I am doing some testing on this branch and I found a broken example: legend_scatter.py now has missing markers in its legend.

An interesting bug with respect to hist(). For regular histograms, I think everything is fine. However, for any other type, if the label is provided through the legend() function, the artist for that histogram is skipped and the label goes to the next available artist. However, it works perfectly if the label is provided through the "label" kwarg to hist().

Note that before this, either approaches worked for histograms.

Another bug (or at least a difference in behavior)... for a dragable legend, if I click but not hold down the button, the legend will move with my mouse until I click again. Note that there is also a separate bug where doing this throws a "TypeError: 'instancemethod' object is not iterable", but that doesn't prevent you from continuing and this bug occurs in master as well.

Note that cherrypicking the recent commit to fix draggable legends fixed both problems I just reported.

The first issue should be solved by commit 1b17994.

Can you provide an example code for the second issue?

I am using a modified version of histogram_demo_extended.py. After investigating a little further, it appears that the problem occurs for any histogram type that is not patch-based.

I am linking to two versions of the script, one that uses the labels kwarg:
http://dl.dropbox.com/u/7325604/histogram_demo_extended_using_labels.py

and one that uses the legend() function:
http://dl.dropbox.com/u/7325604/histogram_demo_extended_using_legend.py

In particular, figures 1 and 3 are impacted.

Ok, I think I have exhausted all scripts I can find that involves legends and I can only find one other issue (and it isn't new). Stem plots cause a problem because they attach both the dots and the line to the list of collections for legends. So, if I do two stem plots in an axes, and then call legend with two labels, the first label gets black dots, and the second label gets the line for the first stem plot.

However, this is nothing new...

I have not done a code review yet, but in terms of backwards compatibility, it appears that we are (mostly) golden.

Regarding the second issue, I think it is related with the current buggy(?) behavior of legend (and the "hist" command relying on this buggy behavior). I posted a related message in the devel-list. The modification of this pull request depends on how this issue is addressed.