:-(

This is very much a bug in VE's test code. It's trying to call ve.ui.Surface.prototype.initialize while the surface is not attached to the DOM, which is explicitly prohibited in the documentation. This causes OO.ui.PopupWidget.prototype.toggle to be called while the popup is not attached to the DOM either, which we technically did not document to be prohibited, but it's clearly meant to be (it calls methods like togglePositioning and toggleClipping which are documented so).

The culprit is this:

QUnit.test( 'special key down: table cells', function ( assert ) {
…
				// Create a full surface and return the view, as the UI surface is required for the insert action
				htmlOrDoc: ve.test.utils.createSurfaceFromDocument( ve.dm.example.createExampleDocument( 'mergedCells' ) ).view,
…
} );

So actually, ve.test.utils.createSurfaceFromDocument appends the surface to the DOM, to #qunit-fixture, which gets cleaned up immediately when the test finishes. The problem comes from the 'afterContextChange' handler, which is async, and runs after that test finishes and the surface is not attached anymore. (This is why the failure is showing up for different tests than the one really causing the problem). I'll note that the ve.ui.Surface never gets #destroy called on it, either, it's just unceremoniously removed.

Change 338011 had a related patch set uploaded (by Bartosz Dziewoński):
ve.ce.Surface.test: Gross workaround for failing tests

https://gerrit.wikimedia.org/r/338011

The above one-line workaround makes the problematic test async, and magically makes everything stop failing, at least locally for me. I'm leaving this assigned to Ed to work on a real fix, or decide that this is enough. :)

Change 338011 merged by jenkins-bot:
ve.ce.Surface.test: Gross workaround for failing tests

https://gerrit.wikimedia.org/r/338011

OK, MW/core and related repos' CI should be back to working. Sorry everyone.

Change 338025 had a related patch set uploaded (by Jforrester):
[Re-apply] Update OOjs UI to v0.19.2

https://gerrit.wikimedia.org/r/338025

Change 338028 had a related patch set uploaded (by Bartosz Dziewoński):
ve.test.utils.createSurfaceFromDocument: Actually attach surface to DOM

https://gerrit.wikimedia.org/r/338028

The issue was still occurring with VE-MW, @Krinkle and I spent more time debugging on #wikimedia-dev.

Krinkle noticed that ve.test.utils.createSurfaceFromDocument is overridden in VE-MW… so I started looking into that function:

[01:05]	<MatmaRex> the cause is that we have a ui.Surface that is somehow not attached to the DOM when it should be, and that breaks things
[01:06]	<MatmaRex> in VE core, the problem was that it was being cleared from #qunit-fixture while things were still ongoing. i fixed that with the setTimeout( assert.async() ).
[01:06]	<MatmaRex> i think right now, the problem is that it does not get attached at all. i'm still trying to figure out how to verify this.

Change 338028 merged by jenkins-bot:
ve.test.utils.createSurfaceFromDocument: Actually attach surface to DOM

https://gerrit.wikimedia.org/r/338028

Change 338025 merged by jenkins-bot:
[Re-apply] Update OOjs UI to v0.19.2

https://gerrit.wikimedia.org/r/338025

I submitted https://gerrit.wikimedia.org/r/338181 "PopupWidget (and similar): Document why it is unwise to show unattached widgets, and emit warnings" in OOjs UI which should make it more difficult to make this kind of a mistake in the future.

The OOjs UI change in v0.19.2 (rGOJU799f500934eb: FloatableElement, PopupWidget: Do positioning from the right in RTL) only exposed a pre-existing bug in VE tests. This was not an OOjs UI bug.