[moetemp]

What changed?

This PR adds buffered writers for Activities and clients: OpenWriter, Publish (synchronous), FlushWriter and CloseWriter. A writer appends every batch_interval (100 ms), at once on force_flush, in batches of at most max_batch_size Records and the Store's batch bytes. A Record over the Record limit fails its publish. A reopen of a closed writer goes on from its last sequence. It retries a Store failure with the same sequence for up to max_retry_duration. A full buffer fails a publish with stream_buffer_full, and a flush that failed for good fails every later call. A writer of an Activity attempt above 1 first sends an empty batch.

Why?

A publish must return at once, and one Store call per Record costs too much. The empty batch lets a retried attempt withdraw the earlier attempt's Records, also when it publishes nothing.

How did you test it?

STREAMS_REQUIRE=redis cargo test -p temporalio-streams --test writer, on standalone and cluster Redis.

• Unit Tests
• Staging
• End to End Tests

[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