events that cannot precede their commit

A live board for staff is a latency problem that is really a truth problem. The obvious implementation is an in-process broadcast: when state changes, publish. It is fast, it is simple, and it makes it possible for a board to display a transition that is later rolled back, or that never committed at all, because publication and durability were two separate events with no ordering between them.

This feed was built the other way. The mechanism is a poll of committed snapshots plus a pure diff: read what is durable, compare with what was last sent, emit the difference. Events can therefore only ever be produced from rows that exist. A transition cannot be observed before it is committed, because an uncommitted transition is not in the snapshot. That guarantee is not an implementation detail to be optimised away later; it is the reason the design was chosen.

The consequences were accepted deliberately. The stream carries a keep-alive while idle, so intermediaries do not sever quiet connections. On connect there is a baseline snapshot and no history obligation — the feed does not promise replay, so a client that needs certainty reconciles with a read. And there is a polling loop running even when nothing happens.

Cost: seconds of latency a broadcast would not have spent, and a client that must know how to reconcile rather than assume it saw everything.

What it proves: where freshness and truth collided, the team chose truth and wrote down what it cost.