Harden checkpointing: error resilience, correctness fixes, and memory safety

[Copilot]

• Analyze CI failures (2 failing tests in Checkpoint > Resume from checkpoint)
• Fix DependencyBasedScheduler.onTaskCompleted to remove the completed task from pendingTasks — prevents tasks restored as COMPLETED from being re-scheduled and re-executed
• CodeQL check: no security alerts
• Code review: no issues

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[sroussey]

@copilot fix build/test errors

[Copilot]

@copilot fix build/test errors

Fixed in commit 2fb92d0.

The root cause was in DependencyBasedScheduler.onTaskCompleted() — it added the task to completedTasks but did not remove it from pendingTasks. In the normal execution flow this is fine (the task is removed from pendingTasks when waitForNextTask() picks it up to run), but during checkpoint restore, onTaskCompleted() is called for each COMPLETED task while they're still in pendingTasks. This caused the scheduler to re-yield those tasks for execution when tasks() started iterating, re-running tasks that should have been skipped.

The fix adds task.config.id === taskId to the existing loop in onTaskCompleted, so the completed task is also removed from pendingTasks. This is a no-op in normal flow (already removed before onTaskCompleted is called) and correctly prunes restored tasks during checkpoint resume.