[jakthom]

Summary

• maintain compact per-connector queue aggregates transactionally so live statistics no longer scan retained envelope rows
• lower the default retention cap from 100,000 to 10,000 and show the effective defaults and retention behavior in the UI
• keep exact traffic counters while sampling expensive source field/raw diagnostics, bounding cardinality, and limiting rich snapshots to requested sensors
• replace idle connector polling with append notifications and reduce UI polling, stream buffers, DOM churn, serialization work, and listener lifetimes
• expose authored and effective connector buffers separately through MCP, scope filtered source-metric reads before rich snapshot work, and add get_latest_payloads

MCP latest-payload contract

get_latest_payloads returns exactly one current decoded payload per configured-source/sensor/PGN stream. sensor_id is the stable Device NAME hex when known and address:<source_address> otherwise. Calls can filter by source_id, sensor_id, source_address, device_name_hex, and PGN. Beacon overwrites one bounded in-memory payload per stream on every message; it does not accumulate payload history, and values reset on restart.

Validation

• go test ./...
• race-enabled tests for queue, stats, connector, UI, store, config, and MCP packages
• go vet ./...
• node --check internal/ui/assets/app.js
• fresh-database migration and rendered connector-form smoke test

Migration notes

Schema migration 5 creates queue_aggregates and performs a one-time backfill from the canonical queue and checkpoint tables. Normal append, acknowledgement, and prune operations then maintain the aggregate row in the same transaction as the underlying queue state.

Retention limits apply to retained rows, including messages still pending delivery; the UI and docs now make that explicit.