interrupt() takes a per-task counter from the scratchpad and returns resume[idx], so a resumed node must call its interrupts in the same order as the first run; reorder, add or conditionally skip one and it silently receives another prompt's answer. That logic sits in Core Types & Runtime Context, which 18 of 26 elements here depend on, 11 directly with no fallback, including StateGraph and the Functional API.

Impact: A node that reorders, adds or conditionally skips an interrupt() call gets another prompt's resume value silently, with no error to point at the mismatch, and the risk travels through every one of Core Types' 11 dependents.

Recommendation (small effort): Call interrupt() unconditionally and in a fixed order within a node, and add a regression test that resumes a multi-interrupt node out of order to catch positional drift before it reaches State Node Spec, StateGraph or the Functional API.

The Pregel class (main.py:449-4190) is the sole entry point into the runtime: all 16 outbound calls it makes reach the checkpoint mapper, channel I/O helpers, task planner, loop and runner — nothing else in the file has a way out. That concentration comes with a duplication cost: bulk_update_state and abulk_update_state are each about 470 hand-kept lines, as are the stream/astream pairs, so a fix has to land twice, as the recent nested checkpoint_ns regression did.

Impact: Every bug fix or behavior change to checkpointing, I/O shaping, task planning or execution has to be applied once for the sync path and once for the async path inside this class, with no shared implementation to catch a missed side.

Recommendation (medium effort): Extract the shared logic behind bulk_update_state/abulk_update_state and the stream/astream pairs into one implementation, so a fix like the checkpoint_ns one lands once, without giving up the single entry point into the runtime.