Portable trajectory retained for exact rollout reuse.
Summary
Types
@type t() :: %Imp.Training.FastSlow.CachedTrajectory{ behavior_logprobs: [number()], cycle: non_neg_integer(), id: String.t(), input_digest: String.t(), output: Imp.Training.FastSlow.Config.json_value(), problem_id: String.t(), prompt_digest: String.t(), response_mask: [0 | 1], response_token_ids: [non_neg_integer()], reward: number(), theta_id: String.t() }