fix: prevent SSE event loss in connectEvents onerror handler

[fallleave001]

Problem

The connectEvents onerror handler unconditionally called settle() +
es.close() on every error, then manually reconnected after 1s. This
created a race condition:

1. A recoverable network glitch triggers onerror
2. settle() resolves the Promise immediately → handleSend proceeds
   to send the prompt
3. es.close() kills the SSE connection → all events emitted during
   the 1s reconnect window (agent_start, message_*, prompt_done,
   agent_end) are lost
4. The UI stays stuck in "waiting for model" until
   reconcileAgentState (15s polling) recovers

Fix

Distinguish between fatal errors (readyState === CLOSED, e.g. 404/500)
where manual reconnect is needed, and recoverable errors (CONNECTING)
where EventSource auto-reconnects on its own.

For recoverable errors we no longer settle() or close() — the 1.5s
timeout in the Promise constructor still ensures handleSend won't
hang. This eliminates the 1s event-loss window that was the root cause
of the stuck UI.

Testing

• Recoverable error: simulate network glitch (offline/online in
  DevTools) during an active run → SSE auto-reconnects, no events lost,
  streaming continues
• Fatal error: return 500 from events route → manual reconnect kicks
  in, UI recovers via reconcile polling if needed
• Normal flow: new session + prompt → streaming works as expected