signed to unsigned (and vice versa)

[laserlight]

If I see it rightly the section 4.7 doesn't exactly define conversion from signed to unsigned. Or if we assume the 'otherwise' part also applies to conversions to unsigned, it therefore is "implementation-defined".

Yes, except that the standard does define what happens if the value "can be represented in the destination type".

Also, does the C++ standard really 'define' the implementation for std::basic_string::npos or is it a concrete implementation where it is defined by assigning -1?

I am not sure if an implementation is permitted to differ as long as the net effect is the same, but the C++ standard does indeed define how npos should be defined, i.e., by initialising it with -1.

implicitly does a static cast but wouldn't necessarily help to get the maximum unsigned integer.

For an integer representation using not two's-complement for negatives the bit representation of -1 could be 10000000000000000000000000000001 (one sign-bit and no complement) and I doubt very much that they would cast that to 0xfffffff.

You might want to read paragraph 2 of section 4.7 of the 2003 edition of the C++ standard again. In particular, the comment on two's complement representation is a good hint as to why you are mistaken.