I recently wrote a CL that adds a text-only button to the status area when running with --enable-features=SingleProcessMash. To my surprise this caused a variety of unit test failures, including:

ShelfViewTest.ShelfRipOff
ShelfViewOverflowFocusTest.Basic
ShelfViewOverflowFocusTest.ForwardCyclingWithBubbleOpen

My code changes the width of the ShelfView, but it does so in a way identical to LogoutButtonTray, which we use in production.

shelf_view.h has a diagram showing that ShelfView::last_visible_index() should match the overflow shelf's first_visible_index(). That isn't true for some shelf widths.

I wrote a test CL to demonstrate the problem:
https://chromium-review.googlesource.com/c/chromium/src/+/1253077

At certain shelf widths ShelfView::last_visible_index_ goes backwards when the overflow bubble appears. According to the diagram in shelf_view.h that shouldn't happen.

I can also see this happening when running chrome, especially if I make the desktop window very narrow.

sammiequon added the overflow focus tests recently:
https://chromium-review.googlesource.com/1125317

manucornet's CL here made some changes to last_visible_index_ computation:
https://chromium-review.googlesource.com/c/chromium/src/+/1150776

manucornet's CL here commented out some overflow-bubble test expectations that are similar to the failures I'm seeing:
https://chromium-review.googlesource.com/c/chromium/src/+/1188848

Commented out expectations:
https://cs.chromium.org/chromium/src/ash/shelf/shelf_view_unittest.cc?sq=package:chromium&g=0&l=909

I think something is wrong in last_visible_index computation, maybe only in tests, but maybe in production as well.

Manu, could you investigate?