Ensure dialog cancel events fired even without interaction

[m-akinc]

The cancel events should not be dependent on the user having first interacted with the dialog. Because this test includes user interaction before sending the close request, it failed to catch a regression in Chromium.

Removing the user interaction (i.e. call to blessTopLayer) so this test covers that scenario as well.

[domenic]

This is not correct. The spec requires a user interaction to either open the dialog, or to interact with the dialog while it's open, if cancel is going to be fired.

There is a slight improvement we can make to the spec, which is being discussed in whatwg/html#10046 . But the fix for that is proving pretty complicated, and will involve significant test changes beyond deleting one line. See https://chromium-review.googlesource.com/c/chromium/src/+/5232387 for my current work in progress there.

[m-akinc]

This is not correct. The spec requires a user interaction to either open the dialog, or to interact with the dialog while it's open, if cancel is going to be fired.

Ah, I hadn't realized that the new CloseWatcher stuff was already part of the official specification (submitted a few months ago). But your description in that PR acknowledges that this was a breaking change:

This change to dialog behavior is a potential compatibility risk, especially since it can cause the cancel event to be skipped if there has been no user activation since the last time it was canceled, or multiple dialogs to be closed at once.

I would just like to raise awareness that this has real consequences. Part of the justification for this change was that only 0.000015% of page loads would have been affected (possibly broken) by this, but what qualifies as an acceptable percentage? It's also worth noting that since you cited that number back in October, it has apparently grown by 37x:

@domenic

[domenic]

The growth is due to a bug where when we remove the use counter, the denominator trends toward 0 (depending on how fast old Chrome versions get out of circulation), so the values swing wildly. There are no valid data in the chart after November.

[m-akinc]

In retrospect, that makes a lot more sense than a steady, drastic increase in a that specific usage pattern. 😅