[BUG] U+FFFD corruption in Cyrillic text — both streaming output and Edit/Write tool

Bug description

Cyrillic characters are silently replaced with U+FFFD (��) in two places:

1. Streaming output — model responses in terminal contain replacement characters
2. Edit/Write tool — files written by Claude contain corrupted characters

Environment

• Windows 11, Claude Code 2.1.85, model Opus 4.6
• File encoding: UTF-8
• Affected: .md files with Russian text

Reproduction

1. Ask Claude Code (in Russian) to edit a .md file containing Cyrillic text
2. Observe the output and the resulting file

Input: контейнера, после, настройки, формулировки, Лучше
Output: конте��нера, пос��е, на��тройки, формулиров��и, ��учше

The corruption is intermittent — it depends on where SSE chunk boundaries fall within multi-byte UTF-8 sequences.

Root cause

Cyrillic characters are 2-byte UTF-8 sequences. When an SSE chunk boundary splits a character between its bytes, each orphaned byte is decoded independently and replaced with U+FFFD.

The SSE line decoder in the Node.js CLI uses TextDecoder without the { stream: true } option, which would carry incomplete byte sequences across chunks instead of replacing them.

Proposed fix

In the SSE streaming decoder, change:

const decoder = new TextDecoder();

to:

const decoder = new TextDecoder('utf-8', { stream: true });

This is a one-line fix. The { stream: true } option tells TextDecoder to buffer incomplete multi-byte sequences until the next chunk arrives, instead of emitting U+FFFD.

Related issues

• #1716 — UTF-8 Corruption Bug (open 10 months, Cyrillic)
• #40574 — Garbled characters since v2.1.86 (CJK)
• #43746 — TextDecoder missing { stream: true } (CJK, detailed diagnosis)

This issue adds a concrete Cyrillic reproduction case and confirms the same root cause applies to 2-byte UTF-8 sequences (Cyrillic), not only 3-byte (CJK).