max / makenotwork

multithreaded: chat message storage and retention The storage half of MT's chat build (wiki livechat-design). No routes and no trait impls yet; this is the layer they sit on. 038 creates chat_messages. BIGINT identity rather than UUID because the reconnect cursor (?after=<id>) needs a total order and chat is log-shaped; this is livechat::MessageId. Three indexes, one per access path that exists: (community_id, id) for backlog replay, (expires_at) for the sweep, (community_id, author_id) for the ban purge, which runs on a hot path at the moment the room is busiest. 039 adds chat_retention_hours and chat_max_messages to communities. 037 gave a community a chat_policy but nothing to build a livechat::Retention from, and the owner settings screen needs somewhere to write. Both NOT NULL with defaults matching Retention::forum_default(), both CHECK-bounded against the crate ceilings, and deliberately no NULL: that would be exactly the "keep forever" sentinel Retention has no way to express. Expiry is computed in SQL as now() + make_interval rather than bound as a chrono timestamp. Not style: the server build unifies sqlx's time and chrono features, so the compile-time macro infers a TIMESTAMPTZ bind as time::OffsetDateTime, and unlike an output column a bind parameter's type cannot be overridden. Computing it server-side keeps all nine statements compile-time checked, which mutations/moderation.rs had to give up. Both halves of retention are one statement each. The age sweep is a global indexed delete, not per-room, since expiry is already on the row. The count cap is one window-function delete across the whole table rather than a query per community per tick, most of which would find nothing. Recompute restamps from created_at, not now(), or an owner touching the setting would silently extend every old message by a full window. 17 storage tests; 311 integration and 178 unit tests pass, clippy and fmt clean.