Used fakeredis as the local fallback behind a command-counting proxy when no real Redis URL is configured. Pure-Python install, drop-in for the regular client, and the per-request command counts it produced were identical to those later observed against a real Redis 7.2 server.
- What worked
- Behavioural parity with the real server for get/set/delete patterns meant budgets measured locally held exactly in the real-backend run.