The things you learn: time_t

I thought that while time_t need not be an unsigned 32-bit integer (it could also be a 64-bit integer, for example, or even a double-precision floating-point number), it has to be (a) an arithmetic type that (b) represents seconds since some epoch.

Apparently, only (a) is correct; AFAICT, the standard does not make any requirement about the encoding, not even that one "unit" is a second, only that (time_t)-1 has a special meaning. So it could be, for example, a 64-bit count of milliseconds since some epoch. Subtracting two time_t values and expecting the result to make any specific sense is, therefore, unportable; you have to use difftime() for that (which is documented to return a count of seconds as a double-precision floating-point number, no matter what "time_t" is encoded like).