[moetemp]

What changed?

This PR adds the crate's writer extension calls to the IDL as an internal WIT interface, crates/streams/nexus/stream-writer.wit.

stream-writer maps by @nexus.proto to the extension.proto messages of OpenWriter, WriteRecords, CloseWriter and CloseWriters. open-writer returns own<writer>, a resource whose only constructor field is writer-id, with write-records (the answer has the first and the last offset) and close. The epoch and the sequence stay in the crate.

The interface is marked @nexus.client-only and has no endpoint. Its doc lines say it isn't a Temporal API and that no server serves it. So nexgen generates only a caller and a Writer class for Activities and clients, over an extension-call transport, and nothing on the Workflow or System Nexus path.

It is a second file in its own package, temporal:streams, which uses the Stream service types. In sdk-python, the Stream bindings are imported while temporalio.contrib.streams initializes. The writer models reference temporalio.contrib.streams.proto.v1.extension_pb2, which can't be reached at that point. A separate file gives the writer its own generated package, which the facade imports once the package is loaded.

tests/wit.rs checks the 100-column rule for both files, the client-only marker and the four operation names.

Why?

Decision S87: the Activity and client handles wrap two generated resources, Stream and Writer, so the writer surface comes from the IDL too. Open for Tim: nexgen describing an in-process call surface, not only Nexus operations.

How did you test it?

cargo test --test wit in crates/streams, and cargo fmt --all --check. With moetemp/nexgen#45, the generated sdk-python writer package passes pyright with 0 errors. Over an extension-style transport, it opened a writer, wrote one Record and closed it against a fake extension.

git fetch origin moe/brook/v4/s17-writer-wit && git checkout FETCH_HEAD
cd crates/streams && cargo test --test wit

Stacked on #32 (moe/brook/v4/s16-stream-resource).

[moetemp]

Closing as stale. This layer is folded into the v5 tip, #41, and the upstream review uses a fresh series instead. Plan: https://www.notion.so/4048fc567738814990b7c796ac28fe3f