Checkpoint and resume cycle Resume is exact-state recovery, not warm-start: checkpoint identity and the remaining WAL must describe onecontinuous durable history. Immutable run identity Running attempt Durable persistence Resume gate fingerprint bind launch append checkpointbarrier snapshot after save: prune≤ frontier recover pass only RunSpec + manifest architecture · contracts· seeds Persistence identity run fingerprint ·journal ID Start or resume new attempt segment Learner + replay replay · optimizer · RNG Coherent capture state + applied WALfrontier Stop or crash at any point SQLite WAL accepted rows beyondcheckpoint Durable write temporary → fsync →replace Checkpoint 2.0 complete state +identity + frontier Validate recovery set schema · fingerprint ·journal · prune floor Restore exact state replay WAL rows by ID Resumed attempt new run segment Fail closed: reject a foreign journal or a checkpoint frontier older than the journal's durable prunefloor. Warm-start copies reviewed model tensors only; it never restores replay, optimizer, RNG, counters orWAL identity.