feat(agents): add an explicit stop op for Sandbox sessions

[tee8z]

A wedged Agent Sandbox session can ignore !shutdown, and the provider had no op to end it.
This PR adds a provider stop op. It suspends the Sandbox, so the controller deletes the Pod with SIGTERM.
The harness still gets its checkpoint window. The normal stop stays !shutdown.

The stop request is the deploy request with "op": "stop".
Before any cluster contact, stop_session() requires provider_config.sandbox, an approved identity_policy, and a verified launch owner.

Observed Sandbox	Response state	Action
None	absent	None
Different agent, owner, or thread	Error	None
lifecycle=ended	ended	None
lifecycle=ending	ending	None
Unknown lifecycle value	Error	None
Live	ending	One merge patch

The merge patch carries the observed metadata.resourceVersion. It sets buzz.block.xyz/lifecycle=ending, buzz.block.xyz/ended-reason=stopped, and spec.operatingMode: Suspended.
On 409 Conflict, stop() reads the Sandbox again, for at most 5 attempts. Then it returns Sandbox kept changing while stopping; retry.

info now returns ops: ["info", "deploy", "stop"]. The field is additive within protocol version 1, so the desktop can gate session controls on stop.

Stack

Position 10 of 20 in the DEV-2884 Buzz stack. Base: #15. Next: #2.

Validation

• cargo fmt --check -p buzz-backend-kubernetes: pass.
• cargo clippy -p buzz-backend-kubernetes --all-targets -- -D warnings: pass.
• cargo test -p buzz-backend-kubernetes: lib 12 passed; bin 177 passed, 1 ignored; wire_fixtures 5 passed.
• New tests:
  • stop_suspends_with_a_cas_patch_and_retries_conflicts
  • stop_requires_sandbox_scope_and_identity_policy_before_cluster_contact
  • parses_stop_with_the_deploy_payload, stop_response_serializes_flat, info_advertises_stop
  • The stop-no-identity-policy wire fixture.

Limits

• The ignored test live_controller_reuses_scopes_and_refuses_lost_workspaces needs the kind-remote-agent-check cluster. It was not run. Its new stop steps were compiled only.
• Temporary glue, removed by the recovery PR:
  • A stop request with sandbox.recovery fails options parsing (deny_unknown_fields). The recovery PR adds the explicit refusal stop does not accept a recovery mode.
  • The live test patches the tombstone with lifecycle::CHECKPOINT and skips the recovery steps.
  • The fixture README does not mention deploy-sandbox-bad-recovery.
• Buzz AGENTS.md requires a human to test the change before it leaves draft.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}