I attempted to convert the example to get rid of _Static_assert's via the following trick:

#define ASSERT_TYPE3(_expr, _type, _c) \
    struct assertType##_c { \
        int x: _Generic(_expr, _type : 1, default : -1); \
    } assertValue##_c = {0}
#define ASSERT_TYPE2(_expr, _type, _c) ASSERT_TYPE3(_expr, _type, _c)
#define ASSERT_TYPE(_expr, _type) ASSERT_TYPE2(_expr, _type, __COUNTER__)

// ...
ASSERT_TYPE(&ptr, const int **);

but apparently CIL is fine with negative bit-field widths. Goblint is not, but the error it encounters is downstream from the actual issue:

Fatal error: exception Invalid_argument("Z.shift_left: count argument must be positive")

Furthermore, the following error probably warrants a separate ticket: Error: __auto_type unsupported for globals.

As for supposed difference between auto type inference and typeof, I believe it's rather straightforward. auto performs lvalue (and other) conversions, whereas typeof yields the type verbatim. Observed difference in behaviour wrt function parameters is explained by the fact that function parameters undergo respective conversions independently of typeof. The footnote in 6.7.3.6 refers to the section 6.9.2 ("Function definitions") which itself defers function parameter type adjustment to 6.7.7.4 ("Function declarators"). It is apparent that the intention of the footnote was just restating that these adjustments need to be respected by typeof.