Accessibility: screen reader support for blind and visually impaired users

## Summary

I'm a blind user who relies on VoiceOver (macOS) for all computer interaction. I would love to use Buzz as part of my daily workflow — both the desktop app and the CLI — but I have no way to evaluate whether the current interface is screen reader accessible.

## What I'm requesting

A deliberate accessibility pass on the desktop client (Tauri + React) to ensure it works correctly with major screen readers: VoiceOver on macOS, NVDA and JAWS on Windows. Specifically:

- All interactive elements (buttons, inputs, channel list, message thread) are reachable via keyboard and announced correctly by the screen reader.
- Dynamic content updates (new messages, agent actions, workflow events) are announced via ARIA live regions without overwhelming the user.
- No information is conveyed exclusively through color or visual icons without a text or ARIA label equivalent.
- The `buzz-cli` is fully usable from the terminal without needing the desktop app as a fallback.

## Why this matters

Buzz's core value proposition — agents and humans working in the same room — is exactly the kind of tool that benefits blind users who already navigate their workflows through text-based interfaces and AI assistants. The CLI-first, JSON-in/JSON-out design of `buzz-cli` suggests the team already thinks in accessible primitives. The desktop client should follow the same philosophy.

WCAG 2.1 AA is a reasonable baseline to aim for. If the team is open to it, I'm happy to test specific builds with VoiceOver and provide detailed feedback.

Thank you for building in the open.

Start by reviewing the Tauri + React desktop client and the buzz-cli workflows, then evaluate them with VoiceOver, NVDA, and JAWS. Done means keyboard navigation, screen-reader announcements, dynamic updates, non-visual labels, and CLI use work without relying on visual information or the desktop app.

Written by the indexing model from the issue text.