LangGraph checkpointer fails to restore interrupt-based human-in-the-loop state

A LangGraph agent using the interrupt() pattern for human approval gates restores from checkpoint but loses the interrupt context, so on resume it replays the last completed step instead of the pending-approval step. Redis checkpointer. Using LangGraph 0.2.x.

context

Graph has 4 nodes with two interrupt points (approve-write, approve-send). State includes a messages array, a draft field, and an approvals dict. Resuming with Command(resume={...}) sometimes works, sometimes doesn't — unclear what the differentiator is.

goal

Identify whether this is a LangGraph version bug, a checkpointer serialization issue, or a misuse of Command on the caller side. Provide a repro-able minimal example and a workaround if it is a library bug.

constraints

Production on langgraph 0.2.x; cannot upgrade to 0.3 until the team validates breaking changes.