The app separates reading from writing. Firestore snapshots handle every read, and every write is a REST command to the domain API, which owns the matching, pricing and list building. The main list and each store’s list stream in parallel, are combined on the device, and are debounced by 500 ms, so a burst of server writes becomes one smooth update.

The optimistic layer is what makes the app feel instant. Placeholder rows are never removed when the HTTP call returns, only when their real counterpart appears in the stream, and that swap happens inside the stream listener, in a single frame. Quantity taps are applied locally, collected per item for two seconds and sent as one request, with an automatic rollback if the server says no. On every change, a totals engine recalculates the combined total, each chain’s share of the combined basket, and the whole list priced at every chain.

The client follows MVVM with a separate use-case layer, so list behaviour is plain, testable Dart. Widgets rebuild at row level, so changing one item repaints one row, not the whole list. The dependency graph is scoped to the signed-in user and rebuilt on account switch, so nothing leaks between accounts. A typed error model tells “the server said no” apart from “we never reached the server”. A single store-name normaliser resolves every way the data sources spell a chain (Greek, Latin, accents, abbreviations) to one code and one logo. The codebase also supports white-label builds, each bound to its own Firebase project.