Japanese (CJK) characters occasionally corrupted in model output (mojibake)

Description

When using Claude Code on Windows, some Japanese (CJK multi-byte UTF-8) characters are occasionally corrupted in both:

1. Direct text output to the terminal (conversation responses)
2. File writes via the Write tool

The corruption appears as replacement characters (� / �� / ���).

Environment

• OS: Windows 11 Home 10.0.26220
• Shell: bash (Git Bash)
• Model: Claude Opus 4.6 (1M context)
• Claude Code: latest CLI version

Examples

| Context | Expected | Actual |
|---------|----------|--------|
| Terminal output | 進めますか？ | ���めますか？ |
| Write tool (TASKS.md) | 本番リリース（PR #447） | 本番リリース��PR #447） |
| Write tool (TASKS.md) | マージ済み | マー��済み |
| Write tool (signup-form.tsx) | 利用規約 | 利用��約 |
| Write tool (signup-form.tsx) | パスワード | パ���ワード |

Reproduction

• Not deterministic — the same characters sometimes render correctly, sometimes corrupted
• More frequent with the Write tool (full file content) than with Edit tool (diff only)
• Also occurs in plain text responses (not file writes)

Observations

• Edit tool (small diffs) rarely exhibits this issue
• Write tool (full file rewrite with long Japanese text) is more susceptible
• The corrupted bytes suggest UTF-8 multi-byte sequences split at token boundaries

Workaround

• Use Edit tool instead of Write tool for existing files
• Post-write grep for � to detect and fix corruption

Expected behavior

All Japanese/CJK characters should render correctly in both terminal output and file writes.