In Go (as in most programming languages) it's not true that the assignment "a = b" implies that "b == foo => a == foo", if a and b are different types.
For example, this C code will print "nope":
float b = 3.5;
int a = b;
printf(a==3.5 ? "yup" : "nope\n");
In practice I haven't encountered any refactoring bugs as you describe, though it is theoretically possible.
I personally don't see it as a big issue.
(In retrospect, Go could've help guide intuition better by using a new keyword like "unset" or something to test for the zero-value of interfaces, instead of overloading nil.)
While I haven't counted and compared, I suspect that in most programming languages, "a=b" being a non-error implies that either a and b are the same type and value, or that a and b are values that, if they are of different and comparable types, will compare equal.
There are certainly popular languages that do it the way you describe, but I don't think most languages do.
While this is true, in all of those other cases you have the benefit of compile-time type checking so cannot call a function on the wrong type.
With nil, you get runtime errors.
The point being argued is that it doesn't matter that go defines this to be the case, the point being argued is that it's surprising, frustrating, and can lead to bugs. Particularly when `bar` starts off as a pointer to a struct, but later is refactored to be an interface type. Code that used to work still compiles, but now encounters a runtime nil panic.