\page \sect{Sharp resolution when the error waveform is unknown} Fix \(0<\alpha<1/2\), a fixed graph with \(d\ge2\), a local signal window \(0<|\theta|\le\theta_*<\pi/4\), and \(\theta=\epsilon L\). Use exactly \(n\) photons launched individually in the same input state, or a number state in that input mode, with unit collection efficiency. The null permits \(\theta=0\) and any admissible waveform of action at most \(a\). The alternative permits \(r\le|\theta|\le\theta_*\) and its own admissible waveform. Define \(r_*(a,n)\) as the infimum separation \(r\) for which a terminal test can keep both worst-case errors at most \(\alpha\). Common sideband evolution may also be unknown. \thm{Theorem 4}{Local minimax separation} In the joint regime \(a\to0,\ n\to\infty\), \begin{equation} r_*(a,n)=\Theta_{d,\alpha,\theta_*}\!\left(a^d+n^{-1/2}\right). \label{eq:minimax} \end{equation} The corresponding coupling resolution is \(r_*/L\). This result is uniform over admissible waveforms with no stochastic calibration; it concerns a single terminal exposure and a specified product probe. The dark offset and photon fluctuations impose independently necessary scales. \prooftext Let \(t_n=\sqrt{2\log(2/\alpha)/n}\). Reject the null if \(|(N_+-N_-)/n|> b_d+t_n\). At the null Hoeffding gives a false-positive probability at most \(\alpha\). On the alternative, the absolute conditional mean is at least \(\sin(2|\theta|)-2a^2|\theta|-b_d\). Since this sine is bounded below by \(2|\theta|\cos(2\theta_*)\), a sufficient condition for the false-negative probability to be at most \(\alpha\) is \[ |\theta|\ \ge\ \frac{b_d+t_n}{\cos(2\theta_*)-a^2}, \qquad a^2<\cos(2\theta_*). \] The same bound follows after arbitrary mixing of the admissible waveforms. For necessity, Theorem 2 gives an exact overlap of the two composite hypotheses for every separation through \(\theta_d(a/C_d)\). Independently, take \(V=K=0\) in both hypotheses. The two pure \(n\)-photon states have overlap \(\cos^n\theta\); diagonalizing their difference on its two-dimensional span gives trace distance \(\sqrt{1-\cos^{2n}\theta}\le\sqrt n\,|\theta|\). Every test with both errors at most \(\alpha\) needs outcome total variation at least \(1-2\alpha\). These two lower bounds give \eqref{eq:minimax} since the sum and maximum agree within a factor two. \fin \thm{Corollary 5}{Unequal edge tolerances} Suppose edge \(e\) has its own action bound \(c_e\rho^{w_e}\) with \(c_e,w_e>0\), \(0<\rho<1\), and independently controllable signed phases. The small-\(\rho\) ambiguity exponent is the \emph{weighted} shortest-path distance \(D=\min_{\pi:u\to v}\sum_{e\in\pi} w_e\). In the same testing experiment, \(r_*=\Theta(\rho^D+n^{-1/2})\) for fixed graph and budgets. \prooftext Divide each edge operator by \(\rho^{w_e}\). Any nonzero endpoint word carries a factor at most \(\rho^D\). Bounding the remaining Dyson series by \(\exp(\sum_e c_e)\) gives a uniform null-amplitude bound \(O(\rho^D)\). On a minimizing path use \eqref{eq:pulse} recursively with each newly introduced edge angle \(h_e=c'_e\rho^{w_e}\). Each edge is used a finite number of times depending only on the path, with compensation angles at most \(|h_e|/2\). Choose fixed \(c'_e>0\) small enough to stay within each budget. The synthesized angle is asymptotic to \(\prod_{e\in\pi} h_e\). The response bound in Theorem 1 applies when there is no direct edge. If a direct edge is present, Duhamel with one insertion instead gives \(|D'_V-D'_0|\le4La\), with \(a\le\sum c_e\rho^{w_e}\); the slope is still bounded away from zero locally. Use the same testing proof. \fin A four-edge path with all \(w_e=1\) has exponent four. A parasitic direct endpoint link of order \(\rho^2\) changes the exponent to two, even when its absolute coupling is much smaller than any adjacent coupling. This makes a bound on the shortcut a meaningful fabrication specification.