[rekram1-node]

Fixes #40690.

Summary

• align OpenAI-compatible Chat stream schemas with the AI SDK's nullish usage and indexless tool-call tolerances
• preserve provider-specific usage fields and prediction-token details in raw provider metadata
• resolve indexless tool deltas using persistent call IDs, parallel positions, and sparse-stream state instead of per-chunk position alone
• treat empty finish reasons as terminal unknown values and reject meaningful content after termination
• intentionally omit Gemini's provider-specific OpenAI-compatible thought-signature extension pending dedicated Google routing support

Source comparison

Compared against ~/development/ai/packages/openai-compatible/src/chat/openai-compatible-chat-language-model.ts and convert-openai-compatible-chat-usage.ts. The native implementation retains its stricter incomplete-stream detection, undefined-vs-zero usage semantics, malformed tool-input isolation, and broader reasoning support where those are safer than AI SDK behavior.

Gemini's extra_content.google.thought_signature extension is intentionally excluded because it is specific to Google's OpenAI-compatible endpoint. Source links and follow-up context are documented next to the tool-call schema.

Validation

• bun typecheck in packages/ai
• 64 focused OpenAI Chat, OpenAI-compatible Chat, and ToolStream tests
• full packages/ai run: 476 passed, 28 skipped; 2 existing OpenRouter recorded fixtures fail because their stored bodies omit the current usage.include request field