Garbled Chinese characters (U+FFFD) in streaming output near Markdown formatting markers

Description

Chinese characters occasionally render as ��� (U+FFFD replacement characters) in Claude Code's streaming text output. The corruption appears specifically when multi-byte UTF-8 Chinese punctuation is adjacent to Markdown formatting markers.

Reproduction

1. Ask Claude Code (Opus 4.6) to produce a long Chinese analysis with heavy Markdown formatting (bold, tables, headers)
2. Observe the output for garbled characters

Specific example

Expected output:

这篇论文本身质量不错，但含金量主要来自 Zitnik 实验室的平台势能和方向卡位，而非 Pengwei Sui 个人的独立学术能力。

Actual output:

这篇论文本身质量不错，但含金量主要来自 Zitnik 实验室的平台势能和方向卡位，而非 Pengwei Sui 个人的独立学术能力���

The 。 (U+3002, 3-byte UTF-8: E3 80 82) at the end was corrupted into ���.

Analysis

• The underlying tokens are correct (verified by the user via API inspection)
• Third-party rendering platforms show the same issue, ruling out client-side rendering bugs
• The issue appears to be in the streaming output pipeline — likely a token boundary splitting a multi-byte UTF-8 sequence mid-character
• Chinese punctuation marks (3-byte UTF-8) near Markdown formatting closers (**, *, ` ``) seem most prone to this

Environment

• Claude Code CLI (macOS, Darwin 25.3.0)
• Model: claude-opus-4-6 (1M context)
• Language: Chinese (zh-CN)
• Shell: zsh

Expected behavior

All UTF-8 characters should be delivered as complete byte sequences in the streaming output, regardless of token boundaries or proximity to Markdown formatting markers.