State-aware hedged requests for serverless GPU inference — return the first valid result and cancel the losers with an audited receipt.