[ODE] trimesh woes

[Jon Watte]

> The problem doesn't occur when I use a box geom instead of a trimesh geom.

I have previously found cases where "small" values will turn 
into denormals, and then they'll be inverted in some test, and 
inverting a denormal turns into badness when you try and compare 
to things.

I forget which specific collider case this was, but I also think 
my suggested patch was rejected because it "impurely" hard-coded 
specific epsilons for float/double. Of course, nobody could 
suggest a better way of figuring out an epsilon that won't turn 
denormal when squared, but that seemed to not matter. I'm sure 
the bug is still in there.

Anyway -- you should trace through the code for a frame, looking 
at where the object gets its forces from, and how those forces 
are decided. This is somewhat cumbersome work, but you can trace 
the actual problem this way. Then develop a patch, and submit it 
to have it rejected.

Cheers,

			/ h+