The test case which is added fails to compile unless one puts the
`parent` field before the `id` field. The cause is explained somewhat by
the change, but I'll spell it out here nontheless.
Previously, the loop in `Compound::make_as_item_iter_statemachine`
assumed that the serialisation order of fields would match their
declaration order. That is not generally true: attributes must be
serialised before element content, because they must be emitted before
the element header is closed.
This change thus splits the generated states into "header" states (for
everything before the end of the element header (think `>`)) and
"body" states (for everything after and including the end of the
element header). After all fields have been processed, we can then
add the data fields of the body fields to the header states so that
they are carried through the generated state machine until they are
needed in the body.
They really clutter the documentation, because they show up in the
module's index. Downside is that users can't inspect the documentation
of those types anymore, but it's not useful anyway: the parts are
private and they just impl some traits.
This allows exactly one arbitrary payload in any element, and is handled
after every other element with a more specific matcher has been parsed.
Both the #[xml(element(n = 1))] meta and its shortcut #[xml(element)]
are allowed and treated the exact same way.
The `xso::text::Base64` struct remains as a shorthand (because frankly,
I find the const names in the base64 crate very unwieldly), but you can
now use any of the base64 engines as codec.
We can do this because we know that `x < y` cannot create a
`TextCodec<T>` for any `T`. This is because `<` is guaranteed to return
a boolean value, and we simply don't implement `TextCodec<T>` on bool.
This allows stateful or configurable codecs without having to express
all configuration in the type name itself. For example, we could have a
Base64 type with configurable Base64 engines without having to duplicate
the Base64 type itself.
(Note that the different engines in the Base64 crate are values, not
types.)
In 1265f4b, we introduced a change which may cause a conflict of type
names when deriving the traits on two different types. While a
workaround existed (use `mod`s to isolate the implementation), that is
ugly.
This commit allows overriding the choice of type names.
