[andystaples]

Warning

Do not merge this PR, or any PR in this stack, individually. The whole stack will be squash-merged only after every stack PR is approved.

Layer 3/6 of the review stack that splits the closed #293. The base is layer 2/6, #358.

Scope

Dispatch teardown for the per-dispatch durable event loop:

• DurableEventLoop installs the async-generator hooks for each dispatch and restores the host's previous hooks afterwards. close() drains, then cancels, pending tasks and closes tracked async generators. close() is idempotent.
• Cooperative cleanup is capped at about 10,000 loop steps per cleanup phase (more for dispatches with many tasks or generators). Each async-generator sweep spends from a budget fixed at sweep entry, so tasks or generators created during the sweep cannot extend it. Truncated cleanup is reported once as a warning without changing the snapshotted response.
• Cleanup errors are collected per handle. The first unexpected failure is kept; a process-control exception such as KeyboardInterrupt or SystemExit takes precedence, even when grouped or raised after other errors. The existing dispatch response is preserved, and background and cleanup errors are logged.
• wait_durable cannot schedule durable work once the loop is closing. The orchestration context also rejects sequence-number allocation during async cleanup, and exiting an entity critical section during teardown does not schedule extra release messages.
• Control-flow exception classification moves from worker into the shared durabletask.internal.exceptions.is_control_flow_exception. Its BaseExceptionGroup handling is unchanged.

Docs

docs/async-orchestrators.md gains a warning callout describing the cleanup cap and truncation, and documents async-generator ownership during dispatch, explicit aclose() / contextlib.aclosing(...), and teardown handling of entity critical sections and process-control exits.

CI

Adds an async-runtime-windows job to .github/workflows/durabletask.yml that runs the asyncio, replay, and dispatch cleanup tests on Windows across Python 3.11-3.15. Layer 4/6 (#361) extends it with test_async_context.py.

Intermediate adaptations reversed in layer 4/6

Async orchestrators still receive the eager task.OrchestrationContext at this layer, so tests annotate contexts as task.OrchestrationContext instead of task.AsyncOrchestrationContext, and operations that will later go through ctx.as_task_context() (for example call_activity(...) and create_timer(...) that need eager handles) are called directly on ctx.

No changelog changes; release notes are added in layer 6/6 (#364).