[moetemp]

What changed?

This PR adds @nexus.one-way on WIT operations. In Python, both Workflow stubs (_<op>(request) and the keyword form) become synchronous and return None. Each stub still calls nexus_client.start_operation(...), so the call goes through every outbound interceptor, and then drives the start coroutine once with send(None). When an interceptor takes the call and completes it without suspending, the stub returns. When the start suspends, the stub closes it and raises RuntimeError.

TypeScript and .NET accept the directive and keep their async stubs. The directive is refused with arguments, with @nexus.long-poll, with @nexus.output-transform, or on an operation that returns a resource. Python refuses a one-way operation with generic models or function fields.

Why?

publish is synchronous in Workflow code, and today each facade must start the async stub as a background task. A Brook plugin buffers AppendRecords and CloseStream in its outbound interceptor and returns a done handle, so the start never suspends.

How did you test it?

tests/wit_streaming.rs checks the stub shape and the refusals. The new Python tests fail on #37. I also ran a live check on a dev server with temporalio 1.34.0: a Workflow calls the stub three times with an outbound interceptor that records the call and returns at once, and it completes with three calls taken. Without the interceptor the Workflow Task fails with the stub's message.

git fetch origin moe/brook/v3/n3-one-way && git checkout FETCH_HEAD
cargo test --features advanced --test wit_streaming
cargo validate markdown

Stacked on moe/brook/v3/n0-base, which merges #30 to #35 onto main. The branch merges each of them again when it changes. The only conflict is the union of the tests that #31 and #33 append to tests/generate_python.rs.

• Unit Tests
• Staging
• End to End Tests

[moetemp]

Reverted out of the stack. The Stream WIT doesn't use @nexus.one-way.