Version: 55.0.2871.0 canary (64-bit)
OS: Windows 10 64 bit

STR:
1. Start the NVDA screen reader.
2. Open this URL in Chrome: https://github.com/nvaccess/nvda/issues/6404
3. Press control+home then g.
4. Press down arrow until you hear something like: "link  fernando-jose-silva commented  link  2 days ago"
5. Press NVDA+tab to report the focus.
Observe: NVDA should say "fernando-jose-silva  link..."
6. Keep pressing down arrow to move through the comment.
Expected: You should eventually reach the bottom of the comment.
Actual: Somewhere within the log output included in the comment, NVDA will say "fernando-jose-silva  link" after reporting the line you moved to.
7. Press down arrow a few more times.
Expected: You should continue to move down the page.
Actual: You have been thrown back to the top of the comment.

Some explanation of what's happening here:
1. The link for the author gets focused in step 4 as a result of you moving the browse mode cursor.
2. As you move forward by line, NVDA scrolls the accessible associated with that line into view.
3. Eventually, the link focused in step 4 gets scrolled off the screen, but it's still focused.
4. Chrome kills the accessible for the focus and creates a new one. It then has to fire a new focus event, since the old accessible died.
5. NVDA gets that focus event and moves the browse mode cursor to that accessible. Even though it's actually the same link and really the focus hasn't changed, NVDA doesn't know this because it's a different accessible.
6. Unfortunately, this means the user gets to cursor around in loops. :(

This doesn't occur everywhere. I suspect the issue is related to the fact that part of the screen stays fixed while the comments area scrolls. However, despite playing with position: fixed, I unfortunately wasn't able to isolate this behaviour in a simple test case, even though the scrolling behaviour looked similar.

Firefox has similar problems with certain CSS stuff causing accessibles to be regenerated, though this particular case doesn't seem to affect Firefox.

Originally reported as NVDA issue: https://github.com/nvaccess/nvda/issues/6405