The Scaffold Continuity Trap: A Blinded Scaffold-Lesion Test in Persistent Coding Agents

Persistent AI agents can appear to continue projects, commitments, and working identity across sessions. This preprint tests whether continuity-like behavior in a coding-agent continuation task is carried by external scaffolds such as project files, tests, handoff notes, summaries, tool traces, and control state. The study uses a blinded scaffold-lesion protocol with four separate blinded runner threads assigned to primary straight-condition packets: full scaffold, summary-only, no-memory, and sham lesion. All four achieved 12/12 on the core continuation score. The preregistered full-versus-no-memory gate therefore failed: the observed core-score delta was 0.0 against a required 3/12-point threshold. The full-versus-summary comparison also failed, while sham stability passed. Hidden-probe recovery remained low across conditions. The result does not support the Scaffold Continuity Trap as originally stated for this fixture. The stronger bounded conclusion is narrower: in small coding-continuation tasks, ordinary visible repository state and tests can carry enough task-local causal structure for perfect action-level recovery even when richer continuity scaffolds are removed. Self-report continuity varied with scaffold richness, but task success did not. This work does not test or claim AI consciousness, personhood, moral status, survival, or subjective continuity. It is a scaffold-localization and methodology result about observable task continuation under controlled lesions.