feat(threads): persist and restore chat history across summarization …

[buptwz]

Closes #2012
Related to #2424

Summary

When the agent's context window fills up, DeerFlowSummarizationMiddleware removes
old messages from LangGraph state. These messages were permanently lost from the UI —
users could only see the most recent portion of a long conversation.

The existing handler in hooks.ts relied on the internal LangGraph state key
SummarizationMiddleware.before_model and a positional index _messages[2],
which are fragile and can break silently across LangGraph version changes.

This PR replaces that with a two-layer mechanism: a file archive for persistence
across page reloads, and a custom stream event for the live streaming session.

What Changed

Backend

• Add MessageArchiveHook (BeforeSummarizationHook) that appends archived messages
  to {thread_dir}/message_archive.jsonl, deduped by message id. Registered
  unconditionally alongside memory_flush_hook.
• Emit a messages_archived custom stream event in DeerFlowSummarizationMiddleware
  via get_stream_writer() before summary generation. Failure is swallowed so
  summarization is never interrupted.
• Add GET /api/threads/{id}/message-archive endpoint that reads the JSONL file and
  returns {"data": [...], "has_more": false}. Returns empty list for threads that
  predate this feature.

Frontend (hooks.ts)

• Remove the fragile onUpdateEvent SummarizationMiddleware block and summarizedRef.
• Handle messages_archived in onCustomEvent: move archived messages into history
  and reset messagesRef.
• Fetch the archive on mount in useThreadHistory so full history is restored on
  page reload.

Design Notes

• No LangGraph state growth — nothing stored in ThreadState or checkpoints
• No dependency on LangGraph internal event key names or message positions
• Follows existing patterns: same shared-filesystem assumption as memory.json,
  same BeforeSummarizationHook protocol as memory_flush_hook, same custom
  stream mode already used by task_running / llm_retry
• MessageArchiveHook writes to disk before the stream event is emitted, so a
  page reload mid-stream will still recover the archive

Verification

# New tests
pytest tests/test_message_archive_hook.py tests/test_summarization_middleware.py -q

# CI regression suite
pytest tests/test_harness_boundary.py tests/test_docker_sandbox_mode_detection.py \
       tests/test_provisioner_kubeconfig.py tests/test_memory_updater.py -q

[CLAassistant]

All committers have signed the CLA.