GPS-denied unmanned aerial vehicles require more than a nominally accurate state estimator: the estimator must preserve observability, reject numerically unsafe updates, expose uncertainty, and remain reproducible across restarts and software revisions. This paper develops a complete mathematical and systems analysis of a consistency-aware visual-inertial estimation architecture derived from the current v2.0.0 software research release of an open GPS-denied UAV platform. The estimator uses a 15-dimensional error state with midpoint inertial propagation, explicit process-noise discretization, first-estimates Jacobians (FEJ), a bounded multi-state constraint Kalman filter (MSCKF) camera window, multi-view triangulation, SVD-based feature-state elimination, normalized-innovation chi-square gating, Joseph-form covariance correction, zero-velocity updates, and fail-closed transactional commit semantics. A new spectral geometry diagnostic is introduced to separate low image residual from informative triangulation geometry. In a reproducible three-view Monte Carlo study using the implementation noise level of 1.5 pixels, a 0.10 m camera span can satisfy the existing reprojection gate while accepted landmarks still exhibit 1.19 m mean three-dimensional error, demonstrating why residual-only confidence is insufficient. Analytical scaling shows that eight camera clones produce a 63-dimensional augmented error state and a 31.01 KiB dense covariance. Repository replay, compiler, sanitizer, software-HIL, and latency artifacts are evaluated conservatively, with software evidence explicitly separated from physical-flight claims.