Stream Announcer

Off by default and switched on by a toggle button whose pressed state is announced: hearing a response as it forms is the listener's choice, and ai-conversation's state-only announcements stay the default. Once on, only complete sentences are announced — the last one is held until the next begins or the stream ends, so a number or a word is never read half-formed. Markdown is read as prose: formatting marks are dropped, links keep their text, and a code block is summarised as its language and line count, since the block itself is there to navigate to. Announcements are paced against an estimated speech rate (`wordsPerMinute`) so the backlog stays in the page, where Pause, Skip and Repeat can reach it, instead of in the screen reader's queue, where nothing can. The region is a plain polite live region present from first paint that only gains nodes — not role=status, which is atomic and would re-read everything, and not role=log, which would add a second transcript to navigate. Skip states how many sentences it will drop in its accessible name, and says how many it dropped.