With this site, you can see for yourself how ready-for-prime-time MathML is, as a method for sending mathematical formulas to browsers, 17 years after its initial release.
My primary browser, Chrome, falls down at the first hurdle, as I already knew — Google has officially given up on MathML. But my secondary browser, Safari, looks pretty good for the first seven of these. It only really starts looking ugly when we get to binomial coefficients, and doesn't get completely unusable until the nested square roots and overbraces. Of course, I tend to use binomial coefficients frequently, and those other things sometimes. Firefox (which I have installed but very rarely use) can do the whole test at an acceptable level — readable if not always perfectly polished — but I don't want to generate web content that only Firefox readers can view properly.
This all makes me sad. I want writing mathematics on the web or in Wikipedia to be as easy as it is in LaTeX. MathJax can do that, but at the expense of unclean markup and occasionally-slow rendering. Instead, MathML offered the promise of clean semantics, fast built-into-the-browser rendering, and high quality, but it just hasn't delivered.
"One cannot write MathML directly, so a preprocessor must be used. But then it doesn't make much of a difference whether to render to MathML or directly to HTML+CSS."
That's just as if you would say assembler is useless because you better write in C and a compiler must be used to translate it into assembler. This is a bit absurd.
And it is precisely MathJax who leads the arguments completely at absurdum. When you look at a formula rendered by MathJax right-click and you get the option "Show Math as: MathML code". Mathjax speaks MathML under the hood. Only the browsers do not use it because it is not implemented.
MathJax doesn't “speak MathML under the hood”, it merely supports MathML as one of the several output formats. Your analogy with an assembly language is incorrect; MathML is a high-level language, like TeX, but without macros and with insane syntax. By the way, there is an assembly language for math: TeX's math list (see Chapter 17 of The TeXbook). Internally, TeX compiles formulas first into a math list; these math lists are then further compiled into horizontal lists. It would actually make a lot of sense if browsers supported something like TeX's math lists.
As for your "under the hood" argument, that is both off-topic (because it's not about browser support, the subject of my post) and an unfair comparison. If you try to use MathML as a server-to-browser communications language, you have to deal with many-to-many issues: many different people trying to write MathML in different ways, many different browsers displaying it in different ways on different systems with different screen sizes and font setups. If you use it as your internal representation under the hood, you know you are only getting the MathML from one source (your own TeX conversion software) and sending it to one destination (your own rendering software) and you can use that knowledge to compensate for any inadequacies in the format.
You do not describe even approximately adequate the considerations of the committee. It seems pointless to discuss with you further before you engaged in some closer study of MathJax and MathML. A good start would be to look at the history of MathML which tells quite a different story from the very start:
"There was a danger during the 1990s that a standard would emerge for mathematical representation on the web that would be based on a TeX-[..]like typesetting language. This would have been disastrous because it would have precluded, or made far more difficult, meaningful computational interaction with mathematical expressions found on the web."
from http://www.mathmlcentral.com/history.html
“As we see it, there are two fundament goals for the MathML standard: … (2) that the standard be one in which it is possible to uniquely represent mathematical meaning in such a way that computer programs can read, parse, and evaluate expressions represented in MathML without ambiguity.” Clueless indeed! As the quote clearly states, MathML is a (bad) standard for computer algebra systems. So why was it incorporated into HTML in the first place? The absolute majority of formulas cannot be assigned a meaningful interpretation that could be useful for computer software. Knuth explicitly speaks against such “semantic” markup, see page 148 of The TeXbook.