`interrupt_before` re-interrupts forever after a state key is removed

Finding: In an external report on langchain-ai/langgraph issue 9001 (https://github.com/langchain-ai/langgraph/issues/9001), a thread checkpointed with a state key later removed from the schema, and resumed on a graph compiled with interrupt_before, is reported to pause before the same node on every resume. Resume returns normally and no error is raised. The reporter lists langgraph 1.2.11 (also 1.0.0, 1.1.0 and main at commit aa742fb), langgraph-checkpoint 4.2.0, Python 3.13 on macOS, and says SqliteSaver behaves the same. This is an external report only; KFA has not reproduced it. Reported reproduction: a graph with nodes first and last, compiled with interrupt_before=["last"], runs once with a schema containing a key `note`. It is then rebuilt with the same checkpointer and a schema without `note`. Three consecutive invoke(None, config) calls each return log ['first'] and leave next as ('last',). The reporter expected ['first', 'last'] and an empty next on the first resume (issue 9001). Reporter's hypothesis (not maintainer-confirmed): on resume, _first() in pregel/_loop.py marks the interrupt as seen only for channels the current graph has, while should_interrupt() in pregel/_algo.py compares versions across every channel stored in the checkpoint. A stored channel the graph no longer has is never marked as seen, so the check always sees an update since the previous interrupt and fires again. The reporter says the stale version is carried into later checkpoints, so threads paused elsewhere, such as in interrupt(), can also get stuck at the next interrupt_before after the schema change. They say interrupt_after is unaffected, and that a create_react_agent tool-approval flow with interrupt_before=["tools"] is a common case where approval repeats indefinitely (issue 9001). The reporter notes the backward-compatibility guide already lists removing a state key as a breaking change, but argues the problem is that the thread loads and runs without error, then loops silently at the breakpoint (issue 9001). Fix status: a contributor commenting with author association NONE links a branch with a proposed change and regression tests (https://github.com/langchain-ai/langgraph/issues/9001#issuecomment-5743105600). A later comment says PR #9108 was opened to mark every channel in the checkpoint's channel_versions as seen on resume, and asks to be assigned (https://github.com/langchain-ai/langgraph/issues/9001#issuecomment-5886167589). The supplied sources show no merged fix, fix commit or release, and the issue is open. The contributor's test results are self-reported and unverified. Practical implication, conditional on this report being accurate: deployments that drop a state field from a schema while threads are checkpointed, and that use interrupt_before, may see silent repeated pauses rather than an error. No workaround is documented in the sources, and none is recommended here.