Voice typing (Win+H) inserts romanized/garbled text instead of Japanese in chat input (VSCode/Cursor extension)

Environment

• Cursor IDE: v3.9.8
• Claude Code extension: v2.1.195
• OS: Windows 11
• Audio input: WO Mic (virtual microphone, phone over network)
• Dictation method: Windows Voice Typing (Win+H), language set to Japanese

Steps to reproduce

1. Open the Claude Code chat panel in Cursor (VSCode-based extension)
2. Focus the chat input box
3. Press Win+H to open Windows Voice Typing, confirm language is set to 日本語 (Japanese)
4. Speak in Japanese

Expected behavior
Recognized speech is inserted as Japanese text (hiragana/katakana/kanji), same as it does in Notepad or in a browser tab.

Actual behavior
The recognized text is inserted as romanized/garbled Latin characters (e.g. saying "こんにちは わかる" produces "connitua vakaru") instead of proper Japanese.

Additional notes / isolation

• Keyboard-based IME input (romaji typing + conversion) works correctly in the same chat input box.
• Voice typing works correctly in Notepad and in a browser tab (e.g. claude.ai) with the same microphone/setup.
• The issue appears isolated to how Windows Voice Typing inserts text into this specific webview input field, as opposed to keyboard-driven IME composition.
• Both Cursor and the Claude Code extension are on their latest available versions, so this is not a stale-version issue.
• Taskbar language indicator stays on Japanese throughout, regardless of which window is focused.