fix(multiagent): add DISPATCHED status to prevent fan-in double-execution on resume (#3633)

[SuperMarioYL]

Description

When a graph node is dispatched (asyncio.create_task) but the process crashes before the node's generator body sets Status.EXECUTING (e.g., Lambda timeout, container restart, os._exit), the node is left PENDING. On session resume, _compute_pending_sources seeds only from EXECUTING nodes — so a dispatched-but-not-started node with no completed ancestor is indistinguishable from a bypassed dead branch. Its fan-in children are marked ready off the completed sibling alone, and the fan-in is double-executed on resume.

This adds a DISPATCHED status to the Status enum, set at the dispatch site immediately before asyncio.create_task. _compute_pending_sources now seeds from both EXECUTING and DISPATCHED, closing the window between dispatch and generator-body start. This completes the partial fix from #3390, which added the EXECUTING seed but left the dispatched-but-not-started gap.

Changes:

• base.py: Add DISPATCHED = "dispatched" to Status enum (additive, non-breaking)
• graph.py: Set DISPATCHED at the dispatch site in _execute_nodes_parallel; update _compute_pending_sources to seed from EXECUTING and DISPATCHED
• test_graph.py: Add TestResumeDispatchedSibling with two tests verifying the fan-in is not double-executed when a sibling was dispatched but not started

Related Issues

Fixes #3633

Documentation PR

No documentation changes needed — this is an internal status enum addition with no new public API surface.

Type of Change

Bug fix

Testing

• Added TestResumeDispatchedSibling with two tests:

  • test_dispatched_sibling_excluded_from_resume_frontier: focused unit test that injects DISPATCHED state directly (right.execution_status = DISPATCHED) and asserts the fan-in join is excluded from next_nodes_to_execute while the dispatched sibling right is included. It pins the resume-frontier computation; on raw master it raises AttributeError on Status.DISPATCHED.
  • test_resume_runs_fan_in_once_for_dispatched_sibling: real-dispatch end-to-end reproduction — a fast-completing left entry point and a right entry point held in the dispatched window (dispatched via the real _execute_nodes_parallel path, blocked before _execute_node sets EXECUTING). The snapshot is captured from left's AfterNodeCallEvent, deserialized into a fresh graph, and resumed; join must run exactly once with both parents' outputs present.

• Both tests are red on master via distinct, valid signals: the end-to-end test fails with an assertion failure (join lands in next_nodes_to_execute → double-execution on resume, because right stays PENDING and is treated as a bypassed dead branch), while the unit test raises AttributeError on Status.DISPATCHED. Both are green on the branch.

• All 83 existing test_graph.py tests pass

• All 181 broader multiagent tests pass (excluding a2a which has a pre-existing missing-dependency issue)

• ruff check and ruff format --check pass

• mypy passes with no issues

• I ran hatch run prepare

Checklist

• I have read the CONTRIBUTING document
• I have reviewed and understand every line of code in this PR, including any generated by AI tools, and I can explain why it works
• My change is focused and reasonably small; I have split unrelated work into separate PRs
• I have added any necessary tests that prove my fix is effective or my feature works
• I have updated the documentation accordingly
• I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.