[veeral-patel]

What

An autoscaling poller pool now reports its current target in two places:

1. On each poll — PollerScalingInfo.poller_target on the workflow and activity poll requests.
2. On the worker heartbeat — WorkerPollerInfo.poller_target, for all four poller types.

Why

Scaling suggestions ride on dispatched tasks and the matcher is FIFO, so a worker's share of suggestions tracks its share of pollers. Growth is proportional to current size, which preserves imbalance between identical workers instead of correcting it — measured at 74/12/14 percent across three identical workers, with a different winner each run.

Reporting the target gives the server the one number it needs to compare a worker against the fleet mean and withhold suggestions that would push it further from its share.

The poll-request copy is consumed inside the matching decision and discarded. Reporting it on the heartbeat too lets a follow-up read it per worker and export it as a metric, making fleet-wide poller distribution visible without instrumenting every SDK.

How

Poll path — PollScalerReportHandle::reportable_target() is read at the poll call site in poll_buffer.rs, alongside the timeout override already derived from the same handle, and threaded through poll_fn into PollOptions.

Heartbeat — the heartbeat closure is built in WorkerHeartbeatManager::new before any poller exists, since pollers are created in a deferred closure once namespace capabilities are known. It therefore cannot capture the scaling handles.

That is the same ordering problem last_successful_poll_time already solves, and this copies its solution: a cell per poller type is created up front in Worker::new_with_pollers, cloned into both HeartbeatMetrics and the poller, and the report handle publishes its target into it on construction and on every change.

Design notes

• reportable_target() returns None for a fixed-size pool rather than its static size: such a pool has no target to balance, and the server leaves that worker out of fairness entirely rather than assuming anything about it. That reads as 0 on the heartbeat, which the proto defines as unknown.
• Nexus is in the heartbeat but not the poll path. poll_nexus_task leaves poller_target unused because the nexus request has no such field, but nexus has its own PollScalerReportHandle and a heartbeat slot, so its target is reported there.

Dependency

Mirrors the proto change in temporalio/api#887. Because the API protos are vendored in-tree under crates/protos/protos/api_upstream, the .proto edits are included here so build.rs regenerates and the branch compiles standalone. Those vendored changes should come from a normal api sync before merge, not from this PR.

Testing

cargo check -p temporalio-sdk-core --all-targets passes and the 15 pollers:: unit tests pass. Not yet exercised end to end against a server — that has been done via the Go SDK (temporalio/sdk-go#2799), which shares the wire contract.

[Sushisource]

De llm-ify

[Sushisource]

Suggested change

	/// The pool's current scaling target, reported to the server on each poll so it can
	/// spread poller-scaling suggestions across a fleet by share rather than letting them
	/// land on whichever worker happens to receive the most tasks.
	///
	/// `None` for a fixed-size pool: it has no target to balance, and the server leaves
	/// such a worker out of fairness entirely rather than assuming anything about it.
	/// The pool's current scaling target. `None` for a fixed-size pool.

[Sushisource]

Suggested change

	/// The poller pool's current scaling target, reported to the server so it can spread
	/// poller-scaling suggestions across a fleet by share rather than letting them land on
	/// whichever worker happens to receive the most tasks. `None` for a fixed-size pool,
	/// which has no target to balance; the server then leaves this worker out of fairness.
	/// The pool's current scaling target. `None` for a fixed-size pool.

[Sushisource]

There are enough of these now that all get passed together that it would be good to roll this up into one struct and avoid passing a bajillion of these at each call site.