fix: isolate RunState checkpoint tool decisions - #4413

Thanks for landing the follow-up. I ran the merged commit 0b93ce8faa27d4631df399fe48856b52a8fd9897 through AgentRunProof using the SDK's public agents.testing.ScriptedModel (no provider/API calls).

Good news: the original sibling-isolation scenario now passes all seven recorded invariants in both Runner.run and Runner.run_streamed:

execution_outcome   PASS
state_transitions   PASS
state_fork_isolation PASS
phase_contract      PASS
tool_linkage        PASS
exactly_once        PASS
stream_parity       PASS

That is a real FAIL → PASS relative to openai-agents==0.20.0.

The same verification found one recursive boundary that appears not to be covered by #4413. With a protected approval nested through Agent.as_tool:

The minimal sequence is:

paused = await Runner.run(outer_agent, "start")
state = paused.to_state()
state.approve(state.get_interruptions()[0])
resumed = await Runner.run(outer_agent, state)

At two or more agent-tool edges, resumed.final_output is None, the same flattened approval remains in resumed.interruptions, and the protected tool has not executed. I reproduced the same result in streamed mode.

So the top-level checkpoint isolation fix is working; the remaining gap is recursive ownership/routing of that approved interruption. I will avoid opening a competing PR. I can provide the focused 2/3-edge regression or validate a follow-up immediately. I am also preparing a content-addressed before/after record so the result is independently checkable; the published 0.20.0 counterexample is in AgentRunProof v0.1.1.