Has fuzzing integration and been scrutinized by a previous review. I've done
version 0.1.2 in some detail and the only changes merged since then are bug
fix PRs addressing those. Mostly worried about inadvertently introduced bugs
in future changes due to awkward interfaces and implicit trust between
different components within the code.

For example, code in struct_item.rs relies on the reader providing only 4
byte aligned buffers instead of reasserting that fact as an internal
invariant. However this invariant is relied upon for an aligned pointer load
later.

Similarly some other C style patterns are prominent. Instead of functions
testing an invariant and constructing a result based off of it, boolean
return and construction by the caller is used. Example:

      self.assert_enough_struct(offset, desc_size)?;

      let desc_be = unsafe {
          &*((&self.struct_block[offset..]).as_ptr() as *const PropertyDesc)
              as &PropertyDesc
      };

This is fine in its current usage as far as I can tell but it's not very
stable with regards to possible future changes.

This also means that manual arithmetic is used for bounds checks which is
prone to missed overflow considerations etc.

A remaining antipattern is that of a byte output buffer: StructItem offers
reading its value as strings or a u32 list. But instead of an iterator over
the backing memory the implementation takes an mutable reference to a byte
slice, manually aligns it to fit the output type, casts it, and writes the
data types &'_ str and u32. I have not found concrete misbehaviour from
this but it seems awkward.