These are the result for the integration tests (green):
https://github.com/BlueBrain/HighFive-testing/actions/runs/4427635111/jobs/7765526499#step:6:35

The issues to consider are:

• Even when loading an empty dataset, since we perform allocation, we need to ensure that the returned object is appropriately empty.
• For multi-dimensional arrays, there's a difference between (n, m, 0) and (0, m, l), e.g. for nested std::vectors the former is a vector with n vectors, each containing m empty vectors. The latter is just an empty vector.
• We might need to consider that the shape of std::vector<std::array<T, 2>> is always (n, 2). Therefore, even if n == 0 the next dimension should be deduced as 2. Currently, we unable to get the shape of compile-time constant dimensions correct, in general. We return 1 instead.
• We currently don't check that in write_raw the buffer isn't the nullptr, nor do we turn it into a no-op. Instead we rely on HDF5 doing the appropriate checks, e.g. if a positive number of elements are to be read into the nullptr.

I have checked around and LGTM. It's a pity we have some duplication in Attribute::read/write and Slice::read/write, but improving that is another story.