Optional Karpenter autoscaling (karpenter-provider-hetzner) as an alternative to Cluster Autoscaler?

[stubbi]

Summary

Would you be open to an opt-in Karpenter integration alongside the existing
Cluster Autoscaler, using karpenter-provider-hetzner?

Disclosure: I help maintain karpenter-provider-hetzner, so I have an
interest here. I'm raising this as a Discussion (not a surprise PR) precisely
because the module already has solid Hetzner autoscaling — I want to check
appetite before building anything.

Why this module specifically

Of the popular Hetzner installers, this one is the cleanest fit for Karpenter:

• The provider already supports Talos worker bootstrap.
• This module already generates Talos worker machineconfigs and ships
  components as Talos inline manifests — so the integration reuses existing
  machinery (data.talos_machine_configuration, the *_manifest pattern)
  rather than adding new bootstrap logic.
• The hcloud CCM you already install provides the hcloud://<id> providerID
  Karpenter needs.

What I'm proposing (not a replacement)

Karpenter is an alternative to the Cluster Autoscaler, not a gap-filler — it
does per-pod instance-type selection and consolidation instead of pre-declared
ASGs. So I'd keep it strictly opt-in:

• New karpenter_nodepools variable; empty by default → zero change for
  current users.
• A guard so Karpenter and cluster_autoscaler_nodepools can't both be enabled
  (both want to own provisioning).
• Follows cluster_autoscaler.tf exactly: data.helm_template to render the
  chart, a generated worker machineconfig as the node userData, manifests
  appended to talos_inline_manifests.

I have a working draft of karpenter.tf + variables ready to share.

Open questions

1. Is an alternative autoscaler something you'd want to carry, or do you prefer
   to keep one autoscaling path?
2. karpenter-provider-hetzner is young — would you want a stability bar (e.g.
   N months / pinned chart, marked beta) before merge?
3. Preference on shape: single shared HCloudNodeClass + per-pool NodePools
   (what I drafted), or a richer per-pool NodeClass mapping?

Happy to open the PR if there's interest. Thanks for the module — it's the
nicest Talos-on-Hetzner setup out there.