Voice (dictate, read aloud) and Draw a picture over /v1, with the app's models scope

The server kept voice and images on /v1 (models scope, which the app's token
already asks for); /me's services and scopes decide what is offered. Rust
sends the recording to /v1/audio/transcriptions, reads /v1/audio/speech back
as a data: URL (CSP media-src data:), and for /v1/images/generations takes
only the id of each of the person's own files from the answer and fetches
/v1/files/<id>/content from the app's own server. The microphone is allowed
only within seconds of the page's own Dictate (voice_arm); WebKitGTK's
media stream is switched on through its settings getter.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

ui/index.html

@@ -84,6 +84,7 @@
           <div class="composer__bar">
             <button type="button" class="tool composer__attach" id="attach" aria-label="Attach files" title="Attach files" data-i18n-aria-label data-i18n-title></button>
             <div class="composer__chips" id="composer-chips"></div>
+            <button type="button" class="tool composer__mic" id="dictate" hidden aria-pressed="false"></button>
             <button type="button" class="round round--stop" id="stop" hidden aria-label="Stop" title="Stop" data-i18n-aria-label data-i18n-title></button>
             <button type="submit" class="round" id="send" aria-label="Send" title="Send" data-i18n-aria-label data-i18n-title></button>
           </div>