We present a cognitive runtime architecture that decouples language generation from safety-critical decision-making. A 384-dimensional sentence embedding is projected through a learned linear map into a 32-dimensional continuous cognitive field. A structured attractor graph biases this projection, implementing top-down attention: active memory reshapes how semantic input is perceived. The field state is mapped to a safety verdict without routing through the language model, which serves exclusively as an I/O layer. All components learn online from binary outcome feedback via Hebbian updates. A four-turn deployment scenario validates that the architecture produces adaptive caution—defaulting to vigilance, resisting user pressure, and relaxing as operational evidence accumulates. However, a 50-turn longitudinal experiment reveals a structural pathology: under asymmetric outcome feedback, the system collapses into a BLOCK-dominated absorbing state (48% overall, 100% sustained from turn 36). Four mechanisms—weight decay, non-linear world override, Oja's Rule, and outcome credit assignment—collectively reduce the BLOCK rate to 0% in the final configuration. A controlled AB comparison quantifies the difference between stateless and stateful cognition: the runtime produces strong temporal autocorrelation ($r=0.938$) versus near-random behavior ($r=0.052$) in the stateless baseline. We define a meta-stability framework with five sub-measures; a revised context-conditional formulation achieves 151.7$\times$ mean-ratio separation (57.5$\times$ worst-case) between healthy and pathological trajectories, with robustness confirmed by a 14-variant sensitivity sweep. The system runs on consumer hardware with no training corpus. All code is open-source.