[J2TeamNNL]

Summary

• New module Live Messages, off by default (id liveMessages). With it on, Database > Live Messages… opens a sheet that takes channels or topics, then a Live tab that lists messages as they arrive: time, channel or topic and payload, plus pattern on Redis and partition, offset and key on Kafka. With it off, no menu item, observer or connection is left behind.
• The tab has Pause and Resume, Clear, a filter field, Copy (selected rows, or all, as tab-separated text, also Command-C), Export as CSV, and a field to send a test message (NOTIFY, PUBLISH, or a produce on Kafka).
• The tab keeps the newest 10,000 messages. Pause freezes the list only, so messages keep arriving and resuming loses nothing up to the cap.
• A subscription always uses connections of its own, opened like the routine debugger's: never the session's query connection, never pooled, no statement timeout. Closing the tab, turning the module off or quitting closes them.
• A dropped connection adds a "Connection lost" row and subscribes again with a backoff of 1 s doubling to 30 s. A subscribe that fails on the first try (a wrong topic, say) ends the tab with the error instead of retrying.
• PluginKit: a new optional protocol PluginLiveMessaging with PluginLiveSubscription and PluginLiveSessionSource. It is a separate protocol, like PluginRoutineDebugging, so it adds no requirement to existing drivers. No kit bump: main is already at 41, one past the 40 that v0.84.0-beta.1 shipped.
• PostgreSQL: runs LISTEN per channel (quoted as an identifier), then reads PQnotifies after polling the socket in 250 ms slices on the connection's own queue. NOTIFY goes through pg_notify($1, $2). PGlite opts out, since a second session opens a different database. Redshift and CockroachDB do not conform.
• Redis: SUBSCRIBE and PSUBSCRIBE on a dedicated connection, read by polling the socket and draining hiredis's reader, never through a hiredis read timeout, which would leave the context unusable. PUBLISH opens a second connection on first use, since a subscribed connection refuses it. Cluster subscribes on one node, which gets every classic PUBLISH.
• Kafka: lists each partition's end offset, then fetches forward per partition with plain Fetch requests and no consumer group, so no committed offset moves. A non-UTF-8 key or payload reads as hex, as in the topic grid.

Defaults I picked: newest rows at the top, so a new message is in sight without scrolling; on Kafka, a reconnect starts at the end again and the gap row marks it.

Docs: TableProApp/docs#157 (draft, merges after this one).

Test plan

• New tests: LiveMessageBufferTests (cap, ids, gap, filter), LiveMessageControllerTests (pause keeps messages, a drop leaves one gap row and resubscribes after the backoff, a first failure ends the tab, stop closes the subscription), LiveMessageExportTests, LiveMessagesModuleResidueTests (no menu item while off, toggling adds and removes it), PostgreSQLListenStatementsTests, RedisLivePushTests.
• Live, against PostgreSQL 17, Redis 7 and Kafka 3.9 in Docker over an SSH tunnel, by loading the built plugin bundles in a small harness that calls subscribe with a session source of fresh drivers:
  • PostgreSQL: a NOTIFY from psql arrived about 200 ms after it was sent, a channel named Audit Log worked, and pg_terminate_backend ended the stream with an error, which the tab answers with a gap row and a resubscribe.
  • Redis: PUBLISH from redis-cli to a channel and to a PSUBSCRIBE pattern both arrived within 100 ms, the test PUBLISH went out on a second connection, and CLIENT KILL TYPE pubsub ended the stream with an error.
  • Kafka, a 2-partition topic: the 3 messages produced before the follow were skipped, new ones arrived with partition, offset and key, and kafka-consumer-groups --list was empty before and after.
• Not checked here: the Live tab on screen. This Mac's shell has no Screen Recording permission, so there are no screenshots. verify.sh plugins was not run locally (low disk); the Kafka plugin was built on its own.

Fixes #3614