Reading the proposed patch, I must agree that it makes more sense in ElementTree to support this as a serialiser feature. ET's tree model doesn't have a notion of prefixes, whereas it's native to lxml.etree.

Two major advantages of putting this into the serialiser are: 1) cET doesn't have to be modified, and 2) it does not require additional memory to store the nsmap reference on each Element. The latter by itself is a very valuable property, given that cET aims specifically at a low memory overhead.

I see a couple of drawbacks:

1) it only supports the case that namespaces are globally defined. The implementation cannot handle the case that local namespaces should only be defined in subtrees, or that prefixes are being reused. This is no real restriction because globally defined namespaces are usually just fine. It's more of an inconvenience in some cases, such as multi-namespace languages like SOAP or WSDL+XSD, where namespaces are commonly declared on the subtree where they start being used.

2) lxml.etree cannot support this because it keeps the prefixes in the tree nodes and uses them on serialisation. This cannot easily be overridden because the serialiser is part of libxml2.

I didn't see in the patch how (or if?) the prefix redefinition case is handled. Given that prefixes are always defined globally, it would be nice if this only resulted in an error if two namespaces that are really used in the document map to the same prefix, not always when the namespace dict is redundant by itself.

Also note that it's good to be explicit about the keyword arguments that a function accepts. It aids when help(tostring) tells you directly what you can pass in, instead of just printing "**kw".