The one remaining test failure with svg.text.css-frames.enabled is a one pixel difference in dynamic-text-06.svg on Android. Reftest log snippet attached.

The failure is random, too. If I change the test https://hg.mozilla.org/try/rev/3bdbdd9c1feb to have a one second delay between each tick of the animation, then I don't get the single pixel failure.

Sorry, with that try run I neglected to set the pref. Changing the delay just makes the failure more obvious: https://tbpl.mozilla.org/?tree=Try&rev=f4d4106b1dda

I guess we're getting rounding errors somewhere in the process of taking the rectangles of each rendered run, converting them into user space and then converting back to the mRect coordinate space, although I didn't actually dig deep enough to find where this was happening. So while I think this is OK for the moment, at some point it'd be worth tracking down where the error is coming in, in case we should be doing the inflation (or rounding differently) at some other stage.