SFM 2.1 Computational Verification Companion II: Verification Before Validation — Synthetic Ensemble Stress Testing, Threshold Sensitivity, and Finite-Graph Audit

This companion reports a corrected synthetic, specification-directed ensemble rerun following operational closure of the finite admissible-mapping family T_a. The frozen experiment generated 2,916 base incidence states across six datum counts, using a balanced 162-cell factor design and three independently seeded replicates. Each base state was paired with an unchanged control and four controlled perturbations: relation-class mutation, orientation reversal, magnitude change, and carrier deletion. This produced 14,580 inter-state evaluations. The exact T_a maximum was proven for every evaluated pair. No mapping search was unresolved, and no generated or perturbed relation magnitude left the declared range [0.5, 2.0]. At the illustrative defaults lambda_min = 0.75, mu = 0.75, and epsilon_P = 1.0, 720 evaluations passed, 6,720 failed, and 7,140 were undefined under the declared carrier and bridge hierarchy. The unchanged control passed in 351 of 2,916 instances, or 12.04%. Among those 351 control-passing instances, relation-class mutation converted 60.11% to non-pass, orientation reversal converted 60.68%, magnitude change converted 74.93%, and carrier deletion converted 100%. A 3,087-point sensitivity grid shows that epsilon_P has a modest effect on the total pass rate, lambda_min does not change the pass count across the tested range, and mu behaves as the expected decision threshold rather than a uniquely privileged constant. The supplied reproducibility package contains the frozen generator, protocol, seeds, exact mapping implementation, regression tests, all pair-level results, sensitivity tables, figures, hashes, and execution instructions. An independent external AI computational reproduction audit regenerated the complete ensemble and sensitivity analysis from source. All eight regression tests passed and every reported aggregate, paired, threshold, and instance-level result reproduced exactly. Companion II was classified as reproduced without material error. This study verifies reproducibility, selectivity, cross-gate interaction, and threshold behaviour within the declared finite synthetic generator. Because the perturbations deliberately target stipulated audit surfaces, it is a formal stress test and not external evidence that the specification corresponds to physical pre-Planck structure. SFM 2.1 Version 1.1 remains frozen and unchanged.