bboxes coordinate system of plt.annotate objects depends on range of xycoords #8475

Bug report

When I plot ax.annotate objects, and then want to get the coordinates of their bboxes using fig.canvas.draw() and annotation.get_window_extent(), then the coordinate system of the bboxes depends on whether the xy of the object is between 0 and 1 (including 0 and 1), even if xycords is always set to 'data'.

In particular, if xycords are between 0 and 1 then the bbox coordinates can be transformed into data coordinates using ax.transData.inverted().transform. This is not true for xycords not in [0,1].

Code for reproduction

from matplotlib import pyplot as plt
fig = plt.figure()
ax = plt.gca()

anns = []
xcoords = [0.5,1,1.5]
y = 0.5

for x in xcoords:
    ann = ax.annotate("Test",xy=(x, y), xycoords='data', verticalalignment='bottom')
    anns.append(ann)            

fig.canvas.draw()

for x, ann in zip(xcoords, anns):
    bbox = ann.get_window_extent()
    (xmin,ymin), (xmax,ymax) = ax.transData.inverted().transform(bbox)
    print bbox
    print x, '==', xmin
    print y,'==', ymin

Actual outcome

Bbox(x0=221.4, y0=147.6, x1=247.24375, y1=160.6)
0.5 == 0.5
0.5 == 0.5
Bbox(x0=388.8, y0=147.6, x1=414.64375, y1=160.6)
1 == 1.0
0.5 == 0.5
Bbox(x0=1.5, y0=0.5, x1=27.34375, y1=13.5)
1.5 == -0.156810035842
0.5 == -0.159050179211

Expected outcome

I would expect all Bbox objects to be in the same coordinate system (figure coordinates?), but they clearly are not.
I would expect a behaviour where the last two lines of the output above become

1.5 == 1.5
0.5 == 0.5

Matplotlib version

'1.5.3' from anaconda2
I am running this in a jupyter notebook with %matplotlib_inline