v0.9.0

• GPUs allocated with Dynamic Resource Allocation (resource.k8s.io, NVIDIA's DRA driver gpu.nvidia.com): pods
  that hold a GPU or a MIG slice through a ResourceClaim instead of nvidia.com/gpu count for their pod, namespace
  and node; the devices each node publishes in its ResourceSlice count as capacity on the Allocation view; pending
  claims show in Pending with what they ask for against what the nodes offer; and without pod labels on dcgm-exporter
  the claim names the pod holding each card, by UUID. Reads v1, v1beta2 and v1beta1, and needs list on
  resourceslices and resourceclaims; clusters without DRA or without that permission work as before (#25).

v0.8.1

• Fix: the tables were wider than the page in a 1440 px window, so Health and Power were out of sight. Text columns
  (pod, node, namespace, model, names, messages) now share the free width and shrink down to their minimum with an
  ellipsis while staying readable, and Health and Power come right after the identity columns on the GPUs and
  Allocation views and before the VRAM columns in the Pods table and drawers. Column widths resized by hand are stored under a new key, since some
  views changed order (#10).
• Fix: when a scrape failed, the rows of the last good one stayed on the page under the error with the status still
  counting the exporters, and nothing said they were old. The status now reads "stale: last good scrape HH:MM:SS",
  the pages built on the scrape (Pods, GPUs, Idle & waste, Allocation) and the drawer tables are dimmed, and the
  pages built on the pod list keep full contrast. When every exporter scrape fails, discovery runs again at the next
  tick instead of scraping the same dead pods until its cache expires (#9).
• Fix: discovery took the node name inside a pod's name for a GPU keyword, so on a node called like gpu-* the
  static control plane pods (kube-apiserver, etcd, scheduler, controller manager) were probed at every discovery pass.
  The node name is now left out of the pod name before matching (#7).
• Fix: on a cluster without exporters the whole cluster was listed again every 20 s (pods, services, nodes), also
  when only a Pod or Node drawer was open. An empty discovery and a failed Prometheus search are now cached for 60 s
  like a successful one, nodes are listed with the discovery instead of with every scrape, and the drawers poll only
  for a pod that can use a GPU (or on a cluster known to have GPUs) and for a node that advertises one (#8).

v0.8.0

• Moved to the freelensapp organization as @freelensapp/gpu-extension: the organization workflows (release with
  provenance, SBOM and checksums, Claude, npm audit and dedupe, trunk and biome upgrades), Playwright integration tests
  inside a packaged Freelens on kind with a fake GPU fixture (integration/fixtures/gpu/), agent guides.
• Supersedes @tal-naeh/freelens-gpu-extension: uninstall the old package before installing this one.