Given the design of Cantarell, you can go as light as you like, and a step or two bolder, before you start to need to do anything special for optical compensations with an extra master. At first it would just be for a few glyphs such as a and e. If you go a lot bolder, you'd need to do a whole master.
You should try starting with keeping the Regular as a master, and doing the lightest and boldest weights. Then see how different it looks (and whether it's a problem) if you drop the Regular master.
That said, in general I would just use two masters, and do corrections at the heaviest end as needed (per Georg's comments). The reason being that for further development, changes and extension, it's just a lot less work if you have two masters instead of three.
And as a bonus, the final file size would be smaller. I know you say that's not a priority, but it can't be a *bad* thing, right?
1) Well, first I will note that for me, if I am doing axis-based design, also targeting variable fonts doesn't make much difference. Variable Fonts have all sorts of capabilities for additional masters or even glyph-specific tweaks to the design space that allow all the same kinds of compensations that one might have made when using axis-based fonts as an upstream source for a large family of fonts.
I'm confused by the fact that you talk about designing a new family in your point #1, but you've also said you are porting Cantarell. This seems contradictory. But it's late at night, so I may be missing something.
But in any case, there is not in general anything inherently massively different about doing axis-based design than just doing design. Mind you, I may have trouble seeing what you might see as differences, because it wouldn't occur to me to NOT do axis-based design if I have at least two weights. Seriously, if I am doing more than one weight of a typeface, I would do a weight axis. If I need a slightly bolder and slightly wider variant of a font, I build a variant with weight and width axes and twiddle the values until they work best. Axis-based fonts are just a basic part of 90% of my type design, and have been for 20 years.
2) In your case, two considerations regarding making the regular weight a master, and having lighter and bolder masters as well:
you are mostly focused on web fonts and size matters. Consider that more masters means a bigger final font file.
I do not know how much you care about compatibility with the existing non-variable version of Cantarell. If it is important that the new variable version be pretty close, you might need an intermediate master.
3) Yes, you can add a width axis later on. BUT, first we need to know: Will you need the later version with the added width axis to be highly compatible in design with the version you create now, without the width axis? Do you need that final version to still be hightly compatible with the original Cantarell?
If you need compatibility between the width-axis font and the later two-axis version, and you aren't constrained by compatibility with the existing Cantarell, then shipping the one-axis version first will put some constraints on the later two-axis version that you would not otherwise have. That would seem like a less than ideal situation, to me.
Back when screens were much lower resolution, and LCD-specific strategies didn't exist yet (because they were dependent on LCDs which were not yet predominant), this idea was pretty obviously true. Screen fonts were specific pixel-based things and very different from print fonts. There just weren't enough pixels (and available sub-pixels) to do serif faces properly.
Early screen-font legibility research easily confirmed this.
Resolutions have improved immensely, and so have font rendering techniques. So the idea that sans serif is better for screen, while clearly true 30 years ago, began to change. Unfortunately, the current truth is not nearly as well known as the former truth. Nowadays I doubt anyone seriously familiar with the topic would make any blanket statement of sans serif superiority for screen use. But memes are hard to kill.
I don't see how this is an interoperability issue. Yes, this font is spaced differently. Why is that a problem? We expect spacing to change when we change fonts.
Also, Adobe has been doing different spacing for French for some characters for at least a decade, and while some people have liked it and a few have not liked it, nobody has freaked out about it. It is not that big a deal.
Graphology is not a science. If you choose to believe in astrology or graphology or the like that is up to you. Just be aware that attempts to validate graphology have consistently failed, over and over, literally hundreds of times.