How would you choose and enforce a safe degree of parallelism for a .NET workload?

In production I would treat the number as a tunable resource policy, not a magic constant. I would observe throughput, p95 and p99 latency, CPU, allocation rate, thread-pool queue length, connection-pool waits, dependency throttling, backlog age, and cancellation. The best point is often before full saturation because the service needs failure and traffic headroom. Limits may need separate budgets for interactive and background work. I would record the chosen value, the constrained resource, and the load-test evidence so later tuning has a baseline. If the bottleneck changes with payload size or downstream health, a fixed number may need workload classes, adaptive admission, or explicit backpressure rather than silently creating more tasks.