bug(kubernetes): sandbox stuck at `Provisioning` with `replicas: 0` has no recovery path (`start`/`stop` both refuse)

Start with the phase gate in start_sandbox in crates/openshell-server/src/compute/mod.rs (~L1784) and the reconcile fallthrough near derive_phase (~L6863), then see how crates/openshell-driver-kubernetes/src/driver.rs suspend_sandbox_runtime_after_dependency_failure (~L4111) sets replicas to 0. Done means a sandbox stuck in Provisioning with an empty backend either gets corrected by reconcile or can be started through the idempotent-restart path, with the workspace PVC kept. The issue offers two directions, so agree the approach with maintainers before coding.

Summary

A sandbox can end up with the gateway's Sandbox record at phase Provisioning
while the backend Sandbox CR is at spec.replicas: 0 with no pod. In that state
there is no supported recovery path:

• sandbox start is rejected — the phase gate only admits Stopped,
  Completed, Starting, a failed-main-process Error, or a provisioning timeout.
• sandbox stop is rejected — it requires Ready.
• The reconcile sweep does not correct the record: with replicas: 0 / no pod, the
  Provisioning case falls through to the reconcile/derive_phase default and the
  stale value is perpetuated rather than corrected.

The only exits are delete + recreate, or out-of-band edits to the gateway's store.

This is a distinct variant of the "unrecoverable phase" bugs #3308
and #4209, which are Error-phase. Here the record is
Provisioning, so the recovery work in #4231 (allow stop from
Error) does not reach it, and start still rejects it. The closest mechanism match is
#2932 (a stale condition perpetuated by reconcile), fixed by
#2933.

Environment

• OpenShell: 0.1.2 (gateway ghcr.io/nvidia/openshell/gateway:0.1.2, supervisor 0.1.2, CLI 0.1.2)
• Driver: Kubernetes (agent-sandbox v0.4.6, CRD sandboxes.agents.x-k8s.io v1alpha1)
• Kubernetes: v1.27.3 (kind)

Reproduction

Non-deterministic (a race). Observed triggers:

1. A node/machine reboot while the gateway is bootstrapping a sandbox — the gateway pod
   restarts during cluster stabilization, and a transient dependency failure fires
   suspend_sandbox_runtime_after_dependency_failure, resetting replicas: 0 while the
   record is still Provisioning.
2. A stop immediately followed by start, leaving the record mid-transition.

Observed state

$ openshell sandbox list
NAME   ...   Provisioning

$ kubectl get sandboxes -n <ns> -o yaml
# spec.replicas: 0
# no openshell.ai/sandbox-runtime-bootstrapping annotation
# (no sandbox pod in the namespace)

$ openshell sandbox start <name>
Error: sandbox must be Stopped, Completed, or a failed main-process Error to start (current phase: Provisioning)

$ openshell sandbox stop <name>
Error: sandbox must be Ready to stop (current phase: Provisioning)

Root cause (pointers; as of e1f3c82ca)

• The record's status.phase is Provisioning but the backend CR is empty.
  start_sandbox rejects it because the gate only admits
  Stopped | Completed | Starting | is_failed_main_process_result | provisioning_timeout
  (crates/openshell-server/src/compute/mod.rs, start_sandbox, ~L1784).
• The reconcile sweep re-derives phase from the driver snapshot; with no pod /
  replicas: 0, the Provisioning case falls through to the _ => phase default
  (~L6564, fed by derive_phase ~L6863), so the stale value is kept.
• suspend_sandbox_runtime_after_dependency_failure
  (crates/openshell-driver-kubernetes/src/driver.rs, ~L4111) is what drives
  replicas: 0 on the dependency-failure path.

Expected

One of:

1. The reconcile loop corrects a Provisioning record whose backend reports the
   sandbox as empty (no pod / replicas: 0) instead of perpetuating it — cf. #2933, which
   made derive_phase let Ready win over a stale Suspended; or
2. start accepts a stale Provisioning whose backend is empty and routes it through the
   existing idempotent-restart path — reusing the persisted identity and re-attaching the
   existing workspace PVC — the same way the Starting branch already does (~L1800).

Actual

No recovery short of delete + recreate. An out-of-band edit of the gateway store
(flipping Sandbox.status.phase from Provisioning to Stopped) does make start
succeed and preserves the workspace — confirming this is a fixable record desync rather
than data loss — but it is not a supported operator path.

Proposed direction

• Extend #4231's recovery approach to the Provisioning-with-empty-backend case (not
  just Error), or
• correct the stale Provisioning in derive_phase/reconcile (precedent: #2933).

Either way, start from this state should reuse the persisted identity and the existing
workspace PVC (idempotent restart), not require a delete.

References

• #3308 — unrecoverable Error, stop/start refuse (same symptom class)
• #4209 — gateway restart → unrecoverable Error (same trigger family)
• #4231 — open: allow stopping errored sandboxes for recovery (in-flight recovery mechanism)
• #2932 / #2933 — stale condition perpetuated by reconcile; fix in derive_phase (closest mechanism)
• #4078 — serialize lifecycle cleanup with sandbox restart (related stop/start race)