fix(google): use responseJsonSchema for Gemini

[Kage18]

Summary

• send Gemini structured-output schemas as generationConfig.responseJsonSchema for the Google API
• keep Vertex requests on generationConfig.responseSchema
• cover generate and stream request bodies, plus a Vertex regression

Fixes #7689.

Kage usage

• Used Kage scoped to packages/google to locate the shared GoogleLanguageModel path and relevant structured-output tests.
• Captured a repo-local memory packet explaining why non-Vertex Google uses responseJsonSchema while Vertex keeps responseSchema.
• Local Kage metrics after refresh: 78 files, 2,158 symbols, 1,369 calls, 564 tests, 100% indexed coverage, 3 memory packets, 82 memory graph entities, 83 evidence-backed edges, 0 stale packets.

Tests

• pnpm --dir packages/google exec vitest --config vitest.node.config.js --run src/google-language-model.test.ts
• pnpm --filter @ai-sdk/google type-check

Note: local verification emitted the repository Node engine warning because this shell runs Node v25.9.0 while the repo declares support for ^18/^20/^22/^24.

[Kage18]

Kage usage for this PR:

This PR fixed #7689 by sending Gemini structured-output schemas as generationConfig.responseJsonSchema for the Google API while preserving generationConfig.responseSchema for Vertex.

How Kage helped:

• Scoped recall to packages/google and the shared GoogleLanguageModel request-building path.
• Kept the relevant implementation, snapshots, and tests together: google-language-model.ts, google-language-model.test.ts, and the structured-output snapshots.
• Helped preserve the important provider split: non-Vertex Google requests use responseJsonSchema; Vertex requests continue using responseSchema.
• Captured the durable rule and regression evidence for both generate and stream request bodies, plus the Vertex regression case.
• Captured the verification commands and the local Node engine-warning context.

Why this matters for future sessions:
A future agent changing Google structured-output request generation can recall the API/Vertex distinction and avoid “simplifying” both providers onto the same schema field.

Full contribution summary and Kage metrics: kage-core/Kage#8

[snouker]

Without any necessary filtering, they have the same "responseJsonSchema" parameter.
Please refer to Vertex AI(linked to gemini-enterprise-agent-platform) for details.

In addition, please modify the convertJSONSchemaToOpenAPISchema function based on ai.google.dev to support more JSON schemas, so as to make this PR more necessary.

Sorry for that I am not a relevant contributor and thus cannot assist you in passing this PR.

[lgrammel]

Thank you for the pull request. Issue #7689 was closed as not planned. Closing this pull request.