Add an end-to-end audit and replay example that proves lineage is usable by operators

The current repo’s strongest architectural claim is that the whole chain from user turn to logit change is inspectable through persisted episodes, rewards, and policy snapshots. The Toulmin analysis rates this as one of the strongest claims, but it also says the paper does not exercise the audit story.

This ticket should make the audit promise tangible by adding a full operator-facing example.

Suggested scope

Add a reproducible demo that reconstructs how a policy change occurred from stored episodes and rewards.
Show how to trace which episodes drove a given logit change.
Add a replay or investigation notebook or script for operators.
Document the difference between policy-level interpretability and base-model interpretability.

Acceptance criteria

An operator can run a replay example locally or in a sample environment.
The example shows episode-to-reward-to-policy attribution end to end.
The docs explicitly scope interpretability to the policy layer and lineage.