This record hosts the third paper in a trilogy on long-horizon human–AI collaboration. Paper 1 introduced a context-engineered collaboration with strong governance and canonical numerics, and Paper 2 formalised the Lean Collaboration Operating System (LC-OS) that governs a long-lived human–AI dyad. This third paper focuses on how such a system actually fails and repairs itself in practice, using twelve internally logged episodes from a yearlong collaboration between one human and a frontier large language model (“Mahdi”) from the GPT-5 family. It develops a six-part failure taxonomy, a set of structural repair patterns, and a minimal tracing toolkit (TraceSpec, ProbeKit, TraceLens) for logging and replaying breakdowns. The record includes the final PDF of Paper 3, which is conceptually linked to but distinct from prior registrations for Papers 1 and 2. OSF DOI: 10.17605/OSF.IO/Z7AQ8