aimdb-mcp tools open a new AimX connection per call, which breaks the serial session #451

Problem

Most aimdb-mcp tools open a fresh AimX connection on every call. Over serial this breaks the session that get_record and drain_record share. A UART never closes, so a board learns about a new client only from a hello, and a hello resets the board's session.

So, on a serial:// endpoint, each of these tool calls either:

• fails, because aimdb-mcp's persistent client already has the tty open (tokio-serial opens it exclusively), or
• if it does get through, sends a new hello, which resets the board's session and discards the drain cursor drain_record was reading.

#450 fixed this for get_record and drain_record only. They now share the persistent client from ConnectionPool::get_drain_client(), and that client is replaced only when its engine has stopped (AimxConnection::is_closed()).

Affected callers

All of these go through ConnectionPool::get_connection(), which always calls AimxConnection::connect:

Each also falls back to its own AimxConnection::connect when the pool is not initialized.

The records resource (tools/aimdb-mcp/src/resources/records.rs) also calls AimxConnection::connect directly. It only reads discovered Unix-socket instances today, so it is lower priority, but it should use the same client.

To reproduce

1. Run a board with SerialServer and a ring record with remote access.
2. Through aimdb-mcp, call drain_record on the ring to open its cursor.
3. Publish some events on the board.
4. Call list_records (or any tool in the table) on the same serial:// endpoint.
5. Call drain_record again.

Expected: step 4 succeeds, and step 5 returns the events from step 3.
Actual: step 4 fails because the tty is busy. If it gets through instead, step 5 returns nothing, because the board's session and its cursor were reset.

Proposed fix

• Route every tool and the records resource through the persistent per-endpoint client: get_drain_client(), take the per-client lock, one RPC per call.
• Keep the same rules as fix: record.get answers at once and no longer drains; MCP get/drain share one session (#444) #450:
  • an error reply (not_found, unsupported, denied, even internal) keeps the session;
  • the client is replaced only when is_closed() reports that its engine has stopped.
• Then remove get_connection() / invalidate_connection() and the per-tool fallbacks, so no path opens a second connection. Consider renaming get_drain_client(), since it would no longer be drain-specific.

Tests

Extend tools/aimdb-mcp/tests/serial_session.rs, which puts a SerialServer board behind a pty and counts the hellos it receives:

• open a ring's drain cursor and publish events;
• call each affected tool once;
• drain again.

Every call should succeed, the drain should return the queued events, and the board should receive exactly one hello.

Related

• aimdb-mcp get_record waits 3 s on a single_latest record with no value #444: the 3 s get_record delay
• fix: record.get answers at once and no longer drains; MCP get/drain share one session (#444) #450: get_record / drain_record on one persistent session