[S] There are two different discussions -- how could namespaces have been designed differently, and, how can we improve the current mess.

[S] What would an XML app do when it comes across an element you've forgotten [to add to the vocabulary definition]. [S] It would be in the same namespace as the parent. E.g. if you had said orderedlist implied a docbook namespace, then a lst element, and a para within that would be ins the same namespace without you having to say anything. Validation using an existing schema would involve first adding the namespace URLs explicitly (e.g. with XSLT, a transformation that would read the namespace file and apply the namespaces to the instance), so that then if you'd forgotten an element that mattered, you'd find the error.

[S] Re losing namespaces from attributes. [Loss of] compatibility in current data models that know about namespaces. Existing vocabularies with attributes in namespaces wouldn't be able to migrate to the new namespace version.

[S] There are some important specs like FpML that (for better or for worse) place very heavy reliance on xsi:type, and the reality is that they aren't going to change in a hurry: there are zillions of financial transactions whizzing around the world that use this attribute.

Incidentally XSLT also makes use of namespaced attributes (e.g. xsl:use-when). They aren't used in every stylesheet, but they serve an important purpose and you can't just get rid of them.