fix(langgraph): regenerate first assistant response

[prajvalrasik]

Summary

Fixes #2109.

Regenerating the first assistant response can send the same number of messages as the checkpoint:

• checkpoint: user message + assistant response
• incoming request: same user message + empty assistant placeholder

Because the message counts are equal, the existing count-based regenerate heuristic is skipped and prepare_regenerate_stream() is never reached.

This change detects the explicit empty assistant placeholder case and routes regeneration from the preceding checkpointed user message.

Notes

I kept the user only path unchanged because it overlaps with the existing SSE-drop recovery behavior where the client can be behind the checkpoint. This PR handles the unambiguous placeholder signal.

Tests

uv run python -m unittest tests.test_prepare_stream_edited_message
uv run python -m unittest tests.test_oss28_sse_drop_recovery tests.test_prepare_stream_interrupt_resume tests.test_resume_input
uv run python -m unittest discover tests -v

Result:

383 tests passed

[prajvalrasik]

Hi maintainers, just checking in on this PR when you get a chance.

This fixes the first-assistant-message regenerate case from #2109 and keeps the existing user-only/SSE-drop recovery behavior unchanged. Local LangGraph integration tests pass. Happy to adjust if you want the detection handled differently.