Or should they denote a glyph's default letter-spacing, in which case kerning classes are an additional layer of spacing refinement?

Yes.
Relying only on kerning would be a nightmare: you'd have to kern every single pair! Plus some software ignores the kerning.

So you set the sidebearings as best you can, accommodating all the possibilities as best you can. Just be warned: this makes the act of spacing a very zen-like effort. :-)

hhp

Open up a font from, say Adobe, and study the sidebearing assignments. That'll go a long way to help your understanding.

As a rule, I kern all possible letter combinations, plus numbers, common symbols and punctuation. BUT…before I kern, I go through all of the combinations in the Metrics Window. A slight nudge here and there can end up substantially reducing the number of kerning pairs required…or sometimes not.

Once upon a time, I created outlines for a font with very unusual serifs (I was going to call the font Twinkletoes) based on hand-lettering from an old chapbook. As it turned out, the only way the letters really looked good together was as they were presented in the book—A to Z, upper and lower. Virtually EVERY other letter combination required kerning. Bummer…

Thanks for the advice, guys!

Total bummer there, oldnick =P

Thanks for this post and the responses, very helpful! Good and easy suggestion by Charles to just check the sidebearings of an existing (possibly similar) typeface. Thanks guys!

Thanks for this post and the responses, very helpful! Good and easy suggestion by Charles to just check the sidebearings of an existing (possibly similar) typeface. Thanks guys!

To make your life a lot easier you should get your head around classes - both metrics and kerning. By doing this you can apply the same values to the similar elements of different characters, for example the left sides of b, h, i, k, l, m, n, p, r. And curved sides like the right of b, o and p. etc. Setting classes up properly in the first place can hugely decrease the amount of time spent spacing and kerning , and also the number of kern pairs. It also helps consistency. If you end up with thousands of pairs you're doing it wrong.