Description

Implements #157

Summary

• Adds a parallel execution (fan-out / fan-in) framework to GraphExecutor, enabling a single node to execute multiple downstream branches concurrently via asyncio.gather()
• Introduces ParallelBranch and ParallelExecutionConfig dataclasses for controlling branch lifecycle, timeouts, and failure strategies (fail_all, continue_others, wait_all)
• Adds detect_fan_out_nodes() and detect_fan_in_nodes() to GraphSpec for static analysis of parallel topology in edge definitions
• Implements automatic convergence detection (_find_convergence_node()) so branches rejoin at a shared downstream node
• Supports per-branch retry via node max_retries, configurable memory conflict resolution (last_wins, first_wins, error), and a 300s default branch timeout

Changed files

• core/framework/graph/executor.py — parallel branch execution, convergence detection, multi-edge traversal
• core/framework/graph/edge.py — fan-out/fan-in detection on GraphSpec
• core/framework/graph/node.py — async memory write support for parallel context
• core/framework/llm/litellm.py — minor adjustment for parallel compatibility
• core/framework/runtime/execution_stream.py — stream support addition

Test plan

• Verify fan-out triggers when a node has multiple ON_SUCCESS edges
• Verify all branches execute concurrently (not sequentially)
• Verify convergence: all branches rejoin at the correct fan-in node
• Test fail_all strategy cancels remaining branches on first failure
• Test continue_others strategy lets healthy branches complete
• Test wait_all strategy collects all results before proceeding
• Test branch timeout fires at configured branch_timeout_seconds
• Test memory conflict strategies (last_wins, first_wins, error)
• Test per-branch retry respects node max_retries
• Verify single-edge paths still execute normally (no parallel overhead)