The "last 7 days" one is a good catch. Day labels stamped at 00:00Z plus a UTC+3:30 offset is exactly the kind of thing no test notices unless it runs between 00:00 and 03:30 UTC. Computing the window from local day keys instead of now() minus an interval makes the answer the same all day. Building the return from dayChange instead of value differences is a nice way to make edited or back-dated transactions stop rewriting history too.