fix: keep established agent runs attached through disconnects

[zenozhg]

Summary

• keep retrying reattach requests for established agent runs after the short initial reconnect limit is exhausted
• stop durable retries for explicit missing/invalid run responses
• replay a terminal run output summary when text events are unavailable, so completed background work is restored instead of ending with an empty done

Reproduction

A long Codex run continued in the backend for about 15 minutes and completed successfully, but the browser stream disconnected. The run ledger and persisted chat contained the final answer, while the reconnect path stopped early or could return only done.

Tests

• pnpm --filter @mindos/web exec vitest run __tests__/api/agent-runs-reattach.test.ts __tests__/ask/ask-concurrent-sessions.test.tsx
• related Web tests: 102 passed
• pnpm --filter @mindos/web typecheck
• targeted ESLint: no errors