1. By their nature, niche segments are likely to be small, right? That means you may well need large total sample sizes to find them at all and once you've found them you want to have enough respondents to profile them and to understand them well (else why find them in the first place). Looking for niche segments isn't for the faint-hearted (or for those with slim budgets).
2. I think you're asking about what segmentation method. I'd avoid k-means, which tends to produce roughly equal-sized hyperspheres. Hierarchical might be a better bet. If you're using conjoint results you might also look into latent class MNL as being a way to find segments of respondents.
3. I suppose it depends on how many segments there "really" are. If you have 3 big segments and a niche segment it's not clear to me how much adding a 5th segment will help you, assuming that structure really exists and you're using a method capable of finding it.
4. I'm not sure what you mean by "type of variable." Some methods are more sensitive to some types of variable scaling than to others, but I don't know that that's what you mean by "type of variable."