Function predict_seed 

pub fn predict_seed(episode_seed: i64, predict_index: u64) -> i64

The per-predict sampling seed: a reproducible mix of the episode’s reset seed and the re-plan ordinal within that episode.

A stochastic policy seeded once per episode cannot be reproduced under interleaving — another episode’s forward advances the same global generators between two of ours. Deriving a fresh seed per predict from (episode_seed, predict_index) makes each forward reproducible on its own, independent of what ran between. The SDK stamps the result on the predict context as predict_seed; the same function is exported to Python as rlmesh.predict_seed, so a model can derive it itself.

FNV-1a over the seed and the ordinal, masked to 32 bits — the narrow range every framework’s seeding API accepts.