v0.2.2

Changed

• Catch up to current sibling workspaces. Path-dep version pins
  bumped to atomr 0.3.1, atomr-infer 0.6.0, atomr-accel
  0.3.1. No source changes were required for the bump itself —
  atomr-agents existing usage of the sibling APIs is forward-compatible.
• Event::AgentTurn carries two new u32 fields, reasoning_tokens
  and cached_tokens, sourced from atomr_infer_core::tokens::TokenUsage.
  Both are #[serde(default)] so existing event JSON deserialises
  unchanged. Surfaces o1-style reasoning-token billing and Anthropic
  prompt-cache / OpenAI cached-input pricing in cost reports.

Added

• Native aarch64-Linux CI coverage. release.yml builds the
  atomr-agents CLI binary AND the Python wheel on
  ubuntu-22.04-arm runners (in addition to the existing x86_64 +
  macOS + Windows targets). Drops the brittle gcc-aarch64-linux-gnu
  cross-compile step in favor of native ARM builds. PyPI wheel
  coverage after this lands: linux-gnu x86_64, linux-gnu aarch64
  (new), macOS x86_64, macOS aarch64, windows x86_64.
• Event::ToolCallStreamed — new variant emitted per detected tool
  call before dispatch (distinct from ToolInvoked, which fires after).
  Lets tracers and UIs surface tool intent in real time. The agent
  pipeline also now returns the full Vec<ParsedToolCall> aggregated
  across iterations in TurnResult.tool_calls (was always empty).
• atomr-infer-testkit::MockRunner adopted as a dev-dependency on
  atomr-agents-agent. Replaces the vendored InlineTextMock shim
  removed in de34c38 (testkit is now published as part of atomr-infer
  v0.6.0).
• atomr_agents_memory::query — re-export module surfacing
  atomr-persistence-query's ReadJournal, EventEnvelope, Offset,
  and SimpleReadJournal. Wrap any Journal with SimpleReadJournal
  to get events_by_tag / events_by_persistence_id /
  all_persistence_ids for lineage and replay.
• Provider runtime back-ends as opt-in features. New agent crate
  features provider-anthropic, provider-openai, provider-gemini
  pull in atomr-infer-runtime-{anthropic,openai,gemini} and re-export
  the *Config / *Pricing / *Runner types via
  atomr_agents_agent::providers::{anthropic,openai,gemini}. Forwarded
  through the umbrella crate as atomr-agents/{provider-anthropic, provider-openai, provider-gemini}.

Deferred / known gaps

• Publish gate. Releasing this version to crates.io requires
  atomr-infer 0.6.0 and atomr-accel 0.3.1 to be published
  upstream first; both are currently at 0.4.0 / 0.1.0 on crates.io.
  Local builds and tests work today against the path deps.
• MSRV regression (pre-existing). Transitive clap_lex 1.1.0
  requires edition2024, breaking cargo +1.78 check. Independent
  of the sibling bumps; either pin clap below 4.6 or raise the
  workspace rust-version.
• Phase 3 candidates not adopted yet: atomr-cluster-metrics
  adaptive routing in org, atomr-infer-pipeline strategy alignment,
  atomr-accel-agents::CpuVectorIndex adoption in embed, atomr-streams
  new operators (keep_alive, merge_prioritized, recover_with_retries,
  conflate, expand), FsmBuilder rewrite of workflow,
  TelemetryBus::subscribe_topic, expect_msg_* matchers in testkit.