Abstract
Agentic coding systems need terminal states below success. The local Codex corpus and related portfolio repositories contain many examples where the correct outcome is not "done" but "blocked," "held," "not run," "publish held," "admission blocked," or "held-out gate failed." This paper develops a failure and blocker taxonomy for agentic software work. It uses aggregate Codex-log evidence from A Longitudinal Corpus of Human-Codex Software Work by A.G. Mauro and C.A. Harris, repo-native closeout artifacts from historical Fractal Governance, Spinoffs, and Switchboard, and public-safe error-class examples rather than row-level private transcripts. The central claim is simple: honest stop rules are evidence-producing infrastructure. A stopped run can preserve source truth, prevent claim inflation, and route the next action to the right owner. A run that forces success can erase the most important fact in the system.