[BUG] Intermittent Korean syllable corruption in AskUserQuestion — wrong valid characters generated (not U+FFFD)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [ ] I am using the latest version of Claude Code

What's Wrong?

When Claude generates long Korean prose in AskUserQuestion parameters (question text + option labels + option descriptions), it intermittently produces wrong-but-valid Korean syllable blocks instead of the intended characters.

Examples:

• 공유 (gong-yu, "shared") → 겵유
• 완전 독립 ("fully independent") → 방독립

Unlike #46863, these are NOT U+FFFD replacement characters. The model generates syntactically valid Korean characters that are semantically wrong — suggesting a token sampling error, not a stream buffer split.
Adding self-review instructions to the system prompt ("check Korean text before submitting") has no effect — the model cannot detect its own incorrectly generated syllable blocks.

What Should Happen?

Korean syllable blocks in AskUserQuestion parameters should be generated correctly without random character substitution.

Error Messages/Logs

Steps to Reproduce

1. Create a skill that calls AskUserQuestion repeatedly with 3–4 options, each containing a 1–2 sentence Korean description.
2. Run the skill multiple times.
3. Observe intermittent Korean character corruption in the generated option text.

Frequency increases proportionally with the amount of Korean text generated per call. Skills generating many consecutive AskUserQuestion calls with detailed Korean descriptions reproduce this far more reliably than skills with short or English text.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.92 (Claude Code)

Platform

Other

Operating System

macOS

Terminal/Shell

Terminal/Shell: cmux (zsh)

Additional Information

_No response_