fix: llms/googleai - handle nullable JSON Schema types in tool parameters

[sbroenne]

Summary

Fix JSON Schema type handling for nullable types in Google AI and Vertex AI integrations.

Discovery Context

This issue was discovered while running agent-benchmark tests against the Windows MCP Server using Gemini 2.5 Pro via Vertex AI. The MCP SDK generates JSON schemas with nullable types (["string", "null"]) for optional parameters, which caused tool calls to fail.

Problem

When tool parameters use nullable types per JSON Schema spec:

{ "type": ["string", "null"] }

The integrations fail with:

tool [0], property [fieldName]: expected string for type, got []interface {}

JSON Schema Spec Reference

Per JSON Schema specification:

"The type keyword may be either a string or an array... An array of strings allows matching against more than one type."

Solution

Added normalizeSchemaType() function that handles both forms:

• String: "string" → "string"
• Array: ["string", "null"] → "string" (extracts first non-null type)

Testing

• Added comprehensive unit tests for normalizeSchemaType()
• Added integration tests for nullable types at all nesting levels (single-level, nested objects, array items, deeply nested)
• All existing tests pass
• Verified with real Vertex AI API calls using agent-benchmark

Files Changed

• llms/googleai/googleai.go - Added normalizeSchemaType() function and updated convertSchemaRecursive()
• llms/googleai/googleai_unit_test.go - Added tests for normalizeSchemaType() and nullable tool parameters
• llms/googleai/vertex/vertex.go - Added normalizeSchemaType() function and updated convertTools()
• llms/googleai/vertex/vertex_unit_test.go - Added tests for normalizeSchemaType() and nullable tool parameters