Deterministic Replay

HyperToken records dispatched actions and supports seeded randomness through mulberry32 and shuffleArray. Supplying the same seed to a shuffle with the same input order can reproduce that shuffle.

That is bounded determinism, not a promise that every session is a byte-for-byte replay. Reproduction also depends on the initial document, action order, action payloads, token IDs, generated IDs, timestamps, agent decisions, and the handling of concurrent remote changes. Unseeded randomness and arrival order introduce additional variation.

History is checkpoint-based, not a full snapshot for every action. Checkpoints are taken every 50 actions and capped at five. Undo restores the nearest available checkpoint and can therefore move back by up to the checkpoint interval; it is not exact action-by-action reverse execution. Compaction removes older CRDT history entirely, so compacted history cannot be replayed.