0.0.14 (2026-02-07)

Fixes

fix: prevent infinite gRPC self-loop in shard routing

When shard_assignments marked a shard as owned by the local runner but owned_shards had not yet been populated (during the acquire phase window), send(), notify(), and interrupt() would route the message as "remote" via gRPC to the local runner itself. The gRPC handler called back into sharding.send(), hitting the same code path, creating an infinite loop that grew memory at ~350MB/s.

Added self-send guards to all three routing methods: when get_shard_owner_async() returns the local runner's address, route locally instead of making a gRPC call. Also handles persisted messages correctly in the self-routed path.

Concurrent acquire_batch/refresh_batch using FuturesUnordered with a concurrency limit of 64, reducing 2048-shard acquisition from 30+ seconds to under 5 seconds.

fix: resolve stale state in workflows after activity calls

self.state in #[workflow] and #[rpc] methods previously captured a
point-in-time snapshot at method entry via ArcSwap::load(). After calling
an activity that mutated state, subsequent reads of self.state within the
same workflow would still return pre-activity values.

Introduces StateRef<S>, a read-only proxy wrapping Arc<S> that implements
Deref<Target=S>. The macro-generated activity delegation methods now refresh
the proxy after each activity commits, so self.state.field always reflects
the latest committed state. User-facing syntax is unchanged.