fix(a11y): announce that a reply is generating, and that one failed

[opensource-joe]

What and why

The assistant's turn is invisible to a screen reader from start to finish.

There is no live region anywhere in the frontend. No aria-live, no role="status", no role="log", no role="alert", no aria-atomic, in any .jsx or .js outside node_modules, across 589 files.

The pending state is the sharpest case. PromptReply renders it as a bare dot-falling animation:

<div className="mt-3 ml-1 dot-falling light:invert"></div>

That is purely visual, so a user who cannot see it is told neither that the assistant is working nor, later, that it has stopped. The error branch has the same problem from the other end: the turn ends and nothing announces that it ended badly.

I want to be clear that this reads as a gap rather than a position. The same codebase already carries 56 aria-label attributes plus aria-hidden, aria-checked, aria-pressed and aria-current, and sr-only is already used in PromptInput and Toggle. The intent is plainly there. This is one thing that was never wired up.

What changed

Two attributes and a string, both in ChatHistory/PromptReply/index.jsx.

Pending gains an sr-only role="status" carrying the text, and the animation is marked aria-hidden:

<span className="sr-only" role="status">
  {t("chat_window.generating_response")}
</span>
<div
  className="mt-3 ml-1 dot-falling light:invert"
  aria-hidden="true"
></div>

The text sits inside the status region rather than the region wrapping the animation, so it announces once on appearance instead of on every repaint.

Error gains role="alert". Alert rather than status, because a failed reply is the one case here that warrants interrupting the user: nothing else on the page changes to tell them the turn is over.

What this deliberately does not do

It does not announce the streamed reply text. Piping a token stream into a live region makes a screen reader re-read the growing partial message on every chunk, which is worse than silence.

Doing that properly means announcing the completed message once, and the completion boundary is not observable from inside PromptReply: this component re-renders on every chunk as reply grows, and is swapped for HistoricalMessage when the turn ends. So it belongs in ChatHistory, in a separate change. Happy to follow up with it if you want it, and happy to leave it if you would rather that stayed as it is.

Verification

npx prettier --check on both changed files	pass
npx eslint on both changed files	exit 0
npx vite build	exit 0, built in 28s

What I have not done, stated plainly: I have not exercised this with a screen reader against a running instance. Producing a real stream needs an LLM provider credential I do not have. The change is two ARIA attributes and a translation string with no behavioural logic, and the reasoning above is about which region type fires when, so I would rather say that outright than imply a test I did not run. If a maintainer with a running instance wants to sanity-check it with VoiceOver or NVDA before merging, that seems entirely reasonable to me.

i18n

The string is added to locales/en/common.js only. i18n.js sets fallbackLng: "en", so the other 31 locales fall back rather than break. Say the word if you would prefer all 32 populated in this PR.