[Bug]: GraphSpec.validate() accepts entry points targeting missing nodes

GraphSpec.validate() validates the main entry_node, terminal nodes, and edge references, but it does not validate the node IDs referenced by entry_points.

As a result, a graph can pass validation even when an alternate entry point targets a node that does not exist. The failure is only discovered later during execution or session resume.

Location

• core/framework/orchestrator/edge.py
• GraphSpec.validate()
• GraphSpec.get_entry_point()

Steps to Reproduce

1. Create a graph with an entry_points mapping such as:

   python复制

   entry_points = {
       "resume": "missing_node"
   }

2. Define nodes that do not include missing_node.

3. Call:

   python复制

   graph.validate()

Expected Behavior

Validation should return a blocking error identifying the invalid entry point and the missing target node.

Example:

Entry point 'resume' references missing node 'missing_node'

Actual Behavior

Validation succeeds because only the main entry node, terminal nodes, and edge source/target references are checked.

Later, get_entry_point() can return missing_node, causing execution or resume to fail outside the validation step.

Impact

Invalid agent graphs can pass validation and fail at runtime. This is especially risky for pause/resume flows and alternate execution entry points, where the configuration error may be difficult to diagnose.

Proposed Fix

Update GraphSpec.validate() to verify that every value in entry_points references an existing node.

Add regression tests covering:

• A valid entry point target.
• A missing entry point target.
• Multiple entry points with mixed valid and invalid targets.
• Preservation of existing validation behavior.

Additional Context

This is related to validation gaps discussed in issue #6090, but this report focuses specifically on missing validation for entry_points targets.