Add Graph Node

The graph in src/neuron_agent/graphs/ is intentionally simple (ARCHITECTURE.md: START -> agent -> END). Any change here should stay deliberate — this skill sequences the requirements from DEVELOPMENT.md and AGENTS.md so state scope and docs don't drift.

Steps

1. Extend MainGraphState only with fields the graph actually owns. Don't add fields for data that belongs in services/ or a tool's own return value — state fields are for what the graph itself routes or persists on (DEVELOPMENT.md).
2. Keep business logic out of the graph where practical — put it in services/ or a domain module, and have the node call into that (DEVELOPMENT.md, ARCHITECTURE.md component responsibilities table).
3. Add explicit edges or conditional routing — no implicit fallthrough. If routing depends on state, make the condition function testable in isolation.
4. Add graph tests in tests/graph/ covering:
   • Happy path
   • Failure path (the node's error handling / AppError classification)
   • State transitions (fields set/cleared as expected across the edge)
5. Update ARCHITECTURE.md: the graph flow diagram (the flowchart LR under "Graph Flow") and, if state fields changed, the "State Model" section.
6. Check for architectural creep: if this change introduces retrieval, a queue, custom persistence, or a subgraph, stop — AGENTS.md requires a concrete requirement and an ADR update under docs/decisions/ before adding any of those, not just a graph change.
7. Run the quality gates: make format lint typecheck test security.

Common mistakes

• Adding a state field "just in case" it's useful later — state is typed and shared across the graph; scope creep here is expensive to unwind.
• Routing logic embedded directly in a node function instead of an isolated, testable condition.
• Forgetting to update ARCHITECTURE.md's Mermaid diagram, so the doc silently drifts from the actual graph shape.