The 'qname' argument is not only used for element names, so calling
the non-namespace part 'tag' is confusing. The XML namespaces spec
calls it 'local part', so I picked 'local' as the new name.

All other callers of add_qname() already check the argument type
prior to calling, so by also checking before adding an attribute
name, add_qname() doesn't need to deal with TypeError anymore.

The diff is actually tiny when whitespace changes are ignored.

This is in preparation of a future change that requires the default
namespace to not be applied to attributes. The cache population code
will become non-trivial then.

This is in preparation of a future change that requires the default
namespace to not be applied to attributes.

Currently there is no real need to do this, since the serialization
used for tags and attributes is identical. But in the future, they
will use different serializations and we want to avoid defining
namespace prefixes for namespaces that are not going to appear
in the serialization.

This is in preparation of a future change that requires the default
namespace to not be applied to attributes.

Because the default namespace is no longer in the mapping, the first
generated synthetic namespace prefix is now always "ns0", while
previously it was "ns0" or "ns1" depending on whether a default
namespace was provided. I updated the expected serialization in
one test case to match this new behavior.

Previously, it returned a mapping from namespace to prefix. But later
the serialization code sorted that mapping by prefix, so inverting the
mapping simplifies things.

Another reason to invert the mapping is that currently there is a 1:1
correspondence between prefixes and namespaces, but in the future it
will be possible for the default namespace to exist both with an empty
prefix and with a actual prefix.

This is not my work: I extracted this test case from the second
iteration of wiml's patch for issue 17088.

https://bugs.python.org/file33125/bug17088_2.patch

I did modify the line wrapping, as the original patch went very wide
(over column 100). I also updated the prefix numbering in the expected
serialized output, since numbering always starts at 0 now.

Unprefixed attributes are considered to not be in any namespace,
according to the XML namespaces spec.

This also fixes issue 17088, since the exception that rejects
non-qualified names when using the default_namespace option is
no longer raised when serializing an attribute name.

When no default_namespace is passed, serialize_qname() returns the
same value regardless of whether is_attr is True or False. So we can
save some time by storing the serialization for both the tag and
attribute in the cache when one of them is computed.