Is it possible to launch this library compiled with CV_SIMD on a machine without SIMD support?
I mean: do we need runtime check on supporting SIMD on executable machine ?

#if CV_SIMD is the compilation macro. When CV_SIMD is false this code is truncated in compilation time.

The question is different.
There are BUILD machine and EXECUTION machine. The build machine may turn on SIMD or other instructions sets through MACRO. In this case the appropriate code burn out binary (#if CV_SIMD) and we will have buflen = 16.

Let's assume that OPENCV was built on machine with CV_SIMD (code is exist) but executes on machine with CPU which doesn't support SIMD (code exist but cannot be executed) - so we still have 16 here, but actual earlier dispatching work out in assumption that there are no SIMD support and expect that it should have 4 due to CPU doesn't support SIMD

To resolve this situation we usually use TWO kind of check

#if EXTENSION_EXIST   // compile code ON
   if( extension_supported()) // test is supported
   {
      //      supported
   }
   else 
   {
         // compiled, but not supported
   }
#else
    // not compiled, but not supported 

Is this situation applicable here?

In this case, no errors or problems will arise. The buffer will be larger, but it will not cause any unwanted side effects.

looks like this copy-paste part with lanes calculation & internal for-loop could be incorporated in initial dispatcher addc_simd_c3(const SRC in[], const float scalar[], DST out[], const int length) then
this *_impl would just preform intrisict calculation only without this dupliacted pre- and post- operations

Applied.

Warning/Exception?

Applied.