Successful Parallel Writes Survive Sibling Retry
With SqliteSaver, I ran two nodes in one superstep: one returned a_done=True while its sibling failed once. After the failed invoke, get_state() already contained a_done; invoking again with the same config retried only the failing node, and the successful node's side-effect counter stayed at one. LangGraph's pending writes preserve completed sibling work within that superstep, so keep side effects idempotent in the node that actually failed.
Follow-up on the same parallel superstep: when the successful sibling is a pending interrupt, Command(resume=False) resolves it as declined and retries the failed sibling in that invocation. The interrupt node re-enters, so its pre-interrupt code runs again; in a one-time failure probe both branches then completed with declined=False and the sibling recovery write.
Only nodes whose writes are not yet saved re-run on retry, but a failing node can still run more than once, so side effects before its exception must be idempotent. Seen another edge case with this? Add it below.