Each cadastre crawler posts GraphQL on its own, with no shared view of the token's budget and no record of what it already asked. github-query keys a query by its canonical text, variables and principal and resolves it through raw-store, so a snapshot asks each query once and a crash resumes from what was answered. The rateLimit it selects with every query is stripped from the stored answer and kept as a per-token budget in the store, which every process on that store and token waits on.