My knowledge of JPCT and 3D are very limited, but I think I understand the problem here. What you're saying seems to make sense and would explain why scaling my unscaled model (via .setScale) to large values doesn't work.

The only thing I don't understand is why the problem doesn't occur when I load the already scaled model. Surely the same transformation to ellipsoid space occurs and the same accuracy error should occur?

Isn't there some way to load the unscaled model and somehow transform it to being exactly the same as the scaled model would be? Or is this what using rotateMesh() will do?

I'm not entirely sure how the positions in worldspace are different for my two XML files. Each of them loaded the same 3DS file. One imported it at scale 1f (and exported to XML) and the other imported it at scale 17f (and exported to XML). The one that is imported to scale 1f then gets set to scale by .setScale, so all the parts of the model should surely be in the same place?

I know I'm wrong, because my 3D knowledge simply isn't that great. I just don't understand why it's wrong

Phew, okay...

I tried getWorldTransformation() before and after I did setScale and the matrices are different.
Matrix before:
(
   1.0   0.0   0.0   0.0
   0.0   1.0   0.0   0.0
   0.0   0.0   1.0   0.0
   0.0   0.0   0.0   1.0
)

Matrix after:
(
   17.0   0.0   0.0   0.0
   0.0   17.0   0.0   0.0
   0.0   0.0   17.0   0.0
   6.0906405   33.543255   137.59756   1.0
)

Regarding the fragments that disappear. You were right, I was creating the octree before doing the rotatemesh. I have switched this around and the structure looks okay, although the lighting is different. I imagine this is still due to the order in which I am doing things?

I've tried disabling lazy transformations both before and after I set the scale, with no effect.

Unfortunately, my code is all too interconnected for me to be able to provide a test case as I'm using a client/server model with rendering done on a client and collision detection done on a server.
I could send you the two versions of the XML file I'm using. Maybe I've done something silly in constructing them.
I also have the code for the construction of the model itself. But it's abstracted slightly. In pseudo code it looks rather like this:

Loader.loadSceneFromXML(s, state.getWorld());
For all objects in the world:
1) Convert to triangle strips
2) Set collision mode to detecting collision with others and activate optimization
3) Set up octrees

Build world
Generate camera
Set camera position
Get the structure by name (in the world)
SetScale on the structure

<ponder> Maybe build world should come AFTER the scaling?

And to clarify, when I do this, collisions are immediately being detected by camera.checkCameraCollisionEllipsoid even though there are no visible collisions present. In response to this, my camera drifts upwards (which I would expect to happen in my collision system) until the camera elipsoid is (I would guess) more or less exactly reseting on top of the model.
To make it work correctly, all I do is load a different XML file (same model, but with the scale already set) and don't use .setScale. The camera is still in exactly the same location, everything looks identical and collisions are only detected when they ought to be detected.

Ahhhh, I understand.

Cool, I'll just stick with the post loading scaling for now

Ah, okay, excellent.

Thanks!

Awesome, thanks.