[lxsaah]

Closes #434. Related to #422 and #438.

Problem

AimX builds late-join snapshots from each record's current value when a client subscribes. WebSocket instead cached the last bytes it sent per (record, topic) in a Mutex<HashMap> on the broadcast path and served snapshots from that. It was a second copy of state the database already holds. With a topic writer, it gained one entry for every distinct topic ever published and never shrank. And it gave WebSocket and AimX clients different initial state.

Change

aimdb-core

• New public OutboundLatest (outbound/latest.rs), built per scheme like OutboundRoutes and using the same route ids. render(id) reads the record's current value (TypedRecord::latest, i.e. buffer peek) and runs it through the link's topic writer and serializer, returning LatestMessage { record_index, topic, payload, metadata }. It keeps no state and needs no with_remote_access(). render_if(id, accept) resolves the topic first and serializes only if accept takes it; a link with a fixed topic is checked before its value is even read.
• ConnectorLink carries the renderer (latest_fn) next to its route factory, built in typed_api.rs from the same serializer and topic writer the live route uses.

aimdb-websocket-connector

• WsSession::snapshots calls render_if with the subscribe pattern for each route the client may read, so non-matching links are never serialized.
• Removed: the snapshot cache (SnapshotCache, DynMapSnapshot), the crate-private SnapshotProvider / NoSnapshot / Snapshot, and server/session.rs. run_broadcast only fans out.
• with_late_join docs describe the new behaviour.

Tests (tests/e2e.rs)

• The fixture gains SingleLatest records (sensors.temp, sensors.humidity, t) and a SingleLatest record with a topic writer (moving). The existing snapshot and golden-frame tests use them and no longer need sleep(100ms), since snapshots read the record directly.
• New ring_records_send_no_snapshot: a ring record's subscriber gets the next live event first, with no snap.
• New topic_writer_snapshot_is_on_the_current_topic: after writing on m.a then m.b, m.# gets one snapshot, on m.b.
• New snapshots_serialize_only_matching_links: a counting serializer shows a non-matching subscribe serializes nothing and a matching one serializes once.
• The snapshot tests in decouple_record_keys_topics.rs, including the permission-filtered one, pass unchanged.

CHANGELOG entries for aimdb-core (Added) and aimdb-websocket-connector (Changed, breaking, wire).

Behaviour changes

• SpmcRing records send no snapshot, as on AimX. A ring has no current value. Records whose subscribers need the current value should use SingleLatest.
• Topic writers: a snapshot is the current value on the topic that value produces now. Topics the record published on earlier get none.
• Exact-topic subscribes still get snapshots on WebSocket. Doing the same on AimX is AimX exact-topic subscribes send a late-join snapshot, like wildcards and WebSocket #438.
• Frame format unchanged.

Not changed

• No public API is removed: the snapshot types were crate-private.
• aimdb-weather-mesh's hub uses rings, but only with link_from, not WebSocket link_to, so it is unaffected.

Merging alongside #437

#437 touches record_query in server/dispatch.rs and the aimdb-core CHANGELOG. A trial merge of both branches applied cleanly.

Tested locally

• cargo test -p aimdb-websocket-connector --all-features (e2e run 3×)
• cargo test -p aimdb-core with the Makefile's four feature sets (alloc, std,tracing, std,tracing,observability, alloc,observability)
• cargo clippy -p aimdb-core -p aimdb-websocket-connector -p aimdb-mqtt-connector --all-targets --all-features -- -D warnings
• cargo clippy -p embedded-check-fixture --target thumbv7em-none-eabihf -- -D warnings
• cargo doc -p aimdb-core -p aimdb-websocket-connector --no-deps --all-features
• cargo fmt --check

The full suite runs in CI.

🤖 Generated with Claude Code