arXiv:2611.13928v1 Announce Type: new Abstract: In this paper, we study sparse multiple testing under a Gaussian two-group model with known one-factor covariance and centered Gaussian signals. We derive a first-order risk expansion for the observed-data Bayes Oracle, including a finite-dimensional loading average for heterogeneous loadings and an explicit limiting constant for homogeneous loadings. Under the stated factor-estimation and signal-calibration conditions, the homogeneous one-factor Oracle has strictly smaller leading risk than its independent-coordinate counterpart; observing the latent factor yields no further first-order reduction. We establish asymptotic Bayes optimality under sparsity (ABOS) for the Maximum Residual Down (MRD) step-down procedure with Gavrilov--Benjamini--Sarkar (GBS) stagewise calibration (MRD--GBS). Uniform control of residual errors along the rejection path permits comparison with a fixed factor-known rule without requiring the rejection paths to coincide. Fixed-residual versions of GBS and Benjamini--Hochberg (BH) attain the same leading risk. We also prove ABOS for the Bayesian Step-Down (BSD) rule when supplied with the true mixture, signal variance, covariance, and loss parameters; this is an Oracle-parameterized result. When the expected signal count grows faster than the logarithm of the reciprocal signal proportion, BSD and the observed-data Oracle disagree on several coordinates negligible in probability relative to the expected signal count, explaining BSD's near-Oracle behavior in companion simulations. The results cover diverging and bounded positive expected signal counts, including every fixed polynomial sparsity rate up to the bounded-count regime, under the stated calibrations. A fixed-dimension simulation illustrates absolute-risk behavior but does not establish ABOS.