emph{(d)} Using $\tanh^2q_*=\sin^2\theta/\Delta_0$ (from $D(q_*)=0$) in $\varepsilon_*=x-\tanh^2q_*\coth\mathsf X\coth\mathsf Y$ gives $\varepsilon_*=[x\Delta_0-(1-x^2)\coth\mathsf X\coth\mathsf Y]/\Delta_0=(x\coth\mathsf X-\coth\mathsf Y)(\coth\mathsf X-x\coth\mathsf Y)/\Delta_0$ with $\Delta_0>0$, and $x\coth\mathsf X-\coth\mathsf Y=(\cosh\mathsf Y-\cosh\mathsf X\cosh\mathsf Z)/(\sinh^2\mathsf X\sinh\mathsf Y)$, $\coth\mathsf X-x\coth\mathsf Y=(\cosh\mathsf Y\cosh\mathsf Z-\cosh\mathsf X)/(\sinh^2\mathsf Y\sinh\mathsf X)$. By the hyperbolic sine rule $\sinh q_*=\sinh\mathsf Y\sin A_{\mathsf X}=\sinh\mathsf X\sin A_{\mathsf Y}$ ($A_{\mathsf X}$ the angle opposite $\mathsf X$), so $q_*<\min(\mathsf X,\mathsf Y)$ unless $A_{\mathsf X}$ or $A_{\mathsf Y}=\pi/2$, i.e.\ $\cosh\mathsf X=\cosh\mathsf Y\cosh\mathsf Z$ or $\cosh\mathsf Y=\cosh\mathsf X\cosh\mathsf Z$ (hyperbolic Pythagoras), which is $\varepsilon_*=0$. Otherwise $e(q_*^2)>0$, $u(q_*)=\varepsilon_*/e=\tau$, and \eqref{eq:q6j-D} gives the simple turn. The integers: $K_+=C-|A-B|+\frac{|m-n|-1}2$ equals $c+b-a+\max(m-n,0)$ if $A\ge B$ and $c+a-b+\max(n-m,0)$ if $A0$ on $[0,q_*]$ and $e\sin p=\sqrt D$), so it may be differentiated under the integral sign, with the endpoint term $p(q_*)\partial_{\mathsf Z}q_*$. In $\cos p=(x-b_0)/e$ only $x$ depends on $\mathsf Z$, with $\partial_{\mathsf Z}x=-\mu$, so $\partial_{\mathsf Z}p=\mu/(e\sin p)=\mu/\sqrt D$ and \begin{equation} \int_0^{q_*}\partial_{\mathsf Z}p\,\dd q=\int_0^{q_*}\frac{\mu\cosh q\,\dd q}{\sqrt{\sin^2\theta-\mu^2\sinh^2q}}=\arcsin\frac{\mu\sinh q}{\sin\theta}\Big|_0^{q_*}=\frac\pi2 . \end{equation} At $q_*$, $\cos p=\varepsilon_*/e=\tau$, so $p(q_*)=0$ for $\tau=+1$ and $\pi$ for $\tau=-1$. Hence $\partial_{\mathsf Z}S_*=\pi/2$ ($\tau=+1$) and $\partial_{\mathsf Z}(S_*-\pi q_*)=\pi/2$ ($\tau=-1$). By \eqref{eq:q6j-tau}: for $\mathsf X>\mathsf Y$ the second factor is negative, so $\tau=+1$ exactly for $\cosh\mathsf Z<\cosh\mathsf X/\cosh\mathsf Y=:\cosh\mathsf Z_c$, i.e.\ on $(\mathsf X-\mathsf Y,\mathsf Z_c)$ (note $\mathsf Z_c>\mathsf X-\mathsf Y$), and $\tau=-1$ on $(\mathsf Z_c,\mathsf X+\mathsf Y)$; for $\mathsf X=\mathsf Y$, $\tau=-1$ throughout; $\mathsf X<\mathsf Y$ is symmetric. As $\mathsf Z\downarrow|\mathsf X-\mathsf Y|$ or $\mathsf Z\uparrow\mathsf X+\mathsf Y$ the triangle degenerates, $\sin\theta\to0$, and $\sinh q_*=\sinh\mathsf Y\sin A_{\mathsf X}\to0$ while $p$ stays bounded, so $S_*\to0$ and $q_*\to0$. Integrating $\partial_{\mathsf Z}$ from the degenerate end of each interval gives $S_*=\frac\pi2(\mathsf Z-|\mathsf X-\mathsf Y|)$ on the first and $S_*-\pi q_*=-\frac\pi2(\mathsf X+\mathsf Y-\mathsf Z)$ on the second. With $\mathsf X=2hA$ etc.\ and $N=1/h$ this is $NS(q_*)/\pi=C-|A-B|$, resp.\ $NS(q_*)/\pi-Nq_*=-(A+B-C)$.