Orchestrates parallel task execution across up to 12 concurrent Sparky subagents using a rolling pool scheduler.

• Parses markdown plan files, extracts task definitions, and launches subagents continuously as slots open without waiting for batch completion
• Maintains canonical file paths and naming constraints across parallel tasks to prevent filename drift and cross-task conflicts
• Validates each subagent result, updates the plan file with completion logs, and immediately schedules the next pending task to keep concurrency maxed
• Performs final integration pass after all tasks complete: reconciles conflicts, runs tests, fixes failures, and ensures the full codebase converges to plan expectations
• Requires explicit agent_type: sparky for all subagent launches; stops and reports if a task needs paths outside its pre-built context pack