[epic] Failures are silent (worst mode for screen-reader users)

[ctoth]

Root cause: failures are silent, which is the worst mode for a screen-reader user

Many error and edge paths end at console.log/console.error or drop work with no user-visible signal. For a blind user, a client that silently does nothing is worse than one that reports an error, because it forces guessing.

Instances already filed

• close() leaves local mode armed; sends are silently swallowed #60 send on dead transport drops silently while the command still echoes
• sendData flow-control wait can hang forever if data channel closes #74 sendData flow-control wait can hang with no timeout/rejection
• WebRTC recovery rebuilds connection without re-signalling; dead connection + concurrent recoveries #72 WebRTC recovery loops without surfacing state

Other observed

• Transport-layer errors (ICE failures, peer errors) end at console.log with no UI surface.
• output.tsx:527-536 announces every line via an aria-live region unconditionally, while client.ts also speaks via speechSynthesis when autoread is on — the same line can be spoken twice, and neither path is gated on the other.

Suggested direction

Route transport/transfer failures to a user-visible surface (the output/error store), and make the announce path and the speechSynthesis path mutually aware so a line is spoken once.