fix: harden orchestration restart recovery

[numnx]

Summary

This PR hardens orchestration recovery across app restarts and restart-edge cases.

• Adds General -> Restart Behavior settings for sprint and invocation recovery policy.
• Applies startup recovery policies before watch-loop recovery so active runs can continue, pause, or cancel coherently.
• Preserves Docker workspace/runtime volumes while allowing restart/cancel policy to remove stale labelled containers.
• Repairs paused sprint projection drift so paused runs do not keep appearing as running after restart.
• Prevents virtual merge-conflict workers from falsely resolving an item when no merge evidence was recorded.
• Keeps checked-out worker branches clean after patch materialization updates refs.

Root Cause

Restart recovery had several independent paths for sprint runs, dispatches, provider invocations, QA rows, leases, Docker containers, and sprint projection rows. Those paths could disagree after a restart, especially around paused runs, interrupted local invocations, preserved Docker volumes, and virtual merge-conflict repair.

Validation

• pnpm run typecheck
• pnpm vitest run tests/backend/services/runtime-startup-recovery-service.test.ts tests/backend/infrastructure/providers/cli/workspace-artifact-service.test.ts tests/backend/repositories/settings-repository.test.ts
• pnpm vitest run tests/backend/services/virtual-worker-service.test.ts
• pnpm run build
• Local rapid restart stress: graceful restart, repeated graceful restart, hard-kill restart, delayed restart.
• Runtime readiness verified at http://127.0.0.1:4444/ready after stress.

Risk / Rollback

Risk is concentrated in startup recovery and virtual-worker merge-conflict handling. Rollback is reverting this PR; existing persisted sprint/task state remains compatible because the new restart policy settings default to continue and sanitize unknown values back to defaults.