\hfill$\square$ \paragraph{Proof of Corollary~\ref{cor:q6j-consequences}.} (i), (ii), (iv), (v) are Theorems~\ref{thm:general-mesoscopic}, \ref{thm:general-critical}, \ref{thm:general-second-order}, \ref{thm:general-bulk-zeros}, \ref{thm:general-rank-airy} and Propositions~\ref{prop:exact-hyperbolicity}--\ref{prop:forbidden-decay} applied to the chain of Proposition~\ref{prop:q6j-class}; $\widehat\Psi_k\to\widehat d_k$ at $h\to0$ because $\widehat\Psi_{f_{\min}}$ has the sign of $d^{f_{\min}}_{mn}$, so Theorem~\ref{thm:general-rank-airy} carries the Darboux sign $\epsilon_{mn}=(-1)^{\max(n-m,0)}$; (R) is Proposition~\ref{prop:q6j-class}(c); the Airy amplitude is Proposition~\ref{prop:airy2-amplitude} with (iii). In (v), $\partial_qD(q_*)=-2e_*^2u_1$ gives $u_1$ from \eqref{eq:q6j-D}. For (vi), under $N\to N/\varsigma$ and $q=\varsigma\rho$ at step $h=N^{-1}$, the (E1) functions in the $\rho$-variable are $e(\varsigma^2s,\varsigma^2t)$ and $v(\varsigma^2s,\varsigma^2t)$, which multiplies $\varepsilon_1,\beta_1,\beta_2,\chi,\lambda_1,\eta,\delta_{\rm fl}$ by $\varsigma^2$ and $\varpi_\tau$ by $\varsigma^{4/3}$ (since $(\varsigma/N)^{4/3}\varpi_\tau=N^{-4/3}(\varsigma^{4/3}\varpi_\tau)$ while $(\varsigma/N)^{-2/3}\zeta_\tau(\varsigma\rho)=N^{2/3}\zeta_{\tau,\rho}(\rho)$); expanding $\coth(2\varsigma\alpha)=(2\varsigma\alpha)^{-1}(1+O(\varsigma^2))$, $\sinh(\varsigma\rho)=\varsigma\rho(1+O(\varsigma^2))$, and $\mu\sinh(\varsigma\rho)=\gamma\rho/(2\alpha\beta)+O(\varsigma^2)$ gives the Racah limits. (iii) \emph{The flux defect, from the closed seed}, by exact symbolic computation. By the symmetry reduction in Proposition~\ref{prop:q6j-class} it suffices to take $m\ge|n|$, since $\widehat\Psi_k/\widehat U_k^{[2]}$ and $\mathsf J_2(0,x)$ are invariant under both involutions. By Theorem~\ref{thm:general-phase-exactness} and the contact identity, the coefficients of $t$ in $\widehat\Psi_k$ and $\widehat U_k^{[2]}$ differ by $c(x)\widehat d_k$ for every $k$, so at $k=m$ (both start from $\widehat d_m$ at $t=0$) \begin{equation} c(x)=[t]\log\widehat\Psi_m-[t]\log\widehat U^{[2]}_m=\tfrac12\mathsf Q-\nu_m . \end{equation} \emph{Seed side.} From \eqref{eq:q6j-seed-sinh}, with $\log[z]=\log z+\frac{(z^2-1)t}6+O(t^2)$ and, for a symmetric multiset, $\sum_{\sigma\in\Sigma_L}\log\sinh(\ell+h\sigma)=L\log\sinh\ell-\frac t2\frac{L(L^2-1)}{12}\frac1{\sinh^2\ell}+O(t^2)$, \begin{equation} \begin{aligned} \mathsf Q=[t]\log\widehat\Psi_m^2={}&\mathsf f(2m{+}1)-\mathsf f(m{+}n)-\mathsf f(m{-}n)\\ &-\tfrac12\Bigl[\mathsf T(m{+}n)\bigl(\tfrac1{\sinh^2\ell_-}+\tfrac1{\sinh^2\ell_+}\bigr)+\mathsf T(m{-}n)\bigl(\tfrac1{\sinh^2\ell'_-}+\tfrac1{\sinh^2\ell'_+}\bigr)\\ &\qquad\ -\mathsf T(2m{+}1)\bigl(\tfrac1{\sinh^2\mathsf X}+\tfrac1{\sinh^2\mathsf Y}\bigr)\Bigr], \end{aligned} \end{equation} $\mathsf f(N)=\frac{N(N+1)(2N+1)}{36}-\frac N6=\sum_{j\le N}\frac{j^2-1}6$, $\mathsf T(N)=\frac{N(N^2-1)}{12}$, \emph{Pullback side.