My two cents are:

(A) It's generally useful to embed the build time into the executable.
I've recently done so myself at work; an HTTP server uses build time
as a lower bound for Last-Modified headers to prevent caching of some
files across version changes.

(B) Constant evaluation seems like a very bad vehicle for anything
non-deterministic. Remember that the results of constexpr calls can be
memoized, and myarray[computed_size()] needs to be consistent across
all TUs to prevent ODR violations. The proper way is to let the build
system compute the timestamp, which is then consistent across all TUs.
This feature should not be in the preprocessor, let alone constant
evaluation.

For what it's worth, I think this feature has merit if you say that
all compilation happens instantaneously for the purpose of constant
evaluation clock time. In other words, you could require that there is
a global "build timestamp" and now() always returns that value in all
TUs. However, I'm not sure if this is implementable, considering that
TUs get compiled independently, and considering incremental
compilation.