I was surprised to see the implementation of CTE recursion (in 2005 at least) performed similarly too inbuilt functions for small string lengths (read napkin test of up to 400) for the last example compared to

SELECT @str2 = SUBSTRING(@Str1,1,@offset-1) + @char + Substring(@Str1,@offset+1,len(@str1))

The elapsed time was comparable

The recursive reported measurable cpu time and 2 worktable scans

The inbuilt did not report anything

Forgive if I am new to this but does that mean that inbuilt functions don't add to reported CPU ?

The inbuilt in a scalar select didn't generate a query plan on execution; I guess it wasn't considered worth caching.