fix(redis): explorer views share one socket, so SELECT + command is not atomic; folder delete can hit another DB #1569

Problem

RedisService.acquire keys sockets by connectionId::role (lib/core/database/redis_service.dart:26). Every Redis explorer for a connection, for any database index, uses the same explorer socket. Views select the database with a separate command before working:

await widget.connection.selectDatabase(widget.database); // SELECT n
await widget.connection.delMany(keys);                    // DEL ... (several round trips)

SELECT is socket state, and the pair is not atomic. Anything else that runs on the socket between the two awaits changes the database the next command hits.

1. Folder delete can delete keys in another database (data loss)

RedisKeysView._deleteFolder (lib/features/redis/redis_keys_view.dart:218-225) selects the DB once, then delMany (lib/core/database/redis_connection.dart:749-761) sends sequential DEL chunks of 200, awaiting each.

Steps: open db0, load a folder with more than 200 keys, Delete folder, then while it runs click db3 in the tree. The workspace replaces the explorer (ValueKey('redis_<id>_db_3'), workspace_panel.dart:363). The old view is disposed but its delMany loop keeps going (no cancellation). The new view sends SELECT 3 + SCAN on the same socket, and the remaining DEL chunks run in db3, deleting any keys there with the same names.

2. The CLI runs commands in a database other than the one it shows

RedisCliWorkspace tracks _currentDb only when the user types SELECT (redis_cli_workspace.dart:487-490) and sends every other command as is (:512). The key editor and keys view in the same explorer re-select widget.database before each operation. After SELECT 3 in the CLI, opening any key moves the socket back to db0. The CLI prompt still says db3, but the next DEL / FLUSHDB runs in db0.

3. delMany itself

• DEL frees large values synchronously and blocks the Redis event loop. UNLINK (Redis >= 4.0) is not used anywhere in the codebase.
• Chunks are sent one after another although the class already pipelines in typesAndTtls.
• On a failure midway the UI does not know which keys are gone: _keys is not updated and the error hides the partial delete.

Fix

• Make "select DB + command(s)" atomic: either one socket per (connectionId, role, db) (no SELECT switching at all), or a per-socket async lock with runInDb(db, () async {...}) that issues SELECT inside the lock. One socket per DB is the simpler model; the CLI then gets its own socket and SELECT there only moves the CLI.
• delMany: UNLINK with fallback to DEL on ERR unknown command, pipeline the chunks, stop when the owner is disposed (cancel token), return the count/keys actually removed and reload the list on partial failure.
• Add UNLINK to RedisConnectionTestFake (redis_connection.dart:817).
• Tests: interleaved SELECT from a second user of the socket during delMany; CLI SELECT 3 followed by an editor action.