Oops, a test case I recently added might be somewhat fragile. I'll fix it (increase relativeMargin) if the problem repeats.

FAIL tests/unit/doc-printer.js
`printDocToString` has linear time complexity at most to print fill() > numWords=20000
    Error: expect(received).toBeLessThan(expected)
    
    Expected: < 41.39885533368215
    Received:   41.98062500008382

Though we expect "fill" to have odd length, but we didn't throw for violations, we even did wrong things in core, so plugins should already made mistakes, we can't force them to follow the strict rule, at least not a good time.

I know I said add assert for "fill", but it should only run in dev mode.

For "stripTrailingHardline", instead of remove "hardline" from parts directly, maybe we can replace it with empty string. Do you think it will work?

so plugins should already made mistakes

Yes, I agree with you, they shuold.

However, maybe the comment misses an important point. We don't (won't) expect fill() to have odd elements. The rule we are working on is that even-th (1-indexed) elements have to be line-like doc.
What I'm worrying about stripTrailingHardline() is, change in behavior.
It's not easy to demonstrate the change simply. New stripTrailingHardline() always output valid fill() when input is valid fill(). The one in main can output invalid fill() for valid fill(). The behaviour will change, anyway.

And I found that new logic might be able to get simpler 😅 . I was somewhat confused and it might confused you. I'll see it later.

I reverted some commits to get this PR to be conservative.
We can discuss whether we let current stripTrailingHardline() go with seiyab@8565e91 or fix stripTrailingHardline(), later.

seiyab/prettier@8565e91 seemsgood, but it should not need call cleanDoc since the parts already "clean"ed.