CaseMark is locking the doors before launch

Most of CaseMark's recent work isn't a new feature you'd notice - it's the unglamorous plumbing that decides whether a public deployment embarrasses you. The team fixed a bug that was silently dropping chat history from matters, added rate limits so nobody can hammer the AI endpoints, and tightened how errors reach users: instead of leaking raw internal messages, the system now only returns a vetted set of safe ones.

There's also a correctness fix that matters to anyone reusing this code. The fork was quietly corrupting structured data when run against a standard database rather than the hosted setup it was built on - the kind of failure that doesn't throw an error, it just rots your records. That's now handled, along with cleaner handling of mid-conversation failures so the chat doesn't freeze half-finished.