I like my C alternative better. I pass growth factors/increments as parameters to the vector macros so that I can affect how it grows on capacity exhaustion during each call and I have macros for creating and closing uninitialized gaps. C++ loses on many potential optimizations by insisting that its types always be in fully well-defined states except inside methods. Moreover the particular GNU implementation of the STL on Linux completely fails to turn certain vectors methods into memsets, memcpies and memmoves where it could, which pesimizes those particular ops by like two decimal orders of magnitude. The insistence on using new/delete based allocators instead of reallocs is a significant pessimization too. Reallocs perform, on average, several tens of percents better than new allocs followed by copies. An even more noticable pessimization is in the compilation times. Including `<vector>` adds good chunks of a second to the build time of an average-length translation unit. In comparison, working with C, even with all my generics included, on top of a good build system makes me feel as if I was working with a scripting language. No lags.
These are fair points and issues like these are why many larger projects have their own custom vector-like classes. One thing I did want to call attention to is your point that C++ best practices require that objects be in well-defined states except inside methods. That's true, although it's not a language requirement, but in many cases you can get the best of both worlds through the use of lambdas. A simple example to give the flavor:
template <typename Function>
void my_vector::munge(Function initialize_gap) {
create_uninitialized_gap();
// Some of our elements now contain uninitialized memory.
// We don't want to expose this to arbitrary code, but
// it's OK to expose this to the function the user
// provided specifically to handle this.
initialize_gap(gap_begin(), gap_end());
// The gap is now initialized, so when we return from
// this method, we'll be in a well-defined state.
}
This is the standard pattern you use for such things in functional languages, and it works very well. The idea is to ensure that intermediate states are only seen by code that explicitly expects to handle them.
Your criticisms of vector are well put, but if anything they're all really problems with a particular implementation of the standard library, not the C++ language. If you make your own libraries you're free to do whatever you want, and template metaprogramming can help you gain even more performance by easily using optimized implementations for specific types, etc.