[BUG] Desktop: Reply opens an inline popup without dictation

## Preflight Checklist

- [x] I have searched existing issues and did not find this specific regression.
- [x] This is a single bug report.
- [x] I reproduced this on the Claude Desktop version delivered by silent auto-update in my environment. The bundled Claude Code CLI version is listed below.

## What's Wrong?

In the **Code** tab of Claude Desktop, selecting text from an assistant response shows a contextual popover. In the Japanese UI, choosing **返信** (`Reply`) now opens a small inline reply editor next to the selected text.

That inline editor does not expose the Dictation control that is available in the main composer. There is no microphone button or other in-app path to start Dictation from the popup.

This is a regression in a speech-first workflow. Previously, choosing `Reply` inserted the selected text as a quote in the main composer at the bottom of the window. I could then add text with Dictation and send the reply from the normal composer.

Now, each quoted reply requires copying the selected text, pasting it into the main composer, and reconstructing the context before Dictation can be used.

## What Should Happen?

Either:

1. `Reply` should again quote the selected text into the main composer, or there should be a setting to restore that behavior; or
2. the inline reply popup should provide the same Dictation capability as the main composer.

## Error Messages/Logs

None. This is a UI behavior and accessibility regression; no error message is shown.

## Steps to Reproduce

1. Open an existing conversation in the **Code** tab of Claude Desktop for Windows.
2. Select part of an assistant response.
3. In the contextual popover, choose **返信** (`Reply`) in the Japanese UI.
4. Observe that a small inline input popup opens near the selected text instead of placing a quote in the main composer.
5. Try to use Claude Desktop Dictation in that inline input popup.
6. Observe that the popup has no microphone button or other in-app Dictation entry point, while the main composer does.

## Claude Model

Not sure / Multiple models. This appears to be a Desktop renderer UI issue rather than a model-specific issue.

## Is this a regression?

Yes, this worked in a previous version.

## Last Working Version

Unknown. The previous main-composer behavior was present before the silent Desktop auto-update installed on 2026-09-14.

## Claude Code Version

2.1.270 (bundled with Claude Desktop 1.52386.6.0)

## Platform

Other — Claude Desktop Code tab (not an API platform)

## Operating System

Windows — Windows 11 Pro 10.0.26200

## Terminal/Shell

Other — Claude Desktop UI; no terminal is involved in reproduction.

## Environment

- Claude Desktop: 1.52386.6.0
- Package: `Claude_1.52386.6.0_x64__pzs8sxrjxfjjc` (MSIX)
- Update: silent auto-update installed 2026-09-14 at 08:52 JST
- Bundled Claude Code CLI: 2.1.270
- UI language: Japanese
- Regression noticed: 2026-09-15

## Additional Information

Related issues, but not duplicates of this regression:

- #86642 documents the previous behavior: selecting text and choosing `Reply` quoted the selection into the prompt field. That issue requests keyboard shortcuts; this report is about the new inline editor removing access to Dictation.
- #90761 is a broader Desktop keyboard-operability tracking issue.
- #90576 requests a Dictation keyboard shortcut. This report concerns Dictation being unavailable specifically in the new inline `Reply` editor.

I could not find this UI change documented in the public release notes or changelog as of 2026-09-15. I am requesting either a way to restore the previous main-composer behavior or Dictation support inside the inline reply popup.

No source files, tests, or entry points are identified in the report. First reproduce the Japanese Reply flow in Claude Desktop 1.52386.6.0 on Windows 11, then locate the Desktop renderer implementation for the contextual popover, inline reply editor, main composer, and Dictation control; done means the selected text can be replied to with Dictation through one of the requested behaviors.

Written by the indexing model from the issue text.