PRODUCTION READINESS: Implement transaction semantics for cluster operations

[sfloess]

Overview

Cluster operations perform multiple state changes without transactions, leaving inconsistent states on partial failures.

Problem

Deploy/start/stop/undeploy update multiple systems (local + etcd + scheduler). Partial failures leave split state.

Examples

See issues: #251, #252, #253

Acceptance Criteria

• All cluster operations atomic or compensated
• Partial failures trigger rollback
• State always consistent
• Lost leadership handled
• Operations idempotent
• Recovery from any intermediate state

Estimated Impact

~5 major operations, 3-4 weeks work, CRITICAL for data integrity