} By the proof of Proposition~\ref{prop:general-fixed-rank-contact}, $\widehat U^{[2]}_m=\widehat d_m+t\,\widehat u_m+O(t^2)$ with $\widehat u_m=\sqrt{2m+1}\,\mathcal Q_{2,m}d_m$, so that, with $\kappa=m+\frac12$, $s=1-x^2$, $\partial_x\log d_m=-(mx-n)/s$ ($d_m\propto\cos^{m+n}\frac\theta2\sin^{m-n}\frac\theta2$) and the data \eqref{eq:q6j-data}, \eqref{eq:general-rho20}, \begin{equation} \nu_m=\Bigl(\frac{\kappa^2\chi}3-\sqrt s\,\rho_2(0,x)\Bigr)\partial_x\log d_m+\frac{\kappa^2((\varepsilon_1-x\beta_1)/2-x\chi/6)}{s}+\frac{x\,\rho_2(0,x)}{2\sqrt s}. \end{equation} \emph{The identity.} In the variables $u=e^{\mathsf X/2}$, $v=e^{\mathsf Y/2}$, $w=e^{\mathsf Z/2}$, every quantity above is a rational function of $(u,v,w)$ with coefficients polynomial in $(m,n)$ ($x$, $\coth\mathsf X$, $\coth\mathsf Y$, $\sinh\ell_\pm$, $\sinh\ell'_\pm$ and $\sqrt s\,\rho_2(0,x)$ are rational, the square root cancelling in $\sqrt s\rho_2$ and in $x\rho_2/\sqrt s$), and \begin{equation} \tfrac12\mathsf Q-\nu_m+\mathsf J_2(0,x)+\tfrac1{12}=0 \end{equation} is an identity of rational functions of $(u,v,w)$ with symbolic $(m,n)$: its numerator, a polynomial in $(u,v,w,m,n)$, expands to zero in exact arithmetic. This gives $c(x)=-\mathsf J_2(0,x)-\frac1{12}$, i.e.\ \eqref{eq:q6j-c} (in the scalar channel $m=n=0$, where $\widehat\Psi_0=1$ and $\mathsf Q=0$, all $\varepsilon_1,\beta_1,\beta_2$ terms in $-\nu_0+\mathsf J_2(0,x)$ cancel directly to $\frac12\lambda_1=-\frac1{12}$). The relation to the Gauss weight is Remark~\ref{rem:flux-gauss}.\hfill$\square$ \section{Verification and a computable certificate} \label{sec:supp-verification} \label{sec:supp-certificate} \label{cor:jacobi-aposteriori-certificate} The proofs are analytic; none rests on numerical fitting. Exact computer-algebra checks reproduce the identities used in the proofs, and high-precision numerical checks illustrate the stated rates against exact $3j$ and $6j$ symbols (Section~\ref{sec:numerics}); the latter enter no proof. For a self-adjoint Jacobi recurrence $a_{k+1}Y_{k+1}+a_kY_{k-1}=b_kY_k$ the Casoratian $a_k(Y_kW_{k-1}-Y_{k-1}W_k)$ of two solutions is constant, so a fundamental matrix $F_k$ has $\det F_k=\mathcal C/a_k$ and $\|F_k^{-1}\|_2=a_k\|F_k\|_2/|\mathcal C|$. Variation of constants then bounds the error of any approximation $\widetilde Z_{k+1}=T_k\widetilde Z_k+\rho_k$ by \begin{equation} \|Z_K-\widetilde Z_K\|_2\le\frac{\|F_K\|_2}{|\mathcal C|}\Bigl[a_{k_0}\|F_{k_0}\|_2\|Z_{k_0}-\widetilde Z_{k_0}\|_2+\sum_{k=k_0}^{K-1}a_{k+1}\|F_{k+1}\|_2\|\rho_k\|_2\Bigr], \label{eq:jacobi-aposteriori-certificate} \end{equation} a bound in finite exact-recurrence quantities that can be evaluated in interval arithmetic; it is not used in the proofs.