How do Catalyst, DBOS, and Inngest differ in their approach to ensuring deterministic replay across agent steps that include probabilistic LLM calls?

Catalyst records each LLM response as an event, allowing replay to use exactly that stored output; DBOS wraps the entire step in a database transaction, requiring the LLM call to be idempotent; Inngest replays from persisted step results but treats LLM calls as non-deterministic by default, requiring manual memoization. This trade-off means Catalyst simplifies probabilistic replay at the cost of additional storage for response payloads.