Shots Are Not Enough: Mask Coverage and Shot Precision in Adversarial Evaluation of Stochastic Quantum Neural Networks

Stochastic quantum neural networks (QNNs) couple finite circuit measurements with random gate masks, so adversarial evaluation must resolve both measurement precision and stochastic-support coverage. We study this interaction in a frozen five-seed ideal-circuit evaluation spanning two datasets, two QNN architectures, three budgets ($B\in\{128,512,2048\}$), and three mask--shot allocations. Larger budgets modestly strengthen attacks, whereas fixed-budget reallocations have small, directionally inconsistent effects. Under matched scientific settings, the validation-selected SPSA configuration is weaker than parameter shift in all 18 predeclared strata per dataset; observed physical costs differ by estimator. An exact 512-mask oracle for the one-qubit model and quantitatively certified operational-$K$ references for the two-qubit model separate shot error from incomplete stochastic coverage. At $B2048$, all twelve calibration rows pass the frozen downstream loss, success, and denominator tolerances; gradient-fidelity shortfalls determine the remaining six-clause certification boundary. A checkpoint-common re-score of all 69,120 saved two-qubit endpoints removes native denominator variation and preserves the headline conclusions across three deterministic reference banks. Thus mask coverage, shot precision, estimator structure, and physical execution cost are distinct axes of stochastic-QNN robustness evaluation.