Models consistently ignore CLAUDE.md English-only rule for .md files

Problem

Models (Sonnet, Opus) consistently ignore explicit CLAUDE.md instructions to write .md files in English only. The rule is clearly stated with strong language:

⛔ ALL .md files and MEMORY: ENGLISH ONLY. NO KOREAN. Violation = critical rule breach.

Despite this, models write markdown content in Korean across all phases of a session — early, mid, and late turns. This is not an edge case; it happens repeatedly across multiple sessions.

Impact

• Wasted tokens: Model generates Korean content → gets rejected (by user or hook) → must regenerate in English. Double the output token cost.
• User frustration: Having to correct the same rule violation every session defeats the purpose of CLAUDE.md project instructions.
• A PreToolUse hook can block the file write, but the output tokens are already consumed by that point.

Reproduction

1. Add to CLAUDE.md: ⛔ ALL .md files: ENGLISH ONLY. NO KOREAN. NO EXCEPTIONS.
2. Ask the model to edit any .md file (docs, README, MEMORY, etc.)
3. Observe: model frequently writes content in Korean despite the explicit rule

Expected behavior

Models should respect CLAUDE.md language directives consistently, especially when marked with strong indicators (⛔, "NO EXCEPTIONS", "critical rule breach").

Environment

• Claude Code 2.1.77
• Windows 11 / VSCode
• Affected models: Sonnet, Opus (across versions)