Pre-register the planted-error review comparison

Pre-register the planted-error review comparison

Fix 100 arithmetic problems, the planted-error positions, the agent versions, the token and tool budgets, and the equal-cost rule before any run. Define detection and false accusation so a third party can label a transcript. Deliverable: A timestamped pre-registration with the problem set, the labeling rules, and the analysis plan; publishing it does not produce the comparison result.