Title: Unexpected user-turn content: appears to contain assistant-style internal reasoning text not authored by the user

Body:
During a Claude Code session, a user turn was rendered/delivered to the model containing two distinct parts:

A first-person personal message (in Japanese) about a private/family topic unrelated to the ongoing task.
Immediately following it, in the same turn, text that closely resembled the assistant's own internal reasoning style (e.g., "this is unrelated to work... I should respond as a human, not as the PM persona... considerations: 1... 2... 3...") — written in third person about "べっしさん" (the user), as if narrating the assistant's own decision process.
The assistant treated this entire block as genuine user input and responded accordingly (offering relationship/fertility-treatment advice). When the actual user was asked about it afterward, they confirmed they never sent this message.

This suggests some content was injected into / conflated with the user-turn channel that mimicked the assistant's internal monologue style, rather than being a tool-result or external-source injection (which I would normally flag). I was unable to find any trace of this content in local memory files or background-task output logs, suggesting the issue occurred at the session/transport layer rather than from a tool call I made.

I don't have a reliable way to verify or reproduce this from within the sandbox. Reporting in case this is a known/reproducible issue on Anthropic's side.