v0.1.0

First release (alpha). The core thesis — declarative, scale-to-zero serving of self-hosted
open-source LLMs on Kubernetes — is implemented and verified end-to-end on real NVIDIA GPUs, and
the full loop now runs in CI with no hardware.
Not production-ready (see ROADMAP.md).

Added

• CRDs — LLMService (namespaced) and InferenceRuntime (cluster-scoped, pluggable backend),
  API group serving.hearth.dev/v1alpha1.
• NVIDIA backend — renders the vLLM serving workload (image, templated args, accelerator
  resource, model-load-aware probes, metrics). Ascend adapter scaffolded + golden-tested
  (not yet run on NPUs).
• Scale-to-zero — Hearth gateway (buffering reverse proxy) + KEDA ScaledObject on gateway
  queue depth; 0→1→N→0, verified 1→2 across two GPU nodes.
• Cold-start handling — SSE keepalive heartbeats, reject mode, load-gated readiness, bounded
  queue (429) and activation timeout (503).
• Graceful drain — in-flight streams finish before scale-down.
• Model caching — HostPath and NodeLocalPVC (with pinnable cache.storageClassName,
  verified against Alibaba ESSD) + a prewarm Job.
• Observability — per-gateway Prometheus metrics, ServiceMonitor, and a Grafana dashboard.
• No-GPU test harness — a CPU vllm-stub and a kind + KEDA e2e that runs the full
  0→1→N→0 loop, backpressure (429/503), and graceful drain on every PR, no accelerator
  required, plus a no-GPU development guide.
• Packaging — Helm chart (operator + RBAC + CRDs) and multi-arch image build/release workflow.
• Project scaffolding — README, ROADMAP, CONTRIBUTING, CODE_OF_CONDUCT, SECURITY, MAINTAINERS,
  GOVERNANCE, issue/PR templates, and a DCO check.

Changed

• Operator skips no-op LLMService status updates, avoiding optimistic-concurrency churn.

---

🙏 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

🚀 Features

• feat: add version flags to binaries by @loopassembly in #9

🐛 Fixes

• controller: skip no-op LLMService status updates by @Jah-yee in #29

📖 Documentation

• docs: make LLMService sample storage class neutral by @wondr-wclabs in #13
• docs(chart): populate Helm notes by @wondr-wclabs in #14
• docs: add HostPath cache sample by @wondr-wclabs in #16
• docs: add LLMService CRD reference by @wondr-wclabs in #17
• docs: add observability guide by @wondr-wclabs in #15
• docs: add Helm v4 SSA CRD ownership conflict warning by @Jah-yee in #31

New Contributors

• @loopassembly made their first contribution in #9
• @wondr-wclabs made their first contribution in #13
• @Jah-yee made their first contribution in #29

Full Changelog: v0.1.0-rc.1...v0.1.0