I think this is a candidate for backport because it fixes a un-workaroundable bug in 2.2.2. I am not 100% on that, if someone else agrees merge away!

I actually am 98.5% sure its not a followed code path because of the new clipping for imshow (#10220). 0% codecov isn't very good ;-)

(i.e. max(x)>1 is always False)

This is (for better or worse) a public function, I think the clipping happens when we set the data so using the 'standard' path this won't be hit, but using im.to_rgba(my_data) can still hit it.

Ah, fair enough.

OTOH, should it get a test if its public, including one to test if this error is properly called? But happy to see it merged if thats gilding the lily...

This is the sort of bugfix that I think does not warrant a test.