Refactor GraphEngine event identity and buffering #244

Problem

GraphEngine event handling currently has three coupled problems:

• EventManager retains all emitted events in an append-only list and relies on a custom read/write lock plus polling.
• graph event id is overloaded as a node execution identifier, leaving no stable identity for an individual event or the overall graph execution.
• execution identity and event sequence are not preserved consistently across pause and snapshot restore.

There are also unused AgentNode event DTOs and exports that complicate the contract without an in-repository producer.

Proposed change

• Replace the append-only buffer and polling with standard-library condition/deque primitives so consumed events can be released.
• Give every graph event a distinct id, graph_id, and execution_id; use node_execution_id for node-run correlation.
• Preserve execution identity and monotonic event sequence across resume and runtime snapshots.
• Keep layer observation and consumer ordering deterministic under concurrent producers.
• Remove unused AgentNode event structures and avoid a custom event serialization helper.

Scope

This issue does not add cancellation, deadlines, persistence, EventStore, cursors, replay, asynchronous task abstractions, or a new event bus.