When would you use lock instead of SemaphoreSlim, and when would SemaphoreSlim be better?

In production I treat synchronization as a design smell to inspect carefully. Locking shared state may be fine for small local data, but long critical sections, nested locks, or locks around I/O create latency and deadlock risks. SemaphoreSlim is useful for local async coordination, but it is not a cross-instance lock and does not replace database constraints or queue design. I choose the primitive based on whether I am protecting state, limiting concurrency, or coordinating a workflow, and I make the scope as small and observable as possible.