[BUG] Dictation captures audio but returns no transcript (macOS desktop 1.52386.0) — regression persists from 1.49585

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code

### What's Wrong?

The in-app microphone button in the composer opens the microphone, audio capture
starts, but no transcript is ever returned. No error is shown in the UI and none
is written to the app logs.

Dictation broke ~2 days ago (around build 1.49585) and is NOT fixed in 1.52386.0
(build 1003ca, 10 Sep 2026).

Same symptom as #93170 (Windows, 1.49585). A separate dictation regression on
macOS in 1.49585 is reported in #93165. No dictation issue has been filed yet
against 1.52386, so this build does not carry a fix.

Most diagnostic detail: dictation stops by itself after ~20s of continuous
speech, exactly as if silence had been detected. The app is therefore processing
an empty audio stream while the microphone is genuinely open.

Measurements:
- App logs: microphone capture is timestamped on every attempt, no errors.
- Activity Monitor: zero data sent during ~20s of dictation.
- Likely same mechanism as #61370: audio buffered locally, WebSocket connects
too late, "Final transcript assembled (0 chars)".

Ruled out (tested without success):
- Cmd+Q and relaunch; full machine reboot
- tccutil reset Microphone (app bundle id)
- With and without AirPods, input device stabilised on the internal mic BEFORE
launching the app (per #80366)
- Workaround from #93170 (keeping another app holding the microphone open)

Works on the same Mac, same microphone, same account, same network:
- claude.ai in Safari
- native macOS dictation
- Microsoft Teams video calls

### What Should Happen?

Speech should be transcribed into the composer, as it did before build 1.49585.

### Error Messages/Logs

```shell

```

### Steps to Reproduce

1. Open Claude for Mac 1.52386.0
2. Click into the composer
3. Click the microphone button in the toolbar
4. Speak continuously for 20-30 seconds
5. Observe: dictation cuts out on its own after ~20s, no text appears,
no error is shown

### Claude Model

Other

### Is this a regression?

Yes, this worked in a previous version

### Last Working Version

Claude for Mac build prior to 1.49585 (~8 Sep 2026)

### Claude Code Version

Claude for Mac build prior to 1.49585 (~8 Sep 2026)

### Platform

Anthropic API

### Operating System

macOS

### Terminal/Shell

Other

### Additional Information

_No response_

Reproduce from the microphone button in the composer using the listed macOS steps, then inspect app logs and Activity Monitor during the 20-second cutoff. The report points to locally buffered audio, a late WebSocket connection, and "Final transcript assembled (0 chars)"; done means speech produces composer text without an error or premature stop. No source files or tests are named.

Written by the indexing model from the issue text.