[BUG] Japanese voice dictation inserts a full-width full stop 。 at the start of each new dictation segment (macOS desktop app, Code tab)

Product: Claude desktop app (macOS), Code tab: the prompt box of a Claude Code session, using the app's own microphone (dictation) button
App version: Claude 2.31226.1 · Claude Code 2.1.296
OS: macOS 26.6.2 (build 25G83)
Keyboard / input source: macOS Japanese IME (com.apple.inputmethod.Japanese)
Dictation language: Japanese
Started after: the most recent desktop app update (the user did not see this before it)

What happened

When dictating in Japanese, a full-width full stop 。 (U+3002) is inserted at the start of the transcription when a new dictation segment begins after a pause or after the recording is stopped and started again.

Concrete example from a message sent today (13 Oct 2026). The user dictated two paragraphs. The sent message was:

Google からメールが届いた。（…first paragraph, ends normally…）見といてもらえる？

。あと、この Google のメールと関係ないんだけど、（…second paragraph…）

The first paragraph starts correctly. The second paragraph starts with 。あと、, i.e. a sentence-ending full stop appears before the first spoken word. The user did not say or type that character.

Expected

The transcription of a new segment starts with the first spoken word. A sentence-final 。 belongs at the end of the previous sentence (where the transcriber already put one), not at the start of the next segment.

Reproduction

1. In the Code tab prompt box, with the dictation language set to Japanese, start dictation and say one or two sentences.
2. Pause, then continue speaking (or stop and restart the recording) and say another sentence.
3. Look at the text in the box: the second segment begins with 。.

It happens every time the user dictates more than one segment; it is not a one-off.

Notes

• It looks as if the sentence-final punctuation emitted for the end of the previous segment is attached to the beginning of the next segment instead of the end of the previous one.
• Possibly related to #100792 and #97323 (pending dictation segments being re-inserted on macOS), since all three involve the boundary between dictation segments. This report is specifically about Japanese punctuation at the start of a segment.