... a few dozen such items, each of which is reasonably costly to compile, ... [perhaps] only one or two ... actually need to be used.

Sorry to be so long in answering.

I understand your point about postponing expensive initializations, and I would be inclined to agree in very extreme cases (e.g., an initialization that involved uploading your entire hard-drive to TheCloud™ via a dial-up ISP link). However, my general inclination, even in cases of fairly expensive initializations or object constructions, is to fail as early as possible, if fail one must. Operations that are more expensive tend to be operations that are more prone to failure. If such an operation is postponed until you are in the middle of another, perhaps critical, operation, error handling and recovery can become much more complex. In general, I think it's best to find out your cans of kerosene are empty before you start out for the South Pole, not on the way back!