Guys, there's no 'right way' with this... Stop arguing like there is. There are pros and cons to either method and they both work just as well. Yes, you could save a little UV space. No, you won't kill the engine with little render mesh N calls or whatever Tomo is thinking. Alternatively, you could copy your little details as many times as necessary and save time modelling and UV mapping while also having your polygons match exactly and have an easier time texturing later.
I've done both. Many times. Both work juuuuuuust fine.
All that's true. (Though it is possible to break the collision detection, that's fairly rare and should be easily fixed.)
Modern model-related slowdowns mostly come from excessive OpenGL state changes (eg render calls) and from running out of GPU memory.
The former has a relatively small effect so only matters for 'many ships', the latter has a very large effect.
The actual number of polygons and vertices is doesn't really matter these days. (As long as it's below the maximum allowed.)
Vertices don't eat much GPU memory, textures do.
What I meant was that using 'many subobjects to save polygons' as a rendering optimisation technique actually has the opposite effect.
If you're using subobjects to make your task of building the model, texturing etc easier then great, do it!
The time you and others spend to create the model and get its gorgeousness into the game is far more valuable than the time the GPUs will spend rendering it.
All details aside, in a capship 'small optimisations' like this don't really matter anyway as there's never going to be very many capships in a mission!
These kinds of tweaks only make a difference if there's lots of them.
It's unclear what point your trying to make. You're listing valid points, and then downplaying them.
1) "The actual number of polygons and vertices is doesn't really matter these days. (As long as it's below the maximum allowed.)"
This number is VERY easily hit with modern ships these days. So any technique that can lower the vertex count will cause fewer head aches later.
2) "What I meant was that using 'many subobjects to save polygons' as a rendering optimisation technique actually has the opposite effect."
How? If you are referring to UV map wastage, there are plenty of techniques that minimize or eliminate wastage (nothing stops you from drawing in the occluded portions where non-manifold polies intersect...) However you do seem to have a better grasp on this than I do, so I concede that I may be missing your point here.
3) "...there's never going to be very many capships in a mission!"
People love to whore out cap ships. The other thing is fighter poly counts are climbing as well. We see missions with more and more ships because FSO makes it possible, so trying to use that as a point is pretty short sighted.
So in conclusion, there still is little to no reason to make your mesh entirely manifold.
EDIT: Apologies for the rude flavor above. I would just change it, but in the off chance you are reading this before I can, i'll just tack it on. My main point is that if we agree that:
1) Multiple sub-objects = slower game
AND
2) Hitting the subobject vert limit more often means you must have more subobjects
Then the only conclusion is that fewer verts and polies = better performance, just not for the reason that most people think.