fix(memory): roll back pooled MySQL connections on release (follow-up to #296)

[fede-kamel]

Summary

Follow-up to #296. Fixes a connection-pool transaction-management bug in the MySQL checkpointer that causes an idle-in-transaction → metadata-lock deadlock (and latent stale reads), and fixes the cold-start adapter integration test that was hiding it.

The bug

_MySQLConnectionPool returned connections to the pool without ending their transaction. With the default autocommit=False, every read (load, exists, list_threads, count, get_metadata, query_by_metadata, search_data) — and the SELECT 1 health check added in #296 — opens a transaction that holds a SHARED_READ metadata lock on the tables it touches. A connection pooled in that state sits idle in transaction, which means:

1. A later DROP/ALTER blocks indefinitely behind the pending EXCLUSIVE MDL.
2. The next borrower reuses the pinned REPEATABLE READ snapshot and can read stale data.

How it was found

The new test_mysql_adapter_concurrent_same_thread_cold_start from #296 never actually ran the pool: it mutated a frozen AgentState (state.agent_id = ...), raising ValidationError before touching MySQL — so CI (unit-only) stayed green and the integration suite errored out early. Once that test is corrected to use the immutable model_copy(update=...) pattern, it reaches the pool and hangs.

Reproduced against MySQL 9.6:

performance_schema.metadata_locks
+----------------+-------------+-------------+
| OBJECT_NAME    | LOCK_TYPE   | LOCK_STATUS |
+----------------+-------------+-------------+
| t_xxxxxxxx     | SHARED_READ | GRANTED     |   <- idle pooled conns (Sleep, trx RUNNING)
| t_xxxxxxxx     | SHARED_READ | GRANTED     |
| ...            | ...         | ...         |
| t_xxxxxxxx     | EXCLUSIVE   | PENDING     |   <- DROP TABLE, blocked ~forever
+----------------+-------------+-------------+

The fix

• Roll back any open transaction when a connection is returned to the pool (_release), so every pooled connection is handed out clean; discard it if the rollback fails. This is standard pool hygiene and fixes both the MDL deadlock and the stale-read window regardless of the autocommit setting.
• Factor the connector error tuple into a shared _connection_errors() helper used by both _is_healthy and _release (a dead-connection rollback() raises mysql.connector.Error, which the previous narrow except would have let escape the finally).
• Fix the cold-start adapter test to use model_copy(update=...) instead of mutating a frozen model.

Validation

• MySQL integration suite (LOCUS_MYSQL_INTEGRATION=1) against MySQL 9.6: 11 passed, stable across repeated runs (previously: 1 hang + 1 error).
• Standalone stress repro of the previously-hanging scenario: 5/5 runs complete.
• Full unit suite: pass, total coverage 93.2% (≥90% gate); added 2 regression tests for _release (rollback-on-return, discard-on-rollback-failure).
• ruff, ruff format --check, and hatch run typecheck all clean.

[luigisaetta]

Nice fix. Rolling back on pool release is the right boundary: it protects reads, health checks, and any future code path that borrows a connection without having to remember transaction cleanup locally.

[luigisaetta]

ok approved, great catch!