v0.2.0

Alpha release completing Hearth's first integrated Ascend hardware-validation cycle and hardening
the scale-to-zero control plane based on those results. Atlas 300I Duo is verified through the
multi-replica 0→1→2→0 lifecycle, while Ascend 910B3 is verified through the complete
single-device 0→1→0 lifecycle. Hearth remains v1alpha1 and is not production-ready.

Added

• A shared Ascend hardware-validation guide covering required images, validation levels, evidence,
  and separate 910B, Atlas 300I Duo, and Atlas 300I Pro result sets.
• Full physical 910B3 evidence for device-plugin scheduling, single-device 0→1→0, cold and warm
  inference, reject mode, bounded-queue backpressure, metrics, drain, self-heal, Helm upgrade, and
  reboot recovery.
• Full Atlas 300I Duo hardware evidence for the integrated 0→1→2→0 lifecycle, inference,
  backpressure, reject mode, drain, cache persistence, self-heal, Helm upgrade, and reboot recovery.
• Validation bounds for accelerator counts, scaling values, runtime ports, and termination grace
  periods.

Changed

• Prometheus Operator resources and RBAC are decoupled from core reconciliation. Hearth now exposes
  stable metrics discovery labels while optional ServiceMonitor and Grafana assets live under
  examples/observability/.
• Runtime metric fields are optional metadata for external observability; backend adapters and
  autoscaling no longer carry an unused runtime-metrics dependency.
• InferenceRuntime family and vendor enums now match the built-in implementations: vLLM on
  NVIDIA or Ascend. Manager RBAC is reduced to the CRD operations used by the passive runtime and
  service reconcilers.
• Ascend device profiles use equal priority, so clusters with multiple installed profiles must pin
  the intended hardware runtime instead of silently preferring 910B.
• Runtime and service examples now live in independently deployable, vendor/device-specific
  examples/ directories, with concise filenames that leave the device model to the directory.
• Prewarm Jobs inherit the runtime's node selector, tolerations, scheduler, and Volcano queue so
  node-local model data is prepared where the backend can run.
• LLMService resources are reconciled when a matching InferenceRuntime changes.
• The Ascend 910B runtime now invokes vllm serve explicitly and uses MindCluster's standard
  accelerator=huawei-Ascend910 node label. A dedicated hardware-validation service example uses a
  60-second drain, the shortest tested value that completed a live 910B3 stream without aborting.
• Ascend 310P examples invoke vllm serve explicitly and pin FP16 for the validated 310P3 path.

Fixed

• Keep retrying unchanged reconciliation failures, reject ambiguous equal-priority runtime
  selections, and reject Volcano queue names unless the Volcano scheduler is selected.
• Apply scaleDownStabilization to both KEDA's zero cooldown and HPA scale-down behavior, with
  validation for the HPA's one-hour limit.
• Build multi-platform images on the build host architecture and fail the docker-buildx target
  when the image build fails.
• Embed the release tag in operator and gateway images built by the release workflow.
• Publish the release chart only after both release images build successfully.
• Align Helm manager pod security and termination settings with the Kustomize deployment.
• Pin Kind and KEDA in E2E workflows, install Helm explicitly, and fail CI on uncommitted module or
  generated-file changes.
• Make repeated Podman scale-to-zero E2E runs replace and clean their temporary image archives.
• Reject unsupported BakedImage cache requests instead of silently rendering them as uncached.
• Reject invalid PVC claim names and model paths that could escape the mounted model volume.
• Prevent accelerator-free prewarm Pods from autoloading PyTorch vendor backends and requiring host
  driver libraries.
• Report the committed SSE 200 for keepalive activation timeouts and avoid classifying canceled
  clients as activation timeouts in gateway metrics.

---

🙏 Thanks to everyone who contributed to this release — the merged PRs and first-time
contributors are credited under What's Changed and New Contributors below.

What's Changed

🔧 Other Changes

• docs(crd): flag unimplemented fields by @ded-furby in #40

New Contributors

• @ded-furby made their first contribution in #40

Full Changelog: v0.1.0...v0.2.0