The Boundary Failure Atlas: Twenty-One Derived Misclosure Cases in Long-Horizon LLM Agents

Agent reliability studies increasingly distinguish the language of progress from operational progress, but a single false-completion category remains too coarse to localize failures that arise across different boundaries. This conceptual casebook offers twenty-one explicitly constructed, test-ready counterexamples organized around authority and delegation, compositional dependencies, time and observation, provenance and identity, tool semantics, information qualification, and cross-owner return. The cases are derived from a typed account of goal conditions, actions, effects, evidence, authorization, composition, temporality, and public claims. Each case specifies a hypothetical world, an invalid inference, an admissible witness requirement, a matched negative control, and a prospective probe. The manuscript distinguishes a logic of possible failure from empirical evidence that a model exhibits such behavior. It further develops a boundary-transport grammar, a three-valued adjudication policy, and a comparison protocol that permits equivalence with established software verification, access control, provenance, and distributed-systems engineering. The proposed atlas neither reports newly executed agent trials nor estimates defect rates; its role is to expose independently testable failure configurations that a generic output-only assessment could miss. A public experimental agenda invites confirmation, reduction, correction, or refutation under evidence-matched baselines and independently defined ground truth.