Detect spec drift on Running nodes for mid-life SigningKey patch (and future validator mode switch)

Read internal/planner/planner.go at buildRunningPlan and buildNodeUpdatePlan, then trace SeiNodeStatus and the rollout-watch task. Verify the existing validate-signing-key and apply-statefulset flow before adding the integration coverage described in the issue. Done means SigningKey drift triggers a guarded rollout, status is stamped, the pod mounts the Secret, and the LLD and migration guide are updated.

Background

PR #136 shipped validator.signingKey.secret.secretName for the single-shot validator deployment use case — deploy a SeiNode with SigningKey set from creation and the controller mounts the Secret on the production StatefulSet pod.

What it explicitly does not support: patching SigningKey onto an already-Running validator. buildRunningPlan (internal/planner/planner.go:621-628) only detects image drift today, so a kubectl patch seinode --patch '{"spec":{"validator":{"signingKey":...}}}' against a Running node is a silent no-op — the pod doesn't restart, the Secret never mounts, the validator never starts signing. Documented in LLD §11.

Why this matters

Primary use case — zero-downtime migration cutover

The single-shot deployment in PR #136 trades cutover downtime (bootstrap-Job sync + StatefulSet catch-up) for implementation simplicity. For arctic-1 testnet this is fine; for pacific-1 the downtime envelope may not be acceptable. The original LLD §8 design was a two-phase cutover:

1. Deploy SeiNode without signingKey → bootstrap + sync as a non-signing observer (zero risk, no downtime on the old EC2 validator)
2. Stop EC2 → wait M blocks → patch signingKey in → pod re-rolls with Secret mounted → seid signs

Phase 2 requires drift detection. Without it, the operator's choices are (a) accept the single-shot downtime, or (b) delete-and-recreate the SeiNode (which deletes the data PVC unless dataVolume.import is wired, costing the bootstrap+sync work).

Future use case — validator mode switch

A separate, larger ask: convert a Running fullNode SeiNode into a validator SeiNode by patching the spec (e.g., spec.fullNode → spec.validator with a signingKey). Same drift-detection mechanic, larger pod-spec surface (mode string in seid config, port configuration, probes). Out of scope for v1 of this issue, but the underlying machinery is identical — listing here so the design accommodates it additively.

Proposed mechanic

Reuse the existing NodeUpdate plan shape (buildNodeUpdatePlan at internal/planner/planner.go:656):

1. Track applied state in SeiNodeStatus. New field — likely Status.SigningKeyMountedSecret string — stamped by observe-image (or a sibling task) when the StatefulSet rollout completes. For mode-switch, an analogous Status.ObservedMode would track the currently-running mode.
2. Extend buildRunningPlan to detect drift beyond image:

   if signingKeyDrift(node) {
       return buildSigningKeyUpdatePlan(node)
   }
   

3. Build a re-apply plan that runs validate-signing-key (already present, gates on Secret correctness) → apply-statefulset (server-side apply triggers the rolling update) → observe-image-equivalent rollout watcher → mark-ready. Stamp the new applied state on completion.

The pod restart is the standard StatefulSet rolling update mechanic — no special "kill seid" step is needed. seid restarts naturally when the pod is replaced and reads the newly-mounted key file at startup.

Acceptance criteria (v1 — SigningKey drift only)

• Status.SigningKeyMountedSecret field added to SeiNodeStatus and stamped by the rollout-watch task on success
• buildRunningPlan detects spec.validator.signingKey.secret.secretName != status.SigningKeyMountedSecret and triggers a re-apply plan
• Re-apply plan includes validate-signing-key so Secret defects are caught before the pod rolls
• Integration test: deploy SeiNode without SigningKey → reaches Running as observer → patch SigningKey in → pod restarts with Secret mounted → SigningKeyReady=True → StatefulSet has the volume + subPath mount
• LLD §8 updated to re-document the zero-downtime cutover flow as supported
• .tide/validator-migration.md updated with both single-shot (existing, downtime) and zero-downtime cutover variants

Out of scope (for v1 of this issue)

• Mode switch (full-node → validator). Same drift-detection mechanic, larger surface; file as a follow-up once SigningKey drift is shipped and we have a concrete customer ask.
• Demoting a validator to non-signing (clearing SigningKey on a Running validator). Currently no field-level immutability blocks unset, but the workflow is unsupported per LLD §11.
• Rotating secretName on a Running validator. secretName is immutable by XValidation: self == oldSelf; rotating consensus key requires MsgEditValidator and is a coordinated on-chain operation, not a controller feature.

References

• PR #135 — merged LLD
• PR #136 — single-shot deployment shipped
• LLD §11 — Mid-life SigningKey patch deferred entry with implementation sketch
• .tide/validator-migration.md — Deferred → Zero-downtime cutover