Using clock() solely for delta values is absolutely a valid approach, as stated. The issue is that clock_t may not be large enough of some systems to safely keep you from an overflow, especially with arbitrary values. Additionally, some systems will include the time children processes were alive in subsequent clock() calls, furthering possible confusion. These are reasons why I would avoid clock() in favor of time(), even though your concerns are absolutely valid.

At the end of the day you have to determine which style of unpredictability you want to work around. Dealing with the times(), clock(), and clock_gettime() class of functions opens you up to managing what the kernel considers time passed, and what is accumulated vs what is not. While using time() can have shifts in time according to upstream NTP servers, as well as daylight savings time.

I would also make the argument that if an NTP server is adjusting your time, it is most likely more accurate than what your internal clock (CMOS or otherwise) was counting, and is worth following.