In-app voice dictation works on FIRST use per session, fails on every subsequent use (Windows desktop / Cowork)

Environment: Claude desktop app, Cowork session, Windows 11. Built-in laptop microphone. Signed in with a Claude.ai account.

Summary: The in-bubble voice dictation (microphone button in the chat input box) transcribes correctly the FIRST time it is used in a session, but fails on every use after that within the same session.

Steps to reproduce:

1. Start a brand-new Cowork session.
2. Click the microphone in the input box and dictate — works perfectly, for as long as I like.
3. Use the microphone a SECOND time in the same session.
4. The mic activates and the audio level bars respond to my voice (so audio is being captured), but NO words are transcribed into the box — it shows "..." and then stops after about 3 seconds.

Reproduces in fresh sessions every time (1st use works, 2nd+ fail).

What I ruled out by testing:

• Microphone / hardware / permission: the level bars move when I speak, so capture is fine.
• Network: same connection throughout.
• Account / app build: a brand-new session's first dictation works.
• Session age/weight: it's not old/long sessions — it's specifically the 2nd+ dictation in ANY session, including a fresh one.

Expected: dictation works on every use within a session.

Actual: only the first use per session transcribes; subsequent uses capture audio (level meter active) but produce no text and time out after ~3s. The recogniser appears not to re-initialise on the second invocation.

Impact: high — voice dictation is my primary input method.

Possibly related: #38753, #38712.