### Check for existing issues

- [x] I searched open and closed issues before filing - [x] [#25553](https://github.com/BerriAI/litellm/issues/25553) is related cache-coherence work for virtual keys but does not cover project objects or `/project/update`

### What happened?

The enterprise `POST /project/update` endpoint updates the project row and its budget or object-permission rows, then returns without refreshing or invalidating `project_id:{project_id}`

Authentication resolves projects through that cache key first. `_run_project_checks` then trusts the cached `blocked`, `models`, and budget values. A warmed project can therefore continue using its previous authorization state after an admin blocks it, restricts its models, or lowers its budget

The stale direction also affects availability. Unblocking a project or widening its model access can continue to reject requests until the management-object cache expires

### Steps to Reproduce

1. Create an enterprise project with `blocked=False` and attach a key 2. Send one LLM request to warm `project_id:{project_id}` 3. Call `POST /project/update` with `blocked=True` 4. Confirm the database row is blocked 5. Send another request be…