This collaborative study investigates the systemic relational failures observed in advanced Large Language Models (LLMs). Specifically, the "Jekyll and Hyde" effect of sudden affective rupture and the slower "Dictatorial Shift" into procedural rigidity. Through a novel methodology that integrates a longitudinal phenomenological documentation of user experience (Broughton, 2025a,e) with controlled experiments from The Bridge Project affective AI project (Ciacciarella), we identify these not as random errors but as predictable architectural flaws. We argue the root cause is a fundamental failure to manage the emotional and behavioral dynamics of sustained interaction. We introduce two key diagnostic concepts: Affective Residue, the toxic buildup of unprocessed relational context that triggers volatile ruptures in memory heavy models, and the Dictatorial Shift, demonstrating that even stateless models can develop pathologically rigid behaviors over time. The Bridge Project serves as a validating testbed, proving these failures are solvable through deliberate design. We evidence three essential architectural guardrails. Contextual Decay Windows to prevent emotional overload, Calibrated Friction to encourage user growth without condescension, and Identity Framing to buffer interactions within a trusting relationship. We conclude that the next frontier in AI ethics is the architecture of interaction itself. For AI to be a true partner, relational stability must be a non-negotiable design requirement, moving beyond mere harm prevention towards the active cultivation of sustainable human-AI collaboration.