This PR introduces a new ProjectTask abstraction for project-level asynchronous work and migrates GitHub repository import onto that platform. It removes the old project-import-specific reconcile path and keeps the existing sandbox/database resource reconciliation model unchanged.

What changed

• Added ProjectTask schema, enums, and migration.
• Added task persistence helpers, reconcile loop, and executor registry.
• Implemented CLONE_REPOSITORY as the first task type.
• Migrated GitHub import to create a task instead of driving import through Project.importStatus.
• Updated project APIs and UI status handling to read task state.
• Replaced sandbox-triggered import wake-up with sandbox-triggered task wake-up.
• Refreshed architecture documentation and added architecture evolution guidance.

Why

• GitHub import is not a resource lifecycle concern; it is project-level async work that happens after sandbox readiness.
• We need a reusable platform for future project-scoped work such as skill installation and deploy.
• This separates resource orchestration from task orchestration and gives us a clearer platform model.

Follow-ups / Known gaps

• Existing migration does not backfill historical import state into ProjectTask.
• Task locking currently still allows RUNNING tasks to be re-claimed after lock expiry; this matters more once longer-running task types are added.
• CLONE_REPOSITORY currently treats sandbox RUNNING as task-ready, but ttyd readiness can lag behind pod readiness.

Validation

• pnpm exec prisma generate
• pnpm exec tsc --noEmit
• pnpm lint
• Remote test DB schema updated via prisma db push