I don't think you need the end_tag_handlers handler though, the start one should be enough. You could also flush at the end of the handler to save memory, if that's an issue (untested).

Flushing in start_tag handler doubles the output  <note chop="SuEy" kung="FoO"></note> but end_tag_handlers => { _all_ doesn't get called at all

so nothing gets flushed until the whole tree is parsed

Is this by design of end_tag_handlers?