It happens because each font has different width, even for the space character. You already know about the whitespace issues with inline-blocks. So, when you set Arial, those whitespaces change their width slightly from the browser's default font (or any other font with different width), which is Times New Roman in my case.
See how drastic the change is when you set the monospace font.
Now, why it happens between the 2nd and the 3rd box and not the 1st and the 2nd one? I'm pretty sure it comes down to rounding pixel values based on the width of the words entered, seems like there is a pseudo sub-pixel rendering present in the background, yet the decimal values get rounded in the final render process. See what happens if you use Arial and print Hell Stack Overflow instead of Hello Stack Overflow - the gaps look the same. So, it's just an undesired coincidence.
Another point that proves this is a rounding issue is the change in the gaps across various page zoom levels. It's fairly common to get these pixel mismatches in the layout when dealing with decimals in HTML. Zooming adds another dividing/multiplication stage, which changes the core values even further, resulting in completely unpredictable behaviour in the final layout.
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…