[gpt-5.6-sol] Random multilingual tokens appear in Chinese responses

What issue are you seeing?

During a long Chinese conversation, unrelated fragments from other languages or scripts occasionally appeared inside otherwise normal responses.
The corrupted text was already present in the raw Codex agent_message event before it was processed by the local bridge, so it does not appear to be caused by message encoding or chunking in the bridge.
I had not observed this issue while using gpt-6-astra. It first appeared after I switched back to gpt-5.6-sol, although I cannot confirm that the model switch caused it.

What steps can reproduce the bug?

This behavior is intermittent and has only been observed with gpt-5.6-sol so far.
1. Start a long Chinese-language thread through Codex app-server.
2. Use gpt-5.6-sol with high reasoning effort.
3. Continue ordinary Chinese conversation and short structured-response requests.
4. After multiple turns, unrelated multilingual fragments may appear.

What is the expected behavior?

Responses should remain coherent Chinese unless another language is requested or contextually required.

Additional information

I switched from gpt-6-astra to gpt-5.6-sol shortly before the first occurrence.
- The app-server and MCP health checks were normal.
- The context window was not close to full.
- No replacement-character or transport-encoding error was found.
- I can provide additional redacted raw event excerpts if needed.

Start with the Codex app-server and capture the raw agent_message events described in the report, comparing gpt-5.6-sol with gpt-6-astra during a long Chinese conversation. Confirm whether multilingual fragments originate before the local bridge and identify a reproducible failure case; done means the source is isolated and coherent Chinese responses are restored or the issue is shown to be model-side.