(Or if we gave a decorated matrix that override setZero...)

Tests aren't happy because we're now trying to hit test things that are invisible....

Instead of re-using paintTransform and clip, we could add a new property that describes the accumulated opacity?

I talked with @goderbauer about that. We'd like to try to avoid adding a new method to the interface if possible, since it's already a difficult interface to implement correctly. But maybe that doesn't matter if you have to do some magic around clipping or transforming...

So at least one failure is trying to tap a button that is invisible. That seems like a bogus test to me.

I started going through and looking at tests. In general, the problem is that tests are asserting that ROs that are not visible have certain paint bounds (usually via doing a hit test on them). Very often, it's enough to add a pumpAndSettle or move a Key object from something that's offstage to its parent.

Sometimes it's harder - a duration has to be modified or a pumpAndSettle added, and that will slightly change some hard coded value. In some cases, it also seems like the test was just relying on the ability to get paint bounds for something that has a zero valued opacity.

Rather than fix up all these callsites, I'm adding a flag so that tests are still allowed to hit test/get paint bounds for things that are transparent.

In particular, some callsites are concerning because they explicitly say "I don't want the animation to finish", but finding the exact point where the opacity is no longer zero but the animation isn't finished is hard.

I'm going to conver this to a debug flag.

Ok. the cupertino context menu failure seems like a real break. I'm going to close this until I figure out how to either fix that or a better approach for this entirely.

An internal global presubmit is suggesting ~200 test failures.

This doesn't work because Offstage needs to be able to measure where offstage things would paint if they were visible - this is used in several places in the framework, and is documented behavior:

Offstage can be used to measure the dimensions of a widget without bringing it on screen (yet). To hide a widget from view while it is not needed, prefer removing the widget from the tree entirely rather than keeping it alive in an Offstage subtree.

It deals with semantics for this by returning early if you try to visit its children for semantics.

I think the only way forward for this is to have a separate method on RO called something like bool paintsChild(covariant RenderObject child), and considering updating the existing ROs that zero out the transform to instead return true in that method.