feat(kubernetes): expose only etcd.replicas to users, keep per-cluster etcd sizing internal #3179

Context

Design proposal cozystack/community#25 folds etcd into the kubernetes app (one etcd per cluster). As part of that work, the app will gain a new etcd values block. The proposal's draft exposes the full block to users:

etcd:
  replicas: 3
  size: 4Gi
  storageClass: ""
  resources:
    cpu: 1000m
    memory: 512Mi

This issue tracks a scoped design decision raised in review by Sergei Makarov (@SerjioTT) (comment): keep the user-facing etcd surface minimal — expose only replicas — and treat everything else as a control-plane implementation detail.

Problem

Per-cluster etcd is cluster infrastructure, not a user surface. The kubernetes chart already has a clear convention for this split:

• Control plane (apiserver / cm / scheduler / konnectivity) — tunable via controlPlane.*.resourcesPreset.
• Infrastructure satellites (kccm, cluster-autoscaler, kcsi-controller) — resources are hardcoded, not exposed through values, and there are no field reports of anyone needing to tune them.

Exposing a full {replicas, size, storageClass, resources} block for etcd would break this convention and add UX complexity for knobs users don't need to turn. resourcePresets for the 4 CP components are already hard enough to reason about.

Proposal

Expose a single knob:

## @param {Etcd} etcd - Control-plane etcd datastore for this cluster.
etcd:
  ## @field {int} replicas=3 - etcd replicas. 3 = HA (Raft quorum). 1 = non-HA/dev.
  replicas: 3

Everything else becomes chart-internal:

• resources — hardcoded to the current packages/extra/etcd default (cpu: 1000m, memory: 512Mi); VPA handles growth (same pattern already used for the satellites).
• size — hardcoded to 4Gi (matches current behaviour; ample headroom for a Kamaji datastore).
• storageClass — derived from replicas, not set by the user:
  • replicas: 3 → local-style SC (etcd replicates via Raft),
  • replicas: 1 → replicated/DRBD-style SC (single member gets storage-layer HA as a safety net).
  • The SC mapping should be platform-level config, not app-level — e.g. via cozystack-values (default-etcd-sc-local / default-etcd-sc-replicated), with the kubernetes app just picking between them by etcd.replicas.

Why not the full block

• Matches the existing satellite convention (hardcoded infra resources + VPA).
• Keeps the Kubernetes CR simple — no new tuning surface users didn't ask for.
• Forward-compatible: the minimal API is a strict subset of the full block. If a real production need to override resources/size/storageClass shows up later, the API can be widened with no breaking change.

Notes

• Default replicas: 3 (safe-by-default HA); dev/throwaway clusters flip to 1 explicitly.
• Surface the etcd footprint in NOTES.txt so consumers see the full per-cluster cost (~1000m/512Mi × replicas).

Dependencies

• Depends on the per-cluster etcd refactor from [proposal] Per-cluster etcd: give each Kubernetes control plane its own datastore community#25 — there is nothing to expose/hide until etcd is folded into the kubernetes app.