MCP Server dies of OOM every few days

Thanks to our monitoring, we found that our SaaS MCP server is suffering from a slow and steady memory leak that leads to an OOM kill every few days — unless the buildup is broken by a release.

When the OOM happens, any in-flight requests are dropped, leading to a surge in 5xx alerts and a poor agent/user experience.

## Acceptance criteria

- [ ] Memory usage is flat/stable through the entire MCP server operation.