[cestercian]

Description

Gemini's OpenAI compatible endpoint sends usage on every SSE chunk, right next to the content and tool call deltas. The stream handling treated any chunk with usage as the final one, so text got dropped and tool calls never showed up.

Now the deltas keep streaming when usage is present, the latest usage wins, and unknown extras like extra_content are passed through instead of breaking things. Normal OpenAI behavior (usage only on the last empty-choices chunk) is unchanged.

Fixes #13417

Checklist

• I've read the contributing guide
• The relevant docs, if any, have been updated or created
• The relevant tests, if any, have been updated or created

Tests

• packages/openai-adapters/src/test/openai-gemini-usage.vitest.ts: Gemini style chunks with usage on every chunk plus content and tool call deltas
• core/llm/openaiTypeConverters.gemini.vitest.ts: chunk conversion keeps content, tool calls and usage together
• openai-adapters vitest suite passes, tsc clean in core and openai-adapters, eslint and prettier clean on touched files

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[cestercian]

I have read the CLA Document and I hereby sign the CLA