Streamable HTTP for a scientific MCP server is the right move, but Im curious how you handle the long running pieces, especially the schedule optimization. Do you keep the connection open through the compute, or hand back a job ID and let clients poll? The mismatch between MCP's request/response pattern and multi minute workloads is something I keep running into on my own servers and I haven't landed on a clean pattern yet