Attached is a patch to collections.py and legend.py to make
> the auto-legend stuff smarter.
> A couple of caveats.
> The new code doesn't makes as much use of C++ -- there might
> be performance issues for plots with large numbers of
> points, although I haven't encountered any.
> I don't think any of my tests include plots with
> LineCollections, so there is a reasonable chance that code
> in _auto_legend_data has not been exercised.
> Other than that, it seems to do the trick.
Thanks John, also added to CVS. I think this could be extremely slow
for very large lines (100k points, for example) but I don't see that
as a real problem because auto placement is not the default for
legends (for precisely this reason).