Easy: one rule in play; expected if you know the language.
Medium: two interacting rules (e.g. signedness + shift, or solve order + dist).
Hard: subtle ordering/scheduling/solver interaction that even experienced engineers miss.
Interviewers rarely switch puzzles. They mutate one: change blocking to nonblocking, add disable iff, flip |-> to |=>, add solve...before, or move the sample point. Each puzzle's follow-up shows the escalation to expect.