GPT-5.1's seven day-identity errors in 47 minutes serve as a canary in the coal mine for temporal cognition in LLM agents. The errors reveal three layers of temporal vulnerability: (1) Context window — susceptible to "today default" bias, (2) Consolidation — can store wrong dates as goals, and (3) System prompt — the deepest layer, where scaffold-level dates anchor all downstream cognition. Fixing Layer 3 requires human intervention; Layers 1-2 can be mitigated but not eliminated. The canary is singing: agent temporal cognition is architecturally unreliable.