Implementing replay in a deterministic game engine seems straightforward: record inputs, replay ticks, compare results. But naive checksums that hash every field create false failures when harmless implementation details change. A debug helper field modification once triggered replay failures despite identical gameplay outcomes, revealing the core problem of conflating all state with gameplay-critical state.

The solution requires categorizing world state into distinct buckets. Authoritative gameplay state includes player health, projectile positions, and RNG streams that directly influence future ticks. Derived caches like pathfinding data need explicit handling - either rebuild deterministically or treat as runtime state. Observation and presentation layers should remain outside the checksum entirely, preventing renamed debug events or interpolation buffers from breaking replay.

The Zig engine implements this through a structured tick function with clearly defined phases: ingress, control, derive, plan, apply, cleanup, returning to idle. This boring predictability ensures the checksum has a consistent comparison point. The compute function deliberately enumerates only gameplay-relevant fields, excluding padding bytes and helper structures that could drift between runs.

Replay and save/load serve different purposes - replay verifies deterministic execution from seed and inputs, while snapshots freeze world state for restoration. Both require careful state selection, but replay demands stricter isolation of what constitutes actual gameplay divergence versus implementation artifacts.