Rather than discussing whether MathML is a failed standard, web or otherwise, I recommend we discuss specific, constructive topics. I suggest the discussion be in the context of MathML where appropriate, not because I want to defend MathML but because it is an existing standard. It is a place to start. If the solutions we reach replace MathML all or in part, so be it. Let's not start by throwing it out but by addressing its problems. We can certainly create a new standard if MathML can't be fixed. Finally, if this is the wrong venue for this topic or any other, please suggest a better one. If there are other parties that need to know about the discussion, please let them know.
Assuming others agree, let’s start with perhaps an important issue. Should Presentation MathML dictate a specific rendering or leave formatting choices up to the renderer. Peter says, "I have the impression people generally expect consistent rendering across browsers. But anecdotal evidence is, well, anecdotal." I would agree with this statement. People do expect this. I believe they get that expectation from TeX but it does make sense. Why would a user want a different rendering in a different browser?
The reason I said "no" to this before was because the MathML spec leaves a lot of rendering decisions up to the implementation. Someone reading the MathML spec should NOT expect all renderings to be the same. In fact, the spec doesn't specify the rendering at the required level of detail. Doing so would be difficult. TeX doesn't specify its rendering in detail either except via the code itself. In other words, the only proper rendering of TeX is that done by TeX itself.
We could create a MathML 4 in which the graphical rendering is specified in writing and in detail. Implementations would be constrained much more than by the current spec. Another way to achieve this goal is to create a reference implementation. This would be the TeX way, or close to it.
We could even map MathML onto TeX somehow and then defer to TeX's rendering. The MathML spec would be annotated by TeX templates (perhaps macros) that serve to define the rendering. The reference implementation would consist of a MathML-to-TeX convertor and the TeX engine itself. Implementations that intend to abide by the MathML 4 spec could use the reference implementation or roll their own.
When I say rendering above, I only mean graphical rendering. When we talk about audio or braille rendering, things are much less clear. The state of the art in MathML-to-speech has certainly not reached a point where everyone can agree. Besides, there is personal taste of the reader and multiple languages to consider.
Ok, I'll stop there and take a breath.
Paul