The ORM is async end to end

[czpython]

Flag-day switch of every ORM session to SQLAlchemy's AsyncSession on the serving loop. The sync pool checkout was the last blocking call left in the process: under load it froze the event loop in pool_timeout slices. Now a checkout wait suspends the task — exhaustion is backpressure, not a freeze.

• Async engine (create_async_engine, pool 20+30) with the sync engine kept only for one-shot processes: migrations, CLI, doctor's plain checks.
• db_session is an async_scoped_session keyed by task. Each request binds its own session and commits at the boundary; durable steps keep the one-transaction-per-step shape through step_session.
• DBOS.launch() runs inside the async lifespan so async steps share the serving loop; schedule reconciliation uses the *_async schedule APIs.
• Model and read surfaces are async def throughout; lazy relationship reads go through awaitable_attrs where a load is intended, and Workflow.subject resolves as an awaitable.
• Test harness: per-test async engine + savepoint-joined session, and an ASGI-transport client on the test's own loop.

1329 backend tests green; pyright error count below the main baseline.

ENG-672.

🤖 Generated with Claude Code