why this change?

See the comment in the relevant commit:

Trying to avoid this compiler warning:

 ../numpy/_core/src/multiarray/stringdtype/casts.cpp:860:12: warning: 'char* strncat(char*, const char*, size_t)' specified bound 15 equals source length [-Wstringop-overflow=]
  860 |     strncat(buf, suffix, slen);
      |     ~~~~~~~^~~~~~~~~~~~~~~~~~~

The previous code had already been using memcpy() instead of strncpy() to avoid a similar warning. The new code extends the logic to strncat(). In any case, the code should be consistent, and use only one set of methods:

• Use strcpy()/strcat() all along, relying on the trailing NUL to find the characters to copy. This would avoid compiler warnings. I know these functions feel less secure, but in this case they are not since we know for sure the destination buf is large enough.
• Use strncpy()/strncat() all along, looking for a trailing NUL while copying, but omitting writing the trailing NUL in the destinationbuf, since it is already initialised by PyMem_RawCalloc(). This results in the above compiler warning, because the trailing NUL is omitted.
• Use memcpy() all along, relying on the already calculated length of the strings. That should be the most efficient option and avoids compiler warnings.

Sounds good 👍

this too