Control Capacity Can Convert Failure into Effort: Endpoint-Dependent Attenuation in a Simulated Governance Control Loop

In a deliberately minimal governance control loop, we test whether a binary terminal endpoint can conceal the continuing burden of one explicitly compound six-parameter cognition perturbation. Holding the controller, dynamics and random seeds fixed, and varying only intervention budget across five levels, terminal failure falls monotonically from 1.000 at budget 0.20 to 0.023 at budget 1.00, pooled over 600 runs per cell across three deterministic seed blocks. A matched analysis converts 193-196 of 200 matched pairs from failure to non-failure with zero reversals in every block; each block contains 199 low-budget failures. At operational threshold 0.800 and the highest budget, 35.7% of runs occupy a recovered rather than persistently-stable class; this percentage is threshold-specific and ranges from 12.8% to 79.5% across the frozen operational sweep. The 214 recovered runs at threshold 0.800 spend a median five total steps below threshold and reach a median maximum excursion depth of 0.028. The perturbed condition also produces greater controller demand and lower minimum stability than intact in 600 of 600 matched pairs. A prospectively hash-frozen 64-subset component factorial (192,000 runs) finds no single component above half of the declared absolute-share allocation; three components account for 96.6% under that normalization. Operational and measurement-only threshold analyses preserve the relevant threshold-dependent directions throughout 0.750-0.850. A paired comparison using the frozen trajectories shows that continuous minimum-stability separation also attenuates but remains large at budget 1.00 (Cohen dz = 1.65) while the failure-risk difference is 0.023. We call this one-model pattern endpoint-dependent attenuation. A trust-pathway ablation fails algebraically under every budget and remains an implementation control, not evidence. The result is a reproducible existence demonstration, not a general perturbation class, prevalence estimate, empirical governance finding, or universal reporting law. This record contains the MB-R9 preprint, editable manuscript sources, bounded core reproducibility materials, run-level robustness data, a public revision note, license mapping, file guide, and SHA-256 manifest. An earlier CoMSES release remains private, unpublished, under review, and without a DOI; it is not this record's version of record and is not relied upon for public availability.