Gravitational Manipulation Theory v2.6 — Descriptor-Irreducible Apparatus Dependence under Out-of-Apparatus Validation

Gravitational Manipulation Theory (GMT) v2.6 advances the series from measuring an information-disclosure standard to predicting its reconstruction-information margin for an unseen apparatus. Using the frozen GMT v2.5 dataset of eight synthetic apparatus classes across three interaction-range regimes (24 cells), this study evaluates a pre-registered two-stage estimator under leave-one-apparatus-out (LOAO) validation. Stage A uses a LASSO-selected relaxed-linear model, while Stage B applies a Gaussian-process residual correction. A documented selector discrepancy was resolved by a dated protocol ruling: the canonical low-dimensional descriptor set is d_low = {N}, while {N, ρ} is retained as a labeled noncanonical robustness branch. The canonical surface achieves RMSE_LOAO = 1.197 dex and R²_LOAO = −0.065. It improves by approximately 7% over a fair fold-wise LOAO intercept predictor (RMSE = 1.291 dex) but remains far below the frozen acceptance thresholds of RMSE ≤ 0.15 dex and R² ≥ 0.80. The per-family models also fail the corresponding validation requirement. Applying the frozen mechanical decision rule therefore returns Outcome C: descriptor-irreducible apparatus dependence under the frozen model, grid, and acceptance thresholds. All 24 observed reconstruction-information margins remain positive, with M_G ≥ 1.22 dex. A bounded Tier-2 diagnostic search finds no model-implied zero-margin contour within N ∈ [2, 20] and x_χ ∈ [−1, 1], with a minimum predicted margin of 0.716 dex. Because the surface does not pass LOAO validation, all Tier-2 extrapolated quantities are explicitly non-certified. The deposited package includes the 16-page manuscript, Figures 1–7, the frozen Master Spec, the dated protocol-discrepancy record, Stage 1–3 reports, frozen input data, configuration files, analysis and figure-generation code, machine-readable results, reproduction instructions, citation metadata, and licensing information. All apparatus are synthetic representative classes; none represents HUST hardware or any other real instrument. This work does not claim experimentally demonstrated gravitational manipulation, universal or physically essential apparatus dependence, true χ-independence, or a physical zero-margin boundary.