That’s actually arguably the source of the problem in a lot of these examples. In some of them, the bad alignment would look good if the text next to it contained more descenders.

This could be compensated for, of course. For e.g. a single checkbox and text, move the text a bit lower if there are no descenders in it. But that doesn’t work if the text is user input, or if there are multiple text boxes in a list: it’s noticeable (and unpleasant) if the baselines of the text are in different places next to their individual checkboxes if - even if they’d all individually look best in that alignment viewed alone.

And if the text is editable, it’s even worse. No-one wants their text to be jumping up and down while editing to maintain optical alignment.

The problem gets even worse if you have mixed language support, and the text might be in languages with ‘tall’ characters (Thai, Arabic, sometimes Vietnamese - and a long list of others) or a different idea of ‘line height’.

Compromises must be made, unfortunately.

Of course, that doesn’t excuse a lot of the examples here, which are of specific designs - but it does mean that a lot more ‘design’ time is required, especially if there are multiple localisations, than feels at-all intuitive to anyone who has not done this.