fix(openai-compatible): handle thinking content arrays

[tobwen]

Background

Some OpenAI-compatible providers (see #13703) return structured content arrays (including thinking parts) in chat responses/chunks, while the current implementation expected string-only content in key paths. This caused validation failures (expected string, received array) and broke generate/stream flows.

Summary

• Add support for array-shaped content in OpenAI-compatible chat response/chunk schemas.
• Normalize known content-part types:
  • text parts -> text output
  • thinking parts -> reasoning output
• Keep compatibility precedence:
  • reasoning_content > reasoning > reasoning extracted from thinking
• Keep schema future-proof by allowing unknown array part types without failing validation.
• Add regression tests for:
  • non-stream content arrays with thinking
  • stream content arrays with thinking
  • unknown content part types in both paths
• Add patch changeset.

Manual Verification

• Ran package tests:
  • corepack pnpm --filter @ai-sdk/openai-compatible test:node -- src/chat/openai-compatible-chat-language-model.test.ts
  • Result: all tests passed (including new regression tests)
• Ran package type-check:
  • corepack pnpm --filter @ai-sdk/openai-compatible type-check
  • Result: passed

Checklist

• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
• I have reviewed this pull request (self-review)

Future Work

• Consider extracting the content normalization helper for reuse if similar parsing logic is needed in other OpenAI-compatible surfaces.

Related Issues

#13703