A table created while the step-down timestamp is set has no stable constituent by design. On a search miss, the layered cursor lazily opens the stable constituent, and that lazy-open path unconditionally treats the deferred open as the follower case — the created-during-step-down state is not consulted there. Once the leader has completed any checkpoint (its own checkpoint is implicitly adopted, so the connection LSN is set), the follower-branch era check runs against the leader-era snapshot and refuses the bind.

The result is that a plain search miss on such a table returns WT_ROLLBACK where WT_NOTFOUND is expected — no role transition needs to be involved at all. Retrying does not converge: each retried transaction is again a leader-era snapshot on the same leader, so the application keeps rolling back until the step-down actually completes.

This is not urgent: there is no crash or wrong data, only a spurious retryable rollback in place of a not-found result.