fix: make codex send replays idempotent

[TheVortex8]

Summary

• make send_agent_message_request idempotent when the same client messageId is replayed
• reject messageId reuse with different text instead of silently starting a new run
• allow Codex terminal events to end a stream even if a fresh turn_started event was not observed first
• add regressions for duplicate-send replay and terminal-before-start Codex behavior

Root cause

On reconnect/replay paths, the daemon could accept the same logical user message twice because messageId was only used for timeline recording, not run idempotency. That could start a second Codex run for the same message.

The Codex app-server provider also buffered all pre-turn_started events except permissions. If the replayed/resumed run surfaced only a terminal event (turn_completed, turn_failed, or turn_canceled), the provider ignored it and could leave the agent stuck in running.

Testing

• npx vitest run src/server/agent/providers/codex-app-server-agent.test.ts
• npx vitest run src/server/daemon-e2e/send-message-id-replay.e2e.test.ts --maxWorkers=1 --minWorkers=1
• npm run typecheck --workspace=@getpaseo/server