Terry:  sorry, I missed this before.

Re 1.  Sure, this seems reasonable, if there's a real sense in which the new numbers are better than the old.  Besides MATLAB, there's also a set of numbers given on Wikipedia that might be considered.  I don't have the domain knowledge to know what's sensible here.

I *would* rather see the inverse transformation keep the full 6 digits of precision, though.  Or perhaps even give the inverse to full float precision.  Without that, the result of roundtripping RGB -> YIQ -> RGB could be significantly (perhaps even visibly) different from the original.  I don't think it's acceptable for the roundtrip error to increase significantly w.r.t. Python 3.2.

Re 2.  For this issue, I don't see a real benefit to adding the tests to the maintenance releases.  No strong opinions, though.